Summary

• This is a condensed version of Part 3 of our 3Q26 rebalance, diving into a selection of individual stock decisions — the full report, including our complete Soitec thesis, test and inspection, memory, and the silicon-photonics supply chain, is available at Convequity.

• Recent blow-ups like Situational Awareness and Archegos reaffirm why our portfolio stays unleveraged and diversified across ~80 names, absorbing a 25% drawdown that would have been fatal in concentrated, levered books.

• The rebalance’s central principle is that technical relevance alone is not enough — we exited positions where compute-native entrants or commoditising software layers eroded the strategic moat.

• We favour bottleneck resolvers over bottleneck causers, with Soitec emerging as the marquee conviction because its Photonics-SOI wafers win regardless of which optical architecture prevails.

• Where architectural outcomes remain unsettled, we prefer businesses positioned across multiple paths — hence increasing Ciena for coherent scale-across connectivity and adding to the under-appreciated test and inspection layer.

• Pre-revenue and geopolitically exposed optionality is deliberately kept small — or held back entirely — a discipline made possible by the ballast of more mature, cash-generative holdings across the AI value chain.

Every Cycle Has a Cautionary Tale

In the latest AI-infrastructure boom-and-correction, Leopold Aschenbrenner’s Situational Awareness fund became the clearest cautionary example. Launched in 2024 by the former OpenAI researcher, the vehicle scaled from roughly $225 million to peaks of $20–45 billion and posted extraordinary returns — 439% net through the first half of 2026 — by concentrating in a narrow set of AI-infrastructure names and amplifying those bets with up to 4x leverage. The July 2026 sell-off turned that combination into a 67% monthly decline, forcing a fire-sale of the public equities book to Citadel and shrinking assets to around $8–10 billion. Even though the fund remained up roughly 80% for the year, the episode showed how quickly overconfidence, limited diversification and leverage can erase gains and threaten permanence of capital.

History repeats. In the 2020–2022 growth-stock cycle, Cathie Wood’s ARKK suffered an 80–85% peak-to-trough drawdown. Even with 30–50 holdings it remained far from true diversification and was heavily exposed to high-conviction, high-volatility names. A starker parallel is Bill Hwang’s Archegos, which concentrated in a handful of stocks and layered on massive swap leverage. When those positions reversed in March 2021 the structure collapsed in days, generating more than $10 billion in bank losses and wiping out the firm. Concentration plus leverage is a recurring recipe for asymmetric downside.

This is why construction matters.

The Convequity portfolio is deliberately different. It is unleveraged. It holds approximately 80 stocks spread across 11 segments that we map across the AI value chain — from semiconductors and optical interconnects through power, packaging, memory, software infrastructure, and adjacent enablers and to enterprise and consumer applications sectors. No single name or narrow cluster is allowed to dominate. Position sizing and rebalancing discipline are designed to keep risk manageable through volatility rather than to maximise beta in a one-way market.

That architecture has been tested this year. We added further AI-infrastructure exposure at the start of the second quarter of 2026. The timing proved constructive: after a –7% first quarter, the portfolio returned 37% in the second quarter and reached a year-to-date peak of +45%. When the July AI-infra correction arrived, gains retraced, yet as of early August the portfolio remains up approximately 8% for 2026.

To quantify the drawdown from peak: a nominal $100 at the start of the year would have stood at $145 at the high point and $108 today. The decline from that peak is roughly –25.5%. A 25% peak-to-current drawdown is material, yet it is a fraction of the 60–80%+ declines experienced by the more concentrated, leveraged books described above. Diversification across dozens of names and the absence of leverage absorbed the shock without forcing liquidations or permanent impairment.

The rebalancing decisions detailed in this Part 3 were taken in the final week of June. More than six weeks have since passed, and the intervening market moves have only reinforced the importance of the portfolio construction principles above. Markets will continue to deliver cycles of exuberance and correction. The managers who survive and compound through them are rarely the ones who ride the purest expression of a thesis with the highest leverage. They are the ones who respect position limits, maintain genuine diversification, and refuse to let temporary over-confidence dictate risk parameters. That is the discipline we apply.

Overview

This is the condensed version of the third and final instalment of our 3Q26 portfolio rebalance series, focusing on which companies are positioned to capture value as the AI infrastructure buildout moves from technological promise toward commercial deployment.

Across the holdings discussed below, we distinguish between technical relevance, commercial readiness and durable value capture. An important technology does not automatically make every participant in its supply chain an attractive investment. This framework informed our exits from businesses with weakening competitive positions, the retention of selected higher-risk optionality at modest size, and larger allocations to companies addressing unavoidable physical bottlenecks.

Ciena (CIEN): Increased as the coherent / scale-across play

Ciena has been increased from approximately 0.4% to roughly 1%. The stock rose only about 7% in the second quarter on relatively light news flow, leaving room for catch-up relative to other AI-infrastructure names.

The core thesis turns on the distinctive value of coherent technology. Coherent optics can do two things that conventional intensity-modulated optics cannot. Given the same bandwidth, it dramatically extends reach — enabling true scale-across architectures that connect AI clusters across campus distances (1–2 km) and multi-site distances (10–30 km or more). Given the same distance, it substantially improves recoverable signal quality, allowing systems to tolerate high insertion loss. The second property is becoming increasingly relevant inside the data center itself. Silicon-photonics-based optical circuit switches (OCS) in particular introduce significant optical losses; coherent technology offers a way to recover sufficient signal quality to make these high-loss fabrics practical. While large-scale commercial deployment of coherent specifically to solve SiPh-OCS insertion loss is still early, the requirement is emerging clearly as these architectures scale. In short, coherent is required both to leave the building and, over time, to help advanced optical fabrics work inside the building.

A simple mental model remains useful. Copper works only over the shortest reaches. Conventional optical transceivers extend distance but still hit limits as data rates rise, because the signal weakens and is distorted by dispersion; once the pulses are too degraded, simple intensity detection cannot recover the information. Coherent optics encode data in both the amplitude and the phase of the light wave. At the receiver a local laser mixes with the incoming signal so that advanced digital signal processing can compensate for the impairments and reconstruct a clean waveform. Ciena’s WaveLogic franchise is the clear commercial and technology leader in this domain. WaveLogic 6 Extreme remains the clear commercial leader at 1.6 Tb/s and is still the only solution in meaningful volume deployment. The company continues to push coherent capabilities closer to the compute layer through Coherent Lite and the Nubis co-packaged optics assets.

We frame the opportunity set as two distinct markets. Optical communication within the data center is becoming a red-ocean arena — crowded, well-understood, and heavily contested among short-reach pluggables, CPO, NPO and related technologies. Scale-across connectivity, by contrast, remains a blue-ocean market: fewer players have deep coherent expertise, the architectural requirement is only now becoming binding, and the market is still largely valued as a mature telecom/DCI business. Ciena’s decades of leadership in coherent systems position it to capture both the rising relevance of coherent inside the data center (as a solution to insertion loss) and the still-under-appreciated expansion of scale-across demand.

The market’s attention and capital have so far concentrated on the shorter-reach optical stack closest to the GPU. That framing is increasingly incomplete. The same physics that forced the industry from copper to short-reach optics is now forcing it from short-reach optics toward coherent once clusters become geographically distributed — and, in parallel, once high-loss optical fabrics appear inside the cluster. The increase in the position reflects that assessment.

Semtech (SMTC): Held, with a question mark on trimming

Semtech is maintained at approximately 1.5%. The company designs TIAs (transimpedance amplifiers), laser drivers and, critically, continuous-time linear equalizers (CTLE) that restore high-speed electrical signal integrity. In the optical chain these components amplify and clean the electrical signal before it enters the optical engine and after photodetection on the receive side. In that sense Semtech plays a role analogous to MACOM, but with particular relevance at the electrical equalization layer.

Amplifiers and equalizers are becoming structurally more important as data rates rise. Electrical signals rapidly approach the physical limits of noise, attenuation and frequency-dependent loss. Without high-performance equalization, even the best optical components cannot maintain clean links. The most strategic piece of Semtech’s portfolio in this context is CTLE.

Leading-edge 224G SerDes IP is tightly controlled by Nvidia and Broadcom. Tier-2 silicon designers that want to field competitive networking or accelerator chips often cannot access that IP on acceptable terms. When Google explored an inference-oriented TPU variant with MediaTek and sought to avoid heavy Broadcom royalties, the required SerDes performance proved extremely difficult to achieve. CTLE was used as a compensating technology — effectively a sophisticated analog patch — to recover enough signal quality to make the link viable. As more competitors attempt to challenge the Nvidia/Broadcom duopoly without paying for their SerDes, licensing advanced CTLE becomes one of the few practical ways to close the performance gap. Looking further ahead, when the industry moves toward 448G SerDes, even the tier-1 players will be pushing physical limits and will themselves require more aggressive equalization. Semtech’s CTLE technology is therefore relevant to second-tier players today and potentially to the leaders later.

The near-term risk remains real. Google is reportedly considering scaling back or exiting the MediaTek collaboration because the project has underperformed. Semtech was tied into that program, so any volume reduction would remove a meaningful near-term demand driver. The stock had already run roughly 100% in the quarter, which makes the risk of a pullback more acute and continues to argue for considering a partial trim.

On the financial side, consensus appears to be pricing roughly 25% durable growth. Our internal view sees potential for near-term growth closer to 50% as adoption of these high-speed analog components scales, settling into a more sustainable ~30% baseline as Semtech moves from a marginal to a more established position in the signal-integrity path.

The team remains split between taking some profit after the large run (and in light of the MediaTek risk) and holding for the longer-term structural adoption story. We are keeping the full weight for now, while remaining open to a modest trim if the near-term catalyst turns more negative or if the valuation stretches further without corresponding evidence of accelerating design-win momentum.

AXT Inc (AXTI): Modest intended exposure to the InP wafer bottleneck

AXT was identified for a small ~0.5% position as the near-monopoly supplier of indium-phosphide (InP) substrates, particularly as the industry begins migrating from 3- and 4-inch to 6-inch wafers. InP wafers remain one of the most acute bottlenecks in the optical supply chain; current supply is estimated to meet only around 30% of demand, leaving a substantial gap that is likely to persist as high-speed pluggable and co-packaged optics volumes scale.

The structural case is clear: AXT controls essentially the entire early 6-inch InP capacity and therefore sits at a critical chokepoint for lasers used in AI optical interconnects. It is both a bottleneck causer and, through its expansion plans, a partial resolver. There are effectively no other listed vehicles that give clean public-market exposure to this specific layer of the AI value chain; the remaining producers are Chinese-listed.

The risks, however, are material and multi-layered. Virtually all of AXT’s production sits in China through its Tongmei subsidiary, and China controls the majority of global indium supply. Although indium is not currently treated as a military-critical material and has not been subjected to the same export restrictions as some other critical minerals, the geopolitical exposure is real and largely hidden behind a U.S. listing. China could change its stance.

On the competitive side, Chinese producers are already manufacturing InP wafers at meaningful scale and are expanding rapidly. The competitive set is mixed: some players have upstream indium or related resource advantages and have integrated downstream into substrates, while others are dedicated compound-semiconductor or substrate specialists scaling under domestic supply-chain policies. Most of this new Chinese capacity remains concentrated on the more mature 2- and 4-inch formats. AXT retains a clearer know-how and production edge in early 6-inch crystal growth and is directing a meaningful portion of its 2026–2027 capacity expansion toward larger-diameter capability. Even so, the longer-term overhang from expanding domestic Chinese wafer capacity remains a concern and helps explain management’s relatively measured expansion posture.

While AXT is a genuine bottleneck exposure, the combination of concentrated China operational risk, rising domestic Chinese wafer competition, and elevated valuation relative to realistic 2028 revenue potential argues for only modest sizing. We therefore intended to establish a small allocation but were unable to purchase the shares, leaving the name as a monitored bottleneck exposure rather than an active holding at this stage.

Genomics: Exits and Retention

CRISPR Therapeutics has been fully exited. Although the position was added only about six months ago, the opportunity cost of holding it has risen as higher-conviction AI-infrastructure ideas compete for capital. The deeper reason is structural. AlphaFold demonstrated that large-scale compute could solve a problem — predicting a protein’s three-dimensional structure from its amino-acid sequence — that many in the traditional healthcare industry doubted could be cracked by compute alone. That advance, however, addresses the downstream “backend” of the problem. The harder and still largely unsolved “frontend” challenge is target identification: knowing which protein, pathway or cellular state is actually worth modulating in the first place. Most complex diseases remain poorly understood at the causal level; we can now predict the shape of almost any protein, yet we frequently still do not know which ones are the true drivers, whether inhibiting or activating them will help, or how the broader biological network will compensate.

The next transformative breakthroughs in this upstream domain appear more likely to come from AI-native groups that treat massive compute and foundation-model development as their primary engine, rather than from companies whose core competence remains earlier-generation wet-lab platforms such as gene editing. In the traditional model, discovery begins in the physical lab: a hypothesis is formed and then tested sequentially through cell assays, cultures and animal studies. Progress is slow and the searchable biological space is narrow. CRISPR is a powerful molecular tool for editing genes once a target has already been chosen, but it does not itself solve the harder problem of identifying which targets matter. AI-native approaches invert this sequence — large-scale models trained on biological and clinical data propose the most promising targets, and the wet lab becomes the validation step rather than the primary discovery engine. In that sense, finding the eventual genomics equivalent of OpenAI or Anthropic before the “ChatGPT moment” is the higher-upside path — and public markets may only gain meaningful access later via secondaries or IPOs. CRISPR does not clearly sit on that compute-driven inflection, and the capital has been redeployed accordingly.

Tempus AI was retained at approximately 0.8%. It survived the review because it is more deeply embedded in the AI-native side of the landscape. The company is building and post-training its own models, but its more durable advantage is the large-scale proprietary clinical and molecular data it collects through downstream partnerships. Even if Tempus is not the single pioneer that ultimately delivers the decisive computational breakthrough in causal biology, its data moat positions it to play a significant role in whatever ecosystem emerges. No extended debate was required; Tempus remains the preferred holding within the genomics and precision-medicine portion of the portfolio.

Marvell (MRVL): Monitored, no action

Marvell was discussed at length as a potential new position and as a case study in competitive dynamics the team initially under-weighted. The market has generally interpreted Nvidia’s roughly $2 billion investment as a straightforward endorsement or partnership. The strategic intent appears more sophisticated.

Marvell is one of the few meaningful alternatives to Broadcom in optical DSP. Broadcom can already leverage its DSP IP inside broader custom-silicon bundles; Nvidia historically lacked an equivalent lever. Backing Marvell gives Nvidia influence over a critical part of the DSP supply chain and helps balance Broadcom’s power in that layer.

At the same time, Marvell has been structurally weak in both custom AI accelerators and AI networking silicon relative to Nvidia and Broadcom, in large part because it lacks ready access to leading-edge 224G SerDes IP. The contrast between Marvell-designed Trainium 3 (widely viewed as disappointing) and the success of Google’s TPUv7 underscored the gap. Nvidia’s response has been to pull Marvell closer into its own orbit: rather than Marvell continuing to develop independent high-speed I/O for networking chips, the path of least resistance becomes designing compute silicon that attaches to NVLink. In effect, Marvell risks becoming a shadow custom-ASIC design partner within the Nvidia ecosystem rather than a fully independent competitor.

The same dynamic extends to scale-up networking standards. The field has largely coalesced around two camps — NVLink (Nvidia) and Scale-up Ethernet (Broadcom). UALink emerged as the alternative backed by the remaining players, with Marvell designated as the primary switch supplier. Once Marvell is financially and strategically aligned with Nvidia, the incentive to champion a competing UALink standard diminishes sharply. The investment therefore simultaneously (1) secures Nvidia greater influence over DSP supply, (2) redirects certain custom-ASIC efforts toward NVLink, and (3) weakens a potential scale-up competitor to NVLink.

We remain sceptical of treating every Nvidia partnership as an automatic positive. The game-theoretic overlay is complex enough that the team elected not to initiate a position at this stage. Marvell stays on the monitored list with a clear structural caution around the long-term economics and strategic independence of being an Nvidia ally.

• Jensen Huang’s reassurance that software companies will become “value-added resellers of AI tokens” is actually a polite death sentence, implying a 90%+ valuation compression from SaaS multiples to VAR multiples.

• With AI collapsing both the cost of building software and the cost of maintaining it, the last structural defense of “buy, don’t build” evaporates — which is why Kirkland & Ellis spending $500 million on internal AI is existential strategy, not merely a productivity initiative.

• Foundation model labs are taking over the high-margin “cerebrum” of the software value chain, pushing software vendors down to the coordination-layer “cerebellum” and toward the low-margin, relationship-heavy work that vendors themselves once pushed onto VARs and systems integrators.

• The radiology precedent suggests AI will reshape and expand the legal profession rather than destroy it: automation makes the routine work cheap, demand grows as prices fall, and human judgment becomes the scarce, premium input — but only for firms that make the transition.

• The deeper cause is the harness flywheel: model labs design their products and models together as one system, while app companies build workarounds for model limitations that the next model release makes obsolete — leaving app companies three options: post-train their own models, climb into the model layer, or accept VAR economics.

Jensen’s Polite Warning

Jensen Huang is a nice guy — so nice that he seems constitutionally incapable of saying harsh things in public. So when he addresses the future of the software industry and the SaaS depression of the past few years, his way of calming people down is subtle. More software engineers will be hired, he says. The software industry is here to stay as the plumbing layer. Software companies will become value-added resellers — VARs — of AI tokens.

Read that last sentence again. It sounds reassuring. It is actually one of the most brutal things anyone has said about the software industry, delivered with a smile. VARs and SIs trade at 10-15x P/E and 0.5-1.5x EV/S. SaaS companies have historically traded at 15-20x EV/S. If Jensen is right, he is politely announcing a 90%+ valuation compression for an entire industry. Amusingly, some SaaS companies like Wix already trade closer to VAR multiples than to classic SaaS multiples. The market may be sniffing this out before the narrative catches up.

The Kirkland Disagreement

I’ve been chewing on this since listening to Factory AI’s founder comment on Kirkland & Ellis reportedly spending $500 million to build AI internally rather than relying on legal AI startups like Harvey. His argument, roughly:

“We’re so used to a world where the moat in software was: ‘I know how to do this and you don’t, so you’re going to pay me because I have the engineers who know how to build this and you simply cannot.’ Going forward, there is going to be nothing that no one can build. Every single piece of software, in theory, anyone will be able to build. But then it comes back to resource allocation. Is it worth your time and energy to build it, or should you go to someone who has already built it? I could go pick up lunch for everyone on the team. I know how to do it. But just because I know how to do it doesn’t mean it’s an efficient use of my time. Just because you can build a lot of these things does not mean you should. If something is not relevant to your core business or core competencies, outsource it.”

The lunch analogy is clean, and it’s wrong — or rather, it misses the big picture from Kirkland’s perspective. For Kirkland, this is not lunch. This is not a resource-allocation optimization at the margin. This is a project they must execute, because the alternative is becoming the Blockbuster that gets steamrolled by Netflix. They know it’s not their specialty. They know they have a low probability of delivering a solution as polished as a dedicated startup’s. They must do it anyway, because in the world that’s coming, the AI stack is the core competency of a law firm. Getting lunch was never going to become the core of Factory’s business. Legal AI is absolutely going to become the core of Kirkland’s.

The deeper point the lunch analogy obscures: when “core competency” itself is being redefined by technology, outsourcing your future is not efficiency. It’s surrender.

The Maintain-vs-Buy Objection — and Why It’s Dying

The strongest counterargument to internal builds has never really been about building. It’s about maintaining. The old rule of thumb in IT was that for every dollar spent building software, you spend three dollars maintaining it — patching, securing, upgrading, complying, fixing the integration that broke when a vendor changed an API. Kirkland’s $500 million is not a one-time cost; under traditional economics, it’s a down payment on a permanent tax.

But this objection assumes AI can build and not maintain, which gets the difficulty ordering backwards. Building is the harder problem — it requires understanding ambiguous requirements, making architectural decisions, and creating something from nothing. Maintenance is largely boring, repetitive, well-specified work: dependency upgrades, security patches, regression fixes, compliance updates. AI agents are spectacular at boring, well-specified work. It is precisely the kind of task where agents already outperform bored human engineers who consider maintenance career poison.

So the economics don’t shift from $1-build/$3-maintain to $0.01-build/$3-maintain. They shift toward $0.01-build/$0.01-maintain. And once maintenance collapses alongside construction, the last structural defense of “buy, don’t build” collapses with it. The total cost of ownership argument that justified an entire industry’s existence quietly evaporates.

What SaaS Actually Was

To see where this goes, it helps to be honest about what the software business actually was. In the pre-AI era, software vendors occupied the most enviable position in the entire technology value chain. They had the highest operating leverage: build once, sell infinitely, at near-zero marginal cost. They built standardized products serving 90%+ of customer demand, and for the messy tail — the customer who needed the weird integration, the on-site deployment, the bespoke workflow — they had two answers: adapt to our product, or here’s the number of a VAR or systems integrator who will hire armies of labor to do the non-scalable dirty work.

Vendors kept the high margins, the scalability, and the distance from tricky customer relationships. VARs and SIs got the residual: low-margin, labor-intensive, relationship-heavy, unscalable work. It’s the same structure as Apple keeping the high-value design and brand while Foxconn runs the assembly lines. In this value chain, software vendors were the brain; VARs and SIs were the limbs.

VARification: The Brain Gets a New Tenant

The brain has a new occupant. Foundation model labs have taken the cerebrum — the center of cognition and the high-value core of the software value chain.

What remains for software vendors is the cerebellum: the coordination layer. The cerebellum matters; without it you lose balance and fine control. But it is not where thinking happens. As models improve, the cerebrum expands while the cerebellum is pushed outward — toward the limbs of integration, configuration, hand-holding, and relationship management.

This is the VARification of software. Model labs inherit what software vendors once enjoyed: standardized products, extreme operating leverage, high margins, and distance from the messy end customer. Software vendors inherit what VARs used to have: fierce competition over a commoditizing offering, deal-by-deal wins, tailored low-margin services, and heavy human touch.

There is a dark irony in the common claim that “AI won’t replace jobs involving human relationships and negotiation.” That statement is true — and it is precisely the low-margin territory that most AI application businesses (the companies building on top of the models) will be forced to occupy. Human touch is where the margin isn’t.

The China Precedent: SaaS Was a Local Maximum

Maybe it should have been this way all along. The premium on software businesses in the West arguably existed for one simple reason: software engineering talent was scarce and expensive. China shows what the equilibrium looks like when that constraint is removed.

China has essentially no enterprise SaaS industry. Most enterprise demand is addressed by internally built software stacks tailored to each company. Three reasons: China has a K-shaped company-size structure where value accrues to ultra-large state-owned enterprises and internet giants, with little room for the mid-sized companies that are SaaS’s natural customers; there is an ample supply of highly capable, comparatively cheap software engineers; and big customers demand deeply customized stacks that are ruinously unprofitable for standardized SaaS vendors to serve.

The uncomfortable conclusion: the golden era of licensed software and then SaaS was a local maximum — a monetization of a specifically American market structure with abundant mid-sized enterprises and scarce, expensive engineering talent. AI removes the scarcity. Every company on Earth is about to have access to what Chinese giants have had for a decade: effectively unlimited, low-cost engineering capacity via AI agents. The Chinese equilibrium — tailored, internal, no external software premium — may simply be the global equilibrium, arriving late to the West.

There is an important caveat. China’s model produced heavy duplication, uneven quality, and zero global enterprise software champions. The reason is fundamental: human engineering capacity — even when abundant and cheap — does not compound or standardize well. Every company reinvents similar systems, quality varies by team, and little reusable infrastructure emerges. AI agents behave differently. Once built, they compound, improve, and can be reused across many organizations at near-zero marginal cost. The AI-powered version of the Chinese model therefore keeps the deep customization that large customers demand while eliminating most of the waste and duplication that human-built internal software creates.

Professional Services: The Legal Deep Dive

Professional services — law firms, consultancies, investment banks, accounting firms — are where this collision gets most interesting, because they run on pyramid structures: many juniors, fewer mid-level managers, few senior partners.

If AI automates junior work, the naive prediction is “fewer juniors.” But outsiders consistently misunderstand what junior professionals actually do. Junior work is not just research, drafting, analysis, document review, client communication, and internal coordination. It is also the apprenticeship system through which the firm builds and transfers its knowhow to the next generation of mid-level and senior professionals. That on-the-job learning is how the firm reproduces its own expertise over time. Automating this layer therefore does more than shrink headcount — it marginalizes, and in extreme cases can destroy, the very mechanism that creates the firm’s future knowhow, while simultaneously reshaping leverage, margins, and firm economics.

For law firms specifically, trace the trajectory. Previously, software was marginal to legal work — document search, case databases. Limbs. Lawyers were the brain. A copilot-stage legal AI first lets firms keep middle and senior staff while shedding junior labor. But as legal AI passes the agent stage and climbs in expertise, middle and senior functions get absorbed too. And here’s the trap: if every law firm has access to the same legal AI, then legal service itself commoditizes, and the firm’s remaining differentiation collapses to landing deals, building trust, and maintaining relationships through unscalable human means — while the AI does the substance.

At which point the client asks the fatal question: why go through the law firm at all, rather than directly to the AI provider whose agent the firm is using? If most legal tasks can be served by a standardized software product, law firms become — there’s that word again — VARs and SIs of legal intelligence, rather than premium, high-margin, prestigious service providers. Humans stay in the loop for responsibility and judgment. But the brain-space allocation inverts: legal agents grow both the limbs and the majority of the brain, with human lawyers retaining a minority stake for accountability, subjective evaluation, and final judgment.

This is exactly why Kirkland is spending the $500 million. It’s not IT procurement. It’s a fight over which side of the VARification line they end up on.

The Radiologist Lesson

Before concluding that lawyers are doomed, consider radiology — the most instructive failed prediction in AI history. In 2016, some of the most prominent voices in AI declared radiologists obsolete. The logic seemed airtight: deep learning matched or beat humans at reading medical images. Why keep paying the human?

A decade later, radiologists are in higher demand than ever, with shortages rather than gluts. Three reasons, each mapping directly onto law.

First, the job was never just “look at image, find thing.” A radiologist integrates images with patient history, correlates across studies, communicates with referring physicians, weighs ambiguity, and signs their name to a diagnosis carrying real liability. AI automated the most legible slice of the work, not the job. Likewise, a lawyer’s value was never merely “find the clause” — it’s judgment, accountability, and contextual synthesis sitting on top.

Second, Jevons paradox: make a component of a service cheaper and total demand often expands. Cheaper imaging meant more scans ordered, meaning more reads requiring a responsible human signature. If legal work that cost $10,000 can be delivered for a few hundred, vast latent demand — small businesses, individuals, underserved markets that could never afford lawyers — becomes economically viable. The market grows.

Third, the human moved up the value chain. Do the arithmetic: if volume grows from 100 cases to 100,000, and humans handle only the final 10% of last-mile judgment, that’s 10,000 cases of genuinely human work — two orders of magnitude more than before — and each unit commands a premium precisely because it’s the scarce, non-commoditized input.

The lesson is not “the machine replaced the expert.” It’s that automation unbundles a profession into a commoditized layer and a judgment layer, collapses the price of the former, and amplifies the volume and value of the latter. Honest caveats: the transition is uneven, practitioners who refuse to move up will be displaced, and the timing of demand expansion is uncertain. But structurally, legal agents are more likely to reshape and expand the legal profession than extinguish it — for the firms that make the transition.

Three Objections, Addressed

Regulation and licensing. Unauthorized-practice-of-law rules, malpractice insurance, and bar regulations legally prevent clients from “going directly to the AI” in many cases, and could preserve law firm economics longer than technology alone suggests. Possibly. But we have reasonable faith in the US legal system’s historical willingness to accommodate innovation — and more importantly, there’s now a forcing function: China. Chinese law firms already operate at incredibly low margins with none of the premium American firms enjoy — the same dynamic as Chinese software engineering. If China moves first on AI-native legal services, competitive pressure will drag the US along, because the alternative is watching an entire professional-services export advantage erode. Regulation buys time. It doesn’t change the destination.

The apprenticeship problem. This is the big one, and I won’t pretend it away: if juniors are automated out, where do future senior partners with judgment come from? Judgment is traditionally trained through exactly the grunt work AI is eliminating. Honestly, nobody knows how this resolves. The likely answer: new juniors collaborate with agents from day one, and the career ladder bifurcates brutally — either you learn fast and jump to the rank of managing agents and exercising judgment early, or you stay in low-level tasks and get smoothed out entirely. The comfortable middle rungs of the ladder disappear. Darkly funny footnote: this is roughly what already happened to junior white-collar workers in China’s hyper-competitive labor market. The pyramid becomes an hourglass, then a diamond, then just the top.

The proprietary data moat. The fashionable rebuttal is that incumbents like Kirkland hold decades of privileged deal documents no lab can access, and this data is the real moat. Data matters. But I think this is currently the single most mispriced belief in the market. People systematically overvalue incumbent data and undervalue newcomer ingenuity. We have a live experiment: GitHub and VS Code sat on the largest repository of code and developer behavior data in human history, plus total distribution — and Claude Code, with none of that, ran straight past Copilot on the strength of a clean-slate, agent-native design. Incumbents have data; disruptors have the clean slate. History says bet the clean slate. Data is a moat only when the game stays the same; when the game changes, data about the old game is inventory, not moat.

From Seats to Outcomes: Who Bears the Risk

The endpoint of all this is that “Software as a Service” finally becomes literal — not service in the deployment sense (subscription instead of license) but in the outcome sense. You don’t buy seats of legal software; you buy resolved contracts. You don’t buy a support platform; you buy resolved tickets.

The obvious objection: outcome pricing means the provider bears the risk when outcomes fail, and enterprises are conservative about risk transfer. True — adoption will take time for exactly this reason. But we’ve run this movie before. The cloud transition was itself a massive risk transfer: infrastructure and operational responsibility moved from the customer to AWS and Azure, and customers who initially found this unthinkable (”our data, on someone else’s computers?”) eventually found it unremarkable. Outcome-based AI is the continuation of the same risk-offload continuum — higher stakes, yes, perhaps crossing thresholds today’s buyers aren’t comfortable with. But we’ll get there, because there’s no alternative: once one provider credibly prices on outcomes, seat-based competitors look like restaurants charging you for the kitchen instead of the meal.

Note what outcome pricing does to the VAR analogy: it either breaks it or completes it. The provider who successfully bears outcome risk at scale captures service-industry TAM at software-industry economics. The one who can’t is just a VAR with extra liability.

What Happens to the Actual VARs and SIs?

If software vendors become VARs, what happens to Accenture, Infosys, and the Big Four advisory arms — the actual VARs? The ironic answer: the VARs get VARified too. They face a pincer. From above, newly VARified software vendors descend into their territory — and those vendors understand harness engineering far better than traditional SIs do.

From below, AI collapses the billable-hours labor arbitrage that is the entire SI business model. To see why this is fatal, it helps to understand how these firms actually make money. The traditional SI and consulting model runs on labor arbitrage: the firm hires large numbers of relatively cheap junior and mid-level staff — often in lower-cost locations — and then bills clients at a much higher hourly or daily rate for their time. The spread between what the firm pays its people and what it charges the client is the arbitrage, and it scales with the number of billable hours the firm can put on an engagement.

AI attacks this spread directly. When agents can perform large parts of the analysis, coding, documentation, testing, and configuration work that used to demand armies of billable humans, the volume of hours that can be billed shrinks sharply. Fewer hours means a smaller base for the arbitrage, and the profitable gap between cheap labor and expensive client billing collapses. The business model doesn’t just get squeezed — its central engine loses fuel. Value-add and margins, already thin, compress further from both directions at once.

Their surviving asset is relationships — and relationships are real. But this is the incumbent-data story again: old guards holding relationships versus newcomers offering vastly better value. Newcomers win most of the time. Relationships delay the rollover; they rarely prevent it.

Why Incumbents Can’t Just Adapt: The Organizational Argument

Everything above says software companies must move up the stack — toward the model layer, or at least toward owning part of the cerebrum. Here’s why almost none of them will manage it.

Look at what’s separating winners from losers at the frontier right now. Anthropic’s surge past OpenAI, with Google and Meta lagging, traces to something unglamorous: technical founder retention and leadership continuity. Anthropic’s technical co-founders stayed, caught the early RL-scaling signals in the Sonnet 3.5/3.7 coding era, moved decisively, and delivered the curve. OpenAI lost its technical founders, and without them at the table, high-conviction technical bets got harder; the company leaned on product and go-to-market — strong levers, but secondary while frontier progress remains the main game. Meanwhile scaling’s reported death was exaggerated, and the labs that kept conviction and kept shipping lapped the doubters.

The lesson for the application layer follows directly. The winning AI software organism is a small, concentrated, visionary team with extreme talent density, willing to rewrite 80% of its codebase every six months as models evolve, where low-level technical insight gets spotted by senior leadership and becomes a company-wide priority within weeks. That organism is nearly the exact inverse of an incumbent software company. This is also why serious AI app startups must understand the model layer deeply — you cannot design a harness that fits the model, and evolve it as fast as the model evolves, from the outside. It’s why “just train your own” is becoming real advice for ambitious verticals: take SOTA open-source bases and your proprietary data and RL post-training, and out-execute closed models on your domain — because if you build something exceptional in a big vertical without model leverage, the labs will verticalize right behind you.

Can incumbents respond? Only with something like Zuckerberg’s play: a dedicated elite lab inside the larger org with complete independence and deliberately obnoxious prestige relative to everything around it. Most incumbents can’t stomach that.

And the mindset gap is visible if you know where to look. Watching Wix’s CEO on 20VC praise Figma’s product craftsmanship — rather than grappling with what Claude Code represents (not well-crafted, moving insanely fast, incredibly ahead on utility) — my honest reaction was: this is a Japanese ICE executive praising the mechanical beauty of engine design while the future of the car becomes batteries, controllers, and motors. The kaizen mindset that made Japanese manufacturers great is the same one that left them behind in the internet era and the EV era. Craftsmanship optimizes within a paradigm. We’re between paradigms. Future software winners will look temperamentally alien to current winners: hyper-agile, harness-obsessed, model-fluent, comfortable treating their own codebase as disposable.

There’s precedent for how rarely this transition succeeds. IT services firms watched the software industry’s superior margins for decades and almost never crossed over. Individual entrepreneurs made the jump — CrowdStrike’s founder went from Big Four forensics to security software — but the Big Four themselves never became software vendors. Genes, focus, and know-how don’t transfer. The same asymmetry now applies one layer up: software companies watching the model layer. They must attempt the climb, because not attempting it is the end of the valuation story. Most will fail anyway.

Scenarios and Signposts

Nobody can time this, but first-principles scenario analysis beats false precision.

In a fast-takeoff scenario — models keep scaling, agents reach senior-professional reliability within two or three years — VARification happens violently. SaaS multiples compress toward SI multiples within a market cycle, internal builds like Kirkland’s proliferate across every industry, and the only defensible positions are frontier labs, compute, and the handful of app companies with genuine model-layer leverage.

In a capability-plateau scenario — models stall at today’s agentic reliability — software vendors get extra years, not a pardon. Even current models make building and maintaining software cheap enough to undermine the case for buying it. Enterprise caution slows the shift, but the destination is the same.

The signposts to watch:

• Net revenue retention and seat counts at big SaaS vendors — seat-based models are the canary.

• Frequency of Kirkland-style nine-figure internal builds announced by non-tech enterprises.

• The first credible large-scale outcome-priced contracts.

• Open-source model proximity to the frontier — this determines whether labs or harness-builders capture value.

• SI revenue-per-employee — the first number to break when VARification cascades downward.

The Investor Takeaway

If this thesis is even directionally right, the sorting rule is simple: ask whether the cerebrum can absorb your function or merely needs it. Companies whose product is work the model can increasingly do itself — workflow orchestration, data entry and retrieval, the connective tissue between intent and execution — are being demoted to cerebellum status: necessary, commoditized, low-margin. Companies the cerebrum depends on but cannot replicate — compute beneath it, proprietary data loops beside it — capture value. Switching costs are the second-order adjustment: they set the timeline, not the outcome — high enough friction buys a doomed vendor years of bond-like cash flows, but it never moves them into the winning bucket.

Most exposed: horizontal SaaS whose product is a workflow wrapper around data entry and retrieval — tools where “an agent could do the workflow” is a sentence that parses. Seat-based pricing anywhere is a slow leak, because agents don’t buy seats. Mid-tier point solutions with weak systems-of-record status get VARified first.

Temporarily protected: true systems of record — core ERP, core banking, EHR — where compliance, data gravity, and catastrophic switching risk buy real time. But protected is not the same as growing; these become bond-like cash flows, and the market will eventually price them that way.

Potentially escaping the trap: the small set of AI-native companies with genuine model-layer leverage — those doing serious post-training on proprietary interaction data, whose harness co-evolves with frontier models, and whose usage generates data that improves their models in a real loop. The test is not “uses AI” (everyone does) but “would a frontier lab verticalizing into this space start from behind?” Very few companies pass.

Clear beneficiaries: the labs themselves, compute and the infrastructure layer beneath them, and — the contrarian one — the consumers of software. Every enterprise on Earth is about to get its software bill structurally repriced downward. The value doesn’t vanish; it transfers to customers and to the cerebrum.

Position accordingly: long the brain, long what feeds it — compute and the infrastructure stack beneath the labs — long the rare app-layer companies with genuine model leverage (proprietary data, real post-training and RL loops that frontier labs can’t replicate from a standing start), long the balance sheets that buy software, and extremely skeptical of anything in between still trading like it’s 2021.The software industry isn’t dying. It’s being demoted. Jensen already told us — he was just too nice to say it plainly.

Postscript: The Harness Flywheel — Why the Agent Era Still Belongs to the LLM

A recent observation from Moonshot AI’s founder crystallizes the mechanism behind everything above — not just that model labs eat the application layer, but why the outcome is structurally predetermined.

The mechanism is a forward loop versus a backward loop. When a model lab builds an AI product, it designs the harness first, then trains the next-generation model to fit it. Harness and model co-evolve by design; every training run tightens the fit. A pure AI product company runs the same process in reverse: it receives a finished model as a black box, probes it to discover its limits, and constructs a harness around what it infers. The lab knows the model’s true capability frontier because it drew that frontier; the app company is forever reverse-engineering it from the outside. One process is design, the other is archaeology — and the archaeologist is permanently behind, because the gap is in information access, not talent or effort. It compounds, too: because the lab designed the harness and trained the model into it, it understands the model’s limits better than anyone, which lets it evolve the harness better than anyone, which informs the next training run. The loop feeds itself.

This is exactly why Anthropic rewriting 80% of its harness code every six months is not a symptom of chaos but the flywheel working as intended. Models evolve, so the harness must evolve — and only the party training the model knows which direction it will evolve before it ships. It also explains Claude’s otherwise puzzling market position. There are two ways a model can reason. The first is monologue reasoning — one long, uninterrupted chain of thought before answering, the test-time-scaling paradigm other labs optimized for: a mathematician at a desk, deriving everything before touching anything. The second is agentic reasoning — short bursts of thought interleaved with action: think, run the code, read the error, revise, check, think again. A mechanic under the car. The two are not just different styles; they are trained differently. Monologue reasoning can be trained on static problems with verifiable answers — no harness required. Agentic reasoning can only be elicited, measured, and reinforced inside an environment where the model acts and observes consequences — inside a harness. The capability doesn’t exist as a training target until you’ve built the loop. Claude was never the champion of raw monologue reasoning; Anthropic bet on the agentic kind instead — a bet that was only available to an organization building harness and model as one system. Anthropic didn’t build a model and find agents inside it; it built for the agentic loop and trained the model into it. The forward loop, made visible.

Extend the ladder and the next rung comes into focus. Language models, chain-of-thought, and agents are consecutive rungs — knowledge, then reasoning, then action. But even the best agent today is amnesiac: it can act, but it cannot accumulate. Every session starts from the same frozen weights, which is why no agent yet replaces the employee whose real value is six months of absorbed context, not raw intelligence. The employee learns into their brain; the agent “learns” into a scratchpad. The next rung is closing that gap: persistent learning, perhaps eventually recursive self-learning — merging the monologue and agentic paradigms and, critically, writing what is learned into the model’s weights rather than bolting it on as external memory. Note what that implies for the application layer. External memory, retrieval scaffolds, and context management are precisely the workarounds app companies sell today as their value-add. If the frontier’s next move is to internalize learning, today’s memory architecture is tomorrow’s swallowed feature. Multimodality, search, and video generation matter for products, but they are pluggable components; the tasks that decide whether intelligence keeps advancing remain training, learning, and the model’s own capability. The cerebrum, again.

None of this means pure-agent startups can’t make money today. They can. Enterprises genuinely need permissions, audit trails, data access, and industry workflows, and decent software companies can be built on those needs. But their durable value comes from customer relationships, proprietary data, domain experience, and control over real-world work — the word “agent” itself confers no moat. Ingenious orchestration is, by definition, compensation for model weakness, and model weakness is a depreciating asset. As models improve, vast amounts of today’s intricate orchestration get swallowed by a single model call, and many agent products that look brilliant right now will turn out to have had very short lifespans.

The formulation that closes this essay is simple: the harness decides what the model attempts; the model decides what it achieves. Orchestration routes intelligence — it cannot add it. When the agent era truly arrives, the winner will still be the LLM.

Which brings everything full circle. VARification is the industry-wide result; the flywheel is the cause, operating one company at a time. Every AI app company’s product is, underneath everything, a bundle of workarounds for things the model can’t yet do. The lab knows exactly which of those workarounds its next model will make unnecessary — it’s training that model right now — and the app company finds out on release day. Each release deletes a chunk of someone’s product. Run that squeeze across a thousand companies and the sum is VARification: the flywheel is not an abstraction beneath the thesis, it is the thesis, happening one product at a time. For AI application companies, the strategic menu collapses to the same three options identified earlier: post-train your own models, migrate up into the model layer outright, or accept — with clear eyes — a future of relationship-based, low-margin, VAR economics. And for enterprises like Kirkland, the flywheel is the final justification for the $500 million: if fit between harness and intelligence is where value lives, then owning your own loop — your data, your post-training, your evolving harness — is the only position that isn’t someone else’s flywheel.

For thirty years, software sold picks and shovels of the mind — because minds weren’t for sale. Now they are, by the token. Everything else is distribution.

• Building on the Part 1 rotation into under-appreciated bottleneck resolvers, Part 2 deepens exposure to power-electronics names positioned for the 800 V DC and solid-state transformer transition, while classifying several pure or heavy SiC players as higher-risk and currently unappealing.

• Photonics conviction has been sharpened by fully exiting Lumentum and concentrating capital in Tower Semiconductor (hybrid-bonding bottleneck) and MACOM (terrestrial high-speed optics plus emerging space optical links).

• Within the expanding neocloud and compute basket, the team is emphasizing asymmetric upside (IREN, Oracle, Bitdeer) and strongest execution (CoreWeave), while deliberately passing on more fully valued or capacity-constrained names such as Nebius.

• Software and thematic positions were refined with an increased allocation to Palantir as the marquee AI-software compounder, a preference for Circle over Coinbase within the crypto basket, and the retention of MP Materials as a long-duration physical-AI rare-earth expression.

• Portfolio spring-cleaning continued with the full exit of Quanta Services and a reduction in Tesla, freeing capital for higher-conviction AI-infrastructure and space-related opportunities while accepting elevated volatility in pure story stocks such as Intuitive Machines.

Some of our readers already know that Convequity has been managing a dedicated portfolio since 1 January 2024 (co-managing it with an institutional client). You can follow the live positions and performance on the Convequity site.

Two weeks ago the team completed an in-depth rebalancing of the portfolio. We reviewed every holding, tested the underlying theses against the technological and market developments of the first half of 2026, and made deliberate adjustments to positioning. The full analysis is published for Premium subscribers in two parts; this Substack note covers the first section of Part 2.

Our long-term structural conviction in the AI buildout remains unchanged. What continues to evolve is where the most attractive risk/reward now sits inside the value chain. After taking profits from some of the strongest earlier winners, we have been rotating capital toward less obvious but higher-conviction bottlenecks and enabling technologies. Notably, our AI Bubble Barometer had already been signaling healthy fundamentals and limited hype across the hyperscaler and neocloud complex even before the sharp pullback in AI infrastructure stocks over the past two to three weeks.

Source: Convequity

The live portfolio dashboard (shown above) reflects the current state of that process: a diversified set of 80 holdings across 11 sectors, with Data Centre still the largest allocation and meaningful exposure also maintained in Semiconductors, Consumer Applications, Space Tech, and Energy & Power. Performance since inception remains strong, even as shorter-term volatility has increased.

In the pages that follow we share concise takes on the names discussed in this portion of the rebalance — beginning with the power-electronics opportunity around 800 V architectures and solid-state transformers, and moving into the evolving photonics landscape. For each company we explain the updated investment case and the portfolio decision that followed.

The complete Part 2 report, including the full set of positions and the team’s detailed reasoning, is available exclusively to Convequity Premium subscribers.

Enphase Energy (ENPH) & SolarEdge (SEDG)

At Convequity we are treating Enphase and SolarEdge as a paired power-electronics exposure within the AI data center value chain. Both companies are being re-underwritten around the same structural shift: the move toward 800V DC architectures and megawatt-scale racks makes traditional copper-heavy power distribution impractical, pushing the industry toward higher voltages and semiconductor-based solid-state transformers (SST). Enphase’s established expertise in bidirectional power conversion originated in its storage systems that must move electricity efficiently in both directions while meeting rigorous safety and grid-interconnection standards. In April 2026, Enphase expanded explicitly into AI data centers with the launch of its IQ solid-state transformer, following SolarEdge into the space. This is an earlier commercial-scale product operating at lower voltages, not a full utility medium-voltage (13 kV) SST. This move signals that both companies see their core power-conversion capabilities as relevant beyond residential and microgrid applications, with SolarEdge viewed as having made the initial move earlier.

The investment case centers on how AI data center power architecture is expected to evolve over the next several years. Utility power typically arrives at medium voltage around 13 kV and must be converted down to the 800 V DC levels increasingly favored by high-density AI racks for efficiency and reduced copper losses. A true next-generation solid-state transformer would perform this conversion directly in a single, compact, and highly efficient unit. Because these SSTs are designed to be bidirectional, they must also support the reverse flow — pushing power from the data center’s 800 V DC systems back up to 13 kV for grid export, on-site storage arbitrage, peak shaving, or providing ancillary services such as frequency regulation and backup capacity.

Data center loads add a further layer of complexity compared with EVs: they must absorb extreme “rocket shifts” in power as large clusters of GPUs cycle on and off almost instantaneously, while maintaining fail-safe circuit protection where a single breaker failure could destroy an entire rack worth millions or billions. This combination of requirements plays to the design know-how of Enphase and SolarEdge. Enphase does not currently offer a commercial medium-voltage SST, and widespread adoption of 800V DC rack architectures remains early. The meaningful technical and commercial inflection for direct 13 kV-to-800 V DC capability is still viewed as roughly two years away, creating a window in which companies with proven bidirectional, grid-tied power electronics experience can build relevant positioning ahead of broader recognition.

A secondary but supportive dynamic is the parallel evolution of the broader grid. Smart-grid technologies — two-way digital communication, sensors, and real-time automation — combined with residential microgrids that allow households to generate and store their own electricity, can free up grid capacity for AI data centers. In a more advanced form, aggregated household generation and storage could even provide peak-shaving services back to large AI loads. Both Enphase and SolarEdge already sit at the center of this residential bidirectional ecosystem, giving them an additional avenue to participate in the system-level flexibility that AI power demand will increasingly require.

From a portfolio perspective, this is a capability bet rather than a near-term revenue story. Following the recent correction, Enphase trades at what we consider attractive multiples of approximately 14x gross profit and 38x free cash flow, with SolarEdge also appearing attractively valued on a similar post-correction basis. The team assessed that the AI data center power angle is only partly appreciated by the market, with both names showing chart behavior reminiscent of other holdings that were “just about getting noticed” at comparable stages. SolarEdge is viewed as marginally ahead in timing, while Enphase is catching up. Conviction is constructive but tempered by the reminder that 800 V DC and SST narratives carry reversal risk similar to other high-profile AI infrastructure themes, with mainstream adoption still likely to arrive later. We are therefore initiating a modest paired position of around 1% across the two names as complementary exposure within the power-electronics segment of the AI value chain. Both companies bring relevant grid-interfacing and bidirectional conversion expertise, and we see the current valuation and notice level as offering asymmetric potential for a small allocation at this stage of the rebalancing.

This framing also highlights a broader dynamic: as AI campuses scale to hundreds of megawatts, operators will increasingly value flexible assets that can both consume and support the grid, an environment in which proven bidirectional power electronics capabilities become structurally more important.

Navitas Semiconductor (NVTS)

At Convequity we view Navitas Semiconductor as the highest-beta, standalone pure-play expression of the gallium nitride opportunity tied to the 800V DC transition in AI data centers. AI clusters are driving unprecedented power demand, with individual racks and campuses now requiring hundreds of kilowatts to multiple megawatts. Delivering this level of power at traditional lower voltages would require extremely thick copper conductors, which quickly becomes impractical due to cost, weight, space, and thermal constraints. Because power equals voltage multiplied by current, increasing voltage allows the same amount of power to be delivered with significantly lower current, enabling thinner conductors, reduced losses, and more efficient distribution. This is why the industry is moving toward higher-voltage architectures such as 800V DC. However, operating at higher voltages and the faster switching frequencies needed for compact, efficient power conversion exceeds the practical limits of traditional silicon semiconductors in terms of voltage handling, switching speed, and thermal performance.

This is where wide-bandgap materials become essential. Gallium nitride (GaN) excels at high-frequency switching with very low losses, making it particularly well suited for the fast, efficient conversion stages required in 800V-class systems. Silicon carbide (SiC), by contrast, offers superior voltage blocking capability, higher temperature tolerance, and greater robustness, making it better suited for higher-voltage or more rugged sections of the power chain. Both materials significantly outperform silicon when power conversion must occur at elevated voltages and frequencies, enabling the smaller, more efficient solid-state transformers and power supplies that next-generation AI data centers will require. Navitas has deliberately pivoted from the consumer charger market — where it faced intensifying competition from low-cost Chinese suppliers — toward this higher-margin AI data center opportunity, where it is increasingly viewed as the default independent GaN supplier, ahead of even larger players on the GaN side of the technology.

To strengthen its offering, Navitas acquired a SiC-specialist team approximately two years ago, giving it credible capability in more sophisticated, non-commoditized silicon carbide devices. While some observers remain skeptical about its ability to deliver world-class SiC performance at scale, the team sees this as a meaningful step toward offering a more complete power semiconductor platform. As a small, focused company making a concentrated bet on the data center power transition, Navitas carries elevated execution risk. At the same time, the upside is highly leveraged to the success of the broader 800V DC and solid-state transformer theme.

From a portfolio construction standpoint, Navitas represents the most volatile and concentrated way to express the power electronics angle of AI infrastructure within our holdings. The position currently stands at approximately 1.47% and has already begun to validate the underlying thesis as the stock has responded positively to growing recognition of the data center opportunity. The team expects the runway ahead to support materially higher highs, though it also acknowledges the elevated volatility inherent in a small-cap pure-play making this kind of strategic shift. We are therefore maintaining the current position size while remaining inclined to add on strength, treating Navitas as the dedicated GaN vehicle alongside our paired exposure in Enphase and SolarEdge.

Infineon Technologies (IFX)

We view Infineon as the most mature and best-prepared incumbent supplier in the emerging 800V DC power ecosystem for AI data centers. When Nvidia first proposed shifting to 800V DC architectures last year, Infineon moved decisively while several peers remained skeptical or slow to respond. This early commitment has translated into a significant advantage: Infineon now offers by far the broadest catalog of qualified and fully characterized components. As a result, it has become the most widely designed-in supplier for solid-state transformers, as system designers can reliably model its parts’ performance across a wide range of real-world operating conditions.

Infineon’s leadership is asymmetric across materials. It holds a strong position in silicon carbide, where it leads the market and where Navitas remains the challenger. In gallium nitride, however, it trails Navitas. While the team considered adding to the position, the key limitation is structural: Infineon is a large, diversified semiconductor company rather than a focused pure-play. This diversification dilutes the upside relative to more concentrated bets on the 800V DC theme. In addition, even under optimistic assumptions for a full 800V DC ramp, the team’s analysis suggests it would not be sufficient to meaningfully absorb the industry’s currently idle SiC and GaN capacity, which sits at roughly 30% utilization due to prior overbuilding in China. This tempers the narrative that a broad 800V transition will quickly resolve the current downturn in wide-bandgap semiconductors.

From a portfolio perspective, we continue to respect Infineon as the most technically ready and broadly adopted player in the 800V DC supply chain. However, its scale and diversification result in more muted risk/reward characteristics compared with purer expressions of the theme. As a result, we are not adding to the position this quarter and continue to favor Navitas for cleaner, higher-beta exposure to gallium nitride within the power electronics opportunity.

Power Semiconductor “Unpredictable Basket”: STMicroelectronics (STM), onsemi (ON), and Wolfspeed (WOLF)

Within the broader power semiconductor ecosystem supporting the 800V DC transition and AI data center infrastructure, STMicroelectronics (STM) OnSemiconductor (ON), and Wolfspeed (WOLF) sit in what the team characterizes as the “unpredictable basket.” These names offer optionality on wide-bandgap (SiC and emerging vertical GaN) technologies but carry elevated execution, cyclical, and geopolitical risks that currently outweigh near-term conviction. The team highlighted a preference for staying on the sidelines, citing a potential thaw in US–China trade relations combined with significant Western idle SiC capacity, which could create unpredictable supply/demand dynamics for pure or heavy SiC players.

STMicroelectronics is viewed primarily as a component supplier for SSTs, with the added angle of a SpaceX partnership. However, the read is lukewarm: a large European integrated device manufacturer (IDM) perceived as less innovative and slower than Infineon in embracing the 800V DC shift. While it maintains a solid SiC portfolio (including Gen4 devices for 400/800V automotive and industrial applications), the combination of geopolitical overhang and the broader SiC capacity glut places it firmly in the unpredictable category. No action is contemplated at present.

onsemi is treated as a higher-risk speculative optionality play centered on its vertical GaN bet — an attempt to extend GaN efficiency advantages into voltage ranges (>1 kV) traditionally dominated by SiC. The company, we described as feeling “behind in everything” and under pressure from Chinese EV competition, has been discussing vertical GaN for roughly two quarters. Yet the team notes zero product, only slides, with no revenue or shippable devices to date. Compounding the caution is onsemi’s still-significant cyclical exposure as a ~$6 billion revenue company with the majority of its business tied to the volatile EV/auto market. It too lands in the unpredictable basket; the team views it as too early-stage and too auto-levered to warrant a position.

Wolfspeed is framed even more cautiously as a highly speculative situation. The bull case rests on its ambition to develop medium-voltage SiC supporting very high voltage ranges (reportedly up to ~20 kV), positioning it slightly ahead of onsemi in high-voltage SiC with a potential “back sheet” technology. However, the bear case is stark: prior management effectively gambled and lost, the underlying technology remains a major challenge (large, temperature-sensitive dies requiring substantial cooling), and any meaningful product is still approximately two years from market. Current valuation appears dominated by speculators and traders. Even allowing for a potential long-term “Oklo-style” re-rating narrative, the team sees the timing as wrong and places Wolfspeed in the unpredictable basket alongside the other two.

Bottom line: While all three names provide exposure to the SiC/GaN power conversion tailwinds relevant to AI infrastructure and 800V architectures, the combination of execution shortfalls, cyclical leverage, geopolitical uncertainty around idle Western capacity, and (in Wolfspeed’s case) legacy restructuring scars leads the team to classify them as higher-risk and currently unappealing. They are monitored for meaningful de-risking on products or macro clarity but are not part of the active power-electronics basket at this stage.

Monolithic Power Systems (MPWR): On-Board Power Management in the AI Server Stack

Monolithic Power Systems occupies the most downstream position in the AI data center power conversion flow among the names discussed. While NVTS, Wolfspeed, STMicroelectronics, and onsemi focus on upstream wide-bandgap devices and modules — delivering high-voltage conversion from the 800V DC bus or grid-level inputs down to intermediate voltages — MPWR supplies the final-stage voltage regulators and power-management ICs that sit directly on the server and GPU boards. As racks transition from legacy 48V DC distribution toward 800V DC architectures, this on-board layer becomes increasingly critical for precise point-of-load regulation (the final stage of power conversion in electronic systems), efficiency optimization, and board-level safety circuit-breaking. MPWR is therefore a complementary rather than competing part of the same ecosystem.

Its differentiation stems from a fabless model that has proven advantageous amid overbuilt analog capacity. This agility has allowed MPWR to consistently lead in supporting Nvidia’s aggressive power roadmaps (Hopper through Blackwell to Rubin), where larger IDMs have been slower to respond. The result is strong design-win momentum on the most power-hungry and expensive GPU boards, translating into rising content per rack even as overall industry growth moderates.

Valuation remains elevated at 40x EV/GP and 105x EV/FCF. The stock is up approximately 50% year-to-date but has been essentially flat over the past three months, reflecting a lack of fresh catalysts beyond broad AI infrastructure momentum. While the team views consensus growth assumptions as somewhat conservative, MPWR is now treated more as a “beta” play that rides industry expansion rather than delivering outsized share gains. The current stance is therefore to maintain the existing ~0.49% position without adding.

Tesla (TSLA): Near-Term De-emphasis in Favor of SpaceX

The team has reduced the Tesla position from 1.81% to 1% and is de-emphasizing it in the near term. Attention and more tangible, hard-numbers catalysts are shifting toward SpaceX’s potential IPO and earnings trajectory, while Tesla continues to face a still-weak EV demand environment. The reduction reflects a deliberate rotation of conviction toward the higher-visibility SpaceX opportunity rather than a outright negative view on Tesla’s long-term optionality.

The remaining exposure is justified by several high-conviction, longer-dated catalysts. These include the potential for an eventual Tesla/SpaceX merger, the scaled deployment of Optimus, and — most relevant to the AI infrastructure theme — Tesla’s custom AI silicon roadmap. While the team remains skeptical of the original Dojo project (whose core design team was reportedly disbanded, with a redesigned “Dojo D2” still approximately two years away and prior bullish calls such as Morgan Stanley’s “$1 trillion Dojo” viewed as unrealistic), we are more constructive on Tesla’s “Terafab” ambitions and the AI5/AI6 chips targeted at terrestrial demand. These could evolve into a meaningful revenue stream alongside Optimus.

A notable data point supporting the silicon thesis is that Elon’s companies have reportedly secured roughly 20% of Nvidia’s Vera Rubin supply, equating to around two million chips. Jensen Huang has publicly highlighted Elon Musk as one of the best-positioned leaders to extract maximum value from TSMC’s constrained advanced-node capacity. This underpins the decision to maintain a reduced but still meaningful stake rather than exit entirely.

Implication: Tesla is trimmed to 1% with conviction rotated toward SpaceX in the near term, while preserving exposure to the longer-term AI silicon and robotics optionality.

Lumentum vs Tower: Shifting conviction in photonics

Over the past two quarters, Tower Semiconductor has gradually overtaken Lumentum as our main photonics holding. We have now exited Lumentum entirely and redeployed the capital into Tower.

Lumentum is being reduced because it modestly lagged the portfolio despite the broader CPO rally. Structural pressures are building: high-power CW laser demand remains soft as CPO is not yet ready for volume deployment, traditional EML and CW laser share is shifting toward more aggressive Chinese suppliers, and the migration toward silicon photonics structurally reduces the need for indium phosphide content. While Lumentum retains the strongest laser technology and could continue to participate, the magnitude of upside now looks more modest as the name becomes better understood by the market.

Tower, by contrast, sits at the actual bottleneck of the CPO transition: high-yield hybrid bonding of photonic ICs to electronic ICs. While TSMC leads on the electronic side, it has historically struggled with photonic integration and bonding — precisely Tower’s decades-long strength from high-yield CMOS image sensor to memory buffer bonding. As the industry confronts real CPO yield challenges, control of the hybrid-bonding step is becoming a decisive advantage.

Tower also benefits from the silicon-photonics transition, which reduces heavy indium-phosphide consumption, and from displaced CPO players exiting foundries that lack hybrid-bonding capability. In addition, Tower is already space-qualified with proven radiation heritage, positioning it better for emerging space optics opportunities.

Tower now carries our highest conviction in photonics, with multiple converging drivers across CPO and silicon photonics. It is the photonics name we see carrying the most upside into the late 2020s.

MACOM (MTSI): Increasing conviction in space-enabled photonics

MACOM is being increased as a high-conviction name at the intersection of terrestrial data center optics and emerging space-based optical communications. The company supplies critical analog components — including transimpedance amplifiers (TIAs), laser drivers, and detectors — that are essential for scaling optical bandwidth in both environments.

In terrestrial AI data centers, MACOM plays a key role in enabling the transition to higher-speed optical links. Its TIAs and modulator drivers are critical for supporting the move to 1.6T and 3.2T connectivity, whether in pluggable modules or co-packaged optics (CPO) architectures. As data rates per lane continue to increase, the performance requirements on these analog components become more stringent, and MACOM has established a leading position with early availability of 400G+ per lane solutions.

The space opportunity adds a differentiated and potentially higher-upside dimension. Space-based optical systems are expected to follow a faster bandwidth scaling trajectory than terrestrial networks, moving more rapidly toward 6.4T and 12.8T per link. This creates earlier and more substantial demand for high-performance InP-based analog components. Additionally, inter-orbit link architectures — which involve longer distances and more demanding link budgets — are likely to require higher levels of MACOM content per terminal and support higher average selling prices compared to shorter intra-cluster links.

MACOM is well positioned to capture this opportunity due to its existing space qualification heritage and vertically integrated indium phosphide capability. The company already supplies space-qualified analog components for free-space optical programs, and its ability to deliver both high reliability and leading-edge performance is particularly valuable in space, where hardware cannot be easily maintained or replaced.

While MACOM has already performed strongly, the combination of its critical role in terrestrial high-speed optics and its favorable positioning for space bandwidth scaling and inter-orbit architectures supports a higher weighting in the portfolio. The position is being increased as part of the ongoing reallocation within the photonics complex.

Closing Note

This excerpt covers the first section of our Portfolio Rebalance Part 2, focusing on the power-electronics opportunity around 800 V architectures and the evolving photonics landscape.

The full Part 2 continues with detailed analysis on SiTime, ASM International, BE Semiconductor, Palantir, the crypto basket (Coinbase, Circle, Robinhood), the neocloud complex (IREN, CoreWeave, Oracle, Bitdeer and others), Intuitive Machines, Alphabet, MP Materials, and more.

Both Part 1 and the complete Part 2 are available exclusively to Convequity Premium subscribers. Part 3 will follow shortly.

You can access the full reports and the live portfolio here: convequity.com

SpaceX — now SPCX following its record June 2026 Nasdaq IPO and the merger with xAI — is no longer a launch company with a satellite side hustle. It is the ultimate AI data-center and energy play, and the valuation debate reflects it: priced near $1.75 trillion at IPO, trading toward $2.4 trillion in its first weeks, with a credible path that Musk himself frames as a 10x. That sounds absurd until you work through the physics, the engineering, and the cost models. So let’s do exactly that.

This is a condensation of our six-part deep-dive on SpaceX and the burgeoning space economy. The full series — with detailed valuation models, the interactive DCF, and in-depth supply chain and technology deep-dives — is available to Convequity subscribers. What follows is the argument in full force.

Why Terrestrial Solar Hits a Wall, and Space Solar Doesn’t

Start with the bottleneck everyone in AI is screaming about: power. Natural gas turbines are being deployed at speed, but turbine-blade shortages have created multi-year backlogs. At the same time, battery storage is being added at scale to renewables to capture excess midday solar generation and discharge it during nighttime and cloudy troughs when output collapses. With peak US power capacity around 1,000 GW running at roughly 50% average utilization, storing surplus energy and releasing it when needed can roughly double the amount of usable, dispatchable power available in the near term. Useful — but still nowhere near enough for the scale of demand now emerging.

Rethink it from first principles and the picture clarifies. Even China — the most aggressive nuclear investor on Earth — will derive less than 20% of its electricity from nuclear. The overwhelming majority of new supply will come from solar and wind. Yet terrestrial solar faces hard physical limits: the atmosphere absorbs enormous quantities of high-energy radiation, and the day-night cycle imposes an unavoidable intermittency penalty.

The efficiency gap is stark. On Earth, dominant monocrystalline silicon cells operate around 21%, while even the best commercial heterojunction (HJT) technology has only reached a practical ceiling of ~24.7% — as demonstrated in Risen Energy’s record module. In space, advanced gallium arsenide (GaAs) multi-junction cells achieve 30–35%+ efficiency under AM0 conditions by stacking multiple junctions that capture a much broader portion of the solar spectrum. The main barrier has always been their significantly higher manufacturing cost compared with conventional silicon technology.

The Musk convergence — a 10x advantage. Move panels to orbit and the generation side alone delivers ~5x more output per panel: no day-night cycle, no atmospheric absorption, and multi-junction architectures capturing a broader spectrum. Layer on the elimination of terrestrial storage (continuous orbital sunlight means no batteries, no overbuild) and you save roughly another 2x at the system level. Compound them: space solar is ~10x more effective than terrestrial on a per-watt-delivered basis.

If China’s solar cost is ~$0.25/W, then excluding launch, effective space solar is ~$0.025/W or lower. The economics are transformative — if the launch problem can be solved.

And here is the elegance of it: the entire value chain now runs through Musk’s companies. Starship is methane-fueled, adding incremental natural gas demand that reinforces our structural thesis on US gas producers (EXE, AR, CRK, EQT). More powerfully, cheaper space solar makes large-scale orbital compute viable, which will generate substantial revenue for SpaceX through xAI and other frontier AI labs. That revenue accelerates Starship development and further cost reduction, which in turn makes orbital power even more economical — creating a self-reinforcing loop with Starship and Raptor V3 at its center.

Starship: a generational leap. SpaceX already crushed LEO launch costs from $25,000/kg to ~$2,500/kg via Falcon 9. Starship goes further — both stages fully reusable, powered by Raptor, the first full-flow staged combustion (FFSC) engine ever built. In FFSC, all of the fuel and oxidizer are pre-burned into gas in dedicated turbopumps before reaching the main chamber. This allows more complete combustion of the propellant, delivering higher thrust, better reliability, and true full reusability across both stages. Starship is the most efficient, most performant rocket humanity has ever built: 200+ tons to LEO, targeting below $100/kg.

For context: China’s total tonnage to space in 2025 was ~300 tons; global was ~3,000 tons — of which SpaceX delivered 2,500, including 2,100 tons of Starlink alone.

The killer-app problem. Starship is so powerful it risks being economically unjustified without a use case. SpaceX faced this exact problem after Falcon 9; the answer was Starlink (30,000 satellites at 550 km, soon direct-to-cell). One Falcon 9 carries ~25 Starlink sats; one Starship carries ~500. If Starship succeeds, SpaceX runs out of things to launch. It’s the iPhone before the App Store — and the killer app is now obvious: space-based AI data centers powered by space solar. Because a single leader drives it across vertically integrated companies, it progresses faster and more coherently than nuclear, wind, or gas.

Why SpaceX needs China — the Tesla precedent. Recall Tesla in 2017: the Model 3 was a product triumph but a production nightmare. American high-volume manufacturing carries a brutal ratio — building a line costs ~5x the annual revenue it generates. China broke the deadlock. Giga Shanghai went from groundbreaking to customer delivery in under a year.

SpaceX faces the same production scaling challenge with solar. To power orbital data centers at meaningful scale, SpaceX needs solar manufacturing capacity at rates the U.S. industrial base cannot deliver alone. The Feb 2026 merger — SpaceX at $1T, xAI at $250B — gives Musk 50%+ control across both companies. With unified authority, he can now pursue the same China manufacturing strategy that rescued Tesla, importing or assembling high-efficiency solar panels at the volume required. If 1 GW is already difficult on Earth, 10 GW and 100 GW are near-impossible without it. xAI has the demand; SpaceX now has both the need and the control to solve it.

The technology roadmap. GaAs multi-junction cells will not be used for mass-scale space solar — current space-grade pricing of $50–100/W makes it uneconomic at volume. The realistic path runs through heterojunction (HJT) silicon in the near term and perovskite-on-HJT tandems later.

HJT silicon already offers ~25–26% efficiency at a fraction of GaAs cost, bringing the target per-watt economics discussed earlier within reach at scale. Perovskite tandems could push efficiency above 30% with dramatically better energy-to-mass performance. Durability under space conditions remains the main open question, with in-orbit validation still early.

Only two Chinese firms currently produce HJT at commercial scale: Risen Energy and Huasun. For Chinese manufacturers, scaling durable HJT is now an engineering and capital problem rather than a fundamental scientific one. Musk-linked teams visited several Chinese solar suppliers earlier in 2026. Given ITAR restrictions on direct imports of finished space-grade panels, the most likely route is importing Chinese manufacturing equipment (such as from Suzhou Maxwell and others) to build capacity domestically, potentially through licensing or joint ventures.

The irony writes itself: China builds chip fabs on Western lithography; the US builds solar fabs on Chinese equipment. The question is not whether this works — it’s when and how. The hurdles are political and commercial, not technical. (Recall it was regulation, not physics, that delayed Starship for years under Biden.)

What an Orbital Data Center Actually Looks Like

This is not a gradual extension of today’s satellite industry. SpaceX has filed with the FCC to launch over 1.2 million satellites — more than every satellite humanity has ever orbited, combined. They form a three-tier constellation functioning as a single distributed compute fabric:

• VLEO (500 km) — 800,000 satellites. The edge layer. ~1,000 TOPS each, ~20 ms latency to ground. Real-time inference.

• LEO (1,000 km) — 300,000 satellites. The regional layer. ~5,000 TOPS each. The bulk of training.

• MEO (2,000 km) — 118,000 satellites. The cloud layer. ~20,000 TOPS each. Long-running workloads and storage.

Each Starlink V3-derived satellite weighs ~2 tons, and nearly half — ~850 kg — is the compute payload: radiation-hardened silicon, liquid cooling loops, high-efficiency solar power conversion. These aren’t comms satellites with compute bolted on. They’re flying data centers, with the bus built around the silicon. Compute is a space-rated Dojo 2 variant, 12–24 chips per satellite, stitched together by inter-satellite laser links at 100+ Gbps per link — and in space there’s no practical channel limit, so aggregate throughput climbs into the terabits.

The launch math is what makes it tractable: at 2 tons each, one Starship delivers 50–75 satellites per flight (leaving margin for deployment systems and operational conservatism). On legacy launch systems, the deployment timeline would exceed the satellites’ useful life before the constellation could be completed. With Starship, the full constellation could be deployed in roughly nine years — reaching 10% by year three, 50% by year six, and 100% by year nine.

The critical milestone is 2028: Musk’s deadline for Tesla and SpaceX to each deliver 100 GW of annual solar-cell capacity. Even with 100% acceleration, the fastest HJT reaches commercial maturity in ~two years. A 40 MW data center in space could cost as low as $8.2M versus $167M on Earth — the latter saddled with 2–3 years of supply-chain bottlenecks.

The feedback loop. Tesla and SpaceX converge into a unified industrial ecosystem: SpaceX unlocks limitless energy, xAI unlocks intelligence, Optimus executes physically, Tesla chips supply compute. The missing piece was semiconductor fabrication — now being addressed by Tesla’s Terafab plans unveiled in March 2026. Close this loop and you reach an age of abundance, an economy measured in output rather than currency. A lunar base is almost certainly necessary intermediate infrastructure — no atmosphere, one-sixth gravity, regolith rich in silicon and aluminum, with electromagnetic mass drivers eventually launching satellites without rockets. SpaceX has accordingly shifted near-term priorities away from Mars toward xAI’s space DC buildout.

The solar cost trajectory:

• Today’s Chinese cells, lifted to orbit at no cost: ~$0.025/W delivered (a conservative upper bound assuming no further cell improvements).

• Mature HJT cells (~$0.06/W on Earth, ultra-thin and rollable) with a 10x space multiplier: ~$0.006/W delivered.

• Perovskite tandems have the potential to go significantly lower still.

At a 100 GW/year buildout rate, annual capital expenditure would be roughly $25 billion using today’s cells, ~$6 billion with mature HJT, and as low as ~$600 million with mature perovskite — a more than 40x spread. The solar cell roadmap is not a side detail. It is the central economic lever.

Cooling — the counterintuitive part. Space is cold, but a vacuum has no air or fluid to carry heat away, so chips in orbit cannot cool through convection the way they do on Earth. Terrestrial data centers pay dearly for the air they take for granted — a 1 GW AI cluster draws ~627 MW for the silicon and another ~373 MW (nearly 40%) for cooling, power conversion, and networking. In space the physics inverts: heat can only be rejected by radiating it into the ~3 K background of the universe. The catch is that radiation moves heat far less effectively than convection, so orbital cooling requires large surface area. The architecture is elegant: each satellite is a flat panel, with the sun-facing side generating power and the shaded side radiating waste heat into space. The same geometry that makes generation efficient also makes thermal rejection efficient. Starship lifts power source, radiator, and structure in a single piece of hardware.

Launch frequency. Musk predicts 300–500+ GW/year of space compute deployment within ~five years, with a theoretical 1 TW/year ceiling. 100 GW requires ~5,000–11,000 Starship launches annually — up to one per hour. Each launch consumes ~1,000 tonnes of methane; even at aggressive cadence, gas and LOX demand stays low-single-digit percentages of US production. A lunar factory with mass drivers could ultimately deliver up to 1 PW/year. Tesla’s AI5 (Q2 2027) and AI6 (Q2 2028) chips align with this; AI6 is the intended core chip for orbital DCs, and Dojo 3 has been rebuilt around AI6 clusters in a standard NPU style after earlier custom designs proved unscalable.

Communication — fiber optics without the fiber. Optical Inter-Satellite Links (OISLs) are the nervous system. Starlink already runs 9,000+ laser terminals at up to 200 Gbps — the largest coherent optical network ever built. The 1.2M-satellite filing implies 4–8 terminals each, 5–10 million units over nine years — a number that shatters every existing aerospace optical supply chain and forces photonics toward semiconductor-class volume economics.

The hard part was never the lasers — it was the aim. Two satellites can have relative velocities of up to 15 km/s. Even a tightly collimated laser beam still spreads to roughly 20 meters wide after 1,000 km, while the receiver remains a sub-microradian target (an extremely precise angular aiming requirement smaller than one millionth of a radian). Fast-steering mirrors adjust >1,000x/second with predictive software aiming where the target will be. This Pointing, Acquisition, and Tracking (PAT) is why OISLs stayed a lab curiosity for forty years. SpaceX’s contribution was industrial, not scientific: building PAT cheaply enough, in large enough numbers, to actually work.

The zero-dispersion advantage shapes the entire supply chain. In vacuum, all wavelengths travel at exactly the same speed, eliminating chromatic dispersion and pulse smearing. Terrestrial fiber requires expensive coherent transceivers to compensate for these effects, whereas space-based links can rely on simple, low-power direct detection. Most inter-satellite links operate over short distances in tight clusters, allowing the majority to use cheap electro-absorption modulated lasers with basic receivers. Only the minority of long-distance vertical links require more sophisticated coherent gear. The implication is significant: shifting incremental AI data centers to space does not automatically expand the photonic market — in some segments it may reduce demand for high-end terrestrial components. We’ll explore this in more detail later.

Lasers offer structural advantages that make space a superior medium for data movement. They support far higher bandwidth than traditional radio systems while also outperforming terrestrial fiber in key respects. Because vacuum has no dispersion or nonlinear effects, high data rates can be achieved with simpler and lower-power components than those required in fiber. Lasers are also inherently more secure: even after natural beam divergence over distance, the spot size remains narrow enough that interception or jamming requires precise positioning within the beam path. In addition, optical spectrum is unlicensed, removing regulatory barriers to dense deployment. Most importantly, light travels through vacuum roughly 47% faster than through fiber, meaningfully reducing latency between compute nodes. In this sense, orbital data centers benefit not only from location but from a physically superior interconnect fabric.

Cluster Geometry: What Most Analysts Get Wrong

Most analyses assume orbital data centers would function as a loose, widely dispersed constellation spread across a large orbital shell. In reality, the architecture is built around tight orbital clusters. Each cluster operates as a self-contained data center, with satellites flying in formation within a relatively small volume — roughly a 100 × 100 mile box. Just as you would not build a large GPU training cluster with nodes scattered across different cities, you would not spread compute satellites thinly across an entire orbital plane.

This has major implications for link distances. Because satellites within the same orbital tier fly in tight formation, most connections between them are relatively short — typically 20–160 km. Unlike traditional satellite constellations, there are no long-distance backbone links spanning thousands of kilometers within the same orbital layer. Instead, the longest connections are the vertical links between different orbital tiers: from VLEO to LEO (roughly 500 km), LEO to MEO (roughly 1,000 km), and VLEO to MEO (up to around 1,500 km), with actual path lengths reaching 500–2,000 km depending on the angle.

The supply chain consequence follows directly: the vast majority of optical terminals by unit count are short-range, low-power, direct-detection links that can use cheap EML transmitters and integrated silicon-photonic receivers. Expensive, high-power coherent terminals — which dominate traditional space laser economics — are only required for the minority of long-distance vertical links between tiers.

A VLEO satellite typically carries 5–7 terminals. It has 1–2 Earth-facing RF phased arrays for the revenue-generating link to ground, since optical links are unreliable through clouds and atmospheric turbulence. It also carries 1–2 higher-cost optical terminals for vertical links to higher orbital tiers, which require the most sophisticated pointing, acquisition, and tracking. The remaining 2–3 terminals are ultra-low-cost optical links for intra-cluster communication, where satellites fly in close formation with near-zero relative velocity, allowing 100 Gbps links with just a few milliwatts of power.

LEO clusters, which handle the bulk of AI training, require 4–6 intra-cluster terminals per satellite to support 400G–1.6T aggregate bandwidth for all-to-all gradient synchronization. MEO, which hosts longer-running workloads and storage, carries the most expensive vertical terminals — typically $5,000–15,000 each, compared to just $50–200 for intra-cluster links.

By unit count, intra-cluster links (20–160 km) in LEO clusters make up roughly 60–70% of all terminals and are the lowest cost. Vertical cross-tier links (500–2,000 km) in MEO account for 15–25% and are the most expensive. Earth-facing RF terminals, used only on VLEO satellites, represent 10–15% and belong to an entirely different technology stack.

The volume-shift framework. The common mistake is viewing the space DC optical market as purely additive. The opposite mistake is assuming it will wholesale replace terrestrial infrastructure. In reality, existing data centers represent trillions in sunk capital and will continue operating. The real substitution occurs at the margin of growth.

Hyperscaler IT load is projected to surge roughly 6x by 2035, requiring over 100 GW of new capacity. Much of this faces severe terrestrial constraints — grid queues, permitting, water, and land availability. Space data centers act as a relief valve. GPUs that would have been racked in Texas will instead sit on satellites, and the optical interconnects that would have used fiber transceivers can instead use free-space laser terminals.

For core photonics components — lasers, modulators, detectors, drivers, and TIAs — much of the demand is therefore migrated rather than newly created. The investment thesis rests on three secondary effects: space-qualification pricing premiums, shifts in product mix, and genuinely new demand categories.

Three new demand categories emerge with no terrestrial equivalent. PAT (Pointing, Acquisition, and Tracking) systems have no ground-based counterpart, creating an opportunity for companies like STM. Free-space optics hardware — including telescopes, coatings, and filters — replaces simple fiber ferrules with far more complex assemblies, benefiting players such as Coherent. Inter-orbit links also represent entirely new demand, as there is no equivalent in terrestrial networks.

At the same time, one category faces structural decline: fiber. With vacuum as the propagation medium, Corning’s hyperscale data center fiber business encounters a long-term headwind. Another category shrinks: coherent DSP complexity. As chromatic dispersion disappears in space, demand for advanced DSP ASICs from companies like Marvell and Broadcom is reduced.

The bandwidth demands of intra-cluster links are dictated by the compute layer, not the transmission medium. Whether a connection spans 2 meters of fiber or 80 km of vacuum, the GPUs still require the same high-speed, low-latency interconnects for all-to-all gradient synchronization. As a result, the intra-cluster optical roadmap in space closely follows the terrestrial trajectory: 1.6T per link expected around 2027–28, scaling to 6.4T by 2030–32, and 12.8T thereafter.

At speeds scaling toward 6.4T and 12.8T per link, the choice of optical technology becomes important. These high aggregate bandwidths can be achieved either by using many lower-speed lanes or by pushing much higher data rates per lane.

At 100 Gbps per lane, silicon photonics combined with advanced CMOS drivers and integrated TIAs is generally sufficient. This is good news for cost, because the highest-volume part of the constellation — the short intra-cluster links — can likely rely on silicon-based solutions. In this regime, MACOM’s InP-based laser drivers and TIAs are less essential, as silicon CMOS can meet the performance requirements.

However, at much higher speeds such as 800 Gbps per lane, the technical demands change significantly. Driving lasers and amplifying received signals at these extreme data rates requires components that can deliver high voltage swings, maintain linearity, and operate efficiently at very high frequencies. Silicon CMOS approaches its physical limits in this regime, while InP-based components — where MACOM has a strong position — offer substantially better speed, linearity, and power efficiency. On a satellite, where power and heat are tightly constrained, this makes InP laser drivers and TIAs a practical necessity for the highest-speed links rather than just a performance upgrade.

At 6.4T per link (using eight 800G lanes), each Space DC terminal requires 16 MACOM InP analog ICs, representing roughly $130–320 of content. At 12.8T (using sixteen 800G lanes), this doubles to 32 ICs and $260–640 per terminal — 5–20x more than the $15–40 typically found in a current 800G terrestrial transceiver. This means MACOM stands to generate significantly higher revenue per terminal in the Space DC architecture than in conventional terrestrial deployments.

This dynamic makes MACOM unique within the volume-shift framework. While most photonics companies face flat or declining content as AI infrastructure migrates from terrestrial to space, MACOM’s content per link increases with each bandwidth generation. Higher speeds drive greater reliance on its InP technology, causing revenue per terminal to scale superlinearly with bandwidth.

This advantage is reinforced by developments on Earth. As the terrestrial industry shifts toward co-packaged optics (CPO), demand for discrete InP laser drivers declines because the driver function is integrated into silicon. However, high-performance TIAs remain difficult to integrate at very high speeds and are needed in greater numbers due to higher bandwidth density per switch ASIC. At the same time, the Space DC opportunity creates a high-content, somewhat captive market: because satellites are severely power-constrained, InP’s efficiency advantage becomes a hard requirement, making substitution with silicon impractical. As a result, MACOM’s space business is expected to more than offset any decline in terrestrial laser driver content, while also benefiting from growing TIA demand on Earth.

The Optical Supply Chain: Winners, Shifts, and Newcomers

Current near-term winners (2026–2029) are those already flying radiation-hardened, vacuum-qualified components. SpaceX doesn’t issue open RFPs and wait — heritage is paramount.

MACOM (MTSI) — Top pick. The only company benefiting from all three dynamics: near-term heritage (TIAs and drivers in Mynaric CONDOR and classified programs), superlinear content growth to 6.4T–12.8T, and relevance across both intra-cluster and inter-orbit terminals regardless of who designs the terminal. Its vertically integrated InP fab in Lowell makes it one of only two or three firms globally that can supply III-V analog at 200 Gbaud/lane. Current market pricing doesn’t fully reflect future space revenue.

STMicroelectronics (STM) — Hold, reframed as a PAT and power-management play. Every inter-satellite terminal needs a 1,000+ Hz PAT control loop. STM’s 28nm FD-SOI is the perfect fit — radiation tolerance without full rad-hard cost, automotive-grade pricing for million-unit constellations. STM loses the analog driver/TIA sockets to MACOM at higher speeds, but the PAT socket is permanent — a chip that doesn’t exist on Earth and can’t be displaced by any photonic-integration trend.

Coherent (COHR) — Hold. Free-space optics and inter-orbit amplification. Genuinely incremental: precision optics (telescopes, coatings, filters) across all terminals, plus EDFAs for the longest links. End-to-end vertical integration (crystal growth → polishing → coating → assembly) is unmatched. But the EDFA opportunity is narrower than it appears — most intra-cluster links don’t need amplification. Real near-term wins, not transformative at corporate scale.

Hamamatsu (6965.T) — Buy. InGaAs APDs for weak-signal links (inter-orbit relays, demanding intra-cluster links at 6.4T+). 70%+ market share from decades of III-V detector R&D competitors can’t copy. High content per terminal ($30–80 at 8–16 channels), but the smallest addressable terminal count here. Its NKT Photonics acquisition makes it a credible second EDFA source. Size accordingly.

Lumentum (LITE) — Optionality. Supplies the EML at the heart of every intra-cluster terminal — structurally protected because silicon cannot emit light. But this is largely shifted volume. Benefits from space pricing premiums and inter-orbit demand; the space DC alone doesn’t justify a re-rating.

Tower Semiconductor (TSEM) — Buy. SiPh PIC integration platform, winning on radiation tolerance and heritage. Volume largely shifted, but value per PIC rises (absorbing PAT detector arrays and control logic) and the disappearance of coherent DSP complexity improves yield and margin.

Corning (GLW) — Structural loser. Fiber demand declines as compute migrates from fiber-intensive terrestrial DCs to zero-fiber orbital ones. Not a short — 5G and FTTH provide other drivers — but the exponential-AI-fiber narrative weakens materially.

Inter-orbit communication — the one truly new market. Unlike intra-cluster and vertical links, inter-orbit communication has no terrestrial equivalent and therefore generates purely incremental demand for the photonic supply chain.

In a terrestrial data center, the edge, training, and storage tiers sit close together and are connected by fiber over short distances. In the Space DC, these tiers sit at different orbital altitudes, separated by hundreds to thousands of kilometers. Rather than attempting direct long-distance laser links between tiers, the architecture uses Starlink satellites as relay nodes. A VLEO satellite links to a nearby Starlink satellite, which routes traffic through the mesh to another Starlink satellite near the target tier, keeping each individual hop under roughly 1,000 km.

This relay architecture creates demand for additional high-performance optical terminals on Starlink satellites that did not previously exist. A next-generation relay satellite may need 6–8 terminals operating at 1.6T–6.4T each to handle inter-tier traffic such as gradient updates, model weights, and inference routing. If the Space DC generates 100 Pbps of inter-tier traffic, roughly 15,000 relay terminals would need to operate simultaneously, implying 50,000–100,000 deployed terminals across the constellation. At $2,000–10,000 of photonic content per terminal, this represents a $100 million–$1 billion cumulative opportunity — moderate in absolute scale but 100% incremental, with non-displaceable content for every major supplier.

The newcomer thesis — and this is the part to internalize. The companies best positioned for 2030–2035 may not be today’s space-laser incumbents. A Mynaric CONDOR Mk3 delivers roughly 100 Gbps in a 30 kg package costing around $500,000 — the wrong product for a space data center that needs 6.4T–12.8T in a sub-5 kg terminal at well under $5,000. Meanwhile, terrestrial AI is already solving the core problem at data-center scale and economics. Companies such as Ayar Labs, Celestial AI, Lightmatter, and Ranovus are building high-bandwidth photonic interconnects optimized for the cost, power, and density requirements of large AI clusters. If compute moves to orbit, their technology can move with it — the main additions required are radiation tolerance and a free-space optical front end with pointing, acquisition, and tracking. The architectural DNA flows from the data center to space, not the other way around. Incumbents are well placed to capture the early phase when volumes are low and heritage matters, but the market is likely to transition toward high-volume photonic integration players as scale increases. Do not extrapolate incumbents’ early heritage advantage into a durable long-term moat.

Tower Semiconductor: The SiPh Foundry for the Space DC

Tower operates what is arguably the most important foundry platform in silicon photonics. Its PH18 process has become the de facto manufacturing standard, with Ayar Labs, Lightmatter, and Ranovus all building their ecosystems around it. The lock-in is structural: unlike digital CMOS, where logic can be resynthesized onto a new process node, photonics has no abstraction layer. Every optical element’s performance is defined by precise physical geometry, so switching foundries requires a full redesign from scratch.

GlobalFoundries positioned its 300mm Fotonix platform as a competitive alternative, citing 2.25x more die per wafer. However, the ramp has been slow, the PDK remains thinner, and the wafer-size advantage is inherently weaker in photonics. Waveguides cannot be scaled down like transistors because light has a fixed wavelength of roughly 1–2 μm. Advanced lithography improves dimensional control, but it delivers only incremental gains rather than the step-function economics seen in digital nodes.

At OFC 2026, Tower effectively closed the competitive argument by announcing a SiGe BiCMOS + Silicon Photonics platform on full 300mm wafers. Tower now matches GlobalFoundries on wafer format while retaining clear advantages in ecosystem maturity, customer lock-in, and process heritage. In the space context, the gap is even wider: Tower has published radiation performance data (total ionizing dose, single-event latchup, and proton irradiation), while GlobalFoundries has published essentially none for Fotonix. GlobalFoundries should be removed from the Space DC supply chain thesis.

Tower’s new SBC18H6 node delivers Ft/Fmax of 325–450 GHz, enabling stable 112 Gbaud PAM4 — the threshold required for 1.6T and 3.2T links. Critically, it reverses silicon photonics’ traditional power disadvantage. Where SiPho previously consumed over 30% more power than InP EMLs, the new platform is now approximately 30% more power efficient. The SiGe die is hybrid-bonded directly atop the photonic die in 3D, with low-parasitic interconnects that narrow the integration gap with monolithic InP — not by matching InP’s material advantages, but through superior packaging.

The manufacturing economics are transformative: 300mm silicon wafers, yields above 85% (versus 40–60% on InP specialty nodes limited to 100mm wafers), and roughly one-third the cost of equivalent InP solutions. This represents a structural cost reset rather than an incremental improvement.

Tower controls an estimated 80%+ share of high-end SiGe BiCMOS capacity and has already contracted the majority of its output to Broadcom and Marvell through 2028. It is investing over $900 million to expand capacity by approximately 5x. With SiPho penetration at 800G expected to exceed 85% by 2028, and 400G-per-lane becoming a requirement for SpaceX’s large-scale deployment in 2028–2030, Tower’s combination of high frequency performance, low power, high integration, low cost, and mass-production scalability positions it as the default architecture. InP EML remains short on frequency and high on cost, while TFLN remains unproven for space environments. Tower is the silicon photonics foundry for the Space DC.

The Financials: Where Valuation Meets Physics

SPCX fundamentals. 2025 revenue: $18.7B, up 33%. Starlink contributed $11.4B (61%) and is the profit engine. Adjusted EBITDA: $6.6B; GAAP net loss: $4.9B (driven by aggressive AI, Starship, and orbital capex). At ~$2.4T market cap against ~$30B NTM revenue, the implied forward P/S is ~80x — elevated, but the core business already generates substantial adjusted profit. The IPO wasn’t needed to fund operations; it raised $75B as dry powder for the capex ahead. Segments: Starlink $11.4B (61%), Launch $4.1B (22%), AI $3.2B (17%) — the AI segment absorbed ~$12.7B in capex.

ROI of the terrestrial DC (TDC) business. Q1 2026 AI capex hit $7.7B — a $30.9B annualized run rate. xAI’s infrastructure team executes exceptionally (first 60k H100s in ~120 days), though model training lags — Grok sits firmly tier-1 but a step behind the tier-0 frontier set by Mythos, partly because xAI’s high-pressure, rapid-iteration culture is less optimal for frontier research than the deliberate cultures at Anthropic.

While SPCX front-loaded capacity, most hyperscalers underbuilt — creating excess SPCX capacity Musk monetized via three-year deals with Google ($30B) and Anthropic ($45B): $75B contracted against ~$20B capex, or $25.8B/year. Anthropic’s ARR exploded from $9B to $40B by Q1 2026; Google sold 2M TPUs to Anthropic before realizing its own teams lacked compute (Gemini now at 9M MAU). The Convequity AI Bubble Barometer shows hyperscaler forward ROIC on AI capex at ~32%. H100 hourly rates have spiked again this year — remarkable for a three-year-old chip.

The root cause is an industry-wide bottleneck: TSMC limiting logic and CoWoS (mild); DRAM/HBM/NAND (large and growing); EML/InP lasers (large); and most importantly power — 40 of 100 AI DC projects cancelled for lack of secured power. Regulatory friction is the biggest hurdle, then power access. Gas can’t serve 10 GW → 100 GW; ground solar technically can but tariffs block panel imports. This is precisely why space DC (SDC) becomes the bottleneck-solver: no regulatory friction, falling payload costs.

Space DC economics. With fully reusable Starship V3: 100 GW needs ~10,000 launches at ~$5M each = $50B, or $500M/GW for launch. Solar at aggressive $0.20/W adds ~$200M/GW.

We model first-gen AI1 on a Starlink V3 chassis: 150 kW, NVL72-like rack, non-rack components ~$500–700k/satellite. At 0.15 MW/satellite, that’s 6,667 satellites/GW (using the $500k lower end) → ~$3.3B/GW non-AI-chip cost. Against ~$20B/GW for terrestrial non-IT plus $1B+ annual opex, these look remarkably cheap.

Under conservative assumptions, IT hardware accounts for roughly 80% of the total cost per GW in a space data center, compared to about 60% in a terrestrial facility. This shift occurs largely because the architecture relies on merchant NVIDIA GPUs, which carry very high gross margins of 75–80%. As a result, the computing hardware itself becomes the dominant cost driver.

If Tesla successfully delivers its own space-optimized ASICs on schedule at an estimated $10,000 per kW ($10 billion per GW), the overall cost per GW of space data center capacity would fall significantly — roughly halving to around $16.5 billion per GW.

To cut further you need your own fab — memory makers now run ~80% gross margins on HBM, the largest cost component in high-end accelerators. A Rubin-class GPU’s 288 GB of HBM4 at $55/GB = $15,840 in memory alone, dwarfing the $1,500–2,000 compute dies. Vertical integration could bring an AI ASIC to $3,000–4,000 per chip.

With a custom fab halving chip cost ($5B/GW), total 1 GW cost drops to $11.6B.

TerraFabs (announced March 2026) represent Musk’s push to vertically integrate advanced chip manufacturing at unprecedented scale. The long-term target is 1 TW of annual production — more than all existing global fab capacity combined, with roughly 80% intended for space deployment. Even under conservative assumptions, the initial phase could support on the order of 100 GW of compute, aligning with SpaceX’s broader deployment plans.

This is classic Musk: set an extraordinarily ambitious goal where even partial success creates enormous value. The strategic intent is to collapse the traditional fabless → foundry → OSAT cycle into a single, tightly integrated operation, enabling much faster iteration. Reports suggest the program may draw on process IP or technology from Intel, which is reportedly cash-strapped and seeking both a major customer and a way to offset development costs. Musk has reportedly passed on Samsung’s leading-edge node, possibly because he already has access to it at highly favorable terms.

Execution challenges remain significant. The program relies on unproven innovations such as 450mm square wafers and large-scale deployment of High-NA EUV tools, both of which involve long lead times and technical risk. Memory supply is another constraint, as major players have limited incentive to license advanced processes. Despite these hurdles, the core thesis is that even a meaningfully scaled TerraFab capability would structurally lower the cost and accelerate the deployment of custom space ASICs — a critical enabler for the broader Space DC economics.

Valuation & Modelling the SDC Business. The pre-IPO SOTP pointed to ~$1.75T; the IPO priced there. The market now implies $2.0–2.4T.

A 1 GW SDC site approaches $20B/year revenue (versus ~$10–15B for a 1 GW TDC, because SDC over-provisions power by just 5–7%, supporting 900 MW+ of compute versus ~660 MW terrestrially). Baseline: 7-year life, 5% revenue decay for three years then 15%, $20B first-year revenue, $33B/GW capex (conservative — ignoring up-to-10x solar cost cuts), $250/kg payload (the high bar), 10% WACC.

This yields ~$50B in added market cap per GW deployed, cross-checking against a 50% IRR — not unreasonable given TDC’s 30%+ ROIC. Upside levers: selling models (not just raw compute) if xAI reaches SOTA, and custom ASICs, which lift the figure to $64B/GW.

The trillion-dollar question: how many GW per year? Launch ceases to be the constraint after Year 4 (2029, possibly 2028). The next bottleneck is solar manufacturing.

A combined 200 GW of solar-panel capacity from Tesla and SPCX is clearly a frothy number, given how hard it is to manufacture solar cells in the U.S. at low cost and on time. Direct access to Chinese equipment, production lines, engineers, and know-how would make a 2028 timeline far more achievable — but whether Beijing permits it is an open question. Even assuming SPCX achieves just 10% of 200 GW — 20 GW per year — that still implies nearly $1 trillion in market cap added annually.

Orbital real estate adds a modest constraint: AI1’s dawn-dusk sun-synchronous shell (600–800 km) supports only ~1 GW given 4 km minimum separation, so SPCX will need additional orbits — some with slightly lower solar efficiency but higher density.

Path to $20 Trillion. Musk has indicated that SpaceX can deliver a 10x return. Only substantial success in the Space Data Center (SDC) business can realistically achieve that scale. Reports suggest Musk has already secured roughly 20% of NVIDIA’s upcoming Rubin production — around 1 million GPUs — underscoring the seriousness of the bet.

Reaching a $20 trillion market cap at a 15x P/E would require approximately $1.33 trillion in annual profit. Assuming 30% operating margins on the SDC business, this implies roughly $4.43 trillion in revenue. At $20 billion of revenue per GW of deployed capacity, around 222 GW of space compute would be needed. This figure could be lower if high-margin applications (such as Cursor) are layered on top of the raw infrastructure, increasing average revenue per GW.

If both Tesla and SpaceX can each deploy on the order of 100 GW per year, this level of capacity becomes achievable within a little over two years under optimistic assumptions.

An alternative, more conservative framing assumes that Starlink, Launch, and Cursor together are worth around $2 trillion. In this scenario, the SDC business would need to deliver nearly the entire remaining $18 trillion of value on its own. Using a more conservative revenue assumption of roughly $11 billion per GW — effectively treating SDC as pure infrastructure without significant contribution from application-layer margins — the required deployed capacity rises to approximately 360 GW, or about 36 GW per year over a decade.

While ambitious, this remains plausible when viewed alongside other strategic levers, including custom silicon, TerraFabs, in-house solar manufacturing, and AI model economics. The path to $20 trillion appears credible even if only some of these factors deliver meaningful results.

Cursor. SPCX agreed to acquire Cursor for $60B — a bargain even before synergies (revenue heading to $6–10B by year-end). Written off by many after Claude Code overtook it, the Cursor team proved highly agile, post-training open-source models (DeepSeek, Kimi) into its own Composer — now holding some of the best coding-model RL know-how outside the major labs (Composer 2.5 on Kimi 2.5 is the signal). The bigger prize: accelerating Grok to catch Claude Code and Codex on both foundation capability and agent harness.

The DCF. A standard DCF misses the nuance, so we built a Special View (Base/Bear/Bull scenarios feeding a revenue path) linkable to the DCF View — available in the SpaceX Valuation Model for subscribers.

The valuation range for SPCX is extraordinarily wide and depends almost entirely on execution in the Space Data Center business. Under the base case — assuming around 60 GW of annual solar deployment within a decade and solid long-term margins — the model points to very substantial upside from current levels, with 2040 revenue reaching several trillion dollars.

In a more optimistic scenario with significantly faster solar scale-up (approaching 200 GW within ten years), the valuation becomes extremely large. While the precise output relies on a chain of aggressive assumptions, the asymmetry highlights the magnitude of the opportunity if SpaceX executes at the upper end of its ambitions.

Even under more conservative assumptions — slower Starship cadence and more limited SDC adoption — the model still shows meaningful upside, suggesting the investment case remains attractive even with tempered expectations.

Current sell-side consensus implies roughly $160 billion in revenue by 2028. After backing out Starlink, Launch, and Cursor (estimated at around $30 billion combined), the Space DC business would need to contribute only about $130 billion — equivalent to roughly 6 GW of deployed capacity at prevailing revenue-per-GW assumptions. This level appears achievable even through terrestrial data center deployments alone in the early years.

The Bottom Line

SPCX is the ultimate AI data-center and energy play, with optionality in coding agents and foundation models on top. Its sustainable long-term profitability hinges on one unique capability: delivering AI infrastructure while everyone else faces mounting bottlenecks in regulatory approval, power, and chips.

There is no doubt they can deliver Starship V3 and successors. There is no doubt they can make space DC satellites work. The binding constraint is ultimately semiconductor manufacturing and design — Musk will need to go deep into logic and memory foundry, 224G SerDes and beyond, or partner with Broadcom or Nvidia.

The AI energy crisis is not merely a constraint on data-center growth — it is a forcing function for the next energy revolution. Terrestrial solutions, nuclear or renewable, face fundamental limits of scale, timeline, or geography. Space solar, enabled by Starship’s dramatic cost reduction, offers a path to genuine energy abundance.

Within a decade, the largest AI training clusters may operate in orbit: powered by sunlight unfiltered by atmosphere, cooled by the infinite heat sink of space, connected via laser links, unconstrained by the finite capacity of terrestrial grids. The infrastructure we build in the next five years will determine the possibilities of the decades that follow.

This is a condensation of Convequity’s six-part series on SpaceX and the space economy. The full series — covering the complete space solar roadmap, the orbital data-center architecture, the cluster-geometry and optical-supply-chain frameworks, the component-by-component winners and losers, the Tower Semiconductor deep-dive, and the full financial analysis — is available to subscribers, along with the interactive SPCX valuation model (Special View + DCF) and our AI Bubble Barometer at barometer.convequity.com

If you found the depth here worth your time, the models are where the real work lives. Subscribe for the complete series and the live valuation framework at convequity.com

The market still largely prices SPCX as a satellite broadband and launch business. That lens is becoming outdated.

What is taking shape is a vertically integrated platform capable of delivering frontier AI infrastructure at a structural cost and speed that terrestrial players cannot easily replicate. The advantage is not incremental. It stems from differences in power economics, regulatory friction, and the ability to compress the traditional semiconductor supply chain.

The Economics of Orbital Compute

A single GW of terrestrial AI data center capacity typically supports around 660 MW of usable compute servers once power and cooling overhead are accounted for. In space, that ratio improves significantly. With only 5–7% over-provisioning required, a 1 GW Space Data Center (SDC) site can support over 900 MW of compute.

Our modeling suggests that under conservative assumptions, a 1 GW SDC deployment generates roughly $20–21 billion in first-year revenue. Against a fully loaded cost of approximately $33 billion per GW (including launch, solar, structure, and merchant GPUs), the NPV of one GW of deployed capacity is in the region of $50 billion in incremental enterprise value. Even after applying a 50% IRR threshold, these economics remain attractive relative to terrestrial AI infrastructure returns.

These figures improve further if SPCX shifts from merchant GPUs to custom ASICs and captures more of the stack through TerraFabs. In that scenario, cost per GW can fall meaningfully, lifting value per GW deployed.

Solar Is the Real Scale Lever — and China Changes the Math

The binding constraint on rapid SDC expansion is not launch capacity over the medium term. It is solar panel production at the required scale and cost.

Our base case assumes SPCX reaches 60 GW of annual solar production within a decade. The more aggressive internal target of 200 GW combined between Tesla and SPCX is ambitious, but the physics and manufacturing reality point in one direction: China holds the overwhelming majority of global solar manufacturing capacity, know-how, and supply chain depth.

SPCX does not necessarily need to manufacture solar cells domestically at scale. It can import finished high-efficiency panels (particularly HJT) and perform final assembly and integration in the U.S. or a third country — an approach SpaceX already uses for certain components at Starbase. This route materially de-risks the timeline. Even capturing just 10% of the 200 GW target (20 GW per year) would still imply nearly $1 trillion in incremental market value annually under current modeling.

The dynamic with Chinese solar manufacturers is therefore not a peripheral detail. It is central to whether SPCX can move from single-digit GW deployments to the tens of GW per year required to make the upper end of the valuation case credible.

The Path to Much Larger Outcomes

Starlink and the launch business provide a solid foundation. They are not, however, sufficient to justify current valuations on a standalone basis, nor do they explain the upper range of possible outcomes.

For SPCX to deliver returns consistent with the more bullish scenarios embedded in the current price, the Space Data Center business must scale aggressively. At $20 billion of revenue per GW, roughly 222 GW of deployed capacity would be required to support a $20 trillion market cap (assuming a 15x terminal P/E and 30% margins on SDC operations). If applications layered on top of the raw infrastructure — such as coding agents — can lift revenue per GW further, the required capacity falls.

This is a high bar, but it is not obviously impossible if Starship achieves high cadence and solar supply can be secured at scale. The more relevant question is what happens if only some of the key variables (launch rate, solar ramp, custom silicon, model performance) land in the upper half of expectations. Even partial success across multiple levers would still represent transformative value creation.

The Range of Outcomes Remains Wide

Our modeling shows a broad distribution. In a conservative scenario with slower execution, the DCF still implies several times upside. In the base case, the intrinsic value is substantially higher. The upper scenarios are larger still.

This spread is not an artifact. It reflects the binary nature of the core variables: Starship operational cadence, the speed at which orbital compute can be deployed, and the ability to secure solar and silicon at the required cost and volume. The market is currently pricing something closer to a high-growth satellite and launch business. It has not yet fully priced the possibility that SPCX becomes the lowest-cost provider of frontier AI infrastructure at planetary scale.

Read the full analysis

The points above are drawn from our detailed modeling of SPCX’s cost structure, launch assumptions, solar dynamics, and valuation scenarios.

You can access the full report, including the interactive valuation model (Base / Bear / Bull scenarios) and the underlying per-GW economics, here:

→ Read the full SpaceX report on Convequity

We also maintain ongoing updates on key execution variables for subscribers.

Summary

• The AI Bubble Barometer provides a transparent, data-driven framework that compares forward ROIC with market valuations (EV/IC) to assess whether AI-related investments by the major hyperscalers are supported by fundamentals.

• In Week 4, EV/IC eased further to 6.60x while the group’s forward ROIC remains at 3.07x, keeping valuations within the fair value range with no bubble signal.

• Large-scale compute contracts between SpaceX and both Anthropic and Google demonstrate strong, contracted demand for AI infrastructure and highlight the emergence of new players alongside traditional hyperscalers.

• AI adoption and token usage remain heavily concentrated in software engineering and developer communities, with 77% of employees in high-adoption organisations using coding assistants, while other verticals such as legal, accounting, and healthcare are still at earlier stages.

• As token costs decline and agentic capabilities improve, AI usage is expected to diffuse more broadly across industries, though the recent IPO from SpaceX and upcoming IPOs from Anthropic, and OpenAI may create periods of tighter liquidity for AI-related investments.

We have now been publishing the Convequity AI Bubble Barometer for four weeks. In case you have not yet seen it on our website, this report provides a full overview of the framework — including what it aims to achieve and how it has been constructed — along with this week’s update and some high-level perspectives on the current AI market landscape.

The Barometer is designed to offer a transparent, data-driven way of assessing whether AI-related valuations among the major hyperscalers are supported by underlying fundamentals. Each edition compares forward returns on invested capital with prevailing market multiples to identify potential divergence between economic reality and investor pricing.

You can access the full interactive dashboard at any time here: barometer.convequity.com

AI Bubble Barometer — Week 4 | 12 June 2026

This week’s update shows a further easing in market valuations while underlying fundamentals remain stable. You can view the full interactive dashboard here: barometer.convequity.com

Key Metrics (as of 12 June 2026)

• ROIC / WACC: 3.07x in Q1’26 (up from 2.97x in Q4’25)

• ROIC – WACC Spread: +21.9% (up from +20.9%)

• EV / IC: 6.60x (down from 7.10x last week)

• Aggregate EV / Ann. 2-Yr Forward NOPAT: 50.60x (down from 54.0x)

ROIC and WACC are updated quarterly, while EV/IC and the NOPAT multiple move weekly with market prices.

Signal

At 6.6x EV/IC, with the basket generating ~20% NTM revenue growth and ~32% forward ROIC, valuations have cooled further and now sit comfortably within the fair value range. There is no bubble signal at present. The market is paying less for the same (or slightly improving) fundamental returns, which is directionally positive.

Week 4 of the series. Updates are published every Friday.

Commentary

Recent deals between SpaceX and two of the largest AI developers underscore both the explosive demand for compute and the emergence of new, non-traditional players in the AI infrastructure supply chain. In early May, Anthropic secured exclusive access to SpaceX’s Colossus 1 data center in Memphis, committing to roughly $1.25 billion per month through May 2029 — a potential contract value exceeding $45 billion. The arrangement provides Anthropic with over 220,000 Nvidia GPUs and more than 300 MW of power, primarily to support inference workloads for its Claude models. Shortly afterward, Google disclosed a separate agreement worth approximately $920 million per month from October 2026 through mid-2029, securing access to around 110,000 Nvidia GPUs in SpaceX facilities for a potential total of $29–30 billion.

These contracts are notable not only for their size but for what they reveal about the current state of AI infrastructure. They demonstrate that demand for large-scale, reliable compute continues to outpace traditional hyperscaler capacity in certain segments, prompting AI labs to turn to alternative providers. They also highlight SpaceX’s strategy of monetizing underutilized or newly built infrastructure at attractive economics, while simultaneously advancing its longer-term ambitions in orbital AI compute. For the broader market, these deals represent real, contracted revenue rather than speculative projections — a reminder that while AI capex remains extremely high, it is increasingly backed by multi-year commercial commitments.

To many people outside the technology and investment industries, the sheer scale of spending and excitement around AI can still feel somewhat abstract. Beyond the everyday experience of using ChatGPT or similar tools, it is not always obvious where the value is being created. The reality today is that AI token consumption and productive usage remain heavily concentrated in developer and engineering communities. In high-adoption organisations, 77% of employees now use coding assistants. Software engineering currently shows some of the highest adoption rates and the most measurable productivity gains of any vertical, with studies indicating average productivity improvements of 15–20% for developers using AI coding tools, and gains reaching 35–40% on certain greenfield or lower-complexity tasks. Marketing and sales functions also show relatively advanced adoption. In contrast, verticals such as legal, accounting, healthcare, and many areas of professional services are still at earlier stages, often limited to narrower use cases.

This concentration helps explain why the current wave of investment can feel disconnected from everyday experience. At the same time, the recent SpaceX IPO, introduces a significant liquidity consideration for public markets. A valuation in the region of $2 trillion makes SpaceX the largest IPO in history by a wide margin and could absorb substantial capital from both institutional and retail investors. Looking further ahead, potential IPOs from Anthropic and OpenAI would likely exert additional pressure on available liquidity. These events, combined with the sheer scale of capital still required to expand AI infrastructure, suggest that the funding environment for AI-related companies may become more selective in the medium term, even as underlying demand remains robust.

Taken together, these developments reinforce two themes that run through the AI Bubble Barometer: the enormous capital intensity of the current build-out and the growing gap between companies that can secure large, contracted offtake agreements and those that cannot. While the recent SpaceX deals provide tangible evidence of commercial demand, they also illustrate how quickly the competitive and capital landscape is evolving. As the cost of tokens continues to decline and models develop stronger agentic capabilities, usage is expected to diffuse more broadly — from coding into legal research, accounting, marketing, healthcare, and eventually more consumer-facing applications. The current concentration of activity should therefore be viewed as an early phase rather than the end state.

Methodology

Purpose and Philosophy

The AI Bubble Barometer is designed to provide a quantitative, regularly updated framework for assessing whether the market’s valuation of AI-related investments by the major hyperscalers is supported by fundamentals or is showing signs of excess.

While investor sentiment around AI is often polarised (“bubble” vs “not a bubble”), there has been no transparent, repeatable metric that directly compares the market’s implied return expectations (via EV/IC) with the actual economic returns being generated on the capital being deployed. The Barometer fills that gap by focusing on the single most important long-term driver of shareholder value: Return on Invested Capital (ROIC).

Rather than relying on historical (trailing) ROIC, which can be distorted by the early-stage nature of AI build-outs, we estimate a forward-looking ROIC over a two-year horizon. This forward ROIC is then compared with both the cost of capital and the market’s current valuation multiple (EV/IC).

Core Framework

The Barometer compares two sides of the same equation:

• Fundamental side: Two-year forward ROIC (annualised) versus WACC.

• Market side: Enterprise Value / Invested Capital (EV/IC).

When these two measures move in tandem, valuations are broadly consistent with fundamentals. When they diverge, a signal emerges:

• EV/IC rising while ROIC/WACC is flat or falling → Potential bubble formation (market paying more for the same (or lower) economic returns).

• EV/IC falling while ROIC/WACC rises → Improving attractiveness.

We present both the ROIC/WACC ratio and the ROIC – WACC spread, but the ratio is the primary comparison tool because it maps directly onto the EV/IC multiple under steady-state assumptions.

Calculating the Denominator: AI-Related Invested Capital

The denominator of forward ROIC is the capital actually deployed into AI infrastructure.

• We take each hyperscaler’s total capex over the trailing 24 months.

• We apply an 85% haircut to estimate the portion attributable to AI (data centres, servers, networking, power, etc.). This factor is derived from management commentary, segment disclosures, and the clear acceleration in AI-related spending visible in recent quarters.

• The resulting figure represents the AI-specific invested capital deployed over the past two years that is expected to generate returns over the next two years.

We use a two-year window (rather than one year) to smooth out quarterly lumpiness in capex timing and to better align the capital base with the revenue it is expected to produce.

Estimating the Numerator: Two-Year Forward AI NOPAT

1. Starting Point – Remaining Performance Obligations (RPO)

Most hyperscaler revenue is contracted and therefore appears in Remaining Performance Obligations. We aggregate RPO across the group and then make two key adjustments to isolate AI-related revenue expected over the next 24 months:

• Duration scaling: Not all RPO will convert to revenue in the next two years. We scale each company’s RPO down to a two-year horizon. For most names we apply company-specific factors informed by earnings-call commentary on average contract lengths. AI workloads often have longer durations than legacy cloud workloads (e.g., Amazon has disclosed multi-year AI commitments averaging ~5.5 years). We therefore haircut total RPO accordingly.

• AI-specific haircut: Even within the scaled RPO, not all revenue is AI-related. We apply further haircuts based on management commentary:

  • Microsoft, Amazon, Google: 75%

  • Oracle: 90%

  • CoreWeave: 100% (entire business is AI-focused)

This produces an estimate of committed AI revenue expected over the next two years.

2. On-Demand Revenue Uplift

Not all revenue flows through RPO. A meaningful portion of AI-related demand is on-demand / pay-as-you-go (unplanned token consumption). We therefore apply an uplift factor. Our base case assumes ~75% of revenue is committed and ~25% is on-demand, which produces a 1.33x uplift to the committed AI revenue figure. Alternative scenarios (more or less on-demand) are also modelled for sensitivity.

3. Special Case: Meta

Meta does not disclose meaningful RPO (its business is predominantly advertising). We therefore use a different but transparent approach:

• Take Meta’s trailing 24-month revenue.

• Apply the consensus analyst CAGR for the next two years.

• Apply a 60% factor to estimate the portion of that revenue that is attributable to AI-related activities (recommendation engines, ad targeting, content moderation, infrastructure supporting AI workloads, etc.).

This is admittedly the most judgmental part of the model. However, because Meta is a very large contributor to aggregate AI capex, excluding it would materially distort the group picture. We therefore include a conservative estimate rather than omit the company entirely.

4. From Revenue to NOPAT

Once we have an aggregate two-year forward AI revenue figure for the basket, we convert it to Net Operating Profit After Tax (NOPAT):

• Apply a 34% weighted-average operating margin for the group (derived from segment-level margins and the mix of cloud vs other AI-related activities).

• Apply a 21% effective tax rate.

• Add back the tax shield from 100% bonus depreciation on AI capex (Trump-era policy allowing immediate expensing of qualifying AI infrastructure). The tax shield equals trailing 24-month AI capex × 21%.

This produces a tax-shield-adjusted two-year forward NOPAT.

Why a Two-Year Forward ROIC?

We deliberately calculate ROIC over a two-year horizon (rather than annualising a single year) for three reasons:

• It smooths quarterly volatility in both capex and revenue recognition.

• It better matches the capital deployed in the trailing 24 months with the revenue it is expected to generate.

• It provides a more stable signal for comparing against the slowly moving WACC and the weekly-moving EV/IC.

The resulting two-year ROIC is then annualised for presentation and comparison purposes.

WACC Calculation

We calculate a group-level WACC using market data as of the most recent quarter-end:

• Cost of equity: Risk-free rate (current 10-year Treasury yield) + beta × equity risk premium. We use Damodaran’s implied ERP (updated monthly) and each company’s 5-year beta.

• Cost of debt: Derived from the average credit rating of the group.

• Weights are based on market values of equity and debt.

Important consistency note on off-balance-sheet financing: We deliberately focus only on the capital and debt that will generate returns over the next 24 months. Future data-centre SPVs and associated debt will finance capex that produces revenue beyond our two-year window. By aligning the invested capital, the cost of debt, and the revenue forecast to the same 24-month period, the model remains internally consistent.

Deriving the Signal – ROIC/WACC vs EV/IC

Once we have the annualised two-year forward ROIC and the group WACC, we calculate:

• ROIC / WACC ratio

• ROIC – WACC spread (in percentage points)

We then compare these to the current EV/IC multiple of the basket.

The mathematical link is well-known: under steady-state assumptions,

EV/IC ≈ (ROIC – g) / (WACC – g).

Our framework therefore has internal coherence between the fundamental return metric and the market valuation metric.

EV/IC Thresholds

The current EV/IC thresholds are calibrated specifically to the basket’s present characteristics:

• ~20% NTM revenue growth

• ~32% forward annualised ROIC

• WACC of ~10.5–11%

• Capital-intensive nature of the businesses

Current thresholds (as of Q1 2026 data):

These bands are not fixed. They will be reviewed and updated whenever there is a material change in the group’s expected NTM growth or forward ROIC (typically each earnings season).

Key Assumptions & Limitations

• The 85% AI capex haircut and the various RPO/AI-specific factors are informed by management commentary but remain estimates.

• Meta’s AI revenue attribution is the most judgmental input.

• On-demand uplift assumptions can be stress-tested; we show base, high, and low cases on the dashboard.

• The model deliberately focuses on a two-year horizon to maintain internal consistency between capital deployed, revenue generated, and the cost of capital.

• Off-balance-sheet financing for future capacity is excluded because it does not yet contribute to the revenue being measured.

The Barometer is therefore best viewed as a transparent, rules-based framework rather than a precise forecast. Its value lies in the consistency of the methodology over time and the clear signals that emerge when market valuations and fundamental returns diverge.

Summary

• AI Optical Engines, Simplified: POET’s Optical Interposer replaces slow, expensive manual alignment with wafer-scale, pick-and-place manufacturing — the key step that lets high-volume players like Foxconn build the optical engines powering 800G, 1.6T, and future AI clusters.

• Lumilens Changes Everything: The new May 2026 partnership gives POET a hybrid InP + SiPho platform on the same base, plus an initial $50M order (framework to $500M+). This creates real revenue, turns the SiPho shift into a tailwind, and opens doors to both pluggable transceivers and next-generation NPO/CPO — all outside Broadcom’s closed ecosystem.

• High-Conviction, High-Risk Bet: We added POET at the start of 2Q26; shares are already up ~190%. At a $2.5B market cap it remains a public-market venture play — substantial 5–10× upside if the Lumilens orders convert, but zero margin of safety if production or execution slips.

Update – May 25, 2026

Since publication on April 20, two material developments have occurred. First, POET announced a strategic supply and joint-development partnership with Lumilens on May 14, including an initial $50 million purchase order for Electrical-Optical Interposer (EOI) engines (with a framework that could scale to $500 million+ over five years). This expands POET’s platform into a hybrid InP/SiPho architecture on the same wafer-scale base — directly addressing the SiPho headwind discussed below. Second, on April 27 Marvell (which had acquired Celestial AI) cancelled all remaining purchase orders with POET, citing alleged breaches of confidentiality obligations related to order and shipping details. The loss removes one near-term revenue path but is more than offset in scale by the new Lumilens order.

We added POET to the Convequity portfolio at the start of April 2026 (beginning of 2Q26). The shares have already risen approximately 190% since our entry.

Technology and Market Position

To understand why, it helps to zoom out. With Moore’s Law reaching its limits at the individual chip level — and to a lesser extent, even the package level — meaningful performance gains increasingly come from scale. As a result, the biggest AI companies are building enormous computing clusters, sometimes hundreds of thousands of chips working in concert. All those chips need to exchange massive amounts of data at extreme speed, and that data movement is now the limiting factor. Traditional copper wiring simply can’t handle the required speed and distance without consuming too much power and generating too much heat. The industry is rapidly shifting to optical (light-based) connections, creating a severe shortage of the specialised components needed to make them work. The bottleneck is so serious that even Nvidia has invested billions to secure supply.

Many investors now know Lumentum (LITE) as a key player here (check out our first report on LITE and our second piece on the broader optical landscape). LITE is the dominant supplier of the high-performance laser chips that power optical connections, holding roughly 50–60% of global market share, and through its CloudLight acquisition also builds complete optical modules. It’s a broad, vertically integrated play on AI-driven optical demand. POET sits in a different, complementary part of the chain. Rather than competing with LITE on lasers or finished modules, POET has built a proprietary platform that solves the most expensive and hardest-to-scale step: assembling all the tiny components into a working optical engine.

This is where POET’s real innovation lies. Historically, building an optical engine — the subcomponent at the heart of an optical transceiver that handles the conversion between electrical and optical signals — meant taking individual parts like lasers, detectors, waveguides, and electronics, each manufactured on separate substrates, and painstakingly aligning them by hand under power - far removed from the mature, automated processes of logic and memory fabrication. Technicians would literally turn on a laser and nudge parts into position until the light coupled properly. This “active alignment” was slow and expensive, but for earlier transceiver generations running at 10G, 25G, or 100G, it was workable. Human hands could meet both the precision and the volume requirements at those speeds.

That changed with the move to 400G and 800G. The tolerances tightened to the point where manual assembly could no longer achieve the required precision, and the surging demand driven by AI meant human throughput couldn’t keep pace with volume requirements either. The industry was forced toward machine-automated active bonding — and with 1.6T now ramping and 3.2T on the horizon, fully automated bonding is not merely preferred but essential. This applies at every level of the assembly chain. At the transceiver level, companies like Innolight and Eoptolink have had to invest heavily in automated bonding and testing equipment from the likes of Keysight and ASMPT. The complexity is such that only a handful of transceiver makers can deliver each new generation at volume and acceptable yield in the first year or two — Innolight effectively held a near-monopoly in the early quarters of the 800G ramp and still commands over 60% market share, earning gross margins of 40–50% that rival those of high-value-add semiconductor companies. But at the optical engine level, where the components are smaller and the alignment tolerances tighter still, the manufacturing challenge is even more acute — and this is the level at which POET operates.

POET’s patented Optical Interposer changes the game. It’s a silicon wafer with waveguides and mechanical guides etched directly into it. Lasers, detectors, and electronics — sourced from specialist suppliers — are placed onto this wafer using automated pick-and-place machines, much like conventional chip assembly. No manual alignment, no powering up during assembly. Everything snaps into place passively at the wafer level. Think of it as building with Lego rather than hand-carving each joint. The result is smaller, lower-power, significantly cheaper optical engines that can scale to the speeds AI demands: 800G, 1.6T, and eventually 3.2T.

POET operates a fab-lite model. The Optical Interposer wafers are fabricated by SilTerra, an 8-inch silicon foundry in Malaysia, using standard CMOS processes, and the optical engine assembly, testing, and packaging then take place at contract manufacturers in Penang — primarily Globetronics and NationGate — where POET has installed its own wafer-level equipment. The completed engines are supplied to transceiver makers like Foxconn (via its subsidiary FIT), Luxshare, and LITEON, who integrate them with the DSP, driver electronics, housing, and fibre connections to produce finished transceiver modules. What makes this arrangement strategically significant is who those partners are. Companies like Foxconn are among the world’s largest electronics manufacturing services providers, with unmatched capabilities in high-volume assembly — but they have historically been locked out of the optical transceiver market because building transceivers required deep photonic expertise, particularly around the optical engine, which represents over 50% of the transceiver’s cost and contains all of the most technically demanding alignment work. POET’s interposer changes that equation. Because the optical engine arrives as a fully integrated, pre-aligned module on a standardised substrate with accessible electrical and optical interfaces, the remaining transceiver assembly looks much more like conventional electronics manufacturing — precisely the kind of work Foxconn already excels at. In effect, the interposer absorbs the optical complexity so that downstream assemblers don’t have to. The value-add and cost shift onto the interposer itself, but in doing so, POET opens the transceiver market to an entirely new tier of manufacturer and broadens the industry’s production capacity at exactly the moment the AI-driven bottleneck demands it.

It is important to understand where POET fits — and where it does not. POET’s technology is designed for the pluggable transceiver market: discrete optical modules that slot into the front panel of a switch or server, each containing its own optical engine, lasers, and electronics. This remains the dominant form factor in data centres today and is the architecture behind the 800G and 1.6T ramp. But there is a parallel architectural track — co-packaged optics (CPO) — where the silicon photonics engine is integrated directly alongside the switch’s ASIC inside the package, eliminating the pluggable module entirely. Broadcom is the furthest ahead here, and its approach illustrates why POET is unlikely to play a role in that particular chain.

Broadcom’s CPO platform, built around its Tomahawk 5 “Bailly” switch ASIC and already in mass production, uses what it calls a Remote Laser Module (RLM) architecture. Silicon cannot emit light — this is a fundamental physical constraint — so any silicon photonics engine is inherently “mute” and needs an external indium phosphide (InP) laser to supply it with a continuous-wave (CW) light source. In a traditional CPO design, that laser would be bonded directly to the photonic integrated circuit (PIC), and this is precisely what has made CPO so difficult to deploy at scale: the CW laser is the component most prone to failure in any optical subsystem, and if the laser dies in a bonded CPO configuration, the entire optical engine — and potentially the switch ASIC it is co-packaged with — becomes non-functional.

Broadcom’s solution is to separate the laser deliberately. The InP laser sits in a hot-pluggable module on the front panel of the switch, connected via optical fibre to inject CW light into the silicon photonics engine inside the package where the light modulation occurs, transferring electrical signals onto the CW laser to create the optical signals. If the laser fails, a technician simply swaps the module on-site without touching the PIC or the ASIC. Broadcom has developed dedicated high-density optical fibre connectors at 127 μm pitch, supporting up to 72 fibre pairs, to make this connection low-loss and repeatedly pluggable. The result is a fully vertically integrated CPO system — Broadcom’s own silicon photonics engine, CW lasers supplied by established vendors like Lumentum and Coherent in the hot-swappable RLM form factor, packaging and connector hardware from Foxconn Interconnect Technology (FIT), and the whole platform already validated with over one million cumulative fault-free hours on 400G ports at Meta. There is no gap in this stack for a third-party interposer. What Broadcom sells is a complete switch system — ASIC, photonics engine, and laser ecosystem — and internally there is no motivation to outsource optical integration to a platform like POET’s when the entire point of the architecture is end-to-end control.

This matters for sizing POET’s opportunity. Historically, POET has been more heavily focused on the InP/EML side of the pluggable transceiver market. As the industry shifts toward greater adoption of silicon photonics (SiPho), this created a meaningful headwind because SiPho architectures integrate more components at the process level and can reduce the need for certain discrete assembly steps. However, the May 2026 partnership with Lumilens dramatically broadens POET’s addressable market and future-proofs the platform. Lumilens — which designs its own advanced SiPho PICs and full optical interconnect solutions — has placed an initial $50 million purchase order for POET’s new jointly developed Electrical-Optical Interposer (EOI) engines, with the framework structured to potentially scale to more than $500 million in cumulative purchases over five years. POET can now deliver high-performance hybrid optical engines using either top-tier InP/EML components or advanced Lumilens SiPho PICs — all assembled wafer-scale on the same base.

The joint roadmap starts with 800G and 1.6T SiPho-based pluggable transceivers but explicitly extends into high-density Near-Package Optics (NPO) and Co-Packaged Optics (CPO). Lumilens will integrate the engines into complete modules and sell them to hyperscalers, giving POET firm footing in both InP and Silicon Photonics AI optics paths. The solution is particularly compelling to hyperscalers because the POET-Lumilens Electrical-Optical Interposer platform brings semiconductor-style, wafer-scale manufacturing to optical engines — delivering capital-efficient production, superior yields, and the ability to scale supply in line with exploding GPU cluster growth rather than being limited by conventional labor-intensive assembly. At the same time, hyperscalers are actively seeking credible alternatives to Broadcom’s dominant, vertically integrated stack to avoid vendor lock-in, maintain negotiating leverage, and strengthen supply-chain resilience. It is a smart, measured hedge that creates real orders while preserving the hybrid moat even as InP constraints bite and the broader industry moves toward higher levels of photonic integration — all without relying on Broadcom’s closed ecosystem.

There is also a notable supply chain thread on the other side of the competitive landscape. Marvell, which trails Broadcom significantly in optical DSPs, recently acquired Celestial AI — a startup developing a novel photonic architecture — to help close that gap. POET had been positioned as a core partner, supplying the optical interposer upon which that architecture was being built. However, in late April 2026 Marvell cancelled all outstanding purchase orders from Celestial AI, citing alleged breaches by POET of confidentiality obligations related to order and shipping information. The cancellation removes one previously anticipated near-term revenue path and serves as a reminder of the execution and governance risks inherent in early-stage ecosystem partnerships.

That said, the May 2026 Lumilens partnership more than offsets the setback in both scale and strategic breadth. By enabling POET’s Electrical-Optical Interposer to support advanced Lumilens SiPho PICs alongside its existing InP/EML platform, POET has secured a far larger and more future-proof foothold. The contrast with Broadcom remains instructive: where Broadcom’s vertically integrated CPO stack leaves no room for a third-party integration platform, Lumilens’ open, module-focused approach creates exactly the kind of ecosystem seam that POET’s hybrid interposer is designed to fill — now across both pluggable transceivers and next-generation NPO/CPO architectures. This does not eliminate execution risk, but it suggests POET’s technology is attracting attention from the right nodes in the AI infrastructure ecosystem, with its commercial future now more tightly linked to the Lumilens axis than to any single prior partnership.

Investment & Risk Considerations

In early 2026 the company reached several milestones that moved it meaningfully closer to commercial reality. It raised $375 million across three rounds of equity financing, with each round completed at prices above prior rounds — a noteworthy signal that institutional investors were willing to pay up for the platform at elevated valuations. This capital extends POET’s runway through the critical capacity ramp-up phase and substantially reduces the risk of cash-flow disruption during what will be a challenging period of yield optimisation and production scaling.

On the commercial front, POET received a production order valued at over $5 million from a leading systems integrator for POET Infinity optical engines. The operative word is “production” — this is a mass-production order, not a sample or prototype shipment, meaning a customer has completed qualification testing and committed to deploying POET’s technology in commercial products. For a company that generated roughly $1 million in revenue over the prior year, even a $5 million order represents a step-change. More significantly, the May 2026 Lumilens partnership delivered an initial $50 million purchase order for the new jointly developed Electrical-Optical Interposer (EOI) engines, with the commercial framework structured to scale to more than $500 million in cumulative purchases over five years. These are commitments, not yet cash in the door, and the pace at which they convert to shipped product and recognised revenue will be one of the key variables to track.

The company also announced a collaboration with Quantum Computing (QUBT) to develop 3.2T engines using thin-film lithium niobate (TFLN) modulators, a next-generation material platform prized for ultra-low power consumption and high bandwidth. With the mainstream industry still ramping 800G and only beginning to test 1.6T, a 3.2T development partnership is forward-looking by at least two to three years — but it signals that POET’s interposer architecture is designed to be generation-agnostic and that the company is already positioning for the technology cycles beyond the current one.

Despite these milestones, investors should remain clear-eyed about what they are actually buying. At a market capitalisation of roughly $2.5 billion, POET’s trailing twelve-month EV/Sales multiple is approximately 2,500x — a number so extreme it is effectively meaningless as a valuation anchor. The more instructive figure is the next-twelve-month EV/Sales multiple of roughly 79x, derived from analyst consensus estimates of approximately $14 million in forward revenue — a figure that now looks conservative given the Lumilens order. The valuation itself, however, reflects expectations far beyond the next twelve months, which is precisely what makes it fragile: if near-term revenue disappoints, there is no earnings floor to support the price.

The forward revenue of $14 million is plausible given the announced production commitments, but it remains largely unearned, and any slippage in delivery timelines or yield performance would force a sharp re-rating. In practical terms, participating in POET at this stage is closer to using the liquidity of the public equity market to buy the return structure of a primary-market venture capital investment: if the thesis plays out, the upside is substantial; if it does not, the floor is zero with no margin of safety.

POET shares have risen approximately 190% since the position was added to the Convequity portfolio at the beginning of 2Q26, reflecting strong market reaction to the Lumilens order and the completed capital raise.

To help frame the risk, it is worth laying out a rough timeline of how POET’s story could unfold. The period from 2024 through 2025 was the early stage of narrative formation: cooperation announcements, small-batch orders, and a steady drumbeat of press releases signalling that major players were evaluating the platform. Revenue grew from negligible to $1 million, but heavy R&D and production-line investment meant the company was still reporting substantial losses. The period from 2026 through 2027 is the critical mass-production window. This is when POET has publicly committed to shipping tens of thousands of optical engines from its Malaysia facility. If mass production proceeds smoothly, yield targets are met, and the announced commitments (now anchored by the Lumilens framework) convert to shipped and recognised revenue, the company could plausibly reach $50–100 million in annual revenue, at which point the current valuation multiples would begin to compress rapidly toward something more conventional. Failure to execute in this window — whether due to yield problems, customer delays, or competitive displacement — would likely be fatal to the standalone investment thesis.

The period from 2028 through 2030 represents the broader industry inflection point: as data transmission speeds reach 1.6T and 3.2T, the power consumption, heat dissipation, and instability challenges of traditional pluggable architectures intensify, and co-packaged optics moves from optional to mandatory for large-scale data centre interconnects. The companies that have established volume production and proven reliability by then will capture the lion’s share of what could be a market worth hundreds of billions of dollars in aggregate. For POET, the question is whether it will be among them or whether the window will have closed.

In the next one to two years, analysing this company through traditional financial reports — profits, margins, free cash flow, EPS — is effectively meaningless. What matters are industry-level catalysts: whether POET announces that a Tier 1 customer has passed qualification on its engines, whether it secures further substantive mass-production purchase contracts, and whether it achieves target yields on its Malaysia production line. These are the forces that will either validate the current valuation or expose it.

The competitive picture adds nuance to this timeline. The growing adoption of silicon photonics (SiPho) now represents a clear tailwind rather than a headwind. Thanks to the Lumilens partnership, POET’s Electrical-Optical Interposer platform directly supports advanced SiPho PICs alongside its legacy InP/EML engines — all assembled wafer-scale on the same base. This hybrid capability means POET captures the integration economics in both architectures, exactly as the industry shifts toward greater photonic integration. Even within SiPho, the active bonding cost between the photonic die and CW laser continues to rise with each successive generation, and by the 3.2T era the economics of using an optical interposer are expected to strengthen further. The broader structural trend — rising equipment costs, growing complexity, and an expanding pool of players wanting to enter the transceiver market — still favours a platform that can democratise assembly. POET’s positioning within the Lumilens ecosystem therefore provides a credible and enlarged path to volume adoption, though it does not eliminate the underlying execution risk — the interposer must prove itself in production, not just in partnerships.

It is also worth restating the structural limitation. Broadcom’s vertically integrated CPO roadmap and the relentless iteration speed of cash-rich semiconductor giants mean that POET is not merely racing against its own production timeline but against an industry that moves extraordinarily fast. Optical communication technology iterates on roughly two-to-three-year cycles, and POET’s technology window — the period in which its interposer offers a compelling cost and complexity advantage over both incumbent assembly methods and emerging silicon photonics integration — is most likely concentrated in the 2026–2028 timeframe. If the company has not completed the leap from sample validation to scaled supply within that window, the most probable outcome is a dead end, with acquisition as the only exit.

If the thesis does work, the upside case is considerable. Should POET establish itself as the standard integration substrate for the growing tier of small-to-medium transceiver makers entering the 800G and 1.6T market — the Foxconns, Luxshares, and LITEONs that need POET’s interposer to compete — the current $2.5 billion market capitalisation could have five to ten times upside over a three-to-five-year horizon. That is the return profile of a venture capital investment, and it comes with the corresponding risk: the downside is total loss. From the current “storytelling” phase to real scaled profitability, there are still at least three to five years of road ahead, and every quarter of that road will be contested by well-capitalised incumbents iterating furiously on their own solutions.

Where LITE gives investors broad exposure to optical demand through lasers and modules, POET offers more targeted exposure to the integration layer that module makers desperately need to scale. The risk-reward looks favourable over a longer timeframe because the supply constraints are unlikely to ease any time soon. As more power capacity comes online, more data centre shells get built and connected, and the downstream demand for optical interconnects only grows. POET’s wafer-level approach directly addresses the cost, power, and manufacturing bottlenecks at the heart of that wave — but the gap between platform potential and production reality is where all the risk sits.

Summary

• We added ASM International to our portfolio at the start of 2026 on the thesis that semiconductor scaling is shifting from lithography to materials precision—structurally increasing ALD intensity where ASM dominates.

• The transition to GAA, backside power delivery, and 2nm-class nodes creates a multi-year growth cycle driven by process complexity, not wafer volume. AI is expanding ALD exposure into HBM, advanced packaging, and increasingly logic-like memory.

• 1Q26 confirmed the thesis with force: revenue beat at the high end of guidance, record 33.1% operating margin, accelerating demand across all three logic customers, and 1.4nm pilot revenue already contributing in H2 2026.

• The stock is up 68% since we initiated. At ~80x P/FCF and 85x EV/(FCF-SBC), ASM is not cheap — but complexity tailwinds are multi-year, the competitive moat is unchallenged, and bottleneck dynamics show no signs of easing. We’re holding.

ASM sits at the center of a structural shift

ASM sits at the center of a structural shift in semiconductor manufacturing. As lithography continues to push transistors to ever-smaller geometries, the margin for error in materials engineering is shrinking, increasing the need for atomic-level control over film thickness, uniformity, and interfaces. At the same time, performance gains are increasingly being driven by three-dimensional architectures — first at the transistor level through structures such as GAA and backside power delivery, and then at the die level through stacking, TSVs, hybrid bonding, and broader 3D integration to increase effective silicon density.

Both trends — continued dimensional scaling and the expansion of transistor-level and die-level 3D — rely on highly precise, conformal deposition across complex and high-aspect-ratio structures in logic as well as DRAM, HBM, and NAND. This makes ALD a foundational process for advanced chip manufacturing, a dynamic that plays directly into ASM’s core strengths in ALD and selective epitaxy.

2nm today, 1.4nm next

Leading-edge logic has now entered a node cycle where continued shrink (2nm-class and below) drives a step-up in process intensity, because tighter geometries demand more atomic-level control over film thickness, uniformity, and interfaces. TSMC’s 2nm (N2) entered volume production in 4Q25, and the associated rise in ALD and selective epitaxy intensity is now flowing through ASM’s revenue as ramps progress through 2026.

This cycle also coincides with the industry’s first real move into GAA at scale. Intel’s 18A is its first node to introduce GAA (RibbonFET), alongside backside power delivery (PowerVia), and it’s widely discussed as being in the same “2nm-class” bracket as TSMC N2 even though the naming conventions differ (Intel emphasizes performance/power ambitions, while TSMC often highlights density and manufacturability).

Looking ahead, pilot work for 1.4nm is no longer a 2027 story alone. Management confirmed on the 1Q26 call that pilot-line investments for 1.4nm are already ramping in H2 2026 and delivering meaningful revenue contribution this year. The move from 2nm to 1.4nm is highly accretive: it adds significantly more ALD layers — particularly “performance layers” in the front-end-of-line — plus higher selective epitaxy intensity. Each node transition adds more ALD and selective epi steps per wafer, creating a multi-year growth driver that is tied to process intensity as much as overall wafer volumes.

From FinFET to GAA: why ALD intensity rises

The transition from FinFET to gate-all-around transistors marks a structural shift in how performance gains are achieved. While the previous cycle from 7nm to 3nm was heavily defined by the industry’s investment in increasingly complex and expensive EUV lithography, the next phase is being driven more by transistor architecture and materials engineering.

GAA devices require the gate to fully surround vertically stacked nanosheets, dramatically increasing the need for ultra-thin, perfectly conformal films — a core strength of ALD. At the same time, new elements such as backside power delivery and, at the system level, hybrid bonding and advanced packaging extend the industry’s reliance on atomic-scale deposition beyond the front end.

This shift does not eliminate the need for etch — high-aspect-ratio structures still require precise pattern transfer — but it increases the value of deposition relative to other steps, as device performance increasingly depends on film quality, interface control, and uniformity rather than purely geometric patterning.

As a result, the FinFET → GAA transition has reinforced ASM’s competitive position. The number of ALD layers per wafer rises meaningfully at advanced nodes, process tolerances tighten, and performance becomes more sensitive to materials precision. In short, the industry is moving from a lithography-centric scaling model toward one where materials engineering and atomic-level control play a larger role in enabling each new generation.

1Q26: Thesis confirmation in the numbers

ASM’s 1Q26 results (reported April 21, 2026) provided the strongest operational confirmation yet of the structural thesis outlined above. Revenue came in at the high end of guidance at ~€863 million (+16% YoY at constant currency; +26% QoQ), with equipment sales growing +14% YoY cc and spares & services rising +23% YoY cc as outcome-based service offerings scale.

Profitability was the standout. Gross margin of 53.3% — well above the 47–51% long-term target range — reflected favorable product and customer mix (including accretive China contribution), strong non-China leading-edge product mix, and ongoing structural cost-reduction programs. Adjusted operating margin reached a record 33.1%, up 8 percentage points sequentially. These are not one-off dynamics; management guided gross margin toward the upper end of the long-term range for the full year.

Key takeaways from the quarter and call:

• All three advanced-node logic customers are growing YoY in 2026, with the largest customer remaining the dominant contributor. ASM is actively engaged with all three on both 2nm and 1.4nm programs.

• 1.4nm is already revenue-accretive. Pilot-line investments are ramping in H2 2026. The node adds meaningfully more ALD layers versus 2nm — particularly performance layers in the FEOL — and reduces throughput per layer (more complex geometries = more equipment required per wafer). Molybdenum ALD, already in high-volume production at 2nm, has secured additional process-of-record layers at 1.4nm.

• AI demand is translating directly into capacity expansion across both advanced logic and memory, reinforcing the structural demand driver rather than a one-off inventory build.

• China delivered a strong sequential rebound and is tracking up YoY for FY26, partly driven by pre-buying ahead of potential further export-control tightening. H1 is expected to be stronger than H2 on a sequential basis, though H2 could still strengthen.

• Memory/HBM remains healthy, with solid contributions from HBM-related applications and the ongoing 6F²-to-4F² transition.

• Advanced packaging continues to progress toward commercial wins, with expanding R&D engagements in ALD and selective epi for hybrid bonding interfaces.

Guidance:

• Q2 2026: €980 million ±5% (constant currency) — implying continued strong sequential growth.

• H2 2026: Expected to be higher than H1 overall.

• Full-year 2026: Revenue growth to at least match or beat overall WFE market growth. Longer-term targets (≥12% revenue CAGR through 2030, operating margin path to >30%) remain intact.

The combination of record margins, accelerating revenue, broad customer participation, and visible layer-count upside at 1.4nm significantly de-risks the near-term growth outlook and keeps the structural “ALD bottleneck” re-rating scenario firmly alive.

DRAM, HBM, and the move from 6F² to 4F²

Memory represents ASM’s second structural growth pillar. The transition toward higher-density cell architectures — including the move from 6F² to 4F² — increases three-dimensional complexity within the memory array itself. Tighter geometries, higher aspect ratios, and more demanding gap-fill and interface requirements drive additional ALD steps for oxides, metals, liners, and dielectric structures, particularly in advanced DRAM and HBM.

A second, increasingly important driver comes from the growing logic content around memory. As bandwidth, power efficiency, and signal integrity become critical for AI workloads, the CMOS periphery plays a larger role and is gradually adopting more logic-like process requirements, increasing the use of advanced deposition and selective epitaxy.

This trend becomes even more pronounced in next-generation HBM. Starting with HBM4e, the base die is expected to be manufactured in leading-edge logic fabs rather than traditional memory fabs. As bandwidth demands rise and compute die area becomes constrained, more controller and interface functionality is being relocated into the HBM base die, significantly increasing its complexity. In effect, HBM is evolving from a pure memory stack toward a tightly integrated memory-logic subsystem — a shift that expands ASM’s exposure to advanced logic-style ALD intensity within the memory ecosystem.

Demand & uncertainty

Near-term demand is concentrated in leading-edge logic/foundry and advanced memory, supported by AI-driven capacity expansion and the ramp of 2nm-class nodes and HBM. The 1Q26 results confirmed that this demand is materializing with broad customer participation and accelerating momentum. Power, wafer, and analog have been in a multi-year downturn but now appear to be stabilising, with management expecting gradual recovery through the remainder of 2026.

Looking further out, robotics and industrial automation could create a new demand cycle for power and analog devices, although the timing and scale remain uncertain.

The key swing factor for the remainder of 2026 is spending mix — not just total WFE, but where customers allocate that spend. Management has indicated that both revenue growth and gross margin will depend on the balance between leading-edge logic/foundry, advanced DRAM/HBM, and any recovery in mature-node markets. Leading-edge logic and advanced memory tend to carry higher ALD intensity and stronger margins, whereas a mix shift toward more mature-node or lower-intensity applications would dilute both growth and profitability. The 1Q26 gross margin of 53.3% illustrates what is achievable when mix tilts favorably.

China adds another layer of uncertainty. Export controls have reduced visibility and created opportunities for domestic competitors to build capability and accelerate learning, even though ASM believes its technology position remains strong. The H1-weighted China demand pattern (partly driven by pre-buying) introduces some sequencing risk, although management noted H2 could still strengthen.

While near-term visibility can fluctuate, the structural demand drivers tied to GAA adoption and next-generation node ramps remain intact. The 1Q26 confirmation of 1.4nm revenue already in 2026 suggests the primary risk is mix timing rather than end-demand.

Advanced packaging: an emerging option

ASM is also seeding advanced packaging opportunities, where the shift toward chiplets and HBM is making packaging itself an increasingly important enabler of system-level performance. As scaling within a single monolithic die becomes more complex and costly, performance gains are increasingly being achieved by partitioning functions across multiple dies and reconnecting them within a package. This approach allows designers to increase compute density, improve bandwidth, reduce power consumption, and mix process nodes, effectively extending system-level scaling beyond what transistor shrink alone can deliver.

This shift is raising the importance of interface quality, thin-film control, and surface preparation. It is driven not only by die-to-die hybrid bonding, but also by the growing complexity of the package substrate itself. Advanced package substrates are becoming more logic-like, with higher interconnect density and, in some architectures, the use of TSVs to connect stacked dies. These structures require high-precision, highly uniform films and tight surface control to ensure electrical performance, signal integrity, and bonding yield.

Management confirmed on the 1Q26 call that R&D engagements in advanced packaging are expanding, with progress toward commercial wins. While still early-stage for ASM, the increasing materials intensity within advanced packaging suggests this area could become a meaningful incremental growth vector over time.

Summary — Positioned for complexity

The common thread across all of these trends is complexity. Features are shrinking, geometries are becoming more 3D, and performance margins are tightening. That environment increasingly demands atomic-level deposition to maximize performance, power efficiency, and yield. This is happening across both leading-edge logic/foundry and advanced memory — precisely where ASM is strongest.

Optionality: solar as a long-term materials opportunity

One underappreciated long-term option for ASM is advanced solar manufacturing. Solar panel production relies heavily on thin-film deposition processes such as PVD and CVD, and Applied Materials was once active in this market before exiting roughly 15 years ago after intense price competition from Chinese equipment suppliers pushed the industry toward commoditised, high-volume manufacturing.

That dynamic could shift as the technology frontier moves beyond conventional silicon cells. Next-generation architectures such as heterojunction (HJT) and, further out, perovskite and tandem designs place greater emphasis on materials precision rather than geometric scaling. Unlike logic or memory, solar cells do not require ever-smaller features, but higher efficiency and longer lifetimes depend on ultra-thin, uniform interface layers, precise thickness control, and defect-free barrier films — requirements that align closely with atomic-level deposition. As these technologies evolve, solar manufacturing begins to resemble advanced materials engineering rather than low-cost assembly.

ASM has not highlighted solar as a target market, and any meaningful revenue contribution would likely be many years away. However, if the industry shifts toward higher-efficiency tandem or perovskite-based designs, the manufacturing challenge moves closer to ASM’s core competencies. With AI driving structural growth in electricity demand — and industry leaders continuing to position solar and battery storage as the long-term backbone of global energy — advanced solar manufacturing represents a credible, if distant, optional growth vector.

Valuation

ASM is up 68% since we added it to our portfolio at the beginning of the year. The stock has effectively grown into the bull case we laid out at the end of 2025—what seemed like an optimistic scenario has become the base case as order momentum and backlog growth confirmed the structural demand thesis faster than expected.

At current levels, ASM trades at roughly 80x P/FCF and 85x EV/(FCF-SBC). By no stretch is this cheap. But we’re continuing to hold because the complexity tailwinds that justify premium multiples are not one-quarter phenomena—they’re multi-year structural shifts in how chips are built. Gate-all-around, backside power delivery, advanced packaging, and high-NA EUV patterning all drive incremental ALD intensity per wafer, and these transitions are still in their early innings.

The question isn’t whether ASM deserves a premium—it clearly does as a bottleneck supplier in the most capex-intensive buildout the semiconductor industry has ever seen. The question is whether multiples can expand further from here as the market begins to price in a longer runway of above-trend growth. Given that ALD step counts per wafer are still rising with each successive node, and that ASM’s competitive moat in single-wafer ALD remains largely unchallenged, we think the stock can continue to compound steadily from here even without aggressive multiple expansion. We’d revisit our position if order growth decelerates meaningfully or if competitive dynamics shift—but neither appears imminent.

Summary

Precision timing has shifted from commodity infrastructure to a system-level constraint in AI, and SiTime sits at a critical coordination layer as compute, networking, and interconnect complexity scale non-linearly.

CED is the clear near-term growth engine, with MEMS-based timing content expanding across GPUs, NICs, switches, optics, and advanced packages as AI systems fragment into more clock-sensitive domains. Q1 2026 results confirmed this trajectory emphatically, with CED surging 158% year-over-year and marking its eighth consecutive quarter of triple-digit growth.

The pending acquisition of Renesas’s timing business transforms SiTime from a resonator-and-oscillator leader into a full-stack timing platform — adding clock generators, jitter attenuators, network synchronizers, and distribution buffers that complete the system coordination layer exactly where AI demand is exploding.

Mobile timing represents a genuine architectural inflection, driven by Apple’s internal modem platform and a likely second wave as Qualcomm upgrades its RF stack for 6G and satellite, pulling MEMS timing into both iOS and Android ecosystems.

SiTime’s moat is structural rather than cyclical, rooted in decades of proprietary MEMS process, resonator physics, analog design, and system-level integration that quartz-centric competitors have failed to replicate — and now deepened by ownership of the complete timing stack from resonator through system coordination.

Despite extreme headline multiples, revenue growth is accelerating toward triple-digit rates, and the stock’s valuation reflects category dominance in a structurally expanding space rather than speculative excess.

Precision timing is one of those technologies that sits quietly in the background until systems become large enough, fast enough, and parallel enough that coordination itself becomes the bottleneck. In modern AI infrastructure, a “clock” is not about measuring seconds; it is about synchronization. Elon Musk has likened large-scale AI training to conducting an orchestra of 100,000 musicians, each required to start, stop, and stay in perfect harmony within milliseconds. AI training and inference workloads are similarly distributed across thousands of chips operating in parallel, where every step depends on data arriving at exactly the right moment. When timing slips, GPUs wait, networks stall, and expensive compute goes unused. Traditional CPU-centric workloads are largely sequential and can absorb small timing mismatches. AI systems cannot.

This is the context in which precision timing has moved from commodity plumbing to a system-level enabler of performance — and why SiTime sits at an increasingly critical layer of the AI infrastructure stack. The company’s relevance is not driven by a single product cycle, but by the structural way AI systems are being built: larger, faster, more distributed, and far more sensitive to synchronization errors.

SiTime’s end-market exposure and growth drivers

Before diving deeper into AI and mobile, it is useful to anchor SiTime’s revenue model around how the company thinks about demand. SiTime sells timing across three end-market buckets: communications, enterprise, and data center (referred to as CED); mobile, IoT and consumer; and automotive, industrial and defense. The near-term growth engine is clearly CED, where timing content expands as AI clusters scale and networking complexity explodes. Mobile and consumer is the second axis: less visible, often customer-name restricted, but increasingly important as smartphone timing architectures shift away from quartz and toward integrated, programmable timing. Automotive, industrial and defense provides additional breadth, durability, and product pull-through, especially as timing becomes more mission-critical and safety-critical across environments that are harsh, vibration-heavy, temperature-variable, or GPS-compromised.

Why quartz-based timing breaks down in AI systems

For decades, electronic timing has been dominated by quartz. Quartz works well in relatively static, low-frequency, low-parallelism environments. But AI data centers stress timing in ways quartz was never designed for. As bandwidth increases, jitter budgets tighten. As systems scale, vibration, thermal gradients, and electromagnetic interference rise. As architectures fragment into chiplets, modules, cables, and fabrics, the number of clock domains multiplies.

Quartz struggles in this regime. It is physically larger, mechanically fragile, slow to stabilize, and poorly suited to semiconductor-style integration. Its performance degrades under shock, vibration, and temperature variation — precisely the conditions present in dense AI racks running at extreme power levels.

MEMS-based timing scales differently. MEMS resonators are smaller, lower mass, more resilient, and manufactured using semiconductor-grade processes. This allows tighter control over performance, better environmental stability, and far greater flexibility in how timing is packaged and deployed. As AI systems become faster and more parallel, these advantages compound rather than diminish. This is the structural shift SiTime is exploiting.

Where timing demand is really coming from in AI data centers

Within an AI data-center rack, timing demand still scales most visibly on the networking and interconnect side, but the compute complex itself is changing in a way that matters for timing suppliers. GPUs and CPUs increasingly integrate basic clocking functions on-die, which can cap legacy, discrete quartz timing content per chip. That does not mean timing becomes less important — it means the bar for usable timing rises.

Advanced packages introduce new constraints. As GPUs and CPUs move toward chiplets, HBM stacks, and 2.5D/3D integration, on-die clocks face rising noise, thermal gradients, and cross-die interference. Quartz-based timing is poorly suited to this environment: it is bulky, mechanically fragile, difficult to co-locate near hot silicon, and incompatible with tight package-level integration. As a result, quartz becomes the limiting factor inside advanced packages rather than the solution.

MEMS-based timing scales differently. Today, MEMS timing is still predominantly board-level, with only a modest share — roughly on the order of 10% — appearing in tightly integrated or package-adjacent designs. But that penetration is structurally early. As packages become denser and clock domains proliferate inside and around the package boundary, the need for cleaner reference clocks that can tolerate proximity to high-power silicon increases. This creates natural insertion points for MEMS-based oscillators and clocking solutions that can sit closer to the silicon without the mechanical and environmental penalties of quartz.

Around the compute complex, timing demand continues to explode. Networking is where SITM will experience surging demand. NICs and SmartNICs require low-jitter clocks to coordinate traffic and offload networking functions. High-radix switches and routers multiply SerDes lanes as bandwidth scales. Optical modules, retimers, and active cables all require clean, stable clocks to move data reliably at extreme speeds. Storage systems and accelerators add further clock domains as data paths fragment and parallelize.

The transition from 800G to 1.6T optical modules illustrates the broader pattern. Bandwidth does not simply double; links are split into more lanes running at higher baud rates. Each lane tightens jitter budgets and often introduces additional clock domains for DSPs, SerDes groups, and laser drivers. The result is more timing components per module and higher performance requirements per component, pushing average selling prices up alongside unit volumes.

The same combinatorial effect appears elsewhere in the rack. AI switches scale radix and speed simultaneously, driving more clock domains across fabric, ports, and control logic. Accelerator cards increasingly resemble distributed systems, combining GPUs, HBM, NICs, and retimers, each with its own timing requirements. As bandwidth and parallelism increase, timing does not scale linearly with system size — it multiplies.

Where SiTime sits in the NVIDIA AI stack

This is what differentiates SiTime from more narrowly exposed AI beneficiaries. Its products are not tied to a single silicon function or a single system choke point. They are required across multiple layers of modern computing systems — wherever coordination, synchronization, and signal integrity matter.

In practice, the NVIDIA AI stack is where these dynamics are most visible today, because NVIDIA defines the performance envelope of large-scale AI infrastructure. As GPU density, fabric speed, and cluster size increase, the entire surrounding system — servers, networking, optics, and interconnect — becomes more timing-sensitive. That is where SiTime’s products are pulled in.

In NVIDIA-based AI data centers, SiTime components appear adjacent to GPUs and accelerator cards as external reference clocks; on server motherboards supporting PCIe and memory subsystems; next to NICs and DPUs coordinating east-west traffic; inside switches and routers synchronizing high-radix fabrics; within optical modules, retimers, and active cables enabling 800G and 1.6T links; and across storage and networking systems that must remain phase-aligned under load.

Critically, these components are discrete, independent devices — not integrated into NVIDIA’s GPU or switch ASICs — allowing timing performance to scale independently as system requirements evolve. As NVIDIA increases GPU density, network speed, and fabric complexity, the demand for external, high-performance timing increases in multiple places at once.

This multi-angle exposure is structurally different from investing in GPUs alone, switch ASICs alone, or optical components alone. SiTime monetizes synchronization wherever traffic, parallelism, and coordination increase — across compute, networking, optics, and interconnect.

From heartbeat to full orchestra conductor: the Renesas acquisition and the complete timing stack

SiTime’s opportunity spans three layers of the timing stack, each with a distinct role: resonator, oscillator, clock generator, and system coordination.

Resonators are the fundamental heartbeat. Historically this has been quartz. SiTime’s Titan platform targets a long-term shift toward semiconductor-style MEMS resonators that can be integrated or co-packaged with silicon. Titan is strategically important but not near-term material, with meaningful revenue expected later in the decade.

Oscillators combine a resonator with active circuitry to produce a stable clock signal. This is SiTime’s core revenue engine today and the broadest market by volume. Oscillators appear almost everywhere — across AI data centers, communications, aerospace and defense, automotive, and industrial systems — and are often the entry point into customer designs.

Clock generators sit higher in the system architecture. They take a reference and synthesize or distribute multiple clocks across a system. These parts define how systems are synchronized and are chosen early in the design cycle. While lower in unit volume than oscillators, clocks carry higher ASPs and are growing faster, driven by AI networking and system complexity. Clock design wins can pull through oscillator revenue over time, typically with a 12–18 month lag.

The pending acquisition of Renesas’s timing business dramatically expands SiTime’s position in this upper layer. Before the deal, SiTime owned best-in-class MEMS resonators and oscillators with limited clocking IC capability. After the deal, SiTime will control the entire timing stack end to end — full clock generators including synthesis, distribution, buffers, jitter attenuators, and network synchronizers, plus complete system coordination solutions purpose-built for GPUs, switches, NICs, optics, and more.

This matters because AI systems do not just need a heartbeat — they need an orchestra conductor. As clusters scale to 1.6T-and-beyond optics, chiplets, denser fabrics, and exponentially more clock domains, the value shifts from individual timing components toward integrated solutions that coordinate the entire system. The Renesas deal positions SiTime to sell higher-ASP, broader-scope design wins at exactly the layer where AI infrastructure complexity is accelerating fastest. It widens the moat by making SiTime the only company that can offer a complete semiconductor-based timing platform from resonator physics through system-level clock coordination — a combination no quartz-centric competitor or analog incumbent can match.

Mobile timing: a real inflection, with two adoption paths

In mobile systems, SiTime’s Symphonic clock generator applies the same consolidation logic seen in data centers: replacing multiple discrete quartz clocks with a single programmable device serving modem, RF, and connectivity subsystems. Management does not name customers, but third-party teardown analysis makes the picture clear. The iPhone 16e uses SiTime MEMS-based timing solutions in its RF/modem/connectivity platform, whereas prior iPhone generations relied predominantly on quartz-based timing. That matters less because of the specific model and more because it signals an architectural break: a new cellular platform where timing is being redesigned rather than inherited.

The mechanism is straightforward. When Apple uses Qualcomm’s modem, Apple inherits Qualcomm’s RF platform assumptions, reference designs, qualification stack, and supplier ecosystem — including timing. The RF platform is “Qualcomm-shaped.” Once Apple builds its own modem, that constraint disappears. Apple owns the timing budget end to end and can optimize for board area, power efficiency, integration, environmental stability, and yield. That is exactly the environment where SiTime wins: fewer parts, more programmability, tighter integration, and better stability under real-world conditions.

It is important to keep scale and timing in perspective. The 16e is not the highest-volume iPhone, and the flagship iPhone line today remains largely anchored to Qualcomm-based modem platforms. However, this is not a speculative transition. Qualcomm’s CEO has publicly acknowledged that Qualcomm’s share of iPhone modems is expected to fall to roughly 20% in a future iPhone generation, implying that Apple’s internal modem platform is intended to become dominant rather than experimental. Platform transitions of this kind do not need to be instantaneous to be investable. What matters is directionality. Once timing architecture changes on one SKU, it establishes internal design precedent. As Apple expands its internal modem footprint over time, attached timing choices can scale with it.

The second adoption path is incremental and potentially larger: Qualcomm itself pulling up the timing stack. As cellular moves toward higher bandwidth, more carrier aggregation, tighter power envelopes, and ultimately 6G and direct-to-satellite connectivity, radios become less tolerant of phase noise, jitter, and drift. That increases the economic value of upgrading from collections of discrete quartz devices toward more integrated, higher-performance timing. A Qualcomm shift to a more cutting-edge TSMC node and a more integrated RF platform architecture is a natural moment when component choices get revisited: reference clocks, jitter cleaning, clock distribution, and BOM consolidation. When Qualcomm adopts integrated MEMS-based timing solutions as part of that next-generation platform refresh, the impact is not confined to iPhones — it scales across the far larger Android ecosystem that inherits Qualcomm platform decisions.

The 6G and satellite angle reinforces the same point. Higher frequencies and wider channels tighten jitter budgets. More complex scheduling and multi-radio concurrency increases the number of clock domains. Satellite links introduce Doppler and intermittent correction, which increases the value of stable holdover and timing resilience. The result is that timing becomes a system constraint rather than a commodity input across the handset market. That is why mobile and consumer is increasingly promising for SiTime even if management chooses to play it down: the timing architecture is being forced upward by physics and standards, not by marketing.

The moat: why SiTime is hard to replicate

SiTime’s moat exists because high-performance MEMS timing is not just a materials problem, an analog problem, or a software problem — it is all three simultaneously. The company has spent decades building proprietary MEMS processes, resonator designs, analog architectures, and internal simulation tools that do not exist off the shelf. Multiple large analog vendors and startups attempted to replicate this and failed.

The Renesas acquisition deepens this moat further. Owning the full stack from resonator through system coordination means competitors must now match SiTime not just at the component level but across the entire timing architecture. A quartz oscillator vendor cannot credibly offer integrated clock synthesis and distribution. A clock IC vendor without proprietary resonator technology cannot match the performance floor that MEMS physics provides. SiTime becomes the only vertically integrated timing platform company in the industry — a position that compounds in value as systems demand more tightly coordinated, higher-performance timing solutions.

This integrated capability allows SiTime to do what quartz-centric competitors cannot: scale timing performance as systems become denser, faster, and more distributed, while simplifying designs rather than complicating them.

Q1 2026 results: growth acceleration confirmed

SiTime’s first quarter of 2026 confirmed that the secular tailwinds described throughout this report are translating into financial results at an accelerating pace. Revenue reached $113.6 million, up 88% year-over-year and well ahead of consensus expectations. CED — the AI and networking-driven segment — grew 158% year-over-year, marking its eighth consecutive quarter of triple-digit growth. That consistency matters: it suggests durable infrastructure buildout rather than one-time inventory stocking.

Gross margins expanded meaningfully as product mix shifted toward higher-performance, higher-ASP solutions — exactly what one would expect as timing content per system increases and MEMS displaces quartz in more demanding applications. Non-GAAP operating margin reached 28%, demonstrating that operating leverage is beginning to emerge as revenue scales against a relatively fixed cost base.

Management guided Q2 2026 revenue to $140–$150 million, implying over 100% year-over-year growth at midpoint, and raised full-year 2026 expectations to at least 80% revenue growth excluding the Renesas acquisition. Book-to-bill remained strong, channel pull-through was healthy, and management commentary pointed to continued momentum across AI inference infrastructure, 1.6T optical deployments, and broader networking upgrades. The trajectory suggests that CED growth could sustain triple-digit rates through much of 2026, with mobile and automotive providing additional layers of growth as those design wins mature.

Valuation and risk framing

SiTime trades at extreme headline multiples by any conventional semiconductor metric. At roughly 100x EV/gross profit, the stock prices in sustained category dominance and a long runway of above-market growth. That multiple looks less irrational when set against the actual growth trajectory: revenue is accelerating toward triple-digit year-over-year rates, the addressable market is expanding as timing content per system rises, and the Renesas acquisition adds an entirely new revenue layer with higher ASPs and deeper customer embedment.

The core question is not whether SiTime is “cheap” — it is not, by any traditional lens — but whether the growth rate and durability of the opportunity justify the premium. A company growing revenue 80–100%+ annually, with gross margins expanding toward 65%, a clear path to 30%+ operating margins at scale, and structural moat advantages that deepen rather than erode with time, warrants a valuation framework rooted in long-duration compounding rather than near-term earnings multiples.

Risks remain non-trivial. Customer concentration is notable, with a meaningful share of revenue flowing through a small number of distributors and end customers. The Renesas integration must be executed cleanly to capture the full-stack synergies. AI infrastructure spending, while structurally growing, can be lumpy quarter to quarter. And at current multiples, any deceleration in growth — even a temporary one — would likely compress the stock meaningfully.

In short, SiTime is not just selling clocks. It is selling coordination — and with the Renesas acquisition, it now owns the entire coordination stack from atomic heartbeat to system-level orchestration. As AI pushes computing toward ever larger, faster, and more interconnected systems, precision timing becomes a quiet but indispensable layer of the stack — and that is where SiTime has positioned itself.

Summary

This report introduces and applies a single high-alpha investment heuristic: identify bottlenecks in the AI supply chain, distinguish the causers (constrained suppliers whose scarcity defines the bottleneck) from the solvers (rare vendors capable of resolving it), and invest in the solvers — companies that benefit from simultaneous volume growth, pricing power, operating leverage, and multiple re-rating.

In Wave 1 (2023–2024), bottlenecks migrated sequentially — from NVDA’s GPU supply to TSMC’s CoWoS packaging to HBM (where SK Hynix held near-exclusive allocation) to physical data center power and cooling. Wave 1 was defined less by clear causer/solver pairs than by a chain of scarcity that rewarded whoever sat at the constraint. Companies like NVDA, TSMC, and Vertiv all extracted extraordinary pricing power from their positions within the bottleneck. Vertiv’s case was especially instructive: starting at 12x earnings with low-single-digit growth, AI-driven ASP increases flowed through a fixed cost base to drive ~2.5x FCF growth, compounded by a P/E re-rating to 25–30x — the four-factor compounding the heuristic predicts — yet this was bottleneck beneficiary economics, not bottleneck resolution.

In Wave 2 (2024–2025), the causer/solver dynamic sharpened. NVDA became the bottleneck causer — Blackwell’s ~9–12 month delay (packaging respins, NVL72 reliability issues, SFU constraints) left the industry starved for next-generation compute. AVGO-designed custom ASICs emerged as the solver: Google’s TPUv7 shattered the “ASICs can’t train” consensus, powering Gemini 3 Pro past GPT-5 and cutting Anthropic’s token costs ~3x. MSFT’s Maia (MRVL) and AMZN’s Trainium underscore the cost of choosing the wrong design partner.

In Wave 3 (2025–2026), the bottleneck has migrated to optical interconnect. The causer is the InP supply chain — small-diameter wafers from suppliers like AXTI, low yields, and 18–24 month capacity lead times that cannot scale to meet explosive 1.6T transceiver demand. The solver is the SiPho supply chain — led by TSEM (the only foundry qualified at 1.6T) leveraging SOI wafers from Soitec and standard 300mm CMOS infrastructure to plan ~5x capacity expansion in 2026. AVGO, TSEM, and SiPho-enabled transceiver makers are positioned as the structural winners of the next infrastructure wave.

The Bottleneck Heuristic

In our previous report, AI Value Chain Framework — Part 1, we introduced the canal and dams analogy to illustrate how demand-supply imbalances migrate across the AI value chain — as one bottleneck eases, pricing power shifts to the next constrained segment. This report asks a complementary question: why do certain companies within bottlenecked segments generate 5–10x equity returns while others in the same segment barely outperform? The answer lies not in identifying the bottleneck itself but in identifying the rare vendor asymmetry within each bottleneck — the specific companies that can resolve the constraint while competitors cannot.

One high-alpha heuristic when it comes to investing in AI is to figure out what the bottlenecks are and who can address them. This is not a novel concept — bottleneck analysis is rooted in Eliyahu Goldratt’s Theory of Constraints — but applying it systematically to the AI investment landscape yields outsized returns because most investors are fixated on the demand side (who is buying AI, which models are winning) while neglecting the supply side, which is where scarcity-driven pricing power and margin expansion actually live.

Each major technology wave has required its own infrastructure buildout — fiber for the internet, cell towers and spectrum for mobile, hyperscaler data centers for SaaS. But AI breaks the pattern in scale entirely. Training a frontier model today costs north of $1bn in compute alone, and the enabling infrastructure runs far deeper than any prior cycle: purpose-built AI data centers, tens of thousands of GPUs and custom accelerators, high-bandwidth memory, ultra-fast interconnects, advanced chip packaging, and power generation and cooling systems measured in hundreds of megawatts. The entire value chain — from electricity generation to semiconductor fabrication to liquid cooling — is being stress-tested in ways it has never experienced.

This counters the prevailing paradigm where innovation in the world of bits and stagnation in the world of atoms dominated US industries for the past five decades. Since the 1970s, American industrial capacity in physical manufacturing has been systematically hollowed out, while software and internet platforms became the engines of GDP growth. The physical supply chain became Asia’s problem. US companies designed chips; TSMC manufactured them. US companies branded servers; Foxconn and Quanta assembled them. This offshoring of atoms worked when demand grew predictably at low single digits. AI is a sudden, violent shock to this paradigm — it requires a reacceleration across chips, packaging, memory, networking optics, power delivery, cooling, data center construction, electrical transformers, gas turbines, and raw materials. The breadth of physical infrastructure touched is historically unprecedented for a single technology wave — arguably more analogous to wartime industrial mobilization than to any prior tech cycle.

This creates an abundance of bottlenecks because existing slow-moving physical industries are categorically unprepared for a sudden, multi-fold demand increase. Lead times measured in weeks stretch to quarters; lead times measured in quarters stretch to years.

The heuristic is simple. If there is a bottleneck, it signals that demand has exceeded the industry’s structural capacity, that pricing power is shifting to the supply side, and that vendors who control scarce capacity can extract significant economic rent. If no vendor can ramp production quickly, we observe industry-wide ASP hikes with only mild volume growth. Customers then incubate alternative suppliers or engineer architectures that bypass the bottleneck — both responses that take 12–24 months, during which incumbents enjoy extraordinary economics.

But the best risk-reward arises when a rare breed of vendor can address the bottleneck while others cannot. This asymmetry creates explosive revenue growth, margin expansion, and multiple re-rating simultaneously. The vendor captures volume growth (market share gains), ASP expansion (pricing power from scarcity), operating leverage (fixed costs amortized over a larger base), and valuation re-rating (the market recognizes a “low-quality” business has transformed into a high-quality one).

Wave 1: The GPU and Data Center Bottleneck (2023–2024)

The tsunami of AI demand first hit GPU supply head-on. NVDA’s GPGPU became the primary bottleneck — predictable in hindsight but underappreciated in real-time because the prevailing narrative in late 2022 was still a semiconductor downcycle. The irony is that the demand wave arrived at the exact trough of the traditional cycle, with NVDA’s supply chain in partial capacity reduction.

As the only viable AI chip supplier at scale — AMD’s MI250 was generations behind in software ecosystem maturity, Intel’s Gaudi was lagging, and custom ASICs were nascent — NVDA commanded extraordinary premiums. Gross margins expanded from roughly 57% in FY23 to 75%+ in FY24, implying a shift from approximately 1.3x to 4x COGS markup on data center GPU products. This is an almost unheard-of margin structure for a hardware company shipping physical products at scale. The pricing power was a function of monopolistic scarcity in a category experiencing hyper-growth.

The second bottleneck was TSMC’s CoWoS advanced packaging capacity. TSMC had ample 5nm logic node capacity, but CoWoS — essential for integrating HBM stacks with the GPU die — was originally sized for a niche market at roughly 10–15K wafer starts per month. The H100 ramp required multiples of this. TSMC’s management deftly leveraged the position, raising CoWoS pricing and commanding higher blended ASPs, while its operating margins expanded to a record above 47% by 4Q24.

In memory, SK Hynix enjoyed the first wave of HBM demand almost exclusively. Micron had de-prioritized HBM development in the prior cycle — one of the most costly strategic missteps in semiconductor history — and by the time it had HBM3E samples qualified, SK Hynix had locked up priority allocation with NVDA. Samsung was behind on HBM yield and performance, with thermal and yield issues delaying qualification by multiple quarters.

At the cloud layer, providers who had secured GPU inventory monetized the bottleneck through surging hourly pricing and multi-year take-or-pay contracts. NVDA deliberately cultivated neo-clouds — CoreWeave, Lambda, Applied Digital, Crusoe Energy — granting them priority allocation to diversify its customer base beyond the hyperscaler oligopoly and preserve pricing leverage. CoreWeave grew from approximately $30m in revenue in 2022 to over $1.9bn by FY2024, essentially arbitraging the GPU bottleneck. Microsoft Azure was the biggest structural winner, having committed to large-scale H100 procurement as early as mid-2022 before ChatGPT made AI mainstream, and its revenue growth re-accelerated from mid-teens to 30%+ year-over-year.

As GPU clusters stood up, the bottleneck migrated upstream to physical infrastructure. The power and cooling industry had grown accustomed to low-single-digit growth and 6–12kW per rack power density. When Hopper arrived demanding 32kW per rack and Blackwell’s NVL72 hit at 120kW per rack, the industry experienced a 4x to 10x step-function demand shock. High-power-density racks, PDUs, UPS systems, and especially liquid cooling systems were in immediate, severe shortage.

Certain players — most notably Vertiv (VRT) — saw massive share price returns rivaling NVDA’s, appreciating roughly 8–9x from early 2023 to peak 2024. The mechanics were instructive: VRT started at 10–12x forward earnings with 30% gross margins and low-single-digit revenue growth. AI demand drove sudden growth acceleration and ASP increases that flowed directly to the bottom line through operating leverage. If revenue grows 40% and FCF margins expand from 10% to 18% (as incremental gross profit drops through a fixed cost base), absolute FCF grows roughly 2.5x. When the P/E then re-rates from 12x to 25–30x, the compounded share price impact is 5–6x. This four-factor compounding — volume, pricing, operating leverage, and multiple re-rating — is exactly what the bottleneck heuristic predicts.

A structural divergence also emerged between server OEMs and ODMs. Vendors like HPE and Dell, who outsource manufacturing and won hyperscale orders through aggressive pricing concessions, saw revenue growth but margin compression. The ODMs behind them — Quanta, Wistron, Foxconn — who perform the actual value-adding work of engineering design, thermal testing, signal integrity validation, and system integration, saw their capabilities become more valuable as AI pushed architectures toward dramatic complexity. The engineering value-add commanded increasing premiums, and Taiwanese ODMs saw meaningful margin improvement through 2024.

Wave 2: The Blackwell Bottleneck and Execution Gauntlet (2024–2025)

Massive GPGPU demand left NVDA both empowered and exposed. The strategic logic was straightforward: if customers remained GPU-starved for too long, they would invest more seriously in alternative architectures, and once those alternatives gained even a narrow foothold, they could expand over time. NVDA’s dominance depended on keeping the performance-per-dollar frontier moving fast enough that no competitor could close the gap.

This created something like the innovator’s dilemma in reverse. In Christensen’s classic formulation, incumbents fail because they rationally refuse to cannibalize their own profitable products, leaving the door open for cheaper, “good enough” alternatives. NVDA faced the same threat landscape but responded in the opposite way — racing to obsolete its current generation before competitors could catch up. Jensen Huang’s vision was that Moore’s Law at the chip level was slowing, but “system-level Moore’s Law” could continue if you co-optimized silicon, packaging, interconnect, cooling, power delivery, and software simultaneously. The ambition was enormous, but the execution proved extraordinarily challenging.

Blackwell had significant problems across multiple dimensions. Starting with the chip package: Blackwell was NVDA’s first chiplet-based GPU design, departing from the monolithic die approach of every prior architecture. The B200 consisted of two GPU compute chiplets connected via a high-bandwidth die-to-die link, packaged together with HBM stacks on a single substrate using TSMC’s novel CoWoS-L technology. CoWoS-L was still in early qualification at TSMC when Blackwell taped out — it used a local silicon interconnect bridge to provide ultra-high-bandwidth die-to-die connectivity, offering substantial advantages in package size, signal integrity, and bandwidth density, but was significantly less mature than CoWoS-S. The initial tape-out revealed signal integrity issues at the LSI bridge that caused yield problems, forcing both NVDA and TSMC to engineer a workaround involving mask revisions and process adjustments that consumed approximately 4–6 additional months. During this period, TSMC also had to simultaneously ramp CoWoS-L production from engineering-sample scale to high-volume manufacturing.

Beyond packaging, Blackwell’s system-level architecture represented a quantum leap in complexity. To deliver a scale-up domain of 72 GPUs (NVL72), NVDA designed a rack-level system consuming approximately 120kW — effectively 10x the power density of a previous-generation CPU rack. Nearly every power delivery component had to be redesigned for higher current and thermal loads. Liquid cooling became an absolute necessity, meaning all traditional air-cooled data centers were incompatible with NVL72 deployments — customers had to retrofit facilities or build entirely new ones, adding 6–12 months to deployment timelines. For networking within the rack, traditional passive copper cables could not reliably deliver the required bandwidth over NVL72’s longer node-to-node distances. NVDA adopted Active Electrical Cables — copper cables with built-in signal re-timing and equalization chips, where Credo (CRDO) was a major AEC supplier — as a compromise between passive copper and optical interconnects.

After manufacturing and engineering issues were resolved, a new problem emerged: operational reliability. In NVL72, all 72 GPUs share a single NVLink domain, meaning a fault in any individual GPU can propagate and take down the entire rack — a blast radius 9x larger than the previous NVL8 architecture. Early field data showed significantly more frequent job interruptions than expected. NVDA implemented a redundancy scheme where 8 out of 72 GPUs sit idle as spares, effectively wasting 11.1% of purchased compute capacity. For racks costing millions of dollars, this was a non-trivial economic penalty that narrowed Blackwell’s effective perf/$ advantage.

Then there was the chip itself. NVDA had compressed its product cadence from two years to annual, creating execution risk. While Blackwell’s tensor core compute roughly doubled versus Hopper, the Special Function Units — which handle transcendental math operations critical for softmax computations in attention layers — did not scale proportionally, becoming an on-chip bottleneck. Some workloads saw only 1.3–1.5x effective throughput gains, well below the 2x that raw FLOPS specifications suggested. NVDA had to address this with the GB300 SKU, which further complicated the product lineup.

With all these compounding issues, the practical reality was that the entire AI industry predominantly used Hopper for production workloads from 2023 through late 2025. Blackwell’s first meaningful production deployments did not achieve steady-state operation until 2H25, roughly 9–12 months behind NVDA’s original timeline. NVDA’s stock still performed well in absolute terms — Hopper continued to ship in massive volumes — but the execution challenges created a window of vulnerability that competitors exploited in what we term Wave 3.

Wave 3: The Custom ASIC Breakout and the Optical Chokepoint (2025–2026)

The TPU and Custom ASIC Ascendance

Google’s TPU has emerged as the dark horse of the AI chip landscape. We had been calling out custom ASICs’ potential as far back as 2023, and we noted TPUv7’s (Ironwood) impressive specifications — which on paper rivaled Blackwell in key training benchmarks and surpassed it in areas such as PPAC and scale-up networking — as early as 1Q25. But the consensus narrative dismissed custom ASICs as “inference-only” solutions. It was not until late 2025, when TPUv7 was fully deployed at scale and successfully trained Gemini 3 Pro, that this narrative cracked.

Gemini 3 Pro’s performance leapfrogged GPT-5 on multiple authoritative benchmarks — not a narrow victory, but a decisive one that reclaimed Google’s position as a tier-1 competitor in the frontier model race after a humbling 2023–24 period. Perhaps even more significant was that Anthropic’s Claude Opus 4.5/4.6 was also successfully trained on TPUv7 infrastructure, demonstrating that TPUv7 was not merely a Google-internal solution and delivering a roughly 3x reduction in Anthropic’s token costs compared to equivalent NVDA-based compute. A 3x cost advantage at the frontier is not incremental — it is a structural economic shift that changes the build-vs-buy calculus for every major AI lab.

This inflection occurred precisely when NVDA’s Blackwell deployment had been postponed by nearly a year. OpenAI’s GPT-5.2, the first major frontier model trained on Blackwell hardware, delivered performance that trailed Claude Opus and merely matched Gemini 3 Pro — a deeply disappointing outcome for what was supposed to be NVDA’s triumphant next-generation platform.

Discussion around custom ASICs — particularly those designed by Broadcom (AVGO) — has subsequently heated up dramatically. AVGO’s design partnership model, where it provides silicon design expertise, SerDes IP, packaging knowhow, and advanced interconnect technology while the hyperscaler provides architecture specifications and workload requirements, has proven to be the winning template. Google’s TPU has used AVGO as its silicon design partner for multiple generations.

The critical divergence was in design partner selection. Microsoft and Amazon made what we believe will prove to be a costly strategic error by opting for lower-cost alternatives. Microsoft worked with Marvell (MRVL) on Maia, while Amazon developed Trainium with a mix of Annapurna Labs engineering and external IP from MRVL, Alchip, and GUC. The rationale was straightforward — AVGO’s services are expensive — but this penny-wise, pound-foolish approach led to predictable consequences: postponed deployment schedules, lackluster system performance, and lower effective perf/$ when accounting for delays, respins, and suboptimal utilization.

Microsoft’s Maia 100 has seen limited internal deployment, and Azure’s AI training infrastructure remains overwhelmingly NVDA-based. Amazon’s Trainium 2 achieved roughly 80% of H100 performance at the chip level, but compensating for that gap at the cluster level required significantly more networking and power investment, resulting in a less capable system with higher total costs, immature software, and low reliability — exactly the dynamic Jensen Huang points to when he says buying something cheap doesn’t always mean cheap.

AMZN’s broader strategy of going cheap and in-house has compounded the problem. To drive adoption of its underperforming Trainium clusters, AMZN invested heavily in Anthropic and positioned it as the flagship user. Anthropic ultimately found that Trainium didn’t work and doubled down on TPU instead, prompting AMZN to invest another $50bn in OpenAI in an effort to fill idle Trainium capacity. Now, to address Trainium’s networking deficiencies, AMZN appears to be turning to NVDA’s NVLink chiplet — effectively going back to one of the two suppliers it tried to circumvent, and arguably choosing the least attractive option. NVLink is a closed ecosystem with full lock-in, and NVDA has little incentive to share its most advanced interconnect IP with a competitor’s compute die.

NVDA’s NVLink C2C I/O die is designed so that the highest-value part of the system — high-speed SerDes and networking — is delivered entirely by NVDA, with the customer having zero control. This gives NVDA an elegant way to compete with AVGO in custom ASICs, but it raises an obvious question: compared to simply buying NVDA’s off-the-shelf GPGPUs, what is the ROI of going custom? You get slower time to market, weaker integration, partial control over compute, and zero control over networking. And as AI architectures increasingly merge compute and networking, NVDA retains the ability to recapture on the networking side whatever margin it concedes on compute.

Meta, SoftBank, OpenAI, X.ai, ByteDance, and other major AI players have more recently started custom ASIC projects with AVGO as the design partner, learning from Microsoft’s and Amazon’s missteps. Custom ASIC programs have long development cycles — typically 2–3 years from engagement to volume silicon — so we have yet to see the ramp from these newer engagements. The pipeline, however, is enormous, underpinning AVGO’s thesis as a secular winner. AVGO’s management has guided to a custom AI accelerator revenue opportunity of $60–90bn by FY27 (ending Oct 2027), a figure the market initially dismissed as aggressive but now appears increasingly credible.

The Structural Logic of Custom ASICs

Our long-held belief has been that custom ASICs should represent an important and growing share of the AI chip landscape. For inference, the case is almost self-evident: a general-purpose GPU carries silicon area and power overhead dedicated to capabilities unnecessary for a specific, well-understood inference workload. A custom ASIC can eliminate this overhead, dedicating maximum die area and power to the exact compute primitives needed. The theoretical perf/$ advantage is 3–5x, and this gap only grows as inference workloads become more standardized. If model architectures evolve more slowly — which appears to be happening as the transformer proves remarkably durable — then the risk of an ASIC becoming obsolete before full amortization declines, further improving ROI. To counter this, NVDA must progressively reduce the “general purpose” die area of its GPUs, increasing specialization — a trajectory that converges toward something that looks increasingly like a custom ASIC, raising the question of why customers should pay the NVDA premium for a chip losing most of its general-purpose advantage.

But the paradigm-changing development is that TPUv7 proved custom ASICs can compete for training as well. Training imposes much higher requirements: the chip must provide not only ASIC-level compute efficiency but also sufficient flexibility for new model architectures, training algorithms, and parallelism strategies. Both the silicon and software stack must simultaneously deliver ASIC-like efficiency and GPGPU-like programmability — a combination considered a contradiction until TPUv7 demonstrated it empirically. Google achieved this through heavy investment in XLA, JAX, and years of co-optimization between hardware and model training teams, creating a hardware-software co-design flywheel that is extremely difficult to replicate.

This further threatens the NVDA thesis. If custom ASICs can deliver competitive training performance at substantially lower cost, then NVDA’s TAM is not merely being nibbled at the inference edges but is being attacked at the high-margin training core. NVDA’s response must be to double down on system-level differentiation — and one of the most consequential areas of that response is optical interconnect.

The Optical Interconnect Bottleneck

As AI clusters scale to tens of thousands and eventually hundreds of thousands of accelerators, the networking fabric becomes as important as the chips themselves. A cluster of 100,000 GPUs achieving only 30% MFU due to communication bottlenecks effectively delivers the same useful compute as 30,000 GPUs at 100% MFU — at over 3x the cost and power.

One of TPUv7’s key competitive advantages is its networking architecture. Google designed its TPU pods with a high ratio of optical connections per TPU and utilized custom optical circuit switches to deliver reconfigurable, efficient scale-out networks. Google’s OCS technology allows network topology to be dynamically reconfigured to match the communication pattern of specific training jobs, dramatically reducing the “network tax” on training throughput. This was a major factor in TPUv7’s real-world performance exceeding what raw FLOPS alone would suggest.

Google’s large optical transceiver orders in 1H25 surprised the industry and immediately pushed 1.6T transceiver supply into deficit. The industry had sized its 1.6T ramp for gradual uptake starting in late 2025, but Google’s orders alone consumed a disproportionate share of supply. In 2H25, after TPUv7’s resounding production success, Google doubled down with even larger orders. Other hyperscalers pursuing custom ASIC projects followed Google’s lead, recognizing that a high optical transceiver-to-chip ratio is essential for competitive cluster performance. New AVGO-partnered custom ASIC designs often specify even more aggressive ratios than TPUv7, reflecting the lesson that skimping on networking bandwidth is a false economy.

NVDA’s GTC 2026 announcement confirmed its own optical pivot definitively. In addition to traditional copper NVLink for scale-up within the rack, NVDA now supports optical cable connections within the rack and positions optical interconnect as the critical enabler for NVL576 (576 GPUs spanning multiple racks) and beyond. This represents a fundamental architectural shift — an implicit concession that copper, even with AEC, cannot scale to next-generation bandwidth densities. Before the official announcement, NVDA had already placed emergency orders with Innolight and Eoptolink in late 2025.

The EML Laser Bottleneck

With demand from Google, NVDA, and the broader custom ASIC ecosystem converging simultaneously on advanced optical transceivers, the supply chain has been overwhelmed. The bottleneck has localized to the EML (Electro-absorption Modulated Laser) chip, fabricated on Indium Phosphide (InP) wafers — the only proven commercial solution supporting the combination of 2km reach and 200 Gbps per lane modulation required for 1.6T transceivers.

There are only three major producers at the required scale and performance: Lumentum (LITE), Broadcom (AVGO), and Sumitomo Electric. All three are struggling to meet explosive demand — at least 5x growth from early 2025 to today. The supply response is severely constrained by InP physics and economics: small-diameter substrates (3–4 inch vs. 12-inch for silicon), inherently low yields, slower epitaxial growth processes, and 18–24 month lead times for new capacity. There is no ASML equivalent for InP — equipment is semi-custom and produced in low volumes. EML manufacturers enjoy substantial ASP increases (reportedly 20–40% year-over-year) coupled with only mild volume scaling, as every available production lot is spoken for. This is the classic bottleneck dynamic the heuristic predicts.

SiPho: The Bottleneck-Solver

The key investment opportunity is Silicon Photonics (SiPho). SiPho disaggregates the optical functions within a transceiver, moving modulators, photodetectors, and waveguides onto a silicon-based photonic integrated circuit while using a much simpler external laser source. This reduces InP laser chip demand per 1.6T transceiver by roughly 10x — the complex EML is replaced by a simpler continuous-wave laser, with modulation performed on the silicon PIC.

SiPho-based transceivers come with trade-offs: approximately 30% higher power consumption and shorter reach (500m–1km vs. 2km for EML). For many hyperscale deployments, however, these trade-offs are acceptable, especially within campuses where most links are under 500m. The critical advantage is that silicon PICs are fabricated on standard 300mm wafers using mature CMOS-compatible process nodes, with high yields and no special equipment requirements — scalability fundamentally different from InP.

The biggest winner from this SiPho transition is Tower Semiconductor (TSEM). TSEM operates the industry’s most mature and highest-performance SiPho process node, refined through years of development and customer qualification well ahead of other foundries. Importantly, only TSEM has demonstrated SiPho working reliably at 1.6T data rates; other foundries, including GlobalFoundries, have encountered persistent reliability and performance issues preventing 1.6T qualification. TSEM’s moat is substantial — silicon photonics manufacturing requires nanometer-level waveguide geometry control, low-loss optical-electrical coupling, and reliable germanium photodetector integration, capabilities that take years to develop and are not easily replicated. TSEM’s lead is probably 2–3 years over the nearest competitor.

2026 appears to be a breakout year. TSEM’s latest SiGe+SiPho solution demonstrates 400G per lane modulation for next-generation 3.2T transceivers, delivering approximately 30% power savings and better signal integrity compared to InP EML at 400G per lane — a reversal of SiPho’s traditional power disadvantage. If TSEM transitions these specifications into volume production, it will put immense pressure on InP EML leaders, who are not only capacity-constrained but locked into 200G per lane technology that may hit a physics wall at 400G.

The capacity expansion asymmetry is stark. InP EML makers continue to struggle with expansion — small wafers, low yields, limited equipment, long qualification cycles — with available capacity fully locked through 2026 and into 2027. TSEM, in contrast, plans approximately $900m of investment in 2026 targeting roughly 5x its current SiPho capacity, executable on a fundamentally faster timeline because it uses standard 300mm CMOS infrastructure. Transceiver makers who partnered early with TSEM on SiPho 1.6T — most notably Innolight, the world’s largest optical transceiver producer by volume — are positioned as downstream winners with a meaningful time-to-market advantage.

TFLN: The Wild Card

Another alternative route for optical modulators is Thin-Film Lithium Niobate (TFLN). Lithium niobate possesses exceptional electro-optic properties, and by creating thin films on silicon substrates, designers can fabricate compact modulators with potentially 10x better future bandwidth per lane compared to InP EML. TFLN has demonstrated 400G+ per lane capability in laboratory settings with superior power efficiency and modulation linearity.

However, TFLN’s production maturity remains limited. Interestingly, several Chinese vendors — backed by significant government funding — are aggressively pushing TFLN toward mass production, viewing it as a strategic opportunity to leapfrog Western optical component incumbents, similar to how YMTC leapfrogged in NAND. Combined with Broadcom’s recent launch of the industry’s first 400G per lane DSP, we could see Chinese-made TFLN-based transceivers — potentially from Innolight, which has deep Chinese manufacturing ties — entering the market against incumbent InP EML-based 1.6T transceivers on both performance and cost. This is a development that bears close watching, as it could further compress the TAM for InP EML and accelerate commoditization of the optical transceiver market.

Concluding Framework

The AI bottleneck heuristic produces its highest-conviction investment signals when three conditions are simultaneously met: the bottleneck is acute (demand exceeds supply by a wide margin), the bottleneck is durable (supply response takes years, not quarters), and there exists a specific vendor who can resolve the bottleneck while others cannot. When all three conditions are present, the resulting opportunity compounds across four vectors — volume growth, ASP expansion, operating leverage, and multiple re-rating — and can deliver 5–10x returns in compressed timeframes.

Critically, the heuristic sharpened as the AI buildout matured. In Wave 1, bottlenecks cascaded rapidly — from NVDA’s GPU supply to TSMC’s CoWoS packaging to SK Hynix’s HBM allocation to data center power and cooling — but the causer/solver distinction was blurred. Companies like NVDA, TSMC, and Vertiv were not resolving scarcity so much as sitting inside it, extracting extraordinary rent from their positions at the constraint. VRT’s 8–9x appreciation was a textbook case of bottleneck beneficiary economics: four-factor compounding driven by a demand shock hitting a fixed cost base. The returns were real, but the heuristic was descriptive rather than predictive — it explained the returns after the fact better than it identified them in advance.

In Wave 2, the framework gained predictive power. NVDA became the bottleneck causer through Blackwell’s execution delays, and a specific solver emerged: AVGO-designed custom ASICs, validated when Google’s TPUv7 shattered the consensus that ASICs could not train frontier models. The causer/solver pairing was identifiable in advance — the signs were visible in Blackwell’s packaging respins and NVL72 reliability struggles — and the solver’s equity returns were capturable by investors who recognized the dynamic early.

Wave 3 is where the heuristic should prove most actionable, because the bottlenecks are acute, durable, and the solvers are identifiable. In optical interconnect, InP’s fundamental scaling limitations (small-diameter wafers, low yields, 18–24 month capacity lead times) make it a structural causer, and TSEM’s SiPho platform — the only qualified foundry process at 1.6T — is the clearest bottleneck-solver in the current landscape. In memory, the demand arithmetic is staggering: a single Vera Rubin NVL72 rack specifies over 20 TB of HBM, 54 TB of DRAM, and 1.15 PB of NAND, and fleet-level extrapolations imply demand that could consume a substantial fraction of global production capacity. The big three memory vendors’ collective supply discipline has pushed margins to ~80% gross — 20+ points above prior cycle peaks — but the bottleneck-solver may prove to be YMTC, whose 3D NAND stacking achievements sidestep the EUV constraints that limit Chinese logic fabs. In PCBs and passives, M9 resin, 96-layer orthogonal backplanes, and high-end MLCCs are emerging as binding constraints where only a handful of suppliers — TTM, TDK, Murata — can deliver at the required specification. The energy bottleneck is forming but has not yet peaked; when it does, the causer/solver dynamic will likely pit legacy CCGT manufacturers (GEV, Siemens Energy) against lower-tier OEMs, alternative generation technologies, and the infrastructure operators who maintained aggressive procurement pipelines through the uncertainty of early 2025.

The bottlenecks are not static — they migrate through the stack as each constraint is relieved and the next binding constraint becomes apparent. The investor’s task is to identify the solvers before the market fully prices their scarcity value. In Wave 3, the names where causer/solver asymmetry is most favorable include TSEM, AVGO, and select memory, materials, and infrastructure vendors positioned to resolve constraints that their competitors structurally cannot.

In Part 3, to be published on Convequity.com later this week, we will extend this framework into specific stock-level analysis — including the memory supercycle, the evolving custom ASIC competitive landscape, and the energy infrastructure buildout — with the goal of identifying the next cohort of bottleneck-solvers before the four-factor compounding begins.

Summary

A three-decade “nuclear winter” of underinvestment has left the global uranium mining industry with a hollowed-out supply base incapable of responding quickly to surging demand. The transition to next-generation Small Modular Reactors is constrained by a severe shortage of High-Assay Low-Enriched Uranium (HALEU), creating a strategic chokepoint that favors domestic enrichment leaders like Centrus. Among reactor technologies, Sodium-cooled Fast Reactors currently offer the best balance of safety and fuel efficiency, while legacy light-water SMR designs have struggled to deliver on promises of speed and cost reduction. The “nuclear battery” approach led by startups like Oklo provides a faster, simpler deployment model for AI data centers, though China maintains a commanding lead in long-term materials science for molten salt reactors. Uranium itself represents the highest-visibility, lowest-risk play in the sector because all emerging reactor designs—regardless of which vendor wins—depend on a fragile and geographically concentrated fuel supply chain.

For the full three-part series on nuclear energy, please visit Convequity.com

The Reactor Landscape

NuScale appears to be the weakest SMR player by multiple measures. Despite being founded in 2007 and receiving the first NRC design certification for an SMR in the US, NuScale has yet to bring a single operating reactor to completion. Its flagship project with Utah Associated Municipal Power Systems was cancelled in November 2023 after cost escalation pushed the estimated price per MWh from $58 to $89. NuScale’s trajectory raises the central question confronting the entire SMR concept: can modular construction actually deliver faster deployment, lower cost, and reduced overrun risk? The provisional answer is discouraging. Even China’s CNNC—arguably the world’s most efficient nuclear builder—is taking roughly five years to construct its ACP100 “Linglong One” SMR, about the same timeline as a full-sized conventional reactor. For AI data center operators who need new capacity on far shorter timescales, this is simply too slow.

X-energy’s HTGR (the Xe-100) occupies an awkward middle ground. It offers genuine advances over light water reactors—high-temperature industrial heat capability and superior inherent safety through TRISO fuel—but it is not revolutionary. HTGRs lack the fuel breeding and waste-burning capabilities of fast-spectrum designs. On the metrics that matter most for nuclear economics—fuel utilisation, waste management, and coolant safety—sodium-cooled fast reactors are superior in essentially all dimensions, with a substantially longer operational track record. The one area where HTGRs hold a clear advantage is direct industrial heat application for steelmaking, cement, and hydrogen production. Outside that niche, the case for HTGR over SFR is hard to make. X-energy also faces a fuel supply constraint: the Xe-100 relies on expensive HALEU in TRISO form, creating greater supply chain uncertainty compared to SFR designs that can eventually transition to depleted uranium and recycled actinides.

Molten salt reactors remain furthest from commercialisation but may offer the most disruptive long-term potential. MSRs have a fundamentally simpler mechanical design—atmospheric pressure operation, no high-pressure vessel, potentially no solid fuel fabrication—but the binding constraint is materials science: developing alloys that can withstand decades of contact with hot, corrosive fluoride or chloride salts. Kairos Power, the most prominent US player in this space, sidesteps the hardest problems by using solid TRISO fuel with molten fluoride salt as coolant only. Commercial deployment is realistically mid-2030s at earliest. China holds a commanding lead here, having brought the world’s first operational thorium-based molten salt reactor to full power in June 2024. MSR development is fundamentally a materials science problem demanding large numbers of skilled researchers engaged in painstaking experimentation—a domain where China’s sustained investment over two decades gives it a significant advantage.

SFR: The Strongest Risk-Reward Balance

Sodium-cooled Fast Reactors emerge as the most attractive balance of technological risk and technological superiority. Fuel breeding and waste burning are not incremental improvements — they are paradigm-shifting capabilities that fundamentally alter the economics, sustainability, and waste profile of nuclear power.

TerraPower, backed by Bill Gates, takes a more conventional SFR approach. Its Natrium reactor employs a pool-type architecture broadly similar to Russia’s proven BN-800, with an added molten salt thermal energy storage system that repositions nuclear from pure baseload into a dispatchable technology complementing renewables. This is commercially and politically savvy framing, but the risk is one often seen in well-capitalised ventures: abundant funding and prominent backing leading to feature creep and slower execution. TerraPower received its construction permit for the Natrium demonstration plant in Kemmerer, Wyoming in 2024, targeting commercial operation in the early 2030s.

Oklo, backed by Sam Altman with investment from Chris Wright (now US Secretary of Energy), represents the most aggressive Silicon Valley-style approach. Its Aurora design eliminates external sodium piping entirely, instead using heat pipes embedded directly in the reactor core to conduct thermal energy to the power conversion system. Think of it like electronics cooling: a laptop uses passive heat pipes rather than pumped liquid cooling because the thermal load is small enough. Oklo’s reactor is the equivalent—small enough that passive heat pipe technology suffices, with no pumps or external piping. The result is a dramatically simpler design with fewer failure modes and a form factor suited to mass production. Oklo’s “nuclear battery” model—fuel loaded at installation, 20 years of operation without refueling, then core module replacement—is well-suited to AI data center requirements and distributed applications. Data center operators can add reactors incrementally as demand grows rather than committing to a single massive plant. Oklo’s first NRC application was denied in 2022 for insufficient technical detail; a revised application was accepted for review in 2024.

With the SFR designs covered, we’ve now surveyed the full landscape of advanced reactor contenders. The two tables below consolidate the field into a single reference—first on technical fundamentals, then on execution reality.

A few patterns worth noting: four of the five companies depend on HALEU fuel, a supply chain that barely exists outside Russia. The fast-spectrum designs just discussed (Natrium and Oklo) are the only ones with breeding potential—a long-term advantage, but one that lies years beyond initial operation. And on execution, no company has progressed beyond site preparation or non-nuclear construction. The entire sector remains pre-revenue and pre-proof.

For the full three-part report please visit convequity.com

The Fuel Play: Uranium as an Investment

The nuclear fuel supply chain represents the highest-visibility and arguably lowest-risk segment of the nuclear investment landscape. The thesis is straightforward: regardless of which reactor design wins, all of them need uranium fuel.

Several structural factors create secular bullish dynamics. The US lacks domestic spent fuel reprocessing, making it entirely dependent on primary mine supply and enrichment services. The US has inadequate domestic enrichment capacity and has historically relied on Russian services through TENEX—a strategic vulnerability where every plausible policy response (diversifying to Western enrichers, building new facilities, stockpiling, sanctioning Russia) tightens the uranium market. Raw uranium supply is geographically concentrated in Kazakhstan (43% of global output), Canada, Australia, Namibia, and Niger—a narrow and fragile base where disruption in any single jurisdiction would be difficult to offset quickly. And the entire mining industry has been in structural depression since the mid-1990s, with new mines requiring 7–15 years from exploration to production.

The roots of this crisis trace to a perfect storm. No new US reactor construction was initiated for roughly three decades after Three Mile Island. The “Megatons to Megawatts” program (1993–2013) flooded the market with weapons-derived uranium—approximately 500 metric tonnes of weapons-grade HEU converted into civilian fuel, meeting roughly half of US reactor requirements. This cheap supply arrived precisely when demand was evaporating, reactor decommissioning was accelerating, and arms treaties were reducing weapons production. No economic incentive existed to explore, develop, or invest in new uranium supply.

The tide has now reversed. Megatons to Megawatts ended in 2013. The decommissioning wave has run its course. China is building reactors at a pace not seen since France’s 1980s buildout. Western nations are planning SMR deployment. And AI energy demand has created political urgency around clean baseload generation. The supply-demand reflexivity loop that depressed the industry for three decades is reversing—and the supply side simply cannot respond quickly.

HALEU: The Critical Chokepoint

Most existing reactors use fuel enriched to 3–5% U-235. Many next-generation designs require HALEU, enriched to 5–20%. Without HALEU, these reactors cannot operate, and historically the only commercial supplier was Russia’s TENEX.

Centrus Energy (LEU) achieved the first production of HALEU on US soil in approximately 70 years in November 2023, at its Piketon, Ohio facility. But its demonstration cascade produces only ~900 kg/year. By 2030, demand could exceed 150 metric tonnes annually. Centrus’s current capacity provides less than 2.5% of even the near-term requirement. The company’s modular centrifuge design is technically scalable within ~42 months given adequate funding, but faces a chicken-and-egg problem: Centrus needs offtake contracts to justify building cascades, while reactor developers hesitate without guaranteed fuel. The US government is attempting to break this deadlock through ~$2.7 billion in dedicated appropriations for domestic enrichment expansion.

Investment Strategies

Physical Uranium. Sprott Physical Uranium Trust (SPUT) offers straightforward commodity exposure. The trust purchases and stores physical U₃O₈ without selling or offering in-kind redemptions—a structural feature that typically produces NAV discounts in bearish periods and premiums in bullish ones, while also serving as a meaningful demand source in the spot market.

Uranium Mining. The Sprott Uranium Miners ETF (URNM) provides broad sector coverage. Among individual miners, Kazatomprom (NATKY/KAP) is particularly compelling as the world’s largest producer with the lowest extraction costs via in-situ leach mining, high-quality deposits, and—critically—continuous operations and institutional expertise maintained throughout the nuclear winter. The geopolitical risk of Kazakhstan (landlocked between Russia and China) is the primary discount factor, though much of Kazatomprom’s quality may already be reflected in valuation.

Uranium Enrichment. Centrus Energy (LEU) is the only surviving domestic enrichment operator with production know-how, functioning centrifuge technology, and continuity of skilled labor. It is a key beneficiary of the DOE’s $2.7 billion enrichment expansion initiative. General Matter—a hard-science startup from Founders Fund applying the Anduril/SpaceX model to the nuclear fuel cycle—would be the preferred disruptive play, but it remains private.

Risks to the Thesis

Two developments could fundamentally reshape the uranium value chain if they materialize. Thorium fuel cycles, if MSRs achieve commercial viability, would make the entire enrichment industry progressively less critical—but commercialisation is realistically 15–20+ years away. Fast-spectrum breeder reactors could extract 60–100× more energy from the same quantity of mined uranium and could theoretically power the US for centuries using existing depleted uranium stockpiles alone—but this requires an entirely new reprocessing infrastructure representing decades of investment and hundreds of billions of dollars, in a domain that remains politically sensitive due to proliferation concerns.

At Convequity, we think about AI investment opportunities through a simple analogy: imagine the entire AI value chain — from energy generation through to the end user — as a long canal, divided into segments by a series of dams.

Demand is rampant. But the supply side is brutally capital-intensive with long lead times. That mismatch creates demand-supply imbalances that appear at different times and with different magnitudes across the chain. The dams represent these bottlenecks.

Right now, the most acute bottleneck sits at the uppermost segment: Energy & Power Distribution. Over the past year this dam has begun to lift — new supply is gradually coming online — but it’s far from fully open. Prices remain elevated. When supply does flow more freely and the water level in this segment falls toward balance, that water hits the next closed dam downstream — Data Centres — and the cycle repeats. Rising water, fresh imbalances, another round of price surges.

The real world is messier than a neat sequence of dams opening one after another. Major data centre backlogs are already building while energy bottlenecks are still substantial. The analogy isn’t meant as a literal timeline. Its value is as a conceptual framework for how the centre of gravity of pricing power ebbs and flows along the value chain, rather than segments rising and falling in isolation.

Two mechanisms make this interconnectedness explicit:

The backwater effect works upstream. With the upstream dam already open, water piles up against the closed downstream dam and pushes backward. In practice, this explains why a shortage in data centre capacity can keep energy component demand and prices artificially elevated even after new energy supply has unlocked — lower than if the energy segment’s own dam were still closed, but above a fully balanced level.

Pressure and seepage works downstream. Even with a dam closed, the sheer pressure of rising water forces subtle leaks into the next segment. This appears as early, softer price signals and demand reaching downstream players before the bottleneck fully opens.

We’re seeing both mechanisms play out clearly today. The Energy & Power Distribution dam is roughly one-third open — slowly unlocking new MW and GW supply. Meanwhile, the Data Centre dam remains firmly closed, with massive order backlogs building in colocation, power, cooling, networking, and racks. This closed dam is pushing backward to keep energy prices elevated upstream, while forcing seepage downstream into semiconductors and optical components.

Our Framework: Segmenting the Value Chain

For investors, the segmentation needs to strike a balance: granular enough to surface which areas look attractive, but broad enough that adjacent segments carry meaningfully distinct exposures and demand-supply dynamics rather than moving in lockstep.

We group companies that share similar market dynamics, industry exposures, and business models into the same segment. Within each segment, we’ve created subsegments where dynamics diverge meaningfully — for example, Networking & Optical has notably different characteristics to Servers, even though both live inside the Data Centre segment.

Our full AI Value Chain framework spans eleven segments and over forty subsegments (see table below):

1. Energy & Power Distribution — Nuclear, natural gas & midstream, renewables & storage, power equipment & electrical infrastructure, utilities & grid construction

2. Data Centre — Colocation & operations, networking & optical connectivity, servers & compute hardware, power & cooling infrastructure

3. Semiconductors — EDA & design IP, foundry & packaging, GPUs & accelerators, network & custom silicon, memory/HBM, wafer fab equipment

4. Cloud, AI Models, & Access — Hyperscale cloud, GPU/neo cloud, edge & CDN, independent & hybrid cloud

5. Data (SW + HW) — Storage hardware, databases & data platforms, observability & monitoring, data protection & governance

6. Agentic Execution – SaaS Decoupling from Seats — AI & agentic platforms, consumption & engagement platforms, transaction & commerce infrastructure, activity-indexed SaaS

7. SaaS – Still Seat First — HCM/HR & payroll, enterprise productivity & collaboration, vertical/specialised SaaS. Not part of the value per se but it is crucial to identify companies falling into this category in order to avoid them.

8. Consumer Applications — E-commerce & marketplaces, ride-sharing & gig economy, streaming & entertainment, travel & experiences, consumer fintech & vertical consumer

9. Physical AI — Autonomous vehicles & EV platforms, surgical & medical robotics, industrial & logistics robotics, semiconductors/sensors enabling physical AI

10. AI-Based Defense — AI-driven autonomous systems & drones, large-cap defence primes with AI integration

11. Space Tech for AI Data Centres — Satellite constellations & connectivity, launch & space infrastructure providers, direct orbital compute/lunar/in-space data plays

This structure lets us compare relative value between companies within segments, and compare median multiples across segments to identify which look expensive and which don’t — cross-referenced with knowledge of order backlogs and supply chain dynamics to assess whether elevated multiples are justified.

This article focuses on Segments 1 and 2 — Energy & Power Distribution and Data Centre — with some natural overlap into Segment 3 (Semiconductors) where the optical interconnect supply chain is concerned. These are the segments where the demand-supply imbalances are most acute today, where capital is flowing fastest, and where we believe the most asymmetric opportunities currently sit for investors willing to look beyond the obvious large-cap names. Future publications will work through the remaining segments of the value chain.

How We Screen for Mispricing

For screening, we use a tailored version of the Rule of X (an adaptation of the old Rule of 40), combined with our preferred multiple: EV/(FCF–SBC).

The standard Rule of X, as developed by Bessemer Ventures from studying SaaS investor returns, weights growth 2.3× more than profit:

Rule of X = (NTM Rev Growth × 2.3) + FCF Margin

We adapt it by subtracting stock-based compensation as a percentage of revenue. But here’s the critical step most people miss: applying a uniform 2.3× growth coefficient across all company types is naïve. A dollar of revenue growth is not worth the same across businesses with fundamentally different cost structures.

Think about it at the gross margin level alone. A dollar of growth for a software company at 75% gross margins is worth far more than a dollar for a hardware integrator at 15% gross margins. Bessemer’s 2.3× was derived from SaaS companies carrying 70–90% gross margins and simply cannot be applied wholesale to subsegments with structurally different economics.

What we ultimately want to capture is how investors value the next incremental dollar of revenue. Incremental dollars are valued through the incremental profit they generate. So for each subsegment, we estimate the incremental operating margin by assessing the split between variable and fixed costs. A subsegment with high fixed costs has a higher theoretical incremental margin — meaning each additional growth dollar becomes more valuable because a larger share drops to profit.

We then adjust for risk by dividing by a subsegment-appropriate cost of capital, and finally apply a “realism discount.” A theoretical incremental margin of 60% rarely means 60 cents of every growth dollar falls cleanly to operating profit. In practice, some portion gets absorbed by reinvestment or lost to inefficiency during rapid scaling. Established subsegments with long-standing operational expertise receive a more generous factor; newer or more volatile subsegments receive a more conservative one.

The end result is a growth coefficient specific to each subsegment’s cost structure, risk profile, and operational maturity. When we cross-check each company’s tailored Rule of X against its valuation multiple, relative value becomes significantly clearer. Two companies might both trade at the same multiple and post similar standard Rule of 40 scores. But after applying this process, one may carry a much higher tailored Rule of X — making it materially more attractive on a risk- and quality-adjusted basis.

It’s important to remember this is a screen, not a verdict. After diving into any highlighted company, it may become clear there’s a good reason the market prices it where it does.

Energy & Power Distribution: The Dam That’s Slowly Opening

When ChatGPT dropped in late 2022, hyperscalers kicked off a massive capex cycle. They quickly realised it wasn’t just about GPUs and facilities — they needed enormous amounts of power. The explosion in energy demand against limited supply created the first major bottleneck in the AI value chain, a serious constraint from 2024 through 2025/26.

The nuclear hype cycle. Nuclear initially emerged as the hottest option — carbon-free, reliable 24/7 baseload power, exactly what always-on AI infrastructure needs. The Trump Administration’s pro-nuclear executive orders in early 2025, including streamlined NRC licensing and eased environmental reviews, added fuel. Companies like Oklo — backed by Sam Altman and developing compact fast reactors designed as deployable nuclear batteries (up to 75 MWe per module, with 20-year fuel cycles and no refuelling) — became poster children for the thesis that modular nuclear could power AI campuses incrementally as demand grows. Upstream, Centrus Energy (LEU) emerged as the critical chokepoint: most advanced SMR designs require HALEU (High-Assay Low-Enriched Uranium), and Centrus operates the only NRC-licensed enrichment facility in the United States capable of producing it. With Russia’s TENEX sanctioned and domestic production covering less than 3% of even near-term HALEU requirements, virtually all fuel for the next-generation nuclear fleet flows through Centrus’s facility at Piketon, Ohio.

But by late 2025, reality set in. Deployment timelines proved stubbornly long — most new reactors and SMRs won’t deliver meaningful GW-scale supply until the 2030s, far too slow for the immediate data centre crunch. Meanwhile, SMRs in the 50–300 MW range began to look insufficient as hyperscaler campus designs scaled toward multiple gigawatts. Oklo’s revised NRC application was accepted for review in 2024 but remains a 2028+ revenue story at best. Nuclear stocks pulled back sharply as investors took profits on overvalued pre-revenue names and faced a lack of near-term catalysts — though the long-duration thesis (that AI’s power appetite will eventually exceed what gas and renewables can deliver) remains intact for patient capital.

The solar reality check. Attention swung to renewables, especially solar paired with battery storage. The catch is massive overbuild. To reliably power a 100 MW data centre around the clock, you can’t just install 100 MW of panels. You need enough to carry the full live load during daylight while simultaneously charging batteries at high speed for the 17+ hours of darkness. In practice, supporting a steady 100 MW baseload often demands building 400 MW or more of solar capacity. Factor in weather variability and the overbuild requirement grows further.

Space-based solar: the long-shot wildcard. It’s worth noting a far more ambitious approach to solar that sidesteps virtually every constraint facing terrestrial deployments. Elon Musk has outlined ambitions for launching upwards of 1.2 million satellites into orbit to serve AI’s power demands — effectively moving the entire energy generation problem off-planet. The logic, while audacious, is internally coherent. In orbit, there is no night: satellites in the right configuration receive uninterrupted sunlight 24/7, eliminating the overbuild and battery storage problem entirely. Waste heat radiates directly into the vacuum of space, removing the cooling bottleneck that plagues terrestrial data centres. There is no grid to queue for or build out — power is generated and consumed in the same location. Laser inter-satellite links through vacuum beat terrestrial fibre by roughly 47% on latency (light travels faster through nothing than through glass), which matters enormously for distributed AI workloads. And critically, Starship’s payload economics — dramatically reducing the cost per kilogram to orbit — make the launch cadence required for a constellation of this scale at least theoretically tractable in a way it never was before.

To be clear, this remains a long-duration, speculative thesis — we’re talking about orbital-scale infrastructure that doesn’t exist yet and faces enormous engineering, regulatory, and economic hurdles. But it’s worth tracking precisely because it addresses not just one but virtually all of the bottlenecks simultaneously: power generation, cooling, grid access, and interconnect latency. If even a fraction of this vision materialises in the 2030s, it could reshape assumptions about where the ultimate ceiling on AI compute actually sits. For now, terrestrial solutions — natural gas, ground-based solar with storage, and nuclear — remain the only game in town.

Natural gas as the pragmatic bridge. Gradually, both hyperscalers and investors realised natural gas could be the best short- to medium-term compromise. Not carbon-free, but roughly 50% less CO₂ than coal, and far more readily available. The top shale gas producers — EQT, Antero Resources (AR), Expand Energy (EXE), and Comstock Resources (CRK) — are the most direct beneficiaries. Several have already secured direct offtake agreements with hyperscalers like Microsoft, Google, Amazon, and Meta for dedicated supply to data centres, bypassing traditional intermediaries and giving them strong demand visibility through 2028 and beyond. CRK is particularly interesting as a contrarian play: its Haynesville acreage sits close to Gulf Coast demand, giving it a structural transport cost advantage, and in early 2026 its Western Haynesville site was selected for a major NextEra/Japan-backed power generation hub with potential offtake of up to 1 Bcf/d — a direct “speed-to-power” catalyst for AI.

The real bottlenecks, however, aren’t gas supply itself — much of it comes as a byproduct of oil production — but rather the midstream pipeline infrastructure to move it where it’s needed, and the generation equipment to convert it into electricity. Gas turbine lead times have stretched to three to four years amid surging orders, with GE Vernova and Siemens Energy the dominant manufacturers. Both are sitting on massive backlogs, and their ability to ramp production capacity is one of the key constraints determining how quickly natural gas translates into actual megawatts for AI data centres.

Throughout 2026, projects like the Blackcomb Pipeline are coming online to unlock takeaway capacity from the Permian, where producers are currently paying others to take associated gas because there’s nowhere to send it. Once that gas can flow efficiently to demand hubs, it can be purchased by power generators serving AI data centres. Rather than creating a glut, this infrastructure unlock should reduce market distortions and help lift overall gas pricing — a tailwind for producers, and a signal that the energy segment’s dam is slowly opening further.

Bitcoin miners as a bridge. Another interesting dynamic easing the near-term energy gap is Bitcoin miners (Bitdeer, Corz Scientific, Cipher Digital etc.) pivoting to AI. These operators already secured substantial power contracts, land, substations, and cooling infrastructure — often in remote locations with cheap or stranded energy. By repurposing their facilities for AI workloads, they’re effectively bringing online ready-to-use megawatts much faster than building greenfield data centres from scratch. Some run “mullet mining” (AI in the front, Bitcoin in the back), while others fully transition.

The longer-term energy mix is projected to shift gradually. Natural gas is expected to peak near 50% of AI data centre energy demand around 2027, then gradually decline as renewables with storage rise toward 40% by 2030 and nuclear edges up to 20% once reactor restarts and the first commercial SMRs add meaningful capacity after 2029–2030.

Behind-the-meter generation is increasingly how the largest campuses will be powered. Hyperscalers are building their own on-site gas turbines and renewables rather than drawing from the public grid — both because the grid can’t deliver power at GW scale fast enough, and because pulling gigawatts from local grids would inflate electricity prices for surrounding communities.

Bloom Energy (BE) is arguably the most immediately deployable solution in this space. Rather than waiting years for a utility to upgrade transformers and build new transmission lines, Bloom connects to existing underground natural gas pipelines — infrastructure that already has massive capacity headroom in most industrial areas — and generates electricity on-site using solid oxide fuel cells. It can deliver 50–100 MW in 90–120 days, collapsing deployment timelines from years to months. The technical advantages compound: its fuel cells produce DC power directly (AI chips run on DC, so traditional data centres lose significant energy through AC-to-DC conversion at every stage), run at roughly 60% electrical efficiency, and pack 100 MW into roughly one acre versus the hundreds of acres a firm solar-plus-battery installation would require. A $5 billion Brookfield partnership announced in late 2025 — structured as project finance — solved Bloom’s historical capital intensity weakness without dilutive equity raises, contributing to a record backlog and its first GAAP operating profit in early 2026.

That said, these behind-the-meter facilities are legally obligated to eventually connect to the broader grid, which means demand for grid buildout remains strong regardless. Companies like Quanta Services (PWR) and Siemens Energy — manufacturing and installing the substations, transformers, and transmission equipment needed to fulfil those connection requirements — benefit from both sides of the equation.

Data Centres: The Firmly Closed Dam

This segment — everything from the shell and colocation through to the power, cooling, networking, and servers inside — is currently the most visible bottleneck in the entire AI value chain.

AI workloads simply cannot run effectively in most existing data centres. Training and inference at scale require extreme power densities of 50–100+ kW per rack, compared with the 5–15 kW traditional facilities were designed for. They demand liquid cooling because air cooling cannot handle the heat from thousands of high-power GPUs working in parallel. Retrofitting legacy buildings is extremely difficult or outright impractical. Hyperscalers are forced to build largely new, purpose-built facilities from the ground up.

Colocation demand is exploding. New shells must be constructed and energised before anything else can happen — and the traditional leaders in this space occupy quite different positions relative to the AI buildout.

Equinix (EQIX) operates as the quintessential “carrier hotel” — its value proposition is interconnection density. Facilities sit in major metros at network peering points where carriers, cloud providers, enterprises, and content delivery networks physically meet and exchange traffic. Customers pay a premium for connectivity rather than raw power: the ability to cross-connect to hundreds of networks, use Equinix as a direct on-ramp to AWS, Azure, and GCP, and minimise latency between distributed services. Average power per cabinet is historically low because the model is optimised for many tenants in modest footprints, not a handful of massive GPU clusters.

Digital Realty (DLR) operates a different model — wholesale colocation at scale. Its facilities tend to be larger, often located in suburban or edge-of-city locations where land is cheaper and power more abundant. DLR’s customers are typically hyperscalers and large enterprises leasing entire halls or buildings rather than individual cages. It’s less about connectivity density and more about delivering large, contiguous blocks of power and space. That makes DLR structurally closer to what AI training requires — bulk power in big footprints — though even its legacy facilities weren’t designed for the 50–100+ kW per rack densities that today’s GPU clusters demand.

Neither is truly purpose-built for AI-scale density, but they face different challenges. DLR needs to push power and cooling far beyond its historical norms but at least operates the right form factor — large, power-oriented campuses. Equinix faces a more fundamental mismatch: AI training clusters don’t need to peer with hundreds of networks in a downtown building; they need concentrated megawatts in a purpose-built shell. However, Equinix should be considerably better positioned once inference fully matures and dominates total compute cycles. Inference workloads are smaller, more distributed, and highly latency-sensitive — they benefit from Equinix’s global metro footprint and interconnection density rather than requiring the concentrated GW-scale power that training demands. But that’s a later chapter in the AI story.

For now, the colocation providers capturing the most immediate value are those purpose-built for extreme power density — which is why converted Bitcoin miners (discussed earlier) hold such a structural speed advantage, bringing AI-ready capacity online in months rather than years while traditional players navigate permitting, construction, and fundamental facility redesign.

Power & Cooling providers (Vertiv, Schneider Electric, Modine) are sitting on multi-year backlogs. Their equipment carries lead times of up to four quarters. Much of the installation must happen on-site, adding friction. NVIDIA’s increasingly scale-out architectures — Hopper to Blackwell to Rubin — are amplifying the challenge, with each generation requiring higher power delivery precision, more sophisticated cooling, and tighter integration.

Networking & Optical may be the most acute bottleneck of all. The throughput needed for thousands of GPUs to communicate in a single cluster has rendered traditional copper cables unusable beyond a few metres. Rack-to-rack communications now rely heavily on optical interconnects. The bottleneck is already severe, and it will intensify when the industry moves to 3.2T optical modules — at those speeds, copper will only be viable for roughly two metres, meaning optical will be required even for intra-rack GPU connections.

Compounding this is a massive global shortage of indium phosphide (InP), the critical semiconductor material used to produce the lasers inside these modules. AXT Inc (AXTI) is effectively the pure-play upstream supplier here — one of very few companies manufacturing InP wafers at scale. Demand is surging, but wafer production capacity takes years to build, creating a genuine physical chokepoint that no amount of capital spending downstream can resolve quickly. Soitec sits on the other side of the architectural divide: it supplies Photonics-SOI substrates for silicon photonics platforms, which represent the industry’s primary path to scaling production of optical transceivers using mature CMOS-like manufacturing. As frustration with InP supply constraints pushes more optical integration onto silicon photonics architectures, Soitec’s addressable market expands — and it trades at a fraction of AXTI’s valuation relative to current earnings.

Further down the stack, POET Technologies (POET) is attempting to solve the problem one layer deeper — at the optical engine level. Its Optical Interposer, built on standard bulk silicon rather than expensive SOI wafers, uses passive alignment to snap laser and detector components into place via automated pick-and-place rather than the manual active alignment that breaks down at 800G+ speeds and AI-scale volumes. If the platform delivers on its promise, it offers a structural cost and scalability advantage for optical engines from 800G through 1.6T and into 3.2T — precisely the trajectory the industry is on.

The push toward CPO and NPO. Beyond simply scaling transceiver speeds, the industry is pursuing a more fundamental architectural shift: moving the electrical-to-optical (E/O) conversion as close to the switch’s ASIC as physically possible. Today’s dominant architecture uses pluggable transceivers mounted on the front panel of a switch, connected to the ASIC via long PCB traces. At 51.2T switch bandwidths and beyond, this model is hitting a wall — the SerDes power required to drive electrical signals across those traces consumes an increasingly absurd share of the total switch power budget, while signal integrity degrades with every millimetre of copper.

Co-Packaged Optics (CPO) integrates the optical engine directly into the ASIC package itself, eliminating the longest and lossiest electrical paths entirely. Near-Packaged Optics (NPO) places the optical engine immediately adjacent to the ASIC on the same substrate or board — not quite inside the package, but close enough to dramatically shorten the electrical reach. On-board optics (OBO) represents yet another permutation, mounting optical engines directly on the switch PCB rather than in front-panel cages. Each approach involves different tradeoffs around thermal management, serviceability, manufacturing yield, and cost — and the industry has not converged on a single winner. Broadcom, Marvell, and Nvidia are all developing variations, while startups like Ayar Labs pursue optical I/O at the chiplet level.

The practical impact is significant: CPO/NPO architectures can reduce I/O power consumption by 30–50%, enable higher-radix switches (more ports per switch, which matters enormously for AI fabric topology), and ultimately allow tighter, more efficient GPU cluster interconnects. Pluggables will remain dominant through 2028 — they’re field-serviceable, well-understood, and the supply chain is mature — but CPO/NPO begins meaningful ramp in the 2027–2028 window, particularly for purpose-built AI switches where power budgets are most constrained and the economics of custom integration justify the engineering complexity. This is another reason Soitec’s Photonics-SOI business has strong structural tailwinds: silicon photonics is the natural substrate for CPO/NPO engines, and every permutation of the architecture requires more of it closer to the compute.

There’s also the emerging demand from “scale-across” — where training runs and inference workloads are federated across multiple geographically distributed data centres, because single-site clusters are hitting hard ceilings on available power and permitting. This drives explosive demand for inter-data-centre coherent optical links spanning tens to hundreds of kilometres.

Servers (SMCI, Dell, HPE) are not the current bottleneck. These companies can scale assembly in factories rather than on-site, so they’ve kept up so far. But any energy unlock will dramatically increase demand for NVIDIA GPUs, putting fresh pressure on foundry capacity and downstream on server integration. Even if energy and GPU supply constraints ease simultaneously, the optical interconnect bottleneck could act as an independent cap on how quickly clusters can actually scale — thousands of GPUs are useless if they can’t communicate at the required throughput.

In canal terms, the Data Centre dam is firmly closed, and water levels are rising fast. More energy will not relieve this pressure — it will accelerate shell buildouts, which will pour even more demand into power & cooling, networking, and eventually servers. The imbalances inside Data Centres are only going to intensify.

Key Themes for Investors

Without getting into specific valuations (which shift week to week), here are the structural themes we think matter most:

Speed-to-power is the scarcest commodity. Any company that can collapse deployment timelines — whether that’s Bloom Energy generating on-site power from gas pipelines in 90–120 days, or converted Bitcoin miners bringing AI-ready capacity online in months — holds a premium position in this market.

Physical scarcity trumps capital scarcity. Some bottlenecks simply cannot be solved by throwing more money at them. InP wafer capacity takes years to build. HALEU fuel production has a single licensed domestic facility. Gas turbine manufacturing lines can’t be spun up overnight. In a market where hyperscalers are willing to write almost any cheque, the binding constraint increasingly isn’t capital — it’s atoms. Companies that control physically scarce inputs or manufacturing capacity hold pricing power that grows as demand intensifies around them.

The optical layer may gate the entire buildout. More attention goes to energy and GPU supply, but the optical interconnect layer — constrained by indium phosphide scarcity, the shift to 3.2T modules, scale-across demand, and the architectural migration toward CPO/NPO — could ultimately determine the pace of AI infrastructure expansion more than any other single factor. The convergence toward silicon photonics as the scalable substrate for these next-generation architectures has implications for which upstream suppliers see durable, expanding demand versus those facing eventual displacement.

Bottleneck resolution in one segment accelerates pressure downstream. Investors who buy the energy unlock thesis need to think about where that unlocked supply flows next. More available MWs mean more shells get energised faster, driving even higher demand for specialised power & cooling equipment, optical interconnects, and eventually servers and GPUs. The opportunity doesn’t disappear when a bottleneck eases — it migrates.

Training infrastructure today, inference infrastructure tomorrow. The current buildout is overwhelmingly oriented toward training — concentrated, GW-scale campuses optimised for raw power throughput. But as models mature and inference begins to dominate total compute cycles, the infrastructure requirements shift: smaller, more distributed, latency-sensitive deployments at the network edge. Companies poorly positioned for the training era (such as metro-dense interconnection hubs) may find themselves ideally positioned for the inference era — and vice versa. The sequencing matters for entry timing.

Integration complexity creates durable moats. Companies that function as prime contractors for complex on-site systems — integrating and qualifying deep supply chains under their own brand — hold a different kind of competitive advantage than pure component suppliers. That integration role is hard to replicate and tends to see backlogs grow rather than shrink as supply elsewhere improves.

The grid-bypass thesis is underappreciated. Behind-the-meter generation isn’t just a workaround — it’s becoming the default architecture for GW-scale campuses. Companies positioned in on-site generation and the legal-obligation grid connections that follow are exposed to demand from both sides.

At Convequity, our focus is on identifying the bottleneck resolvers — companies that don’t just benefit from AI demand in a general sense, but that actively relieve specific chokepoints gating the pace of the entire buildout. Bitdeer and other converted crypto miners resolving the speed-to-capacity bottleneck by repurposing existing power-dense infrastructure for AI. Bloom Energy collapsing deployment timelines from years to months by bypassing the grid entirely. Soitec offering the scalable alternative substrate path that relieves the optical transceiver industry’s dangerous dependence on constrained InP and EML supply chains. These are the kinds of asymmetric positions we look for — companies sitting at the narrowest point of a widening funnel, where the structural mismatch between surging demand and physically constrained supply creates durable pricing power and outsized growth.

This is Part 1 of our AI Value Chain overview. In the coming weeks, we’ll be publishing deeper dives into specific segments and individual names. If you want the full framework with current valuations and Rule of X rankings, that’s available to Convequity subscribers — reach out to service@convequity.com.

It has been a little over a week since Citrini Research released their masterpiece on the looming Global Intelligence Crisis — a real eye-opener that forced many of us to reconsider the terminal velocity of the AI-driven economy. Since then, we at Convequity have been working to build a quantitative bridge across the landscape they described.

The result of that exercise is the Agentic Displacement Model.

We want to be clear: this report is intended as an intellectual thought extension rather than a definitive blueprint. The global economy is a system of staggering complexity; predicting its exact trajectory over 25 years with high confidence is nigh on impossible. However, we believe this model serves as a vital eye-opener for understanding the structural tension between firm-level efficiency and macro-level demand.

We hope the projections found within this model are wrong. The results — even when using what we consider moderate inputs — forecast a future of structurally elevated profits alongside a massively widening wealth gap and a decimated middle class.

We present this not as a certainty, but as a framework for debate. To that end, we have included a link to the model itself, where you can select different inputs and calibrate the parameters to create your own alternative 25-year forecasts. We welcome any and all pushback regarding the robustness of the model’s logic and the calibration of its variables.

Macroeconomic Cohesion & The Logic of the Model

The structural tension we describe is rooted in a single, central premise: all demand ultimately terminates at the consumer. Firms do not generate demand — they relay it. Every B2B revenue chain, no matter how long, ends at a consumer spending income they earned as a worker. When household income contracts, every link in those chains feels it, regardless of where in the chain a firm sits.

For the more traditionalists among our readers, we have ensured this logic remains cohesive with standard national accounting. Whether viewing through the Income Method (the shift from wages to profit shares) or the Spending Method (C + I + G + (X - M)), the logic holds. In our framework, Investment (I) is treated as “derived demand”— rational firms do not invest in capacity without a consumer base to purchase the output — and Government (G) is a relay mechanism dependent on the tax base harvested from labor income and corporate profits. While the model simplifies these into a global “value-added” flow, the fundamental mechanics of GDP remain intact.

A recurring theme emerges: at the firm level, the economics of AI adoption look spectacularly positive—profit more than triples at peak. At the macro level, these same dynamics risk hollowing out the demand base on which all firms depend. The Agentic Displacement Model explains why these two truths coexist and why the firm-level signal dominates decision-making until the macro damage is irreversible. We recognize this is a bold simplification of a multi-trillion dollar system, and we welcome rigorous critique from those who see alternative stabilizers in the traditional GDP components.

Executive Summary

The Agentic Displacement Model describes a structural economic trap where individual firm rationality leads to collective macroeconomic collapse. It posits that while AI adoption creates spectacular short-term profits for companies, it simultaneously erodes the consumer demand base upon which all businesses ultimately depend.

The Core Thesis: The “Tragedy of the Commons”

The model is built on the principle that all B2B value chains eventually terminate at a human consumer spending earned income.

• Firm-Level Logic: It is privately rational for a firm to replace expensive human labor with low-cost AI agents to maximize margins.

• Macro-Level Reality: When all firms do this, aggregate household income collapses. Since one firm’s cost (wages) is another firm’s revenue (consumer spending), the mass displacement of workers destroys the market’s ability to purchase the very goods AI is now producing more efficiently.

Key Model Mechanics

The model uses several calibrated variables to track this transition over a 25-year horizon:

• Labor Categorization:

  • M (Revenue-generating): Initially grows as firms chase AI-driven opportunities, then declines as AI matures.

  • N (Non-revenue/Back-office): Rapidly displaced (25%+) within the first 5 years.

  • B (Blue-collar): Resilient until Year 3, then declines as robotics scales.

• Demand Erosion (D_macro): Demand doesn’t drop 1:1 with wages. It is governed by a 2-year lag (as workers exhaust savings/UI) and a demand amplification factor (σ = 1.58). This means every $1 of lost wages eventually destroys roughly $1.58 of economic demand due to Keynesian multipliers and “precautionary pullback” from remaining workers. This impact is filtered by the spending loss rate (k), where k * c. This accounts for income replacement (n) like unemployment insurance and the spending behavior (c) of the affected income cohort. k starts at 0.23, meaning only 23% of displaced wages initially leave the economy, but this grows as temporary buffers like savings and UI are exhausted.

• The Solopreneur Trap: The model rejects the idea that displaced workers will simply become “AI entrepreneurs.” It argues that a flood of “solopreneurs” leads to supply-side saturation (driving prices to zero) and demand-side closure (selling to a shrinking customer base).

The Three Phases of Displacement

Conclusion: The Paradox of Success

The model’s final state is a “chilling” paradox: Firms remain individually profitable even as the economy at large shrinks. By Year 25, labor’s share of revenue has collapsed from 55% to roughly 5.6%. While business owners and asset holders capture more of the “pie,” the pie itself is permanently smaller, creating a world of extreme inequality and potential social instability.

I. The Model

To access the model visit Agentic Displacement Model but bear in mind to fully understand the model it would be best to read the report here first.

Baseline (t = 0)

All quantities are expressed as cost units where baseline revenue R₀ = 100.

Labour share of revenue: 55%. Profit margin: 20%.

A 55% labour share may appear high relative to any individual firm’s income statement. At the firm level, intermediate inputs — materials and services bought from other firms — show up in cost of goods sold, making labour look like a smaller share of revenue. But when you consolidate the entire economy into a single firm, those costs cancel out: one firm’s input cost is another firm’s revenue. What remains is value-added, split between labour income and capital income. This is equivalent to GDP. Labour’s share of US GDP has been roughly 55–60% in recent years, making 55% a well-calibrated baseline.

Labour Categories

M — Revenue-generating white-collar (M₀ = 20). Sales, marketing, product development, client management. These workers directly drive revenue. M grows to 25 by Year 3 as firms hire into the AI-amplified opportunity, holds through Year 5 as the market saturates, then declines as falling revenue reduces headcount needs and AI handles M-type tasks directly. Floors at 5. (Full rationale: Appendix D.)

N — Non-revenue white-collar (N₀ = 20). Admin, operations, back-office, HR, finance, IT support. First to be displaced by AI agents. Reaches 6.7 by Year 3, floors at 2.0 by Year 5. The floor of 2.0 — 10% of baseline — represents the irreducible human roles: agent oversight, exception handling, regulatory compliance requiring legal personhood. (Full rationale: Appendix D.)

B — Blue-collar / physical labour (B₀ = 15). Manufacturing, logistics, maintenance, facilities. Unaffected until commercial robotics becomes viable at Year 4, then declines as AI in robotic form factors scales. Floors at 1.5 by Year 13. (Full rationale: Appendix D.)

Revenue

γ_t is the supply-side revenue multiplier. It reflects the firm’s growing capacity to generate revenue as AI agents make revenue-generating employees (M) more productive. Because each M worker now produces more, the firm hires additional M workers and total revenue-generating capacity rises. γ_t ramps from 1.0 to γ_peak = 1.5 at Year 5 and holds thereafter — AI agents maintain this elevated capacity even as other roles are cut.

D_macro(t) is the demand-side multiplier. It captures whether consumers can actually afford to buy what the firm can now produce. As displaced workers lose income, aggregate household spending falls, and D_macro(t) declines from 1.0. After Year 5, γ_t is flat — the firm’s production capacity stops growing. From that point on, all revenue movement comes from D_macro(t): consumer demand shrinking beneath a fixed supply capacity.

Demand Erosion

D_macro is the fraction of the firm’s original addressable demand that still exists at time t. It starts at 1.0 (no displacement, no demand loss) and falls as displaced wages work their way through the economy into reduced consumer spending. The formula asks a simple question: if wages have fallen by Δ, how much of that lost income would have been spent on goods and services the firm sells? The answer is governed by four things — (1) how an initial spending loss cascades and compounds as it propagates through the economy (σ), (2) what fraction of displaced wages translates into an initial spending loss, after accounting for replacement income and the cohort’s spending behaviour (κ), (3) how much aggregate wage income has been lost (Δ), and (4) when the spending contraction materialises (τ). When D_macro reaches, say, 0.85, the firm’s revenue ceiling has permanently shrunk by 15%, regardless of how much AI-driven capacity it has added.

Spending amplification factor (σ = 1.58). σ decomposes as σ = μ_K + ψ, capturing two distinct channels through which an initial spending loss is amplified as it propagates through the economy. σ is a pure amplification factor: it takes the initial spending loss (determined by κ) and outputs the total spending contraction that results.

• Keynesian spending multiplier (μ_K = 1.40). The Keynesian multiplier is above 1, reflecting the fundamental mechanism that spending cascades amplify an initial shock. A displaced analyst stops buying lunch at a nearby sandwich shop, the shop cuts an hourly worker’s shifts, that worker cancels a gym membership, the gym reduces front-desk hours, and so on — each successive round is smaller than the last because businesses absorb some lost revenue through lower margins rather than cutting costs dollar-for-dollar. That is what “multiplier” means: the cumulative spending contraction exceeds the initial spending loss that triggered it. Empirical estimates for spending multipliers in conditions with economic slack — precisely the conditions that mass displacement would create — consistently fall in the range 1.2–1.8 (Auerbach & Gorodnichenko 2012 estimate 1.5–2.0 in recessions; Blanchard & Leigh 2013 estimate 0.9–1.7 with a central tendency around 1.5; Chodorow-Reich 2019 finds local fiscal multipliers of 1.7–1.9; Ramey 2019 surveys the literature at 0.6–1.5). We use 1.40, a moderate value reflecting that AI displacement is gradual rather than a sudden demand cliff, but that the affected economy has significant slack.

• Wage-suppression spillover (ψ = 0.175). Non-displaced workers experience modest wage compression as displaced workers flood adjacent roles, increasing labour supply and giving employers leverage to suppress pay. This is a purely supply-side mechanism — distinct from the demand-driven cascade captured by μ_K — and affects even workers in sectors not directly exposed to AI displacement. The resulting income loss for non-displaced workers reduces their spending, compounding the initial demand shock. The effect falls in the range 0.10–0.20; we use 0.175.

Combining: σ = μ_K + ψ = 1.40 + 0.175 = 1.575 ≈ 1.58. σ is a pure amplification factor: it takes the initial spending loss and outputs the total demand contraction that results. (Derivation: Appendix A.)

Spending loss rate (κ = 0.23 in Year 1). κ translates displaced wages into the initial spending loss that σ then amplifies. Two factors govern this translation: how much of the displaced income is actually lost to the household, and how much of that truly lost income would have been spent. κ captures both:

• Income replacement (η = 0.65). Displaced workers do not lose all their income. Unemployment insurance, government transfers, early pensions, savings drawdowns, and partial re-employment replace a significant share. Empirically, η falls in the range 0.60–0.70 for the affected cohorts in the early years, meaning 60–70% of lost wages are offset by other income sources. The net income actually lost — the income that could potentially feed a spending contraction — is (1 − η) = 0.35.

• Cohort spending rate (c = 0.65). Not all truly lost income would have been spent. c captures the fraction of each dollar of net income loss that the displaced cohort would have directed toward consumption — the share that actually enters the spending economy as a demand impulse before any multiplier amplification. For the white-collar cohort, which skews heavily toward upper-income households, Consumer Expenditure Survey data shows total expenditures of approximately 65% of before-tax income for top-quintile households. This is lower than the economy-wide average because high earners direct a larger fraction of income toward savings and investment — a well-established empirical regularity confirmed across decades of CEX data. c = 0.65 means that of each dollar of income truly lost to the household, 65 cents would have entered the spending economy as consumption. The remaining 35 cents would have been saved or invested, and its loss reduces asset accumulation but does not directly contract consumer demand. c replaces the need for a separate marginal propensity to consume parameter — it is the cohort-specific spending rate, grounded in observable expenditure data for the income quintile most exposed to AI displacement.

Thus κ = (1 − η) × c = 0.35 × 0.65 = 0.2275 ≈ 0.23 in Year 1. This means that for every dollar of wages displaced, only 23 cents of consumer spending is initially lost — and it is these 23 cents that enter the amplification cascade captured by σ.

η is the sole time-varying component of κ, and its trajectory is the critical driver of the model’s escalating severity. η depends on transient buffers with finite lifespans. UI benefits expire (typically within 6–12 months). Savings are drawn down. Early pensions erode against inflation. And re-employment becomes progressively harder as displacement widens across the economy — there are fewer remaining roles to absorb the displaced, and the roles that do exist face wage compression from oversupply.

Some displaced white-collar workers will find new employment, but often by moving downmarket — accepting blue-collar roles at lower salaries, shifting to hourly or shift-based work, and competing for positions well below their prior income level. This is already reflected in the model at two points. First, η already accounts for it: the realistic opportunities for white-collar workers to outcompete incumbent blue-collar workers are limited, and the income replacement those roles provide is sharply below prior wages, which keeps η from recovering meaningfully. Second, the spillover pressure — displaced white-collar workers flooding into blue-collar labour markets and suppressing wages for workers who were never directly displaced by AI — is captured in the wage suppression component (ψ) within σ. The downstream demand effect of that crowding is already in the multiplier, not in κ.

These are not permanent cushions; they are shock absorbers with a defined lifespan, and as they deplete, more of each displaced dollar shows up as actual spending loss. c, by contrast, is structural — it describes the spending behaviour of the income cohort before displacement, a fixed characteristic of their position in the income distribution, not of their post-displacement circumstances. The time-varying work is done entirely by η.

In the short run, savings buffers, UI benefits, and transfer programmes hold income replacement high (η ≈ 0.70, κ ≈ 0.20). Over time, those cushions exhaust. By the later years of the model horizon, η drifts toward 0.40, pushing κ toward 0.39. We use κ = 0.23 as the Year 1 value and allow it to rise over the horizon. (Full time-varying κ_t schedule: Appendix E.)

Cumulative wage displacement (Δ_{t−τ}). Δ_t = W₀ − W_t measures total wages lost across all three labour tiers at time t. The lag τ = 2 years reflects the delay between job loss and spending contraction — displaced workers exhaust savings, unemployment benefits, and job-search efforts before their spending declines materially. Demand erosion at time t therefore reflects displacement at t − 2, creating a window in which firm-level benefits are visible before macro-level costs arrive.

Baseline revenue (R₀). R₀ anchors the demand loss as a fraction of the firm’s original revenue base, converting the dollar value of lost spending into a share of starting revenue. In the early phase, R₀ may actually rise: AI-augmented output is more productive per unit of cost, firms are actively scaling their M investment, and the spending lag means displaced wages have not yet translated into visible demand declines. The model captures this — the erosion is measured against R₀, not against a declining baseline, which means the initial demand loss registers as a small fraction of a growing top line. This is part of what makes the trap invisible: revenue is improving at precisely the moment the displacement that will later undermine it is accelerating.

Together, σ × κ = 1.58 × 0.23 = 0.36 in Year 1: for every unit of wages displaced, roughly 0.36 units of consumer demand initially disappear — rising toward 0.61 as income replacement erodes. κ is the critical variable determining how much spending power the economy retains after job loss. Its moderate initial value is the reason the model does not collapse immediately, and the reason the erosion is slow enough to remain invisible to firm-level decision-makers during the exact period when they are making their largest AI adoption commitments. Its rising trajectory is the reason the damage eventually becomes unavoidable.

The Solopreneur Saturation Trap

A common counterargument holds that κ should be lower — that displaced workers will use AI itself to become entrepreneurs, building apps, digital services, courses, and automation tools that replace their lost income. This argument underpins much of the optimistic narrative around AI democratisation (e.g., Jensen Huang’s framing that AI “empowers” people to create).

The model treats this as largely illusory at scale, for two reinforcing reasons.

The first is supply-side saturation. When billions — not millions — are forced into digital entrepreneurship not out of passion but necessity, the result is not job creation but economic dilution. In capitalism, economic value accrues to scale: CEOs earn outsized compensation because their decisions impact vast customer bases; software developers command high salaries because their products serve thousands or millions. If billions can produce similar digital products using the same AI tools, market overlap and saturation inevitably occur. The average product or app may reach only 100 users, causing the economic value of each creator’s output to diminish dramatically. A handful of solopreneurs will build breakout successes; the vast majority will generate income far below their prior wages — possibly below subsistence level. If anything, the flood of near-zero-marginal-cost digital products accelerates price deflation in the digital goods market, further compressing the income available to each creator.

The second is demand-side closure, and it is the more fundamental problem. Solopreneur revenue is not exogenous — it depends entirely on consumer spending, which is exactly what the model shows eroding. Unless solopreneurship fully replaces all displaced wages across the economy, aggregate demand is contracting. The solopreneurs who do manage to establish themselves are selling into a shrinking customer base at the same time the number of competing sellers is exploding. The trap closes from both sides simultaneously: oversupply compresses prices and market share from above, while demand erosion removes customers from below. Even a successful solopreneur who carves out a niche in Year 1 will find that niche contracting in Year 3 — not because their product worsened, but because their customers are themselves experiencing the same displacement pressures. This is not a escape route from the demand erosion the model describes; it is endogenous to it.

This dynamic means AI-enabled entrepreneurship provides minimal aggregate cushion to κ. Some displaced workers will carve out a reduced living; most will not. The entrepreneurship narrative’s primary macroeconomic effect is to delay policy intervention by sustaining the illusion of opportunity while the demand base continues eroding. The model’s η = 0.65 already generously accounts for partial income replacement through all channels — including any solopreneur income. The solopreneur channel does not materially alter this.

AI Infrastructure Costs

Note that Δ_t⁺ differs from the Δ_t used in D_macro. D_macro uses aggregate net displacement (Δ_t = W₀ − W_t) because consumer demand reflects total compensation flowing to households — salary and benefits that workers can actually spend. A_t uses clamped per-category displacement because AI costs are incurred against each role replaced — in the early years hiring additional sales personnel (revenue-generating) does not reduce the compute bill for replacing IT and HR personnel (non-revenue-generating).

α = 0.10 is the Year 1 value: the all-in cost of running AI agents — inference compute, orchestration infrastructure, human oversight, security and compliance — expressed as a fraction of the compensation (salary + benefits) of the humans they replace. Current per-agent costs are approximately $6,500/year against average displaced-worker compensation of ~$80,000/year (base salary ~$65,000 plus benefits ~$15,000), giving a raw per-agent ratio of 0.081. The Year 1 value of 0.10 includes a buffer for transition friction — parallel running of human and AI systems, integration overhead, and heavier human oversight during ramp-up — that absorbs over the first 12–18 months as deployments mature. Note that compensation, not fully-loaded employer cost, is the correct baseline here: office space, equipment, and facilities overhead are real savings to the firm but do not flow to households as spendable income, so they do not enter Δ and should not inflate the denominator of α.

Because compute costs per inference token fall approximately 100× every 18 months — net of Jevons Paradox, a 30–50% annual decline per agent as agents consume vastly more tokens as costs fall — α declines over the model horizon. Once transition friction is absorbed, the ratio settles to its raw value (~0.08) before tracking compute deflation downward toward 0.04 by Year 20. The declining-α trajectory and its implications for AI industry demand are developed in Section IV. (Full derivation: Appendix B.)

The Complete Formula

Everything developed so far — AI-driven capacity gains, labour displacement, demand erosion, and infrastructure costs — feeds into a single quantity: the firm’s profit at time t.

The five terms are, in order: potential revenue from AI-amplified capacity; the demand-erosion penalty; labour costs; fixed overhead; and AI infrastructure costs. The structural tension of the model is embedded in the relationship between the third and second terms — reducing W_t (which increases profit directly) simultaneously increases Δ (which reduces profit via demand erosion, with a lag). The model’s central finding is that the third term always dominates the second at the individual firm level, ensuring displacement remains privately rational even as it is collectively destructive.

The reason is structural, not behavioural — no amount of foresight or goodwill changes the incentive. The firm captures 100% of its own wage savings when it reduces W_t, but bears only a fractional share of the demand erosion it causes. In an economy of thousands of firms, any single firm’s displacement is a rounding error in economy-wide Δ. The demand-erosion penalty in the second term is levied on the entire economy’s cumulative displacement, not the firm’s own, so the firm’s contribution to its own revenue loss is negligible. This is a classic tragedy of the commons: each firm rationally cuts its own wage bill, but every firm’s customers are other firms’ workers. No individual firm has reason to stop.

The lag τ compounds this asymmetry. Because demand erosion arrives two years after the displacement that caused it, the causal link never appears in any normal planning cycle. The firm sees immediate margin improvement from cutting W_t today. The revenue softening that follows is diffuse, delayed, and practically impossible to trace back to its own hiring decisions. By the time the demand penalty materialises, it looks like a market downturn — and the rational response to a market downturn is to cut costs further, which means more displacement.

The time-varying nature of κ makes this worse, not better. In the early years, high income replacement (η ≈ 0.65) keeps κ low at approximately 0.23, and the demand penalty small — reinforcing the perception that displacement is free. This means that the actual revenue damage barely registers in anyone’s numbers during the exact period when firms are making their largest AI adoption commitments. By the time κ rises toward 0.39 as safety nets exhaust and η drifts toward 0.40, the demand erosion becomes obvious — but by then, displacement is deeply embedded across the economy and there is no path back.

When summed across all firms, the arithmetic flips. Each firm’s displacement is negligible in isolation, but the sum of all firms’ displacement is Δ — the very quantity that drives D_macro toward zero. Every firm’s decision to displace is individually reasonable, and yet it is precisely the accumulation of those reasonable decisions that destroys the demand base they all depend on. No coordination mechanism exists to prevent this, because no firm can credibly commit to forgoing wage savings that its competitors will capture regardless. The model does not require irrationality, short-termism, or ignorance to produce demand collapse. It requires only the ordinary logic of competitive markets.

Parameters

γ_t follows a firm-level adoption trajectory toward γ_peak. κ_t rises over the horizon as η_t declines. W_t and Δ_t are derived from the underlying wage schedules. α declines as AI unit costs fall.

II. Three Phases

Phase 1 — The Agentic Boom (Years 0–5). All three labour categories begin declining simultaneously from Year 0, but at different rates reflecting the sequencing of AI capability. B is displaced most aggressively — falling from 15.0 to 7.0, a 53% reduction — as physical and routine tasks prove most amenable to early automation and commercial robotics. N declines from 20.0 to 14.5, a 28% reduction, as agentic AI begins displacing knowledge workers. M also declines from 20.0 to 16.7, a 17% reduction, as the AI tools that M oversees become progressively more autonomous — reducing the need for human oversight even in the early stages. Revenue and profit surge. Demand erosion is faint — the 2-year lag means early displacement has not yet fed through (D_macro remains at 1.000 through Year 2), and κ is at its lowest (≈ 0.23) because safety nets are fresh and income replacement is high, giving an effective erosion rate (σκ) of roughly 0.36 per unit of displaced wages. The early economics of AI adoption look unambiguously positive. At the firm level, profit more than triples by Year 5 — from 19.9 to 71.6 — the single strongest financial signal any executive has ever seen from a technology adoption. No rational actor reverses course.

Phase 2 — Peak and Early Erosion (Years 5–10). Revenue crests at ~135.7 around Years 6–7, then begins a sustained decline as demand erosion overtakes productivity gains. B approaches and reaches its floor of 1.5 by Year 10 — physical-task displacement is essentially complete. N displacement accelerates — falling from 14.5 to 6.7 — as agentic AI matures and broadens its reach across knowledge-work domains. M continues its steady decline from 16.7 to 13.3. κ climbs from 0.2600 to 0.2925 as unemployment insurance exhausts and income replacement ratios fall, and the Keynesian cascade (μ_K) and wage-suppression spillover (ψ) compound each incremental rise — a one-percentage-point increase in κ produces a 1.58-percentage-point increase in effective erosion. Demand erosion intensifies as the cumulative weight of displaced wages feeds through the lag at increasingly punitive conversion rates; σκ rises from ~0.41 to ~0.46. D_macro falls from 0.955 to 0.879. Revenue falls from its peak but remains 30.5% above baseline at Year 10. Profits actually rise from 71.6 to 84.0 — now more than quadrupled from baseline — because costs are falling faster than revenue. Margins expand from ~53% to ~64%. By Year 10, the firm-level picture remains unambiguously positive even as the macro foundations are visibly cracking. This is the fundamental asymmetry of the model — the period of maximum firm-level optimism coincides precisely with the period of maximum irreversible macro-level commitment to displacement.

Phase 3 — Deep Erosion and New Equilibrium (Years 10–25+). Revenue declines from 130.5, shedding ground each year but remaining above baseline throughout. N reaches its floor of 2.0 by Year 15. M — the last labour category still being displaced — reaches its floor of 5.0 by Year 20. κ reaches its peak (≈ 0.39) as safety nets are largely spent and η falls to its floor of 40%, pushing σκ toward 0.62 — meaning each unit of displaced wages now destroys nearly two-thirds of a unit of demand. The feedback loop — displacement → demand erosion → revenue decline → further displacement — decelerates as displacement exhausts its remaining scope. The economy converges to a new steady state: revenue at ~108.4 (Year 25), still ~8% above baseline but radically below potential; a workforce wage bill of 8.5, down 85% from its original level of 55.0; and structurally elevated profits of ~74.9, still ~276% above baseline. Margins peak at over 70%.

This last point is the model’s central paradox — revenue growth has been permanently arrested and is contracting toward a floor, but the wage bill has fallen far more, so margins remain radically elevated. Firms are individually profitable in a permanently diminished economy.

The distributional consequences are severe. Radically more profit accruing to business owners while working- and middle-class incomes plummet magnifies populism to unprecedented levels. Asset owners — particularly those with stock portfolios — may also benefit, since higher margins could translate to higher valuations. But this is highly uncertain. Declining revenue growth is bad for valuations, and growth has historically trumped margins in driving equity prices. That said, in a world where the broader economy is shrinking and nothing productive is worth investing in, the stock market may become the only viable destination for capital — which could, in a strange ironic way, lift valuations precisely because there is nowhere else for money to go.

The steady state is self-sustaining, but it represents a lasting loss of output, employment, and consumption relative to the pre-AI trajectory. In a counterfactual with 6% annualised growth sustained over 25 years, nominal revenue would have reached approximately 429; instead it sits at 108.4, representing a ~75% shortfall against potential.

III. 25-Year Projections

Recap: D_macro = 1 − σκ · Δ_{t−τ} / R₀ — the fraction of baseline demand that survives after displacement erodes spending. σ = μ_K + ψ = 1.40 + 0.175 = 1.58 is the spending amplification factor, combining the Keynesian spending multiplier with wage-suppression spillover on non-displaced workers. κ = (1 − η) × c = 0.35 × 0.65 = 0.23 in Year 1 is the spending loss rate — the fraction of each displaced dollar that becomes an initial spending loss before σ amplifies it — where η = 0.65 is income replacement (unemployment insurance, transfers, savings drawdowns, partial re-employment) and c = 0.65 is the cohort spending rate. κ rises over the horizon as η declines from 0.65 toward 0.40, pushing κ toward 0.39. Δ_{t−τ} is cumulative wage displacement lagged by τ = 2 years, meaning demand erosion at time t reflects displacement at t − 2. R₀ is baseline revenue.

25-Year Table

To choose your own input values, visit Agentic Displacement Model

Key milestones summary:

D_macro Decomposition — Year 10

Step 1: Base Spending Loss (κ × Δ)

Displaced workers lose wages, but not all lost income translates into lost spending. Some is cushioned by income replacement (unemployment insurance, partial re-employment), and some of the lost income would have been saved rather than spent.

• Δ₁₀ = W₀ − W₁₀ = 55.0 − 21.5 = 33.5 (wages displaced)

• η₁₀ = 0.55 (income replacement rate at Year 10)

• κ₁₀ = (1 − η₁₀) × MPC = (1 − 0.55) × 0.65 = 0.2925

• Base spending loss = 0.2925 × 33.5 = 9.80

This is the first-round spending that actually disappears from the economy.

Step 2: Amplified Withdrawal (σ × κ × Δ)

That base loss ripples outward through two channels:

• The multiplier captures downstream revenue losses: one worker’s lost spending is another firm’s lost revenue, which reduces their payroll, which reduces their workers’ spending, and so on.

• The spillover captures the fact that workers who were not displaced also pull back spending out of job-security fears.

Total withdrawal = 15.48 dollars of demand removed from a $100 baseline economy.

Step 3: Demand Fraction (② / R₀)

Express the total withdrawal as a share of baseline revenue:

15.48 / 100 = 0.1548

15.48% of the economy’s original demand is gone at Year 10 displacement levels.

Step 4: Unlagged Demand Index

D_unlagged(10) = 1−0.1548 = 0.8452

If demand responded instantaneously, firms would be operating at 84.52% of baseline demand.

Step 5: Apply Lag (τ = 2 years)

Demand does not reach firms instantly. Existing contracts, inventory buffers, and sticky pricing delay the transmission by τ = 2 years. The demand firms actually face in Year 10 is the unlagged value from Year 8:

D_macro(10) =D_unlagged(8) = 0.8790

Year 8 had less displacement than Year 10, so the lagged value is higher. The full impact of Year 10’s wage losses will not be felt until Year 12.

Summary

The economy retains 87.9% of its baseline demand in Year 10. The remaining 12.1% has been destroyed by the combination of lost wages, multiplier effects, and precautionary pullback — delayed by two years.

Years 0–5

Revenue rises ~35% from 100.0 to 135.0, implying a ~6% annualised growth rate — a plausible proxy for nominal GDP growth in a booming AI-adoption economy. Profit surges from 19.9 to 71.6, a ~260% increase, while margins nearly triple from ~20% to over 53%. Total wages fall from 55.0 to 38.2, a 31% reduction, yet labour share of revenue drops even faster — from 55% to ~28% — because the revenue denominator is growing while the wage numerator shrinks. The demand erosion signal is barely visible: D_macro has fallen only to 0.955, meaning aggregate demand has contracted less than 5% from baseline. The 2-year lag ensures that even this modest erosion has not yet fully registered in revenue. From the firm’s perspective, this period is pure upside — costs are falling, output is rising, margins are expanding, and there is no discernible macroeconomic penalty. Every incentive points toward accelerating displacement. This is the window during which the trajectory becomes effectively irreversible: by Year 5, the cumulative wage displacement Δ has already reached 16.8 and the total workforce wage bill has been reduced by 31%.

Years 5–10

Revenue continues to climb from 135.0 to a peak of 135.7 at Years 6–7 before declining to 130.5 by Year 10 — still ~31% above baseline. But the character of growth has changed. Revenue gains are decelerating and then reversing while displacement continues apace. Profit rises from 71.6 to 84.0, a further ~17% increase, and margins expand from ~53% to ~64%. Total wages, however, fall sharply from 38.2 to 21.5 — a further 44% reduction within just five years. Labour share of revenue collapses from ~28% to ~16.5%. The demand erosion signal is now becoming material: D_macro falls from 0.955 to 0.879, meaning aggregate demand has contracted roughly 12% from baseline, but the 2-year lag means this damage has not yet fully surfaced in the revenue line. The firm-level picture remains unambiguously positive. Margins have more than tripled from their baseline level and profits are at an all-time high. No rational executive reverses course. This is the fundamental asymmetry of the model — the period of maximum firm-level optimism coincides precisely with the period of maximum irreversible macro-level commitment to displacement.

Years 10–20

The lag catches up. Revenue declines from 130.5 to 113.5, shedding roughly 16% from its Year 6–7 peak of 135.7, though it remains approximately 13.5% above baseline. The AI productivity multiplier can no longer fully compensate for demand erosion, and each year’s revenue decline confirms that the macroeconomic feedback loop is now dominant. Profit declines from 84.0 to 79.9 but remains ~300% above baseline — which is precisely why correction does not occur. Wages fall from 21.5 to 8.5, and labour share of revenue reaches ~7.5%. The workforce wage bill has shrunk roughly 85% from its original level. Displacement has largely exhausted its scope: B reached its floor of 1.5 by Year 10, N reached its floor of 2.0 by Year 15, and M reaches its floor of 5.0 at Year 20. The feedback loop decelerates not because the system self-corrects, but simply because there are fewer remaining workers to displace. The economy is contracting in slow motion, and the primary beneficiaries — capital owners and AI infrastructure providers — face no firm-level incentive to intervene.

Long-Run Equilibrium (~Year 25+)

With M = 5.0, N = 2.0, B = 1.5 (all at floor), wages stabilise at 8.5. D_macro settles at ~0.723 and continues drifting toward a terminal value of ~0.713 as the income-replacement rate reaches its long-run floor of 40%. Revenue at Year 25 sits at approximately 108.4, about 8% above baseline. Profit stabilises near 74.9 — still ~276% above baseline. The economy has avoided nominal contraction below its starting level, but has permanently foregone the growth trajectory it would otherwise have followed. In a counterfactual with 6% annualised growth sustained over 25 years, nominal revenue would have reached approximately 429; instead it sits at 108.4, representing a ~75% shortfall against potential. The workforce wage bill has shrunk by roughly 85%, and nearly all the surplus has been captured as profit and AI infrastructure spending. Labour’s share of revenue has fallen from 55% to under 8%. AI infrastructure costs (A = 0.040) are negligible at less than 0.04% of revenue — the entire AI replacement infrastructure costs a vanishing fraction of what the human workforce cost.

IV. Theoretical Robustness and Macroeconomic Cohesion

The Agentic Displacement Model maintains structural integrity when tested against standard national accounting frameworks, specifically the Income and Expenditure (Spending) methods of calculating GDP. By consolidating the economy into a single “Value-Added” firm, the model moves beyond static accounting to capture the dynamic feedback loops between factor shares and aggregate demand.

Integration with National Accounting Frameworks

• The Income Method Alignment: The model is fundamentally a dynamic application of the Income Method (GDP = Labor Income + Capital Income + Indirect Taxes). It tracks the aggressive reallocation of revenue from labor (W_t) to capital (Profit), where labor’s share collapses from 55% to under 6% by Year 25.

• The Spending Method and Consumer Primacy: While GDP is expressed as C + I + G + (X - M), the model identifies Consumption (C) as the sole independent variable. In this framework:

  • Investment (I) is viewed as “derived demand.” Rational firms only invest in capacity if there is a corresponding consumer base to purchase the output.

  • Government (G) acts as a transfer mechanism dependent on the tax base harvested from W and Profit. While corporate tax revenue may spike during the Agentic Boom, it cannot bridge the demand gap if the spending velocity of capital owners remains structurally lower than that of the workers they replaced.

  • Net Exports (X - M) are implicitly neutralized by the model’s global scope.

The Productivity-Volume Paradox

A critical strength of the model is its treatment of supply-side gains. While AI dramatically increases volume (potential output per person), the model recognizes that output does not inherently translate into sustainable revenue (R_t).

• Supply vs. Demand Mismatch: In Phase 1, the supply-side multiplier (y_t) increases capacity by 50%, initially outpacing demand erosion. However, this gain plateaus while demand destruction accelerates as safety nets (savings, UI) exhaust.

• The Inelasticity of Demand: The model assumes that consumer demand for many goods is inelastic; humans do not consume 100x more products simply because AI has made them 100x cheaper or more plentiful.

• Economic Dilution: When billions are forced into digital entrepreneurship, the resulting market saturation and near-zero marginal costs trigger hyper-deflation. This collapses the “currency” value of the output. Since firms require nominal revenue to service debt and maintain infrastructure, a high-volume/low-value equilibrium still results in a systemic contraction relative to the pre-AI growth trajectory.

Conclusion: A High-Margin, Low-Value Equilibrium

The results of the Agentic Displacement Model present a striking paradox. By Year 25, the economy has avoided a nominal contraction, with revenue settling at 108.4 — marginally higher than the starting baseline of 100. However, this “growth” is a shadow of its former self. In a counterfactual world without agentic displacement, sustained 6% annual growth would have seen revenue reach approximately 429; instead, the model shows a ~75% shortfall against that potential.

The real story, however, lies in the distribution. While the “pie” has grown slightly in nominal terms, the split has moved aggressively in favor of business owners and capital. Labor’s share of revenue has collapsed from a healthy 55% to a mere 7.8%. We are left with a high-margin, low-value equilibrium: firms are significantly more profitable than at the start of the transition, but they operate within a stagnant economy where the middle-class demand base has been effectively hollowed out.

As we noted at the outset, this is not a prophecy. It is a thought exercise intended to extend the “Intelligence Crisis” identified by Citrini. Predicting the trajectory of a system as complex as the global economy is impossible to do with high confidence. There may be stabilizers we have not accounted for — new industries that are currently unimaginable, or radical policy shifts that repair the demand-side closure before it becomes terminal.

Join the Debate

We built this model to be stress-tested. Whether you believe our spending amplification factor σ is too aggressive, or that the income replacement rate η will be bolstered by new social contracts, we want to hear from you.

We invite you to create your own alternative 25-year forecast by selecting your own inputs in the model, accessible via this link Agentic Displacement Model. What parameters do you believe will save the demand base? Post your alternative forecasts and critiques in the comments below.

• Unique bridge: Pagaya (PGY) connects lending partners’ declined or capacity-limited applications to institutional ABS and forward-flow investors, embedding itself between origination and capital markets.

• Why we want to believe: strong appetite for consumer-credit yield, multi-year forward-flow capital (Blue Owl / Castlelake), and low valuation multiples (~2.5× EV/S, 6× EV/GP, 18× EV/FCF).

• What went wrong: 2024 impairments exposed oversized first-loss exposure through the Opportunity Fund and shortcomings in transparency and governance highlighted by Iceberg Research.

• Why it may be behind them: impairments fell sharply in 1H25, fee quality tracked issuance, cash conversion improved, and funding diversified toward whole-loan buyers.

• What to watch: onboarding of legacy banks, sustained low impairments, disciplined FRLPC, continued governance upgrades, and the upcoming 3Q25 results on 10 Nov, which will test whether the recovery is genuine.

EXECUTIVE SUMMARY

Pagaya sits at a critical junction of modern credit markets: an intermediary linking traditional lenders with institutional credit investors. Its network allows banks and fintechs to offload or re-evaluate declined applications while giving private-credit allocators standardized access to consumer-loan exposure. This dual-sided position effectively embeds Pagaya within the funding and underwriting infrastructure of unsecured credit.

Valuation remains compelling relative to growth potential — roughly 2.5× EV/Sales and 18× EV/FCF — but the company’s credibility has been tested. 2024 exposed the fragility of its risk-retention structure and raised questions about the depth of its AI underwriting. Unlike Upstart, Pagaya discloses little about its model architecture or alternative data sources. Its advantage appears to stem more from network scale than from proprietary technical superiority.

Equally material is where Pagaya operates in the credit funnel. Many of its partners are non-bank lenders already aggressive in underwriting. Pagaya’s “second look” therefore captures loans rejected even by fintechs, inviting scrutiny over incremental risk. That exposure was compounded by large horizontal first-loss positions in its Opportunity Fund, which exceeded regulatory minimums and amplified downside in 2024.

The subsequent recovery has been notable. Credit-loss rates declined sharply through 1H25 as older vintages rolled off, while long-term forward-flow agreements with Blue Owl and Castlelake reduced dependence on short-term securitization markets. These agreements, together with improved cash conversion and operational discipline, suggest a more stable funding model. Yet transparency and governance remain work in progress — the key areas investors will reassess at 3Q25 (10 Nov).

HOW PAGAYA SITS BETWEEN BANKS AND CAPITAL MARKETS

Pagaya integrates directly into lenders’ origination systems. When a bank declines a loan application — often because of credit-policy limits or capital constraints — the file can be routed to Pagaya for re-underwriting. Its models evaluate risk using cross-partner data to identify loans acceptable to institutional investors even if they fall outside the originating bank’s risk tolerance.

Unlike Upstart, which sources borrowers and matches them to banks, Pagaya receives applications from lenders and redistributes them to investors. The distinction is structural: Upstart feeds origination, whereas Pagaya redistributes rejected or excess flow. It thereby performs both risk re-assessment and capital matching functions.

Pagaya earns fee-based revenue across three categories:

1. AI integration fees – payments from lenders for embedding Pagaya’s models into their workflows.

2. Capital-markets execution fees – fees from structuring, rating, and distributing ABS deals to investors.

3. Servicing and management fees – ongoing charges for administering securitizations and managing forward-flow portfolios.

These fees scale with transaction volume rather than with balance-sheet exposure, enabling growth without directly carrying credit assets. For lenders, the benefits are threefold:
(1) borrowers gain approvals that would otherwise be denied;
(2) lenders preserve customer relationships while freeing regulatory capital; and
(3) institutional investors gain standardized access to high-yield consumer credit.

In essence, Pagaya’s vision creates a threefold benefit: (1) for the borrower, who gains access to funding while feeling secure in dealing with their familiar bank; (2) for the lender, who builds lasting customer ties without shouldering excessive risk; and (3) for institutional investors in credit, who uncover fresh avenues to deploy capital in consumer lending.

Source: Pagaya investor relations

From the bank’s perspective, selling loans into Pagaya-structured vehicles improves the Capital Adequacy Ratio (CAR) by reducing risk-weighted assets:

CAR = (Tier 1 Capital + Tier 2 Capital) / Risk-Weighted Assets

By removing loans from the balance sheet, CAR improves while lending capacity is preserved. Investors, in turn, are relieved from sourcing, rating, and servicing each pool individually. Pagaya’s infrastructure handles due diligence, documentation, and compliance, allowing investors to allocate capital more efficiently.

In effect, Pagaya functions as a central coordinator for the flow of consumer-credit risk — a data-driven intermediary standardizing origination for banks and distribution for investors.

PAGAYA’S DISTINCT ROLE IN THE FINTECH ECOSYSTEM

Within the broader fintech landscape, Pagaya occupies an uncommon middle layer. Most platforms are positioned either at the consumer interface or at the lender interface:

• Consumer credit marketplaces (LendingTree, Credit Karma) connect borrowers with lenders.

• BNPL providers (Affirm, Klarna) integrate financing at the point of sale.

• Digital banks and lenders (LendingClub, SoFi) originate and often securitize their own loans.

Source: Convequity

Pagaya, by contrast, links multiple originators to multiple capital sources, allowing investors to purchase exposure across diverse asset types and vintages rather than from a single lender’s pool. This structure creates diversification for investors and balance-sheet efficiency for lenders.

Institutional Funding Partners

Blue Owl Capital. A ~$280 bn AUM alternative-asset manager, Blue Owl extended a $2.4 bn forward-flow commitment to Pagaya in Feb 2025. The capital’s insurance-linked nature offers long-duration stability, contrasting with short-term warehouse or ABCP lines. This duration advantage lowers refinancing risk and supports Pagaya’s shift toward predictable, multi-year funding.

Castlelake. In Jul 2025, Castlelake agreed to purchase up to $2.5 bn in consumer loans over 16 months. The partnership further de-links Pagaya from periodic ABS issuance and reinforces its standing among global structured-credit investors.

Oaktree Capital. Though not publicly partnered with Pagaya, Oaktree’s growing activity in consumer finance, including a $250 m investment in credit-card fintech Coign, underscores the renewed institutional interest in unsecured credit — a positive macro backdrop for Pagaya’s platform.

KKR Credit and others. Major private-credit allocators such as KKR, Apollo, Blackstone, and Brookfield are expanding into consumer and asset-backed lending, providing a large and growing end-market for Pagaya’s standardized pipelines.

Lending-Side Partners

LendingClub. Uses Pagaya’s AI infrastructure to extend approvals beyond its in-house model, selling those loans to Pagaya’s investor network to manage risk exposure.
SoFi. Integrated Pagaya’s underwriting system in 2021, leveraging the platform to diversify funding while retaining customer relationships — particularly valuable given SoFi’s banking license and capital-adequacy requirements.

Auto and BNPL Channels

Pagaya is extending its infrastructure beyond personal loans into both auto and point-of-sale (BNPL) financing. In auto lending, partnerships with Ally Financial and Westlake Financial illustrate how Pagaya’s system helps lenders expand approvals while offloading risk to institutional investors. This enables lenders to sustain origination growth without stretching balance-sheet capacity — particularly valuable for banks and finance companies subject to capital adequacy constraints.

In the BNPL segment, Affirm and Klarna represent vertically integrated models that originate and fund their own credit exposure, occasionally securitizing loan pools directly. By contrast, Pagaya acts as an open marketplace that can aggregate risk from multiple lenders and offer investors diversified access across issuers and consumer segments. For BNPL providers that wish to scale funding beyond their own balance sheets, Pagaya’s platform can serve as a complementary outlet, broadening liquidity options without requiring new origination infrastructure.

Together, these extensions — across personal, auto, and point-of-sale credit — position Pagaya as a standardized bridge between consumer origination and institutional capital. The architecture allows investors to gain exposure to diversified consumer credit portfolios without building or maintaining retail origination systems of their own.

HIDDEN RISKS BEHIND PAGAYA’S EDGE

Pagaya’s position appears defensible, but the underlying mechanics deserve scrutiny. The central questions are:

1. How advanced are Pagaya’s AI models?

2. How differentiated is its data access?

3. How much credit risk does it retain?

AI Underwriting Depth

Pagaya markets itself as an AI-first company, yet technical transparency is limited. Unlike Upstart, which publishes details on model evolution and input diversity, Pagaya rarely discusses model design, explainability, or retraining cadence. The available evidence suggests its models primarily rely on lender-supplied application data rather than external behavioral or payroll datasets. If true, its competitive advantage likely stems from network scale — cross-partner data aggregation — more than from algorithmic novelty.

Data Advantage and Network Effect

Even without proprietary data feeds, a growing network of lenders provides incremental informational depth. Each new partner increases cross-sectional visibility into consumer-credit behavior, potentially improving model calibration. Over time, this network effect can create a statistical advantage even in the absence of unique data types. Still, the quality of this advantage depends on consistent data normalization and governance across partners, neither of which Pagaya discloses in detail.

Risk Retention and Structural Leverage

The most material concern is Pagaya’s exposure through its Opportunity Fund, which retains first-loss tranches well above the 5 percent Dodd-Frank minimum. Dodd-Frank permits sponsors to satisfy the rule via vertical, horizontal, or hybrid slices; Pagaya historically chose an almost fully horizontal structure, magnifying upside in benign conditions and losses in downturns.

While this practice supports deal execution and investor demand, it concentrates risk in the lowest tranches rather than distributing it evenly across the structure, rendering Pagaya’s earnings highly sensitive to rising delinquencies and charge-offs. This asymmetry—that yields amplify in favorable environments but morph into a leveraged bet against consumer credit deterioration when defaults climb—warrants careful investor scrutiny when evaluating the durability of Pagaya’s model.

To illustrate the economics, consider the following simplified example based on a $100m loan pool with a 10% coupon, 1% servicing cost (net yield 9 percent), and zero defaults:

Source: Convequity

Residual yield = $2.35 m ÷ $5 m = 47% per year — the first-loss tranche captures all remaining cash flow, creating leveraged returns when defaults are low.

In practice, Pagaya’s Opportunity Fund has held around $870 m of equity tranches, representing roughly 17 percent effective retention on ~$5 bn of securitized loans — far above the regulatory minimum. The following scenarios scale this structure:

Scenario 1 – 0 % Defaults

• Loan pool $5 bn → $500 m gross interest

• Servicing 1 % → $50 m

• Net interest $450 m

• Senior investors (7 %) → $289.1 m

• Residual to equity = $160.9 m

• Equity tranche $870 m → 18.5 % yield

• No principal loss → Profit $160.9 m

Scenario 2 – 5 % Defaults ($250 m principal loss)

• Reduced pool $4.75 bn → $475 m interest

• Servicing $50 m → Net $425 m

• Senior payments $289.1 m

• Residual $135.9 m

• Post-loss equity $620 m

• Yield = $135.9 ÷ $620 = 21.9 %

• Profit/Loss = $135.9 − $250 = –$114.1 m

Scenario 3 – 10 % Defaults ($500 m loss)

• Pool $4.5 bn → $450 m interest

• Servicing $50 m → Net $400 m

• Senior payments $289.1 m

• Residual $110.9 m

• Post-loss equity $370 m

• Yield = $110.9 ÷ $370 = 30 %

• Profit/Loss = $110.9 − $500 = –$389.1 m

These scenarios demonstrate the asymmetry: when defaults rise, the equity tranche absorbs all principal losses, eroding value quickly despite high residual yields. This leverage explains the Opportunity Fund’s volatility — and the $400 m+ cumulative impairments recorded between 2023 and 2025.

Presentation and Disclosure Concerns

Pagaya’s investor slides often emphasize AAA-rated ABS issuance but do not clarify that those ratings apply only to senior tranches. The Opportunity Fund’s equity holdings are unrated, yet Pagaya’s materials occasionally imply a unified credit quality across the structure. Similarly, presentation slides omit the Opportunity Fund’s role, creating a perception that all funding is external when in reality Pagaya (through its LP investors) bears first-loss risk.

Source: Pagaya investor relations

Implications

Such structural opacity makes it difficult for equity holders to gauge true risk exposure. While forward-flows with Blue Owl and Castlelake reduce the need for retained tranches, the legacy portfolio still exposes Pagaya to residual volatility. Upcoming 3Q25 results should provide clearer visibility on remaining retained-equity exposure and impairment trajectories.

WHAT HAPPENED IN 2024 AND WHERE PAGAYA IS NOW

Pagaya’s credit performance in 2024 deteriorated sharply before stabilizing in early 2025, revealing both the fragility and the leverage embedded in its structure. Through FY24, impairments in the Opportunity Fund — which retains equity tranches from Pagaya-sponsored ABS — rose sequentially:

• Q1 2024: $53 m on an average balance of $1,089 m (≈ 4.9%)

• Q2 2024: $72 m on $1,023 m (≈ 7.0%)

• Q3 2024: $82 m on $923 m (≈ 8.9%)

• Q4 2024: $235 m on $764 m (≈ 30.8%)

Cumulatively, impairments reached $242 m for the first nine months (≈ 30% loss rate on an average fund size of $807 m) and $442 m for the full year (≈ 44.6% on $990 m). The data describe a steady quarterly deterioration culminating in an acute Q4 spike.

The rebound was equally abrupt. Q1 2025 impairments dropped to $24 m on $760 m (≈ 3.2%), and Q2 2025 fell further to $15 m on $870 m (≈ 1.7%). In other words, a full-year 2024 loss rate of 44.6% and a Q4 run-rate above 30% gave way to a combined 1H25 loss of $39 m — an apparent normalization. Whether this marks genuine stabilization or accounting-driven recovery remains the central question ahead of 3Q25 (10 Nov).

Over the past three years, the Opportunity Fund has delivered a cumulative return of only 5.8% (3.1% gain in 2024 following a 1.5% loss in 2023), while the Tel Bond 20 Index gained 41.2% and the S&P U.S. Treasury 1–3 Year Index rose 18.6%. These returns, coupled with high fees and redemption suspensions, suggest that limited partners have absorbed the downside of Pagaya’s structure without participating in its fee-driven upside.

Valuation and Impairment Analysis

In Feb 2025, Iceberg Research released a critique arguing that Pagaya’s Opportunity Fund valuations were unsustainably high. Using 3Q24 data (interest income $20.1 m on $923 m balance), Iceberg applied a perpetuity at the 5.3% risk-free rate, implying a fair value of roughly $379 m ($20.1 m ÷ 0.053) — a 59% impairment. It then adjusted the discount rate to 14–15%, consistent with B-rated ABS yields, producing a value near $140 m — an ≈ 85% impairment relative to fund size.

By year-end, the fund’s balance had fallen to $764 m. Re-running the framework on Q2 2025 data ($870 m fund; quarterly interest ≈ $6 m, annualized $24 m) at a 4.75% risk-free rate yields an implied $505 m valuation — still ≈ 42% below the reported balance, but directionally improved.

Despite these figures, investor concern has been muted. Pagaya’s ABS buyers include BlackRock and GIC, lending surface credibility to loan quality. Many of its ABS deals carry AAA ratings, which reinforce investor confidence — even though only the senior tranches merit those grades. This selective disclosure has led many to focus on transaction volume and fee margins (lender fees minus ABS production costs) rather than total embedded risk. Investor presentations often omit tranche-rating details and the Opportunity Fund’s continuing role, further obscuring true exposure.

Source: Pagaya investor relations

Management attributes the 2025 improvement to three structural shifts:

1. Reduced exposure to 2021–2023 vintages with higher cumulative losses.

2. Greater emphasis on long-duration forward-flow agreements with Blue Owl (Feb 2025) and Castlelake (Jul 2025).

3. Less reliance on being the ABS sponsor, thereby removing the 4–5% Dodd-Frank risk-retention.

If forward-flows expand to ≈ 25–50% of volume, retained-equity exposure should decline proportionally, curbing the tail-risk that defined FY24.

Skepticism persists. The late-2024 acquisition of Theorem (~$2 bn AUM) sparked concern that the platform might be used to manage redemptions or obscure leverage linked to distressed assets. Simultaneously, certain lower tranches from Pagaya’s 2021–2024 securitizations remain unrated or downgraded, and evidence suggests Pagaya repurchased underperforming loans from its own ABS pools — preserving headline performance but re-introducing credit risk to its balance sheet.

These practices maintain deal flow and fee income but reinforce the perception that Pagaya prioritizes continuity of issuance over the protection of Opportunity Fund LPs, many of whom are Israeli investors facing protracted withdrawal restrictions.

In net, FY24 reflects progressive deterioration culminating in a severe Q4, while 1H25 shows a sharp improvement driven by portfolio mix and funding realignment. The recovery appears genuine but incomplete; valuation transparency and governance remain the unresolved variables.

FURTHER CONCERNS RAISED BY ICEBERG RESEARCH

Beyond valuation, Iceberg’s February 2025 report focused on governance, questioning both the backgrounds of senior executives and the Opportunity Fund’s management practices.

GOVERNANCE AND TRACK RECORD CONCERNS

Avital Pardo, co-founder and CTO, previously owned an Israeli cheque-discounting firm, Gibui, sold in 2021 to businessman Yonatan Cohen. Within a year, Gibui collapsed; trading was suspended and an external auditor was appointed. Globes reported that Cohen sued the Pardo brothers for allegedly concealing operational problems. Court-appointed trustees later found that Gibui had rolled over delinquent loans to mask defaults — reporting “close to zero credit at risk.”

Sanjiv Das, President of Pagaya, formerly led Citigroup’s mortgage division (2008–2013), during which the bank paid $158 m to settle U.S. government claims of mis-certified FHA loans. Later, as CEO of Caliber Home Loans (2016–2022), Das’s team settled for $17 m with the New York Attorney General over unaffordable loan modifications.

Amol Naik, former COO (2021–2024), was a Goldman Sachs partner and among 17 directors charged by Malaysian authorities in 2019 in connection with the 1MDB bond scandal. Though charges were dropped after Goldman’s $3.9 bn settlement, the episode adds reputational risk to Pagaya’s leadership profile.

Collectively, these histories foster concern that the company’s governance culture lacks the conservatism typically expected in highly leveraged financial intermediaries.

OPPORTUNITY FUND GOVERNANCE ISSUES

Israeli media reports from 2023 describe a wave of withdrawal requests from Opportunity Fund investors, prompting Pagaya to suspend full redemptions and create “side pockets” for illiquid assets. These actions extended redemption timelines well beyond loan maturities. Investors claimed the mechanism was poorly disclosed and that redemption caps were tightened further later on. Some LPs now face recovery periods of over three years, with withdrawal limits of ≈ 8% per quarter.

Allegations from Israeli press (TheMarker, Globes) suggest that fund capital was recycled into new ABS tranches primarily to sustain fee revenue for Pagaya rather than maximize LP returns. If accurate, such actions would represent a clear conflict between corporate objectives and fiduciary duty. Although unproven, these claims highlight how closely the fund’s interests are intertwined with Pagaya’s core business.

ANOTHER CONCERN — STOCK-BASED COMPENSATION INCENTIVES

A further governance issue lies in the structure of Pagaya’s executive equity incentives. The company has ~20 million stock options, of which ~16.5 million are held by executives, versus ~3.3 million RSUs. With 76 million shares outstanding, options represent over 25% of the total float, an unusually high ratio for a recently public company.

While stock options can align management with shareholders, this concentration at senior levels introduces material asymmetry. Options reward volatility and short-term price movement — zero value below the strike, amplified gains above it — thereby encouraging valuation-sensitive behavior such as aggressive guidance or financial engineering. Academic studies link option-heavy plans to higher earnings volatility and greater use of accrual adjustments.

RSUs, by contrast, retain value across price cycles and provide longer-term alignment. Given the size and concentration of Pagaya’s option pool, the current structure tilts incentives toward short-term share-price optimization rather than sustainable value creation. Unless paired with multi-year performance conditions, this design adds to the company’s existing governance risk profile.

FUTURE GROWTH INITIATIVES — EXPANDING BEYOND DECLINE MONETIZATION

Pagaya’s foundation has been “decline monetization”: re-underwriting loan applications that lenders have rejected and matching them to institutional buyers. This function remains central to its AI-driven credit infrastructure. However, management is broadening the model upstream to support marketing and acquisition for lenders through Prescreen and Affiliate Optimizer.

Prescreen leverages Pagaya’s data to identify prospective borrowers likely to qualify for partner-lender products before application. By analyzing existing lender data and engaging consumers via email and direct mail, the tool creates pre-approved marketing lists that increase application volume and conversion rates. Early pilots have produced multiple term sheets with existing lenders, positioning Prescreen as a scalable next-stage product.

Affiliate Optimizer targets large credit marketplaces such as Credit Karma, LendingTree, and Experian. It uses Pagaya’s underwriting and pricing data to help partners refine offer construction and improve lead conversion. The platform’s low-integration design appeals to smaller lenders with limited technical resources, reducing implementation time and expanding reach without additional engineering cost.

Together, these initiatives shift Pagaya from a reactive, post-decline processor to an embedded growth partner within the lending ecosystem. They extend its presence across the origination funnel — from marketing and application to underwriting and funding — increasing partner stickiness and diversifying revenue sources.

Critics view this expansion as evidence of pressure on the core business; proponents see it as the logical evolution of Pagaya’s data network. If executed effectively, these products could deepen integration with lenders, lift FRLPC margins, and strengthen the company’s strategic importance to financial institutions.

The move signals a broader transformation: Pagaya aims to become a B2B2C infrastructure platform for credit distribution and growth, not just an intermediary converting rejections into approvals.

FINANCIALS & VALUATION

NOTES ON FRLPC AND FEE QUALITY ANALYSIS

FRLPC (Fee Revenue Less Production Costs) is management’s preferred measure for underlying unit economics, so it is the most useful starting point for assessing the core of Pagaya’s financial model.

• Fee Revenue comprises AI integration fees, capital-markets execution (placement) fees, and ongoing contract/servicing fees.

• Production Costs reflect the operational work of structuring and rating ABS deals — legal documentation, due diligence, rating-agency coordination, and regulatory compliance.

Management emphasizes FRLPC as a percentage of network volume (FRLPC %). This indicates how much value Pagaya captures per dollar of loan flow and whether the platform is gaining operating leverage. The stated target is 4–5% FRLPC %.

Over the last six quarters, FRLPC % has moved consistently higher. Because this trajectory is central to the company’s story, we examined whether the improvement is driven by genuine efficiency gains and product mix — rather than by accounting mechanics.

Source: Pagaya, Convequity presentation

RELATED-PARTY FEE REVENUE — HIGH, BUT TYPICAL FOR ABS SPONSORS

A significant portion of fee revenue comes from related parties — primarily Pagaya’s securitization and financing vehicles (which it consolidates for reporting) and the Opportunity Fund (which retains first-loss tranches). This is structurally normal for ABS sponsors: the sponsor manages and arranges its programs, and those vehicles pay placement, administration, servicing, and performance fees when work is performed and milestones are met.

Across the last six quarters, related-party fees have ranged roughly $145–$175 m per quarter, representing a majority of total fees. The scale is notable, but it is consistent with a sponsor-managed platform. Even so, large intra-network flows deserve validation that they represent real transactions tied to third-party capital — not circular revenue. We therefore ran two checks:

1) Fee Revenue vs. ABS Issuance Volume

Only the securitized portion of network volume should generate capital-markets execution fees. Comparing fee revenue directly to ABS issuance volume isolates whether fee growth is grounded in actual securitizations or inflated by internal activity (for example, repackaging or repeat fee recognition on re-worked assets).

In FY24 the relationship was volatile — quarterly fee-to-issuance ratios ranged 7–50% amid lumpy issuance and legacy adjustments. That volatility warranted scrutiny. However, from 3Q24 through 2Q25, fee revenue broadly tracked issuance, with record levels in 2Q25 on both metrics. We found no sustained evidence of fees expanding independent of deal flow.

Context:

• FY24: ~$6 bn ABS issuance supported ~$1.005 bn in fees (≈ 17% annual average despite quarterly swings).

• 1H25: ~$3.8 bn ABS (including $2.3 bn in 2Q) drove $601 m fees (ratios ~13–19%, more consistent).

The FY24 turbulence correlates with disclosed restructurings, but the 1H25 alignment — and reported participation from ~120 investors — supports that fee growth has been tied to genuine third-party demand rather than intra-entity churn.

2) Fee Revenue vs. Cash Generation

If related-party fees were merely accruals inside the structure, cash conversion would remain persistently weak. In FY24, operating cash flow was indeed soft ($67 m on $1.005 bn fees; ≈ 6.7% conversion), reflecting timing effects and legacy portfolio adjustments that fueled skepticism.

In 1H25, the picture improved meaningfully, indicating closer alignment between fee recognition and cash receipts:

• FY24: $67 m operating cash on $1.005 bn fees → ~6.7%

• 1H25: $92 m on $601 m → ~15%

• 2Q25: $58 m on $318 m → ~18%

• TTM to 2Q25: $131 m on ~ $1.02 bn → ~13%

We also tracked Days Sales Outstanding (DSO = [Ending Fees Receivables / Quarterly Fee Revenue] × 90). DSO held steady in FY24 at 46–47 days (not deteriorating), then improved to 42–43 days in 1H25. Receivables were up ~17% to $149 m from YE 2024, but this was below the 26% growth in fees — equating to ~3.7% of H1 fees, a healthy ratio. The shift toward forward-flow agreements is also reducing working-capital lags typical in ABS issuance, shortening the cash conversion cycle. Taken together, these factors suggest the related-party portion of fee revenue — still large by design — increasingly converts to cash rather than remaining as accruals.

WHY LARGE RELATED-PARTY FLOWS DO NOT NECESSARILY INDICATE RISK

Pagaya’s securitization ecosystem is consolidated under its umbrella. When a deal closes, the company’s own vehicles pay for structuring and servicing under standard, contractually defined fee schedules, and those funds ultimately derive from outside investors who purchase tranches. Under ASC 606, fees are recognized only when performance obligations are met (e.g., documentation complete, servicing provided). The improving cash-conversion metrics and the fee-to-issuance alignment both support this interpretation.

OTHER LEGITIMATE DRIVERS OF RISING FRLPC %

Beyond operating efficiency, product mix is likely contributing.

• Prescreen (direct marketing engine) can generate incremental integration and data-usage fees by expanding the top of the funnel for lenders.

• Affiliate Optimizer widens distribution via Credit Karma, LendingTree, and Experian with minimal integration effort.

• FastPass shortens time-to-verification in auto workflows.

Management has commented on a growing mix of AI-integration and usage-based fees. These deepen relationships and raise revenue per unit of volume, legitimately supporting FRLPC %.

RELATED-PARTY FEE REVENUE — INVESTIGATIVE SUMMARY

Source: Convequity

Summary points:

• Fees % of ABS: Volatile in FY24 (7–50% quarterly; ~17% annual) but stabilized to 14–19% in 1H25, indicating fee growth is now aligned with issuance.

• Cash conversion: Weak/uneven in FY24, improved to 12–18% in 1H25, suggesting recent fees (including related-party) are increasingly collected as cash.

• Overall message: FY24 warranted scrutiny; however, we see no sustained evidence of inflated or circular fee recognition, and 1H25 normalization is consistent with healthier collections, forward-flows, and product-driven fee mix.

GROWTH & MARGINS

Over the last six quarters, Pagaya delivered ~20–30% YoY quarterly growth. The stock rose roughly ~200% YTD at its peak, initially catalyzed by 1Q25 (Adj. EBITDA $80 m vs. $65–70 m guidance) and reinforced as management raised full-year revenue and Adj. EBITDA guidance (FY ending 31 Dec 2025). The Castlelake forward-flow announcement in July 2025 added momentum. The sharp decline in impairments through 1H25 also eased a key overhang.

In 2Q25, revenue was $326 m versus guidance of $290–$310 m, and GAAP net income doubled versus 1Q25. Operating leverage improved meaningfully: EBIT margin moved to ~17% in 2Q25 (from ~2% a year earlier). Broader fintech strength in 1H25 was a tailwind.

Source: Seeking Alpha

The subsequent ~20% pullback since early September appears tied to profit-taking, valuation reset after a sharp move, insider selling, and caution around sustaining the pace into year-end. Given this reset, and with the stock still screening as inexpensive on FCF and EBITDA metrics, the setup remains reasonable — subject to the core debate on underwriting quality, governance, and residual risk retention.

Source: Koyfin

Our inclination is that Pagaya has moved past the most acute phase experienced in late 2024. The mix shift toward legacy banks should help: many banks still use more traditional underwriting frameworks, where Pagaya can add value while also enabling capital relief under CAR rules. Fintech lenders, by contrast, are typically more aggressive and do not face CAR constraints; SoFi is the notable exception given its bank charter. If 2024’s large impairments reflected taking on applications rejected by already-aggressive fintech models, increasing the bank mix should reduce recurrence risk.

If stabilization holds, it is plausible Pagaya sustains ~20–30% growth for a period, supported by renewed institutional appetite for consumer credit yield. The company is the leading personal-loan ABS sponsor, and it is making meaningful progress in auto and POS channels, positioning it to benefit from private-credit demand. We will reassess cadence and mix following 3Q25 results on 10 Nov, which should provide a clean checkpoint on impairments, FRLPC %, cash conversion, and forward-flow utilization.

DCF VALUATION

Click here to access the DCF valuation - scroll to the far right to see PGY’s valuation sheet.

Source: Convequity

We present the base case with the following assumptions:

• Pagaya has moved past the acute 2024 issues; concerns around the Theorem acquisition’s purpose are ultimately unfounded.

• Underwriting quality improves and impairments remain low as evidenced in the last two quarters; the behaviors that contributed to FY24 losses are not repeated.

• FY27–FY31 CAGR = 25%, reflecting steady, moderately high growth in a large addressable market.

• The thesis does not rely on proprietary underwriting superiority. Disclosure suggests models use similar inputs to partner lenders, and Pagaya does not share engineering progress publicly in the way that Upstart does. We attribute growth to the network — 150+ investors and 30+ lenders — and to the value proposition for deposit-taking banks seeking CAR relief.

• TTM FCF margin ≈ 10% after recent operating leverage gains; terminal FCF margin = 15% (conservative given ~40% gross margin).

• SBC: TTM ~5.7% of revenue (as of 2Q25) looks low for a tech-oriented company, but normalized to gross profit, SBC/GP ≈ 13.6%, which is mid-pack vs. fintech peers (5–21%). We set terminal SBC to 2.5% of revenue to reflect maturation.

Source: Convequity

On these inputs, intrinsic value is ~$150 per share, roughly ~5× the current share price. Using Koyfin, the current price implies P/FCF ≈ 18× (EV/FCF unavailable), while our base case implies an intrinsic P/FCF ≈ 100×. The valuation magnitude reflects the combination of (i) 20–30% forward growth for the next few years, (ii) low SBC, and (iii) an already positive FCF margin (~10%) that scales to 15% in the terminal stage. Healthy FCF margins, low SBC, and sustained growth are the core ingredients driving a high warranted multiple.

Source: Koyfin

We will revisit these inputs post 3Q25 (10 Nov) to confirm the impairment trend, FRLPC %, cash conversion, and any incremental disclosures around forward-flow scale and residual retained-equity exposure.

CONCLUSION

Pagaya operates at a high-leverage junction between bank origination and ABS/forward-flow funding. With older vintages rolling off, forward-flow partners in place, cash conversion improving, and valuation still screening as low on FCF and EBITDA, the setup remains constructive even after the earlier rally. The central question is whether 2024’s first-loss pain, governance noise, and disclosure gaps are now genuinely behind the company. If the answer is yes, a path to sustained 20–30% growth with de-risked funding and greater bank penetration could justify materially higher value creation from here.

The next catalyst is 3Q25 earnings on 10 Nov. We will monitor results closely to validate the stabilization in impairments, the durability of FRLPC %, the alignment of fees with issuance and cash flow, and any incremental detail on retained-equity exposure and governance practices.

TL;DR: Huawei has further doubled down on optical+networking optimization to deliver node & cluster level superiority over Nvidia. On single chip basis, the incoming Ascend 950 reaches parity with Hopper. 960 will be on par with Blackwell. 970 targeting Rubin+ perf.

Chip-level = Behind NVDA T-2

By 1Q26, Huawei will be shipping the next gen Ascend 950PR. 950PR’s compute die seems to be monolithic, as opposed to two compute die stitched together by MCM interposer. The goal is to achieve H100-like compute without the need to use two compute dies. Additionally, 950PR is dedicated for prefill and recommendation use cases which require larger but not necessarily faster memory. To achieve that, Huawei uses its own version of HBM named HiBL 1.0. In contrast, Nvidia’s incoming prefill-optimized Rubin CPX will only be available by late 2026, and it is based on GDDR not something similar to HiBL.

Subsequently, in 4Q26, 950DT for decoding and training will be available. It uses the same compute die with HiZQ 1.0 memory for high bandwidth use cases like decoding and training. Notably, the spec is almost the same as HBM3e, but it is likely that Huawei’s HiZQ doesn’t have the standard dimension size. It uses Huawei’s custom logic die as the base die to offer faster speed and less controller logic on compute die.

Ascend 960 will likely be the chiplet version of 950DT similar to Blackwelll vs. Hopper. Basically the compute and memory size will double, but memory bandwidth will increase slightly higher than 2x.

Ascend 970 is in planning phase but Huawei is ambitious about the roadmap and it wants to double major specs in order to maintain competitiveness against Nvidia.

System-level = NVDA T+3?

On the node-perspective, Huawei has discontinued naming the sclae-up wolrd as node but superpod. This is because it is not only packing 384 chips together but 8192 via high-speed optical interconnect and optimized networking protocols that will allow these chips to behave like one computer but likely not 100% matching the perf of the NV link sclae up perf as we know.

That said, the scale up domain under Huawei’s UnifiedBus 2.0 tech will have very high reliability and networking efficiency vs. scaleup+sclae-out used by Nvidia. Huawei claims to achieve 95% efficiency on 8k superpod vs. Nvidia's 3584 card superpod.

This is not impossible given Huawei’s massive leadership of the rest of the world in terms of optical communication and networking optimization. It simply controls everything and it is able to use customized specs and components to achieve it.

By using optical, Huawei avoids the hefty world of C2C via CoWoS like Blackwell or Rubin which packs 2/4 compute dies together. It is very likely that Huawei will indeed maintain superpod and cluster-level performance superiority over the years to come.

As a result, while NVDA is confined in using chiplet and denser PCB designs to pack more compute dies in one rack, Huawei is able to link more racks together to form a superpod. For 950 Superpod, that’s 128 compute racks, 32 networking racks, totalling 160 racks and 1000m^2 in area size. For 960 Superpod, that’s 176 compute racks 44 networking rakcs totalling 220 racks occupying 2200m^2 in area size.

Compared to NVIDIA's NVL144, which is also set to launch in the second half of next year, the Atlas 950 Super Node Card is 56.8 times larger in scale, has 6.7 times the total computing power, and 15 times the memory capacity, reaching 1152TB; its interconnect bandwidth is 62 times higher, reaching 16.3PB/s. Even compared to NVIDIA's NVL576, planned for 2027, the Atlas 950 Super Node remains superior in all aspects.

Compared to Huawei’s previously launched Atlas 900 (CloudMatrix 384) supernode, the training performance of the Atlas 950 supernode has been improved by 17 times, reaching 4.91M TPS. By supporting the FP4 data format, the inference performance of the Atlas 950 supernode has increased by up to 26.5 times, reaching 19.6M TPS.

The Atlas 960 supernode further amplifies our advantages in AI supernodes. Based on Ascend 960, its total compute power, memory capacity, and interconnect bandwidth are doubled compared to the Atlas 950. Specifically, total FP8 compute power will reach 30 EFLOPS, while total FP4 compute power will reach 60 EFLOPS; memory capacity will reach 4,460 TB, and interconnect bandwidth will reach 34 PB/s. Compared to the Atlas 950 supernode, the performance of large model training and inference will be improved by more than 3 times and 4 times, respectively, reaching 15.9M TPS and 80.5M TPS.

With such a large scale-up superpod, it is obvious that Huawei’s cluster is able to reach a giant size easily.

The Atlas 950 SuperCluster consists of 64 interconnected Atlas 950 supernodes, integrating over 520,000 Ascend 950DT chips from more than 10,000 racks into a unified whole, delivering a total FP8 compute power of up to 524 EFLOPS. The launch date will coincide with that of the Atlas 950 supernode, which is the fourth quarter of 2026.

In terms of cluster networking, Huawei supports both UBoE and RoCE protocols. UBoE carries the UB protocol over Ethernet, allowing customers to leverage their existing Ethernet switches. Compared with traditional RoCE, UBoE networking offers lower static latency, higher reliability, and greater savings in both switches and optical modules. Therefore, Huawei recommends UBoE.

This is Huawei’s Atlas 950 SuperCluster. Compared to the current world’s largest cluster, xAI Colossus, it is 2.5 times larger in scale and 1.3 times more powerful—making it the undisputed most powerful compute cluster in the world! Whether for today’s mainstream hundred-billion-parameter dense or sparse large model training tasks, or future trillion- or hundred-trillion-parameter model training, the supernode cluster can serve as a robust compute foundation, efficiently and stably supporting continuous innovation in artificial intelligence.

Correspondingly, in the fourth quarter of 2027, Huawei will also launch the Atlas 960 SuperCluster, based on the Atlas 960 supernode, further expanding the cluster to the million-card level, with total FP8 compute power reaching 2 ZFLOPS and total FP4 compute power reaching 4 ZFLOPS. Like its predecessor, it supports both UBoE and RoCE protocols, and with the advantages of UBoE, its performance and reliability are further enhanced, with even lower static latency and greater network uptime, making UBoE networking highly recommended. Through the Atlas 960 SuperCluster, Huawei will continue to accelerate customer application innovation and explore new heights in intelligent capabilities.

Numbers aside, how far is it ahead or behind $NVDA?

On single die PPAC perspective, it seems 950DT (for decoding and training) will be quite close to Hopper/Blackwell.

On memory front, by using custom HBM standard whilst NVDA only customize the base die of HBM4, Huawei is ahead. But its sources of DRAM (Swaysure) is slightly behind in DRAM die.

In terms of networking, Huawei is massively ahead of NVDA thanks to Huawei's years of networking know-how and advanced vertical integration across all layers of optical networking.

Lastly, and most importantly, new Ascend chips now support SIMT to improve CUDA compatibility. All things considered, Huawei is now possibly the biggest contender to NVDA in merchant AI chips. Its 8k+ networking tech has been in testing for more than a year, and with iDUV solved, it is a matter of time when large frontier models will be trained on Ascend chips.

950 will be the first chip for Huawei after 910-series which was originally designed in 2019 and wasn’t planned to be used for LLM use cases like today. It will be an very interesting time to see how Huawei’s rearchitected and full optimized AI chip will be adopted by the Chinese companies in quarters to come.

• Given the hype around Palantir, we've decided to revisit the company's TAM via the lens of AI agents.

• From the TAM analysis, we estimate a terminal revenue level for Palantir.

• In our DCF valuation, we attempt to reconcile the top-down, TAM-based terminal revenue with our standard bottom-up approach.

Note: This research note was originally published for subscribers on 21st February 2025. We believe the note is still highly relevant.

The Palantir (PLTR) rocket continues on its ascent following the 4Q24 results. The $828m 4Q24 revenue beat guidance by $59m and the $373m adjusted operating income beat guidance by $73m. Within the US, both commercial and government revenue growth continues to reaccelerate, reaching 64% YoY (20% QoQ) and 45% YoY (7% QoQ), respectively. Moreover, this growth reacceleration is accompanied with adjusted EBIT and FCF margin expansions, reaching 45% and 63%, respectively.

Ontology, Ontology, Ontology

In 2021, we pinpointed PLTR’s ability to build customized ontologies for each enterprise client as a unique and potentially transformative skillset — well before the company’s management regularly highlighted it. In those days, “ontology” was barely mentioned in earnings calls, but fast-forward to the 3Q24 call and it was referenced nine times. Our initial view focused on how ontologies would help customers make better business decisions, without fully predicting how quickly AI would propel PLTR’s growth. While we’ve certainly been fortunate in timing, it underscores that our early understanding of the ontology’s strategic importance was on the mark.

We may sound like a broken record when we discuss PLTR as we dive deep into why ontologies are critical for enterprise LLM implementation, but this really is the core of our long-held thesis. Most organizations still struggle to get generative AI models beyond proof-of-concept demos, but PLTR’s platforms, particularly Foundry and AIP, have reaccelerated adoption by translating cutting-edge AI into tangible, immediate ROI. This is driving the demand for PLTR's services.

Real World Impact

Central to this success is PLTR’s strategy of focusing on fewer but deeper enterprise relationships. Unlike the legion of enterprise SaaS providers looking to sign up every potential customer, PLTR devotes itself to making a smaller set of industry titans as successful as possible. It doesn’t just embed LLMs or advanced analytics — it tailors entire operational workflows, from manufacturing lines (Warp Speed) to insurance underwriting (78 AI agents condensing a two-week process to three hours). Every time an organization sees that PLTR can cut costs or open new revenue streams at that scale, it’s an advertisement for the platform’s real-world impact. It also illuminates that choosing PLTR over DIY approaches provides a clearer and more certain path to a good ROI, as DIY deployments are often too complex due to the lack of expertise at the enterprise.

Meanwhile, the government side remains one of PLTR’s historical strengths, and it’s surging again. The company’s ties to the US Department of Defense and allied agencies have deepened in everything from intelligence data analysis (Maven) to on-the-ground operational systems, boosting public-sector revenue by 45% YoY last quarter. PLTR’s willingness to offer its FedRAMP-accredited infrastructure to other software startups is also forging new partnerships and reducing friction with system integrators who might otherwise see it as a threat. That helps explain why PLTR’s government pipeline continues expanding across multiple branches and commands, and why customers, from the Army Software Factory to new defense tech entrants, increasingly prefer to build on PLTR rather than replicate its capabilities in-house.

The long-term opportunity could be even bigger. PLTR’s potential to sustain — or even accelerate — 30%-plus top-line expansion rests on the idea that, in an AI-driven future, the LLM will become the commodity. The real differentiator is how effectively an enterprise can integrate these models into its data landscape. PLTR has been quietly perfecting that integration layer for two decades, giving it a head start that new entrants may never match.

The TAM

The TAM for PLTR is challenging to estimate, simply due to how many workflows, decisions, businesses, and industries their software can add massive value to. As we riffed on in the recent Notes discussing PLTR, ServiceNow (NOW), and Monday (MNDY), the TAM driven by AI agents alone could be colossal, and given PLTR is already operationalizing agents before any other vendor, revisiting this analysis could offer insights to PLTR's terminal revenue level.

The essence of the following analysis is to estimate the impact of AI agents on the global economy. Agents will be a huge productivity booster, enabling companies to use radically fewer input resources for a given amount of output and to significantly increase output for a given amount of input resources.

In this framing:

• Inputs are the costs associated with labor.

• Outputs are either non-revenue or revenue.

Non-revenue output includes:

• IT functions: Keeping the corporate network running smoothly, ensuring employees can access resources, and securing valuable data and systems.

• HR functions: Managing employee-employer contractual relationships, conducting regular performance reviews, and maintaining positive employee welfare.

• Accounting functions: Ensuring the company’s financial records accurately reflect business performance and comply with accounting regulations.

Revenue output is more straightforward, as it represents the total revenue generated by the business.

As agents empower cost reductions and revenue increases, this will boost profits. If we can estimate the profit increase for an average sized company, we can then scale this profit increase to a worldwide estimate. Then, we can surmise what percentage of this profit increase, or value creation, can be captured by AI software agent providers. Ultimately, this becomes the TAM for PLTR and all other software vendors eyeing AI agent offerings.

As agents are replacing and/or augmenting human employees, it's rational to focus the productivity gains, both non-revenue and revenue-based, at the employee level. This is the framework for the following analysis.

Baseline Company (200 Employees)

The main parameters for this analysis are the estimates for what is the average salary + employee-specific overheads, the average revenue per employee, the average number of employees per company, and the total number of companies worldwide. The end worldwide profit increase estimate varies widely subject to the values assigned to these parameters.

1. Employees

   • Revenue-generating: 40% → 80 employees (e.g., HR, IT, accounting, etc.)

   • Non-revenue: 60% → 120 employees (R&D, product, S&M, etc.)

2. Costs

   • Salary + overhead per employee: $100,000 (includes $85k for salary and $15k for software and services each employee requires for their job)

   • Total labor-related cost: 200 × $100,000 = $20,000,000

   • Other costs (rent, insurance, materials, etc.): $4,000,000

   • Total costs (baseline): $24,000,000

3. Revenue

   • Baseline revenue per revenue-generating employee: $350,000

   • Total baseline revenue: 80 × $350,000 = $28,000,000

4. Baseline Profit

   • $28,000,000 (revenue) – $24,000,000 (costs) = $4,000,000

With AI Agents Deployed

1. Non-Revenue Roles (120 employees)

   • 3× productivity ⇒ 67% cost reduction overall.

   • Old labor cost: 120 × $100,000 = $12,000,000

   • New labor cost: $12,000,000 × (1 – 67%) = $4,000,000

   • Savings: $8,000,000 ⇒ Unfortunately, this saving would come mostly, if not entirely, by replacing humans with agents.

⚠️ Warning: AI agents may lead to significant job displacement in non-revenue roles as companies cut costs by replacing human employees with automation.

1. Revenue-Generating Roles (80 employees)

   • 2× productivity ⇒ 100% more revenue.

   • Old revenue per employee: $350,000

   • New revenue per employee: $350,000 × 2 = $700,000

   • Total new revenue (80 employees): 80 × $700,000 = $56,000,000

   • Increase in revenue = $56,000,000 – $28,000,000 = $28,000,000

New P&L

1. New Revenue: $28m + $28m = $56,000,000

2. New Labor Cost:

   • Non-revenue: $4,000,000 (down from $12m)

   • Revenue roles: still $8,000,000 (no change in headcount or pay)

   • Total labor = $12,000,000

🔔 Alert: Unlike non-revenue roles, revenue-generating employees are less likely to be displaced by AI agents. Sales and other customer-facing roles require human interaction, relationship-building, and nuanced negotiation — areas where AI is more likely to augment rather than replace workers.

1. Other Costs (unchanged): $4,000,000

2. Total Costs Now

   • $12m (labor-related) + $4m (other) = $16,000,000

3. New Profit:

   • $56m - $16m = $40,000,000

• Profit increase over baseline: $40m – $4m = $36,000,000

Willingness to Pay for AI

If a software vendor starts offering AI agents, it may raise its prices by 25%. For instance, a company paying $20,000 per year for software might now be content to pay $25,000. Applied to our scenario, that 25% increase equates to $9m of the $36m additional profit, or additional value, going to the vendor, leaving $27m as the company’s net gain.

Scaling Up Worldwide

• Let’s assume there are 500,000 “mid-sized or larger” companies worldwide.

• If each such company were to invest an additional $9m in AI agents to achieve an incremental $27m of profit, that would be:

  • 500,000 × $9,000,000 = $4.5 trillion

• This $4.5tn would be the theoretical TAM for AI agents, above and beyond what is already being spent on software.

• Obviously, not every company will adopt at full scale simultaneously, but this highlights the theoretical upper bound if agents truly deliver these productivity improvements.

A Simpler Method

The above calculations and estimations are fun, but perhaps we could arrive at a similar TAM with a simpler approach. Let's say the average worldwide salary + employee-related overhead for knowledge workers is $60k, and there are 1 billion knowledge workers. This equates to a labor cost of $60tn, which can be sanity checked by taking the worldwide GDP of $110 trillion and multiplying it by the global labor income share of 52%, which equals $57.2 trillion. Let's speculate that agents can reduce this cost by 40%, generating $24tn additional profit for companies. 25% of this additional profit might be shared with the agent vendors, thus equating to $6 trillion. Thus, here, the TAM for agents rises to $6tn.

PLTR's TAM & Terminal Revenue Potential

These back-of-the-envelope calculations show PLTR's potential TAM is colossal, and this is only for the future world of agents, not even the current state of software. If we say PLTR's TAM is somewhere between these two methods, then the company's TAM could be ~$5tn. If the company can capture just 5% of this agent-based software TAM, then the company's terminal revenue would be $250bn (plus tens of billions for non-agent software).

Valuation

Foreword: Thinking Beyond Traditional Valuation Constraints

At first glance, valuing PLTR at a significantly higher intrinsic value than its current market price—despite its already elevated EV/S multiple—might seem excessive. After all, most investors instinctively apply conventional growth decay models, assuming that any company reaching a few billion dollars in revenue must inevitably slow down. This thinking is often correct. But history has repeatedly shown that a select few technology companies defy this pattern, reaccelerating and compounding growth in ways the market initially dismisses.

The goal of this valuation exercise is not to argue that PLTR is an immediate bargain, nor to suggest that such an aggressive forecast should be taken as a base case. Instead, this is an exploration of what could be possible if PLTR continues on its trajectory as a core infrastructure provider for the AI revolution. If AI agents do become a ubiquitous enterprise necessity, and if PLTR remains the leader in operationalizing them at scale, then a multi-hundred-billion-dollar revenue figure no longer seems outlandish.

Wall Street often struggles to correctly price long-tail growth opportunities. Investors in the late 2000s did not foresee Microsoft’s reacceleration in the 2010s, just as few predicted NVDA’s meteoric rise as AI workloads exploded. By extending our forecast period and testing different growth trajectories, we’re not making a bold call on where PLTR should trade today — we’re simply challenging the assumptions that have led investors to prematurely cap its long-term potential.

This valuation framework is as much an academic thought exercise as it is an investment thesis. We are not claiming certainty — only that PLTR could be one of those rare generational companies that rewrites the rulebook on compounding growth.

DCF Valuation

Usually we apply a 15-year explicit forecast period for our DCF valuations. PLTR appears to be a different beast, however. As we have articulated, the AI opportunity is colossal for PLTR, and if $250bn+ of revenue is eventually achievable, we need to extend the forecast period. If we retained the 15-year period, then reaching $250bn in 15 years would require a few years of ~70% growth (you can test this out by altering the FY27-FY30 CAGR). The funny thing is, we think this is actually possible, or even 1-2 years of 100% growth is feasible, but we've decided to instead extend the forecast period to 25 years combined with a growth reacceleration, reflected with the FY27-FY30 CAGR, to 45%.

Palantir DCF Valuation - 25 year forecast period

Source: Convequity

Why do we think such a reaccelerated and sustained level of growth, for a revenue base of currently nearly $3bn, is feasible, or even likely? Well, this is because currently PLTR's US Commercial revenue is growing QoQ at 20%, which annualizes to a 100%+ growth rate, and appears to continue growth acceleration. In the 4Q24 call, management guided a FY25 US Commercial revenue of at least $1.079bn. This would yield a YoY growth rate for US Commercial of 153%.

The following analysis shows a plausible growth deceleration for US Commercial from there, through to FY30, and also shows the rest of the business growing 20% through FY30. If these growth rates largely play out, then US Commercial will represent 3/4 of PLTR's total business by FY30, at $19bn of revenue, and still growing, at what we expect to be, a high growth rate (40% is our forecast, as shown in the table).

Source: Convequity

As you can see in the DCF model, after the FY27-FY30 CAGR period, we sharply drop the growth. Our thinking here is that growth normally doesn't decline linearly from such high levels. Usually, there is a sharp drop after a few years of high growth. Then, from FY34 to FY39, you can see we've given you an option to select a CAGR. We have defaulted this CAGR to 15%, and our thinking is that after such high growth, followed by a sharp drop, usually you see stellar software companies generate 10% to 20% growth, even from a high revenue base, for several years - Salesforce, ServiceNow, and Adobe are good examples here.

After this period, as per the theory of DCF valuations, we scale down the growth linearly to the sustainable growth rate, which for tech companies, we set at 3% (just above the US long-trend real GDP growth rate of 2%), in Year 25 of the forecast period. In Year 25, of FY49, the terminal value turns out to be $290bn.

This $290bn reconciles with the top-down AI agent TAM approach we showed previously. Potentially, PLTR makes most of its revenue, say $250bn, from providing and running agents, and maybe a few tens of billions of dollars will also be made from non-agent software.

In our view, it’s conceivable that PLTR could reach a mature-state free cash flow margin north of 50%, which would make it one of the most profitable software companies in the world, and probably the most profitable mega software company in the world. In the DCF model, we simply scale the FCF margin linearly from its 40% level in FY24 up to 50% in FY49. Though, this could be an underestimation, because in the last two quarters PLTR has generated FCF margins of 55%+. It is unclear whether this is the new FCF margin level or that there is a good dose of volatility here due to the mix of contract terms. If it is the new level of FCF margin, this boosts the intrinsic value per share even higher.

In regards to the SBC %, a one-time accelerated SBC expense, related to market-vesting stock appreciation rights (SARs), increased SBC % from 20% to 34% in 4Q24. As this is a one-time event, in the DCF model we have set FY25 SBC % to back 20%, and scaled it down linearly to the terminal SBC estimate of 10%.

Based on these parameters and arguments, we arrive at an intrinsic value per share of $186 versus the current price of $106. On the face of it, the valuation seems crazy, indeed. But if you really think about PLTR's opportunity, the current US Commercial growth, and PLTR still being the only software vendor able to quickly and safely implement LLMs in enterprises, and then contemplate the future of AI agents, a $300bn terminal revenue does not seem implausible.

• The valuation overview (Mar-25) identifies high-performing stocks with low multiples using Rule of X (inc. SBC), spotlighting opportunities across sectors like Chinese equities, semiconductors, and consumer tech amid geopolitical and market influences.

• It includes initial thesis recaps and updated DCF valuations for FTNT, S, and PANW, offering a preview of long-term investment perspectives.

• Further detailed analysis and updated valuations will be provided in Parts 2 and 3.

Valuation Overview

The table below is designed to highlight stocks with strong financial performance yet low valuation multiples, sorted by their Rule of X (inc. SBC) as a key indicator of financial health. We’ve percentile-ranked these stocks based on their Rule of X (inc. SBC) and paired this with their valuation percentiles to identify "low-hanging fruit" — stocks with a high Rule of X but undervalued multiples. Note that stocks with a negative EV/(FCF-SBC) have been assigned a multiple of 10,000, resulting in an "NA" percentile rank for valuation.

Source: Convequity - March 2025

Several stocks stand out as attractive opportunities. Chinese stocks PDD (Pinduoduo) and LI (Li Auto), operating in e-commerce and consumer tech respectively, exhibit impressive financial performance with notably low multiples. Despite the People’s Bank of China’s stimulus efforts in late 2024, lingering pessimism and uncertainty around Chinese equities have driven steep discounts. For PDD, the company behind Temu, investor concerns about incoming Trump tariffs are likely weighing on its valuation, yet at an EV/(FCF-SBC) of just 7x, the stock appears to be a compelling bargain.

In the semiconductor space, TSM, KLAC, LRCX, RMBS, ASM, and ASML also present above-average financial performance paired with below-average valuations. This undervaluation may stem from supply chain uncertainties tied to the Trump 2.0 agenda, as well as market speculation about reduced GPU demand following DeepSeek’s innovations, which could challenge the need for expensive LLMs if cheaper alternatives gain traction. While not a semiconductor company, ANET (Arista Networks) shares similar characteristics and can be grouped with these stocks for this analysis, given its role in networking solutions for data centers.

Additionally, consumer tech stocks are trading at lower multiples than their enterprise tech counterparts while still delivering strong financial results. MELI (Mercado Libre), OSCR (Oscar Health), HIMS (Hims & Hers), UPST (Upstart), SE (Sea Limited), and GRAB (Grab Holdings) all appear attractive based on their financials and valuations. The discounts for MELI, SE, and GRAB are likely influenced by their non-US status, a factor that continues to suppress valuations. Meanwhile, OSCR and HIMS may be out of favor due to the heavily regulated industries they operate in — healthcare and telehealth, respectively — where uncertainty under the new U.S. administration adds risk to investor sentiment.

Check out the following charts if you prefer scatter plots for this type of valuation analysis. Due to the number of stocks, it's not possible to attach all the names to each dot. Therefore, click the link for a closer look - you can see the stocks by hovering the cursor over the dot. Note that the x-axis is the Rule of X (inc. SBC) for the EV/(FCF-SBC) multiple and is Rule of X for the EV/FCF multiple.

Click the following link to access the updated DCF valuations. Note that only FTNT, S, and PANW have been updated thus far. In Part 2 and Part 3 we shall have updated all the valuations.

DCF Valuations

Now we will provide a recap and update on the long-term theses for FTNT, S, and PANW.

Fortinet (FTNT)

Reaffirming Our Thesis Amid Cyclical Recovery

As we anticipated in late 2023/early 2024, FTNT has rebounded from its product revenue slowdown, which was exacerbated by supply chain distortions. The 2021/22 period saw an aggressive 40%+ growth in product revenue, driven by pent-up refresh cycles, heightened security spending, and pandemic-induced liquidity. But the revenue recognition that fueled this surge — stemming from a backlog of orders accumulated during those distorted years — created tough comparisons, leading to the slowdown we saw in 2023/24. At its lowest point, growth dipped to 7% in 1Q24, a moment when Wall Street’s short-termism, fixated solely on the next 1-2 quarters, spurred overly bearish sentiment.

We’ve been clear on this cyclical overhang for well over a year. The Historical Price Target chart below highlights how consensus targets fell below $70 in late 2023, with even the highest dropping under $80. Yet, our valuation framework has consistently pegged intrinsic value at ~$100+, with upside potential contingent on long-term execution.

Beyond SASE – Why the Market Still Undervalues FTNT’s Strategy

A major disconnect we continue to see is that investors are still viewing FTNT’s strategy through a narrow SASE lens. But as we outlined in our SASE, SASO, & Beyond report and Themes: SD-WAN & SASE Industry Review 2024 (Pt.1), FTNT's ambitions are far broader.

Their latest Investor Day presentation finally drew a line between their vision and Gartner’s strict SASE definition — a distinction we’ve long argued was necessary. As we outlined in our SASO Impulse nearly two years ago, FTNT’s approach isn’t just about delivering Gartner's definition, and securing the First Mile (from user to first security inspection). They have been investing aggressively to develop converged networking and security that can be deployed anywhere (cloud PoP, on-prem, remote, etc.). They are also investing aggressively in the Middle Mile, which remains a critical and often overlooked bottleneck in enterprise networking.

FTNT’s Push into the Middle Mile – A Key Long-Term Advantage

Most SASE vendors focus primarily on optimizing the First Mile, assuming a cloud-first world where users primarily access SaaS applications. But the reality is that many enterprises still operate a mix of cloud and private applications, with substantial workloads residing in their own data centers. The challenge? Efficiently routing traffic to these on-prem resources has traditionally required MPLS — an expensive, rigid solution. This is where WAN-as-a-Service (WANaaS) presents an opportunity.

The Middle Mile is the critical link between the first and last security checkpoints. The problem is that most ISPs operating in this segment are regional or, at best, bi-continental, meaning packets must traverse multiple carriers, adding complexity, potential congestion, and cost inefficiencies. Unlike the First Mile, which can be optimized by SD-WAN overlays, the Middle Mile requires numerous peering agreements between operators to ensure low-latency and cost-effective traffic routing.

FTNT is actively working toward solving this bottleneck through its Middle Mile strategy. FTNT aims to build a vertically integrated PoP network, similar to how Cloudflare and Netskope have expanded their global presence. By owning and operating more of its Middle Mile infrastructure, FTNT can directly manage traffic flows, reducing reliance on third-party carriers, lowering latency, and improving overall network performance.

This approach not only enhances the user experience but also creates long-term cost advantages. As FTNT scales its proprietary PoPs, it can bypass hyperscaler and colocation markups (~66% and ~50% gross margins, respectively) and optimize network transit at a fraction of the cost of competitors that remain dependent on leased infrastructure. This, in turn, strengthens FTNT’s ability to deliver secure, high-performance networking while structurally improving its gross and FCF margins.

Four-Pronged PoP Strategy

FTNT has a four-pronged strategy to quickly expand their global coverage in the short-term, while positioning themselves for global secure networking superiority in the longer term. Their four-pronged approach includes:

1. Hyperscaler PoPs (Short-Term Expansion Tactic) – FTNT is leveraging Google Cloud’s network to quickly expand its PoP footprint. This mimics PANW’s approach but is only a temporary measure.

2. Colocation PoPs (Equinix & Others) – A middle-ground approach that allows FTNT to deploy its FortiGate ASICs, but with space constraints.

3. Service Provider (SP) PoPs – A unique GTM strategy where FTNT enables SPs to build private SASE offerings using Fortinet’s hardware and software (marketed as Sovereign SASE). This is a vastly underserved segment, giving FTNT a cost advantage over premium-priced alternatives like ZS or PANW. A further note is that this sub-strategy really reduces the latency of the first mile of networking, as typically SPs are already closest to end users (able to serve process requests in under 5ms).

4. Wholly-Owned Proprietary PoPs (Long-Term Goal) – The ideal end state, where FTNT fully controls the infrastructure, optimizing for cost, performance, and integration with its ASIC stack.

FTNT's orchestrator layer is key to enabling FTNT to offer such varied implementations, as it provides 1) a centralized control plane that abstracts away infrastructure differences, 2) automates integration with hyperscaler and SP environments via APIs, and 3) enables multi-tenant policy delegation for service providers looking to offer private SASE solutions.

FTNT's Private SASE is another implementation whereby FTNT provides the hardware and software for enterprises to operate SASE in their own data center for their internal requirements only (in contrast to Sovereign SASE whereby SPs are using FTNT's SASE in multi-tenant scenarios to provide secure networking for their customers).

Number 4 — FTNT’s goal of fully owning its PoPs — is the natural extension of its deep-rooted vertical integration strategy. Ken Xie and FTNT’s engineering leadership recognize the immense value that can be unlocked by bringing network infrastructure entirely in-house. The clearest advantage? A structurally lower cost base that significantly improves competitiveness against SASE rivals reliant on colocation (e.g., Cloudflare on Equinix) or hyperscaler IaaS (e.g., Palo Alto Networks on GCP).

For reference, Equinix operates at a ~48% gross margin, meaning FTNT could save 48 cents on every dollar by bypassing third-party colocation. But the bigger efficiency gain comes from FTNT’s ASIC-driven cost structure.

• ASIC Efficiency – FTNT’s custom silicon is designed for high throughput at a fraction of the power and compute costs of general-purpose x86-based security processing. Relative to PANW’s reliance on software-based appliances running in GCP, we estimate 10x cost savings in compute and energy.

• Proprietary PoPs vs. Hyperscaler IaaS – Public cloud infrastructure carries heavy markups, with hyperscaler gross margins around 66%. By replacing these hyperscaler-hosted PoPs with proprietary ones, FTNT avoids these IaaS costs, translating to a 3x reduction in raw infrastructure expenses.

COGS Breakdown – The Competitive Gap

We estimate that compute and energy costs account for ~70% of total COGS, with the remaining 30% attributed to networking and storage. Applying FTNT’s 10x ASIC efficiency to this 70% share results in a 7x reduction in total COGS from compute/energy savings alone. When combined with the 3x reduction from moving off hyperscalers (due to removing the 66% gross margin), the total COGS savings stack up as follows:

Final COGS Reduction Estimate:

• 7x from compute/energy efficiency

• 3x from proprietary PoPs replacing hyperscaler IaaS

• Net impact: ~10x lower COGS vs. PANW

These structural savings don’t just enhance FTNT’s pricing power — they also translate into significantly higher cash flow conversion and stronger long-term shareholder returns. While most SASE competitors are locked into third-party infrastructure with no path to vertical integration, FTNT’s model allows it to scale more profitably while delivering superior network performance.

The Valuation Perspective – Where We See Alpha

Despite a cyclical recovery and an underappreciated Middle Mile strategy, we don’t anticipate FTNT’s revenue growth reaccelerating past the mid-teens. Instead, our thesis centers on margin expansion, specifically an eventual FCF margin of ~45% (vs. LTM 32%). Management's 3-5 year guidance includes mid-to-high 30s FCF margin, and if you play around with the DCF valuation, it seems as the market predicts FTNT's terminal FCF margin is around 40%. But as FTNT’s proprietary PoPs scale, we see a clear pathway to higher terminal margins, plausibly to 45%.

Our updated valuation model reflects this, placing intrinsic value at $131/share, assuming a 15% CAGR for FY27-FY30 and a structurally improved cost base. At the time of writing, this presents a potential 30% upside.

FTNT stands apart by expanding horizontally beyond the First Mile (unlike Zscaler), to optimize the Middle Mile, where network performance and cost inefficiencies are most pronounced. At the same time, it’s integrating vertically by owning its PoPs and running custom ASIC-powered FortiGates, unlike Cloudflare and Netskope, which rely on colocation providers and COTS hardware.

This dual strategy gives FTNT a structural cost advantage and superior performance, positioning it as a hybrid-optimized secure networking leader rather than just another SASE vendor.

SentinelOne (S)

Unlocking Alpha Through AI-Driven Cybersecurity

Note: S reports 4Q25 after the market close on 12th March 2025.

The valuation gap between S and CrowdStrike (CRWD) remains one of the most compelling inefficiencies in cybersecurity investing today. Both companies share a similar trajectory — starting as endpoint security players, offering MDR to varying degrees, and later expanding into cloud security and identity protection. The primary distinction is that S is roughly three years behind CRWD in GTM maturity, not in technology or execution potential.

Over time, we see S reaching, if not exceeding, CRWD’s profitability profile. SentinelOne has placed a greater emphasis on autonomous software and automation-first security, whereas CRWD’s business model still leans heavily on human labor for identifying and stopping threats. This suggests that in a mature-stage scenario, S should be structurally more profitable, with higher gross and FCF margins than CRWD. Despite this, their multiples are poles apart, creating substantial alpha for investors who recognize SentinelOne’s long-term cost advantages.

The Power of a Unified Data Architecture

S' architectural advantage lies in its foundational data layer, which has allowed it to scale security capabilities more effectively than rivals. Unlike most competitors that rely on third-party SIEMs (e.g., Splunk) or hyperscaler infrastructure, S owns its data ingestion, storage, and retrieval stack — a crucial differentiator in an era where AI-driven security requires seamless access to vast amounts of structured and unstructured data.

The 2022 Scalyr acquisition was a pivotal moment, equipping SentinelOne with a high-performance, flexible, and cost-efficient data lake — later rebranded as DataSet. This strategic move gave S the control needed to pioneer XDR, allowing for:

• Seamless ingestion of third-party data (crucial for enterprise-wide threat visibility).

• Faster queries and analysis, reducing response times.

• A highly scalable, AI-ready foundation, enabling next-gen security applications like Purple AI.

Unlike most EDR vendors, S has built an expansive and unified data layer that allows security AI models to operate effectively across an enterprise’s entire environment. This is what gives Purple AI an edge — it can analyze threats across third-party tools with minimal integration friction, while competitors struggle with fragmented architectures. Ultimately, this advantage can be traced back to S' strategic objective from its early days of being more open and interoperable, which has led to smooth third-party integrations, which in turn has led to third-party data ingestion and compatibility, which has given them an advantage to lead in XDR and now in AI-assisted security.

Scaling AI Beyond XDR – AI SIEM & AI-SPM for GenAI Workflows

SentinelOne is leveraging its data-first strategy to expand beyond endpoint security, evolving into a full-stack cybersecurity platform. The launch of AI SIEM is a prime example. While it builds on DataSet’s existing capabilities, AI SIEM is designed for real-time detection, streaming data analysis, and autonomous investigation, positioning SentinelOne to lead in autonomous security operations.

Key developments:

• AI SIEM has already achieved FedRAMP High, a crucial milestone for U.S. government adoption.

• MSSPs are expanding platform adoption with S, leveraging AI-driven capabilities to reduce costs and improve visibility.

• Enterprise traction is growing, with a major federal agency selecting AI SIEM alongside SentinelOne’s endpoint security for unified threat visibility.

Beyond SIEM, SentinelOne has launched AI-SPM (Security Posture Management) — a solution designed specifically for securing GenAI applications and workflows. Unlike traditional CSPM (cloud security posture management) or SSPM (SaaS security posture management), AI-SPM focuses on protecting AI-driven infrastructures. This includes:

• Monitoring and securing LLM APIs and GenAI-powered applications.

• Preventing prompt injection attacks, unauthorized model access, and adversarial manipulation.

• Ensuring governance, compliance, and risk mitigation across dynamic AI workflows.

Recently, S has also made Purple AI available within AWS Bedrock to enhance AI-driven security operations. Through this integration, SentinelOne taps into Bedrock’s foundation models to power real-time threat detection and response for AI-driven applications, reinforcing security visibility across enterprise AI workflows.

This positions SentinelOne at the forefront of securing enterprise AI adoption, giving it a strategic advantage as LLMs and AI-powered workflows become embedded across industries.

Cloud Security Growth & Differentiation vs. CrowdStrike

SentinelOne’s non-endpoint business has now grown to ~$200 million, reflecting its successful expansion into cloud security. A key differentiator is S’ ability to provide both agent-based and agentless solutions, giving enterprises flexibility in how they secure workloads.

One of SentinelOne’s greatest advantages over CRWD is its deep Linux expertise, dating back to the company’s early days. While CRWD originally focused on Windows-first agent development and only recently began optimizing for Linux, SentinelOne has always treated Linux as a first-class citizen.

This early investment enabled SentinelOne to develop an eBPF-based agent, which is tailor-made for cloud-native security. Unlike traditional agents, eBPF operates efficiently at the kernel level, allowing for deep visibility, high performance, and lightweight enforcement in dynamic cloud environments. This gives S a structural edge over CRWD in cloud workload protection — an area of increasing importance as enterprise workloads shift further into Kubernetes-based and microservices architectures.

MITRE Performance – A Testament to Autonomous Security

SentinelOne’s latest MITRE ATT&CK evaluation further reinforces its autonomous security model. S has achieved 100% detection for five consecutive years, a feat unmatched by any other vendor. More importantly, SentinelOne’s detection and response do not require heavy fine-tuning or manual configuration (check out our recent note for a detailed review of MITRE's latest test for endpoint security vendors).

CRWD, by contrast, relies more on manual tuning and SOC expertise to optimize its detection stack. This reflects a broader contrast between the two companies:

• S’ autonomous-first approach reduces human labor costs and enables out-of-the-box functionality.

• CRWD’s model still relies on significant human intervention, particularly within Falcon Complete MDR, where analysts continuously refine detection models.

As security teams seek more automation and fewer operational burdens, S’ model is becoming increasingly attractive.

Market Dynamics & the Path to Profitability

S reports 4Q25 and FY25 results after the market closes on 12th March. In 3Q25, it appears SentinelOne’s strategic execution is beginning to translate into financial improvements:

• Net new ARR grew 22% QoQ, signaling a return to strong growth.

• Approaching non-GAAP breakeven, while maintaining aggressive investment in R&D to sustain long-term innovation.

• Revenue per employee remains just $285k, significantly below cybersecurity leaders like PANW, FTNT, and CRWD. This is surprising as it suggests that perhaps the core unit economics are not as scalable as some cybersecurity peers. It's possible the price war with CRWD is keeping this metric on the low side. Alternatively, it could indicate substantial upside potential, as we have seen a number of vendors scale their rev/emp as they have turned themselves into a platform play (e.g., Cloudflare, Tenable, Monday).

CRWD’s July 2024 outage has also provided momentum for S, as enterprise customers continue to evaluate their heavy reliance on CRWD. Combined with strong MITRE ATT&CK results, SentinelOne is proving itself as a stable, enterprise-grade cybersecurity platform.

We're keen to see whether progress in net new ARR and margin improvements continues in the 4Q25 results. Another note for investors is that S is still in the midst of an executive transition, having appointed a new CRO in November 2023 and a new CMO earlier in April 2023. This echoes the transitions we saw at FTNT and Cloudflare when they made assertive S&M shifts to expand beyond SMBs into the larger enterprise market. For both, the process was challenging — marked by upheaval and skepticism before top-line results materialized. We believe SentinelOne is on a similar trajectory, and early signs suggest the company is successfully evolving into an enterprise-grade vendor.

Expanding Beyond Endpoint – The Lenovo Deal

SentinelOne’s Lenovo partnership, targeting 30 million PCs over the next few years, could become a very notable revenue stream. While the impact remains uncertain, it reinforces S' strategy of embedding its platform deeper into global enterprises.

Valuation – The Alpha Opportunity in SentinelOne

S' single-architecture data model has positioned it ahead of legacy security vendors, allowing it to compete aggressively in the AI-driven security era.

The thesis for S is the rate of change in EBIT and FCF margins coupled with durably 20%+ growth for the next few years. We expect S' Rule of 40 will ascend steadily above the threshold mark as the company leverages its platform to deliver 20%+ growth while becoming highly cash generative and profitable.

Based on a 20% CAGR for FY27-FY30 and a 40% terminal FCF margin, we arrive at an estimated intrinsic valuation of ~$40, 2x the price at the time of writing.

With improving profitability metrics and an unjustifiably wide valuation gap versus CRWD, SentinelOne remains one of the most overlooked cybersecurity plays in the market today.

Palo Alto Networks (PANW)

Palo Alto Networks – Platformization Done Right

The Platformization strategy, introduced by Nikesh Arora in early 2024, is reshaping PANW’s business model by consolidating security spend across Network Security (NetSec), SecOps, and Cloud Security. The premise? Expand revenue by deeply embedding PANW’s full-stack security offerings into customer environments, gradually eliminating competing vendors. The execution has been aggressive — customers using only parts of a PANW platform gain extended free access to the rest of the suite, ensuring they become dependent on PANW’s ecosystem before renewal.

This approach differs sharply from traditional legacy platformization, where vendors cross-sell subpar products at discounted rates to drive expansion. Microsoft’s security strategy is the prime example — offering broad but inferior solutions that lock in customers through bundling rather than product superiority. CrowdStrike is following a similar playbook in cloud security, leveraging its EDR reputation to sell cloud offerings at low cost rather than through BoB differentiation.

PANW’s model is different. Instead of simply bundling, it has built and acquired genuine next-gen BoB solutions across its three core segments:

• NetSec (mostly homegrown)

• SecOps (a mix of homegrown and acquired)

• Cloud Security (mostly acquired, but with massive post-M&A success via founder empowerment)

This approach ensures PANW remains competitive in new business acquisition rather than just milking its installed base. Unlike Microsoft or CRWD, PANW actively attracts new customers with BoB offerings, forcing internal teams to stay on the innovation frontier.

The results are evident — as of 2Q25, 1,150 customers have been platformized, up 35% YoY. This model is not only increasing ARR per customer but also strengthening PANW’s leadership across three rapidly transforming security markets.

The Three S-Curves of Growth: NetSec, Cloud Security, and SecOps

PANW’s growth is unfolding across three distinct S-curves:

1. 1st S-Curve: Network Security (NetSec) – Traditional but evolving with SASE.

2. 2nd S-Curve: Cloud Security – An emerging frontier, now demanding real-time protection.

3. 3rd S-Curve: SecOps – The most transformational opportunity, primed for AI-driven automation.

Each of these domains is undergoing structural change, and PANW has positioned itself at the forefront of each shift.

1st S-Curve: NetSec – Strengthening the Moat with SASE & Enterprise Browser

PANW’s Strata & Prisma SASE business is evolving beyond traditional firewall-led security, capturing market share through its best-in-class offerings rather than just leveraging its firewall installed base.

• SASE Leadership – PANW has >5,600 active SASE customers, growing +20% YoY, with $1m+ SASE deals up 2.5x YoY.

• BoB Enterprise Browser, a Game Changer – PANW’s Talon Security acquisition gave it the only BoB enterprise browser in SASE, now with 1M+ licenses sold and a clear path to $1bn ARR.

Unlike competitors relying on premium pricing or bundling tactics, PANW’s secure enterprise browser provides:

• Granular security controls at the endpoint.

• Faster, localized security enforcement.

• A superior UX, enhancing SASE adoption.

This move is far more than an incremental product play — it’s a strategic bet on the browser as the enterprise’s new security perimeter (for more information check out our previous report). If PANW can continue embedding security into the browser — as OpenAI is rumored to be doing in consumer AI experiences — this could become a multi-billion-dollar opportunity.

2nd S-Curve: Cloud Security – From Agentless to Real-Time Protection

A few years ago, next-gen agentless cloud security (pioneered by Wiz and Orca) emerged as the dominant paradigm. Vendors positioned quick deployment & compliance adherence as the main selling points. In our Cloud Security Series, we anticipated that eventually this alone would not be sufficient, and that demand real-time protection would emerge.

The industry is now demanding real-time cloud protection — favoring vendors that can deliver dynamic, portable agents for live threat detection and response.

PANW has built exactly this capability in Prisma Cloud, enabling it to address the cloud security trilemma, where historically vendors have needed to balance a tradeoff between comprehensiveness, timeliness, and ease of deployment. Now PANW is leading in both agentless and agent-based cloud security:

1. Comprehensiveness – Covers agent-based & agentless approaches.

2. Timeliness – Leads in real-time threat mitigation (via Cortex & shift-right capabilities).

3. Ease of Deployment – Quickly deployable with agentless coverage, quickly supplemented with lightweight, portable agents that can be deployed in a highly scalable manner.

This real-time, shift-right security model has given PANW the edge in Cloud Detection & Response (CDR) — a critical gap that agentless vendors fail to address.

Like S, PANW is also leading in AI-SPM for GenAI Workloads, positioning themselves favorably as enterprises integrate LLMs & AI-powered applications. This ensures that AI applications — which will become central to enterprise security — are governed and protected natively within Prisma Cloud.

3rd S-Curve: SecOps – The Most Transformative Opportunity

SecOps is the most labor-intensive domain in cybersecurity, with SOC analysts drowning in alerts, struggling to separate signal from noise while attackers automate their techniques at scale.

PANW is solving this with XSIAM, which has now surpassed $1bn ARR. SIEM disruption is already happening, and PANW is leading the charge.

• XSIAM replaces traditional SIEMs (Splunk, Qradar, ArcSight) with a big-data-driven, AI-powered SOC.

• The IBM Partnership accelerates penetration – IBM has exited SIEM, handing its Qradar customer base & IP to PANW.

• Automating SOC Operations – PANW already reduced its own SOC workforce by 33% using XSIAM’s AI automation.

The opportunity here is enormous — legacy SIEM solutions are not built for modern security operations, and PANW is capitalizing on this shift at scale.

Valuation – The Security Platform of the Future

PANW is executing on three simultaneous growth curves, each backed by structural market shifts:

1. NetSec (1st S-Curve) – SASE dominance, fueled by its enterprise browser as a security perimeter.

2. Cloud Security (2nd S-Curve) – Real-time security via dynamic agents & shift-right capabilities.

3. SecOps (3rd S-Curve) – SIEM disruption via XSIAM’s AI-powered SOC automation.

With its Platformization strategy, PANW is securing higher spend per customer while continuing to win new business through BoB solutions.

The biggest risk in cybersecurity platformization is stagnation — but PANW has designed its model to prioritize innovation, ensuring it stays ahead of emerging security trends.

For investors, this means PANW isn’t just another vendor bundling solutions — it is building the future of enterprise security.

However, currently our optimisim for the business and the valuation are not exactly aligned. Based on our base case parameter settings, we see PANW as trading in the fair value range at present. Hence, we're waiting for a substantial correction before adding to our position.

However, there are signs that PANW could surprise to the upside and deliver higher growth than the 14% and 15% expected for FY25 and FY26 (PANW's fiscal year ends July 31st). Potentially, the key sign is the return of growth of Net New ARR, as this is a good indicator of new customer acquisition and upsell success. It is very possible that Net New ARR growth is emanating from the platformization strategy - I guess we'll learn more in the coming quarters.

Instacart (CART) and DoorDash (DASH) are two of the most strategically important players in the evolving landscape of last-mile commerce. While both monetize through consumer and merchant fees and are expanding into retail media, their approaches couldn’t be more different — and those differences may define who emerges stronger over the next decade.

Two Models, Two Philosophies

CART is building deep integrations with grocery retailers, aiming to become the digital operating system for how families plan, shop, and eat. Its model is quality-first: a single shopper picks and delivers each order, ensuring item accuracy and personalized service. CART wants to help families eat better, reduce stress, and stay on budget — and it's investing in AI to power that vision.

DASH, on the other hand, is a logistics machine. Born in restaurant delivery, it prioritizes batching, route optimization, and speed. Whether it's sushi, snacks, or shampoo, DASH doesn’t care what’s in the bag — only that it gets delivered quickly and cost-effectively. Its modular approach separates picking and delivery, making it highly scalable across categories and geographies.

Retail Media: A High-Stakes Margin Game

Retail media — ads embedded into the shopping experience — is becoming a critical profit lever. CART has the edge here. Its tight POS integrations, tools like Carrot Tags and Caper Carts, and SKU-level attribution allow it to offer brands real-time, inventory-aware ad placements that convert. Retail media now makes up ~35% of CART’s revenue and ~90% of gross profit.

DASH, meanwhile, is playing catch-up. Most grocery orders are still fulfilled using prepaid Red Cards and limited integration, making ad targeting less precise. That said, DASH’s sheer order volume and high-frequency user base offer promise. If it can scale its retail media capabilities, it could significantly lift gross margins from ~50% to the 65–70% range.

AI: Agentic vs. Operational Intelligence

CART’s AI ambitions are focused on decision-making and personalization. Imagine telling the app: “I’ve got $150, I want low-sodium, gluten-free meals this week,” and CART builds your cart with recipe pairings, substitutions, and delivery logistics. With Fidji Simo now Head of Applications at OpenAI, and still Chair of CART, there's reason to believe CART could lead in agentic AI for household grocery planning.

DASH’s AI is more execution-focused: optimizing order batching, routing, and Dasher efficiency. It’s also experimenting with autonomous delivery via DashMarts, its own micro-fulfillment centers. These proprietary hubs give DASH control over inventory, layout, and — eventually — automation. While not as flashy as CART’s AI assistant vision, this infrastructure could prove critical for long-term margin scalability.

Strategic Control vs. Agility

CART is doubling down on the grocery vertical. It’s embedding into retailers' systems, powering white-label ecommerce, and improving store-level inventory visibility. But that focus comes with tradeoffs. Competition from Amazon/Whole Foods and Walmart looms large, and CART has shown little appetite for geographic or vertical expansion beyond grocery. Its future is tightly tied to its retail partners’ ability to evolve — and that could prove risky.

DASH, by contrast, is aggressively expanding across categories (flowers, convenience, groceries) and geographies. Acquisitions like Wolt (Europe) and Deliveroo (UK) have extended its global reach. DASH’s willingness to onboard grocers without full integration — using prepaid cards and Shop & Deliver models — means it can scale faster, even if it sacrifices some data fidelity and control.

The Investment Case

Valuation is where the story flips. CART, with ~75% gross margins and 23% trailing FCF margin, trades at just 22x EV/FCF (ex-SBC) — suggesting investors are pricing in prolonged stagnation. Yet online grocery penetration in the U.S. is only ~12%, and if CART can grow at 10% with 30% FCF margins, the stock could be worth nearly double.

DASH, growing faster at 20% with 18% FCF margins, trades at a rich 88x EV/FCF. But its execution track record, scale advantage, and global playbook — reminiscent of companies like CrowdStrike or Datadog — make it the more predictable compounder. If it lifts gross margins into the 60s via retail media and automation, the current valuation may be justified.

Who Wins?

If the future shifts toward intentional consumption — health, affordability, AI-assisted planning — CART is structurally better positioned. It’s building deep moats around personalization, grocery loyalty, and ad targeting.

If convenience remains king — fast delivery, category expansion, and operational excellence — DASH’s scale, speed, and adaptability could outpace CART over time.

In truth, both companies may dominate in different verticals. CART has more upside if execution improves and AI adoption accelerates. DASH offers safer compounding through relentless efficiency and broader TAM.

As an investor, the choice comes down to this: depth or scale?

Jensen Huang recently claimed on the a16z podcast that AI will create jobs, not destroy them. His argument centers around the democratization of software creation through AI tools like ChatGPT, which allow anyone — not just professional coders — to develop applications.

On this point, Jensen is undeniably correct. AI has lowered the barriers dramatically. Yet, does democratization of software creation necessarily translate into job creation?

We've observed a similar phenomenon before, notably with YouTube. It empowered millions of previously excluded creators. But crucially, it didn't translate into millions of stable jobs. Most YouTubers earn minimal income, with only a tiny minority achieving significant economic success. Why? Because capitalism rewards scale — creation does not automatically lead to compensation.

However, AI-driven disruption differs markedly from the YouTube scenario. YouTube represented optional, side-hustle opportunities. AI, especially with GenAI and eventually AGI, threatens core, full-time knowledge worker roles in non-revenue-generating corporate departments.

Consider the IT department. Its final state is clearly defined: networks provisioned, secure, employees granted resource access, and tech support reliably available. Once achieved, maintaining this state requires fewer people due to automation and intelligent AI agents.

HR and finance departments follow the same pattern. Tasks like onboarding, compliance, leave management, accounting, and internal financial reporting will increasingly be handled by autonomous agents, significantly reducing employee headcount.

In short, non-revenue-generating roles face serious and immediate displacement risk due to clearly defined end states that AI can efficiently achieve and maintain. In contrast, the future for revenue-generating roles is far less certain and nuanced, driven by market dynamics, productivity gains, and scalability constraints.

Revenue-generating departments — sales, product development, R&D — may not shrink, but could in fact expand due to AI-driven productivity gains. This theory hinges on fundamental economic principles. AI dramatically boosts productivity, effectively shifting the supply curve outward. For revenue-generating roles, unlike non-revenue-generating roles, there's no clearly defined "final state" or ceiling. Companies continuously seek to increase revenues, meaning the productivity boost from AI incentivizes them to scale up, hiring more employees or deploying more AI agents until marginal returns diminish to zero.

Let's test this theory practically with concrete examples:

Software Developers: Developers produce revenue-generating products (apps, software, services). AI boosts each developer's productivity. Companies are thus economically incentivized to keep hiring developers because each incremental hire, supported by AI tools, continues to deliver attractive marginal returns. Since software developers can work remotely, there are few physical constraints limiting this scaling. In short, software development roles could feasibly grow significantly due to AI.

Teachers: Teachers produce educational content — essentially, a revenue-generating role as they create a "product", which is educational content. As such, AI will enable each teacher to produce more and higher-quality educational content. However, physical constraints like classroom capacity limit how many additional teachers a school can realistically hire. Yet there's nuance here: AI might enable more personalized online education or virtual schooling options, theoretically allowing institutions to "hire" more teachers for digital classrooms — expanding the market beyond physical constraints.

Marketing Professionals: Marketing roles directly contribute to revenue generation by attracting new customers and retaining existing ones. AI dramatically improves marketing productivity through personalized, targeted campaigns and data-driven strategies. Companies thus have strong incentives to expand marketing departments to leverage these productivity gains — investing more in marketing efforts until marginal returns eventually diminish.

However, the split between investing in additional revenue-generating humans and deploying more AI agents is complex and uncertain, even without physical space constraints. If the productivity of AI agents is capped due to constraints like compute availability, companies may be forced to limit their hiring of new employees. Currently, even major cloud providers (hyperscalers) struggle to support widespread, agent-driven workloads at scale. Furthermore, market saturation poses another constraint. Certain markets simply aren’t large enough to sustain an endlessly increasing number of sales reps, even if each is highly productive with AI support. Thus, the precise balance of human-to-agent investment, dictated by the associated point of diminishing marginal returns, will vary significantly by vertical and specific market conditions, making predictions challenging.

Historically, disruptive technologies have led to "job upgrades", not permanent unemployment. For example, machinery once displaced most agricultural labor, yet humanity successfully transitioned from physical to intellectual labor. But GenAI and especially AGI might challenge this historical precedent since they target intellectual labor directly.

If intellectual roles become fully automated, what's the next upgrade? Possibly cutting-edge STEM roles, advanced creative positions, and frontier scientific research. But realistically, the majority cannot quickly upgrade into elite intellectual roles.

When AGI eventually surpasses even our most sophisticated intellectual capabilities, capitalism itself may need reevaluation. Radical alternative economic frameworks — including forms of socialism or universal basic income — could become necessary to preserve societal stability.

In conclusion, Jensen Huang is right about democratization — AI has made creation more accessible than ever. But he may be overly optimistic about net job creation unless our economic structures evolve meaningfully.

Roles in non-revenue-generating functions face serious displacement risk, as AI efficiently achieves and maintains their defined end states. Revenue-generating roles, by contrast, may multiply — but only to the extent that market demand, compute availability, and marginal productivity allow.

And if capitalism continues to reward scale above all else, billions may find themselves participating in creation, but excluded from compensation.

One practical solution remains: investing in AI itself. Today, everyday investors have limited access, typically only able to invest in mature, publicly traded AI-focused companies. Expanding opportunities for early-stage AI investment to the general public, as advocated by figures like Jason Calacanis, could help mitigate the emerging economic disparity driven by AI. Policymakers should explore avenues allowing broader participation in early AI investment, potentially transforming AI's impact from economic displacement into widespread prosperity.

• CVNA’s unique model integrates an online marketplace with physical inventory management, a complex operational challenge that the company has successfully executed.

• If CVNA navigates its current auto loan situation — an outcome we believe is more likely than not — the company will be well-positioned for long-term growth.

• After weighing current risks against the long-term thesis, we see some upside potential in CVNA, although the discount has narrowed in recent weeks.

The extended version of this report is available for Convequity Entry & Premium subscribers at the Convequity website

Overview & Evolution

Founded in 2012, CVNA is transforming the $800bn US used car market by pioneering a fully online car-buying experience. Traditional dealerships rely on in-person sales, which can be time-consuming, opaque, and stressful for customers. CVNA disrupted this space with a no-haggle pricing model, detailed vehicle data, and flexible delivery options.

Early on, CVNA relied on centralized warehouses and third-party logistics, resulting in long wait times and high delivery costs. To address these inefficiencies, CVNA developed its own logistics network and introduced its now-iconic car vending machines. Over time, strategic acquisitions such as ADESA strengthened its wholesale capabilities and supply chain efficiency, enhancing its ability to scale and improve margins.

Following the post-pandemic correction, CVNA shifted from a growth-first mentality to profitability and operational efficiency. Today, its vertically integrated model provides a durable competitive edge in a fragmented market with substantial digital adoption potential.

Market Landscape

CVNA operates in a highly fragmented $800bn used car industry, historically controlled by traditional, regional dealerships that rely on outdated, in-person sales tactics. The company identified a key gap — consumer frustration with dealership interactions — and introduced a digital-first approach emphasizing transparency and ease of transaction.

Beyond retail sales, CVNA has expanded into financing, insurance, and trade-ins, providing a full-stack car-buying solution. More recently, its entry into the wholesale market via ADESA enables it to serve both individual and commercial buyers, further leveraging its infrastructure for incremental revenue streams.

Sales Distribution

CVNA employs a direct-to-consumer model with no physical dealerships. Vehicles can be delivered directly to a buyer’s home or picked up from vending machines, offering flexibility and convenience. In 3Q24, CVNA sold over 108,000 retail units, marking a 34% YoY increase.

When considering distribution, a useful analogy can be drawn between CVNA and Google. Google manages the Android OS that powers smartphones but does not manufacture devices itself. Similarly, CVNA oversees the logistics and infrastructure of buying and selling cars without producing vehicles. Both companies also offer complementary services — Google via its Play Store and ecosystem, and CVNA through financing, insurance, and extended warranties — deepening their customer engagement.

To extend the parallel, conversely, Tesla and Apple represent the vertically integrated model, controlling hardware, software, and customer experience end-to-end while targeting the premium segment of their respective markets. The CVNA-Google parallel highlights an alternative strategy: rather than owning the entire stack, CVNA optimizes logistics and technology to enhance efficiency and scale.

However, key differences exist. Google enables an ecosystem of hardware manufacturers, while CVNA competes directly with dealerships that act as extensions of automakers. Furthermore, Google dominates the smartphone OS market, whereas CVNA operates in a fragmented used car industry, limiting its market share compared to entrenched incumbents. These distinctions underscore the limitations of the analogy.

CVNA's acquisition of ADESA expands its presence in the wholesale market, leveraging ADESA’s auction network to facilitate vehicle sales to commercial buyers, including dealerships and fleet operators. This integration not only strengthens CVNA’s supply chain and pricing flexibility but also provides a crucial safety net for managing inventory risks. When acquiring vehicles from individual sellers, unexpected repair needs can arise, sometimes making refurbishment economically unviable. Instead of taking a full loss or scrapping these vehicles, CVNA can reroute them to ADESA’s auction platform, recouping a portion of the cost while maintaining capital efficiency. This capability helps mitigate financial exposure and enhances overall inventory management.

Competition

CVNA faces competition from traditional dealerships, online marketplaces like CarGurus, and hybrid models such as CarMax and AutoNation that blend physical and digital sales channels.

Emerging startups are experimenting with AI, VR, and AR to enhance the buying experience. However, these solutions typically address only specific elements of the value chain, such as virtual tours or AI-powered pricing, rather than offering an end-to-end digital car-buying solution like CVNA.

Recent failures in the online vehicle retail space highlight the complexities of scaling a pure-play e-commerce model. Vroom collapsed due to poor logistics execution, excessive costs, and a lack of investment in infrastructure. Similarly, Shift’s hybrid inventory approach led to inefficiencies and financial struggles, culminating in its bankruptcy in October 2023. These cases underscore the difficulty of replicating CVNA’s model without significant capital investment and operational expertise.

Competitor Overview:

• CarMax (KMX): Hybrid retailer combining physical dealerships with strong e-commerce capabilities.

• AutoNation (AN): Operates in both new and used vehicle sales with an expanding online presence.

• CarGurus & Cars.com: Online marketplaces reliant on third-party dealerships for transactions.

• TrueCar & Edmunds: Focus on digital retailing via dealership partnerships.

• Manheim (Cox Automotive): Wholesale auto auction leader with strong logistics.

• Tesla & Amazon Autos: Potential disruptors with direct-to-consumer models.

CVNA’s hybrid approach — bridging digital and physical inventory management — distinguishes it from competitors that lean heavily toward either retail or marketplace models. Only CarMax and AutoNation have demonstrated resilience post-pandemic, suggesting that evolving from brick-and-mortar to online is a more sustainable transition than the reverse.

Entry Barriers & Competitive Advantage

CVNA’s success stems from its ability to integrate digital sales with inventory management, a challenge that has stymied many new entrants. Startups attempting to replicate CVNA’s model face steep barriers, including:

• Infrastructure costs: Developing a logistics and reconditioning network requires significant capital.

• Operational expertise: Managing procurement, inspection, and delivery at scale is complex.

• Brand trust: Convincing consumers to buy vehicles online requires a proven track record of reliability.

CVNA has reinforced these barriers through continued investments, particularly its ADESA acquisition, which expands its logistics capacity to handle three times its current vehicle volume. This is great news for investors because going forward, capex as a percentage of revenue and free cash flow, should be significantly lower than in the past 2-3 years. As more ADESA sites are converted into inspection and reconditioning centers, CVNA will unlock additional cost efficiencies and pricing advantages that should flow through to free cash flow.

The company also benefits from a data flywheel effect: as transaction volume grows, its pricing algorithms and risk models improve, further enhancing its competitive edge.

Customer Experience & NPS Considerations

While CVNA promotes a strong Net Promoter Score (NPS), third-party sources paint a more nuanced picture. Comparably, and TrustPilot data suggest a mixed customer sentiment — 67% of TrustPilot reviews are 5-star, yet 19% are 1-star, indicating that a notable minority of customers have a poor experience. However, the actual percentage of dissatisfied customers may be lower, as satisfied buyers are generally less inclined to leave reviews.

Media reports have also highlighted customer complaints, which often stem from the inherent complexities of CVNA’s operational model rather than systemic failures. The company’s rapid expansion, particularly during the pandemic-induced demand surge, likely contributed to logistical bottlenecks and service inconsistencies. As the market stabilizes and macroeconomic conditions improve, CVNA has an opportunity to refine its processes, resolve inefficiencies, and enhance customer satisfaction.

Despite these challenges, CVNA’s customer experience still appears superior to that of its largest competitors, AutoNation and CarMax, based on TrustPilot ratings. This suggests that CVNA’s tech-driven, data-centric approach to car retailing is resonating with consumers more effectively than traditional dealership models. Addressing remaining workflow issues and improving service consistency will be key to further strengthening its brand reputation and customer loyalty.

Risks: Financial Challenges & Auto Loan Exposure

From 18th December to 3rd January, CVNA’s stock dropped 30%, but has recovered well since. The volatility is largely driven by Hindenburg Research’s critical short report, which alleges fraudulent activity and financial mismanagement. As is typical of Hindenburg, the report is structured to paint CVNA in the worst possible light to benefit their short position. While it does raise valid concerns regarding CVNA’s auto loan exposure, it does not present conclusive evidence of fraud or misconduct.

Comparing this to Hindenburg’s prior short report on Supermicro (SMCI) is insightful. In the case of SMCI, Hindenburg provided tangible evidence of minor related-party fraud but weakened its argument with incorrect industry insights, such as claiming SMCI was losing market share to Dell (DELL). With CVNA, the dynamic is reversed — Hindenburg’s fraud allegations lack substance, particularly concerning related-party transactions, yet its observations on auto loan risks merit closer examination.

CVNA’s auto loan portfolio accounts for approximately 40% of its gross profit. The company originates loans and sells them to third parties, thereby transferring most of the risk. Hindenburg points out that 35% of these loans are deep subprime, a category prone to high default rates. It also claims that delinquency rates among CVNA’s prime borrowers are higher than industry averages, suggesting the company may be misclassifying borrowers as prime to secure better financing terms.

One of the more serious claims in the report is that Ally Financial, a major buyer of CVNA’s loans, has reduced its purchases, leaving the company with more loans on its balance sheet than desired. To counter this, CVNA secured an $800m loan purchase agreement with Cerberus Capital. Hindenburg argues that this transaction constitutes an undisclosed related-party deal, as Dan Quayle — CVNA’s director and former U.S. Vice President — is also the Chairman of Cerberus Global Investments, a division of Cerberus Capital.

At first glance, the lack of disclosure raises concerns, but Cerberus is a reputable firm managing over $60bn in assets with strict fiduciary duties to its investors. Given this, it is difficult to see a clear-cut case of fraudulent intent. It is more plausible that the non-disclosure was an oversight rather than deliberate misconduct.

Notably, Hindenburg’s narrative about Ally Financial distancing itself from CVNA’s loan business has been undermined by recent developments. On January 6, 2025, Ally renewed its loan purchase agreement with CVNA, committing to buying $4bn worth of auto loans over the next year. This suggests that CVNA’s newer loan cohorts are performing better than those analyzed by Hindenburg, which likely focused on loans originated in 2022-2023. While the stock has yet to recover, this renewed partnership signals confidence in CVNA’s underwriting.

Macro factors may also alleviate some of CVNA’s loan-related risks. The used car market appears to be stabilizing, with price trends aligning with broader economic conditions. Additionally, if interest rates decline, as some analysts expect, auto financing affordability could improve, reducing default risk. Furthermore, CVNA’s recent operational shifts — prioritizing profitability over aggressive growth — should enhance its ability to absorb potential losses from its loan book.

Hindenburg also casts doubt on CVNA’s rising gross profit per unit (GPU), questioning how the company improved margins despite a 20% decline in used car prices since 2022. However, this criticism ignores a key factor — CVNA has increasingly sourced vehicles through its ADESA wholesale network, allowing it to secure inventory at lower costs. The resulting margin expansion explains much of the GPU improvement that Hindenburg finds implausible.

Ultimately, while Hindenburg’s concerns about CVNA’s auto loan exposure are worth considering, its broader allegations lack strong supporting evidence. CVNA’s willingness to serve the lower-end borrower market—often ignored by rivals like CarMax—reflects a strategic bet on its data-driven risk assessment capabilities. Using AI and machine learning, CVNA can evaluate borrower risk beyond simple FICO scores, potentially making subprime lending more viable than traditional metrics suggest.

The company’s approach resembles the evolution of high-yield bond markets, where “junk bonds” were once dismissed as uninvestable but later became widely accepted as “high-yield” instruments due to improved risk-pricing models. If CVNA continues refining its loan underwriting through advanced analytics, it could transform subprime auto lending in a similar fashion. However, this strategy carries inherent risk, as a deteriorating macro environment could exacerbate defaults.

Our analysis estimates a 3.0-3.6% loss rate on CVNA’s subprime loan book, translating to potential annual losses of ~$136m (details of this analysis are available to Convequity subscribers at the Convequity website). While this is significant, it remains manageable within CVNA’s growing cash flow profile. The renewed partnership with Ally, alongside potential macro tailwinds, suggests that the worst-case scenario outlined by Hindenburg may not materialize. Investors should remain cautious but recognize that CVNA has multiple levers to mitigate risk and sustain its recovery trajectory.

COVID-19: Boom and Near Collapse

From its inception, CVNA experienced rapid expansion, but as its revenue base grew into the billions, its growth rate naturally slowed. The onset of COVID-19, however, triggered a massive disruption in the automotive industry, creating a perfect storm that temporarily reignited CVNA’s momentum. Global supply chain issues restricted the production of new vehicles, dramatically increasing demand for used cars. CVNA, along with other used car retailers, benefited from this surge, with growth surpassing 100% during the peak of the pandemic.

However, as supply chain constraints eased and the balance between new and used car demand normalized in 2022, the market corrected sharply. Used car prices fell, interest rates surged, and CVNA found itself burdened with significant debt. The company’s stock plummeted by 99% from its August 2021 high to its January 2023 low, as investors feared insolvency. Faced with a liquidity crisis, CVNA took aggressive action — restructuring its debt, cutting costs, and streamlining operations to improve efficiency.

This crisis forced a fundamental shift in management’s approach, moving away from an unsustainable “growth-at-all-costs” mentality to a renewed focus on profitability and cash flow. Since 2022, CVNA has successfully transitioned toward a more disciplined financial model, improving margins and restoring revenue growth over the past three quarters.

Valuation & Investment Outlook

With only ~1% market share in an $800bn industry, CVNA has substantial room for growth. If digital penetration in used car sales reaches 10% and CVNA captures 50% of this market, it could achieve ~$40bn in revenue.

Our DCF model with base case parameters values CVNA at $300 per share, reflecting ~ 22% upside from current levels. While risks remain, operational efficiencies, improving macro conditions, and continued digital adoption support a favorable long-term trajectory. (Note: access to the DCF valuation model is available to subscribers at the Convequity website).

Conclusion

Despite past volatility, CVNA remains a leader in the evolving used car marketplace. Its vertically integrated model, strategic acquisitions, and tech-driven operations create durable advantages that are difficult to replicate. While auto loan exposure presents near-term challenges, its long-term growth potential remains compelling.