To sustain this upward momentum, tonight’s report and forward outlook must demonstrate that Platform Revenue can maintain a clear acceleration trajectory into the second half of the year. The institutional bull thesis hinges on Platform Revenue growth approaching the +40% mark by Q4, a rare operational pace that would place Snowflake in an elite growth cohort alongside Palantir and unlock further valuation expansion.

Debate continues regarding competitive pressure from Databricks, which recently reported a run rate acceleration exceeding 1,500 basis points to above +80% year over year. However, headline growth differentials mask critical underlying unit economics, as a material portion of Databricks’ top line surge is driven by low margin pass through token reselling. In an enterprise software environment where organic top line acceleration remains extraordinarily scarce, Snowflake’s mid-30s growth profile reflects durable, high margin platform consumption rather than low margin pass through volume.

For tonight’s Q2 realization, institutional desks are targeting a Platform Revenue beat of approximately +5.5% over guidance, translating to a buy side whisper expectation of +36% year over year against official guidance of +30% and Street consensus of +30.5%. Crucially, unlike Q1 where Cortex Code contributions provided an unmodeled upside surprise, CoCo monetization is now explicitly embedded into baseline guidance, removing a key buffer and raising the bar for operational execution.

Forward guidance parameters leave little room for conservatism. To satisfy buy side models, the Q3 Platform Revenue guide must signal sequential acceleration to at least +31% to +32% year over year, implying an actual Q3 operational realization above +37%. Furthermore, full year FY27 Platform Revenue guidance must be revised upward from the current +31% level to at least +33.5% (~$5.975B). Looking out to FY28, institutional models remain anchored between +33% and +37% growth, well ahead of consensus at +26.5%, reflecting conviction that long term consumption trends remain structurally intact.

However, extensive channel checks across the supply chain over the last 90 days reveal that investor panic regarding Broadcom’s competitive standing at Google is fundamentally flawed. What the market misinterprets as competitive displacement is actually Google bringing on auxiliary design partners to assist its internal COT team which has notoriously struggled with the rollout of the v8t/Zebra fish program, while expanding TPU attach ASIC opportunities. Google’s internal chip design capabilities continue to lag Broadcom by more than 18 months in advanced physical design, leading edge packaging, and IP integration.

Far from being displaced, Broadcom remains firmly locked in as Google’s exclusive volume partner across the next four generations of TPUs (v8, v9, v10, and v11). The street continues to discount the binding, five year agreement signed in early April, which explicitly commits Google to annually increasing its TPU related purchases with Broadcom through 2031. With early IP design work for the 2nm TPU v9 progressing smoothly toward a calendar 2028 ramp and initial design work already underway for TPU v10, Broadcom’s deep IP library and 12 year codesign execution history create a competitive moat that third party design houses simply cannot breach.

For tonight’s July quarter print and October quarter outlook, the operational setup is primed to beat expectations imo. While buyside survey benchmarks are high, looking for consolidated revenue growth of +33% QoQ (+85% YoY) and AI revenues surging over 150% YoY, the core operational inflection is fully supported by shipping data.

The July quarter will reflect the aggressive volume acceleration of Google’s next gen TPU v8i, sustained strength in TPU v7, the initial ramp of Meta’s MTIA Iris ASIC, and robust demand for Tomahawk 5 switching silicon, augmented by a cyclical recovery across non AI legacy semi.

Moving into the October quarter, total FY26 AI revenue is on track to easily clear $56 billion as the TPU v8i and MTIA Iris ramps hit full stride, joined by the first volume shipments of OpenAI’s 1st generation Jalapeno XPU ASIC.

Crucially, AI networking remains severely supply constrained…Tomahawk 5 demand continues to outstrip supply, while next generation Tomahawk 6 (102T switching silicon) is effectively sold out through next year.

The primary catalyst for a violent upward rerating lies in the structural expansion of Broadcom’s FY27 and FY28 financial framework. Management’s current FY27 AI revenue target of “in excess of $100 billion” is widely recognized by institutional desks as a floor rather than a ceiling. Bottom up modeling across Broadcom’s expanding portfolio of six frontier AI hyperscalers—spanning Google TPU, Meta MTIA, OpenAI Jalapeno (1st and 2nd gen), Anthropic TPU, and custom accelerators for ByteDance and Alibaba indicates an actual FY27 AI revenue potential tracking between $130 billion and $150 billion.

Even if CEO Hock Tan maintains his signature conservatism tonight and defers a formal guide bump, any constructive qualitative commentary around backlog durability, order momentum, and FY28 visibility will shatter the bear thesis.

With net positioning far less crowded than blue chip tech peers, short interest providing an embedded bid, and an implied earnings move of 6.3%, paying current depressed multiples for the undisputed tollbooth operator of custom hyperscale compute represents one of the highest alpha risk/reward setups in global technology today.

Anyway as always ill follow up after the print, good luck!

The July quarter (Q2 FY2027) will deliver an in line to slightly ahead result, tracking baseline consensus of $2.7 billion in revenue (+35% YoY), 58.8% gross margins, and $0.93 EPS (+39.3% YoY). But much like Nvidia yesterday, the current print is table stakes, the true inflection lies in the October guide (Q3 FY2027) and management’s multi year commentary.

Q2 FY2027 Baseline Expectations & Setup

• Consensus Revenue: $2.70 billion (+35.0% YoY).

• Consensus Margins: Non GAAP Gross Margin of 58.8%, OM of 36.7%.

• Consensus EPS: $0.93 (+39.3% YoY).

• Operational Reality: Steady execution supported by robust 800G and 1.6T optical DSP demand, accelerating Teralynx 10 (51.2T) switching adoption, and early stage contribution from the AWS 3nm Trainium 3 XPU program as Trainium 2 completes its scheduled wind down.

Q3 Guidance & Datacenter Acceleration

• Q3 Revenue Expectations: Street consensus sits at $3.03 billion, but channel tracking supports a guide extending toward ~$3.10 billion (implying 13%–14% sequential top line expansion).

• Data Center Outperformance: Data center segment growth is set to outpace the broader business at 16%–18% QoQ, fueled by the volume crossover into 3nm Trainium 3, sustained optical DSP momentum, and initial tape out/build milestones for Microsoft’s 3nm Maia program (which already has design work underway for next gen 2nm Maia).

Ik the primary bear debate around Marvell has focused on custom ASIC competition and whether cloud hyperscalers will eventually disintermediate merchant partners. That framework fundamentally misunderstands the broadening “XPU attach” TAM.

The expanded Google partnership is critical operational proof of this shift. Marvell doesn’t need to win the central TPU socket to capture massive content value. By supplying high margin peripheral architectures—SRAM based AI inference offload engines, PCIe Gen5/Gen6 storage controllers, SmartNIC/DPU controllers, and CXL memory interfaces, Marvell effectively taxes the entire data center infrastructure around the accelerator.

CY27–CY28 Earnings Power & Valuation Disconnect which is what the sell side consensus models CY27 Data Center growth at +55% YoY, the compounding convergence of optical DSPs, switching, Trainium 3 volume build, Maia order visibility, and Google attach silicon sets up a far steeper trajectory. This structural visibility pushes projected CY28 earnings power toward ~$11.00 per share (versus the current Street consensus of ~$9.64).

Just as with Nvidia, short term chop or minor program transition friction (like the Trainium 2 to Trainium 3 handoff) is noise. The underlying hardware buildout is expanding from individual accelerators into total data center system monetization, securing Marvell’s position as a core infrastructure beneficiary for years to come.

Like always will update after the print with raw thoughts and if there’s any red flags. Good luck!

Supply chain checks reveal that the July quarter is tracking well ahead of official consensus. A 15% sequential increase in GB300 rack shipments to 17,000–18,000 units, paired with mid teens networking growth, puts total Q2 revenue on track for $94B–$95B against the $92.1B Street estimate.

Looking into October, an additional 13–14% sequential jump in rack shipments to 19,000–20,000 units including the initial 1,000 to 2,000 Vera Rubin racks adds a 5–10% uplift in blended average selling prices. Combined with mid teens networking growth and initial H200 deliveries to China, Q3 revenue guidance should land between $107B and $108B, delivering a clean $2.5B to $3.5B beat over the $104.5B expectation while holding mid-70s gross margins.

Yet a top line beat is no longer a positive catalyst on its own. Over the past four quarters, Nvidia’s guidance has outpaced consensus by an average of 4%, only for the stock to slide an average of 3% over seven days and 5% over thirty days.Near term numbers are table stakes…the equity is trading on macro structural clarity across four distinct fronts.

First, the primary bear narrative of custom silicon taking market share is hitting a structural cost wall. While GPUs and specialized ASICs will move toward TAM parity over time, Vera Rubin delivers an estimated 90% per token platform cost reduction relative to Blackwell Ultra. Dropping the marginal cost of compute by an order of magnitude preserves enterprise customer retention and makes migrating off CUDA economically irrational.

Second, anxiety surrounding recent massive infrastructure financing agreements is misplaced. Management will frame these nondilutive, asset backed third party structures as operational enablement tools designed to remove energy, data center, and balance sheet bottlenecks for underlying demand that already exists.

Third, the long delayed China H200 business represents immediate, unpriced upside. Initial deliveries are finally commencing, translating to roughly $3B of incremental revenue for every 100,000 units shipped.

Finally, recent fieldwork showing memory specification adjustments for Rubin and Vera CPUs is not a signal of softening demand, but pure supply-chain triage. Scaling back unit memory allocations allows Nvidia to fulfill massive compute volume commitments amidst global HBM supply constraints.

The stock has de-rated to historically cheap multiples of ~17x CY27 and ~13x CY28 EPS even as underlying earnings power accelerates. When a dominant monopoly expands its earnings baseline while its multiple compresses, short term post earnings pullbacks are noise and the structural target remains $280…

As always I’ll follow up with this after the earnings as we get more evidence! good luck

To understand why $CRWV is coiling for a structural breakout, you have to examine how market expectations were mispriced. The company entered the quarter with a heavily levered balance sheet (~91% debt ratio), making it a fav target for bears pointing to rising annual interest expenses ($650M–$730M) and a -26% TTM net margin as I've acknowledged time and time again.

But, high financial leverage acts as an asset side trampoline when operational momentum accelerates. CoreWeave has spent the last several weeks systematically dismantling every core bear narrative… I will try to address them one by one on this as we hold a significant stake within the company;

1. CoreWeave inked a landmark partnership with U.S. defense contractor Leidos to deliver secure AI cloud infrastructure for U.S. intelligence agencies and the Department of Defense. This provides bulletproof counterparty quality, embeds CRWV directly into multi year federal appropriations budgets, and creates a strict regulatory compliance moat.

2. Securing a massive 360MW data center footprint in Indonesia proves that $CRWV’s site selection team is executing globally, which if you understand anything about the asian geography, you should understand how important this is. Taiwan and Korea are power constrained and land constrained. Japan also falls in the same category. So this is a strategic entry to control the Asian markets as they’ll need more compute. In the NeoCloud land grab, energized capacity is currency, and locking down 360MW before grid bottlenecks constrain competitors validates its global expansion playbook.

3. Wall Street coverage upgrades have increasingly awarded CRWV premium price targets relative to peers like Nebius. Analysts cite CoreWeave’s near term cost controls and massive scale, while competitors face local permitting delays—such as Nebius’ grueling public hearing over unapproved on site power generation at its 300MW Vineland, NJ facility.

The Macro Credit Umbrella

The macro rocket fuel backing CoreWeave’s bull thesis arrived yesterday when Nvidia and CEO Jensen Huang partnered with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR on a historic $500B AI infrastructure financing initiative.

By organizing private credit heavyweights to underwrite off balance sheet AI datacenters, Wall Street effectively institutionalized AI compute as a primary asset class. For a hyper leveraged NeoCloud like CoreWeave, this $500B private credit umbrella:

1. Derisks debt refinancings and lowers cost of capital spreads.

2. Validates its massive $99.4B RPO backlog.

3. Guarantees an ongoing stream of alternative capital to scale hardware without diluting equity holders.

Shortsellers argue that rapid hardware iterations will force unexpected write downs, their thesis overlooks CoreWeave’s conservative 5 year straight line GPU depreciation model and high secondary market demand for enterprise grade clusters.

MetricConsensus;

• Q2 Revenue Estimate$2.55B – $2.56B (+108% YoY)

• Full Year Revenue Target $12.0B – $13.0B Establishes $18B–$19B exit run rate

• Contracted RPO Backlog $99.4 Billion Secures long term debt service coverage

• Implied Earnings Move 12.0% – 15.5% Sets up high volatility squeeze potential

As newly commissioned megawatts are fully energized and recognized as billable capacity, operating margins are projected to surge from 1.0% in Q1 toward 14.6%. Backed by a $99.4B contracted RPO backlog, debt service coverage ratios improve rapidly as exit revenue run-rates approach $18B–$19B.

The Five Pillars We’re Tracking Tonight

1. Capacity Delivery & Power Time to Market:

2. Revenue Quality & Backlog Conversion

3. Capital Structure & Debt Coverage

4. Unit Economics & Depreciation Defensibility

5. Transition to Software Orchestration

If CoreWeave meets or exceeds its $2.56B quarterly target and reaffirms its full year $12B+ revenue trajectory tonight, this rally will mark the opening leg of a multi year structural breakout.

Three days earlier, we changed ratings across our security coverage adding Tenable to the Focus List, upgrading Qualys, reiterating the consolidators because Chinese open weight models are closing the vulnerability discovery gap with the US frontier while CVE volumes inflect.

I talked about this on Fox business a month ago, the biggest concernSA back then were two things 1) the token costs, and there are a bunch of ways to reduce that as shown by openai and anthropic just this past month and 2) the over regulation, which still is a big concern but after Kimi K3, you cannot really regulate the tech itself, the market has to pick and choose.

Chinese open weight models reaching capability parity is a supply shock and it’s hitting two markets at once…the market for tokens, and the market for vulnerabilities. Everything worth positioning for flows from those two shocks.

The Token Shock

Everyone is talking about the open letter. Almost nobody is asking why it exists why the biggest names in American technology need to publicly defend the “development space” of a category where Chinese labs account for the vast majority of frontier releases.

The answer is a single number, Chinese open source models are approaching 60% of weekly token usage inside American enterprises.

The majority of AI work actually executing inside US companies, tickets triaged, documents summarized, code reviewed, queries routed increasingly runs on open weight models, and the open weight frontier is overwhelmingly Chinese. Not because enterprises are making a geopolitical statement. Because the tokens are 80–90% cheaper and good enough for the job.

Now run the counterfactual the letter is really about. Ban or restrict those models, and every workload running on cheap open tokens reprices to frontier closed rates overnight. Enterprise AI budgets already the most scrutinized line in the 2026 planning cycle, don’t absorb a multifold cost increase. They shrink the deployment instead. The letter is a 60% of the demand curve defending itself.

But to understand how structural the is, you have to understand how the share got this large this fast, understand what changed in enterprise AI this year…adoption left the experimentation era and entered the engineering era. And engineering means routing.

Workloads now get sorted into two buckets. High value scenarios, legal analysis, complex agentic workflows, customer facing reasoning where a wrong answer costs real money keep running on high value tokens from closed frontier models. That business is healthy. But high frequency, low complexity, high volume tasks, the overwhelming majority of enterprise token volume get routed to the cheapest token that clears the quality bar. And the cheapest qualifying token, in category after category, now comes from an open weight model.

The sophisticated buyers got there first. The hedge funds we speak with run exactly this architecture…frontier models for alpha relevant reasoning, quantized open source for the bulk pipeline, routing layers solving for security and token cost at the same time. What the smart money built in 2025, the Fortune 500 is standardizing in 2026.

Hold onto the split, because it’s the key to every incentive in this fight…closed labs dominate token value. Open models are taking token volume. And volume is where infrastructure gets paid.

more interestingly, who signed the letter also matters a lot here. sort the letter’s signatories by business model and it organizes itself into three clean layers.

The chip layer. NVIDIA sells compute for training and inference regardless of whose weights are loaded. A world where 60% of tokens run on open models isn’t a threat to NVIDIA…it’s the demand expansion. Restricting cheap tokens shrinks the total inference market. Jensen didn’t join X to defend a philosophy. He joined to defend the denominator.

The cloud layer. Microsoft’s economics improve with model diversity. Every model a cloud offers is a margin opportunity and open weight models are the best margin opportunity, because the cloud keeps the serving economics without paying a frontier lab’s take. Azure doesn’t care who wins the model war. It cares that the war continues.

The deployment layer. Palantir, Perplexity, Ollama…companies whose product is built on top of tokens, have an inherent drive toward the highest cost performance input available. Their margins are literally the spread between what they charge for the workflow and what the tokens cost. Cheap open tokens are their gross margin.

Chips, clouds, deployers. Three business models, one shared dependency…abundant, cheap, unrestricted token supply. The letter has the moral vocabulary of openness and the commercial anatomy of a supply chain protecting its input costs.

The other side of this trade

Anthropic has called on Congress to restrict Chinese open source model development. Whatever you think of the policy argument and as you’re about to see in Part Two, the security case is not frivolous, the commercial reading is unavoidable…it’s an acknowledgment of exactly the momentum described above.

And here’s the tell that this is structural…Kimi K3 was the catalyst, not the cause. The Chinese open source share of enterprise token consumption was surging before K3 shipped. The trendline was set by GLM, by Qwen, by the entire cadence of open weight releases closing the gap quarter after quarter. K3 just made it impossible to ignore in a single news cycle. Without K3, the incentive to seek restrictions would exist in identical form because the threat isn’t any single model. It’s the routing architecture. Once enterprises build machinery to send volume to the cheapest qualifying token, closed labs are competing against their own commoditization every quarter, forever.

so if we put all these together, one side of the American AI complex is lobbying to restrict the cheap token supply.

the other side just assembled the most valuable signature page in tech history to defend it.

This is the first open commercial conflict within the US AI industry, and the dividing line sits exactly where the token economics say it should… value capture versus volume capture.

The Vulnerability Shock

The capability convergence driving the token shock doesn’t stop at summarization and code review. It extends to the capability that matters most to attackers…finding vulnerabilities.

At the model layer, parity has effectively arrived. GLM 5.2 now ties Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5 on baseline vulnerability discovery harnesses, an open weight Chinese model matching the US frontier on raw discovery capability, with only a specialized agentic harness lifting the frontier clear.

On real world cybersecurity benchmarks, GLM-5.1 is approaching the US frontier cluster from below, and the trajectory across releases is unambiguous, each generation lands closer.

But real world offense is not benchmark performance.

Actual attacks require chaining multiple vulnerabilities into end to end exploits against hardened targets and there, the frontier still leads and the ceiling still holds.

OpenAI’s GPT-5.6 system card, published June 25, documents it plainly, the flagship “can find vulnerabilities and pieces of exploits” but was “unable to carry out autonomous, end to end attacks against hardened targets.” GPT-5.6 roughly doubled its predecessor on Easy and Medium cyber challenges while Elite difficulty success held at exactly 0% for both generations.

AI is not yet an autonomous elite attacker. It is already a tireless, rapidly improving vulnerability discovery engine and the open weight versions run outside any US lab’s safety framework, downloadable by anyone. Discovery scales now. Exploitation scales next. The window between the two is where the security industry gets paid. I wasn’t wrong about cyber security in the beginning of the year, but my take incomplete, now its clear that its very much needed for existing saas companies to survive not for the labs themselves and we have seen that with anthropic having its own security capability and same goes for openai.

And the acceleration has stopped being a forecast. First quarter severity weighted CVE counts inflected sharply, and enterprises are openly preparing for what one CISO described to us as a “tsunami of vulnerabilities” across their IT estates.

but the interesting part here is the demand side of the equation because the more discovery capability means more CVEs. More CVEs, found faster, by more actors collapses the mean time between disclosure and exploitation. Enterprises that patched on monthly cycles against human speed attackers now face machine speed exploitation of a vulnerability count that’s inflecting upward. Every assumption in legacy vulnerability management, quarterly scans, severity score triage, patch windows…breaks at the same time.

that breakage is the demand event. but the most important thing here is how we’re positioned for it…

1) Tenable, $40 target

Vulnerability volume is Tenable’s demand curve, and the platform transition gives this cycle something concrete to convert into. Hexa AI, Tenable’s Claude powered agentic engine went GA May 20 inside Tenable One tiers only, notstandalone VM. That’s the upgrade forcing function for the installed base, legacy scanning customers facing a vulnerability tsunami now have a dated, concrete reason to migrate.

The math, ~30% of customers on Tenable One today, two thirds of a 40,000+ footprint as upsell runway, 2,000+ customers above $100k ACV anchoring the economics, Tenable One at 41% of Q1 new business with ASP uplifts of 6–60%.

The underappreciated piece imo is Tenable’s own telemetry attributes 63% of breach entry points to nonCVE exposures, misconfigurations, identity gaps, attack paths, despite those being only 37% of findings. Hexa works that surface, sitting on an Exposure Data Fabric built from two decades of telemetry, the trusted execution layer that lets an agent act in production without breaking it. In an agentic security market, twenty years of telemetry is the moat that lets you automate what competitors can only alert on.

At 12.8x EV/FCF on CY27, against an FY29 framework of HSD/LDD growth, 28% operating margins, 31% unlevered FCF margins, Rule of 40, the multiple prices Tenable as a legacy scanner in a stable vulnerability environment. It is neither. Risk reward as asymmetric as anything in our coverage.

2) Qualys

The same tide lifts this boat despite real execution issues. The bear case was never the category, it was platform conversion, and that problem is intact still ETM and CSAM at just 11% of LTM bookings, up three points in two years, VMDR still 50% of mix, no multi year financial framework. Until that changes, Qualys captures less of the tailwind than Tenable.

But Underweight stopped being defensible when the core VM franchise sits directly in the demand path. Over 200 million agents worldwide, participation in Glasswing and Daybreak, and a quietly shipped multi agent platform, Val for autonomous patching, Nyra for prioritization, Sid for exposure context, Sara for risk elimination, ROCky as the conversational layer, Agent Grant emerging on identity against the company’s own stat that 74%+ of breaches involve compromised credentials. The financial floor is exceptional 47% EBITDA margins flat since 2022, Rule of 40 at 57, 45% FCF margins against a 26% peer median. When the wave hits the installed base, renewal pressure builds regardless of new product execution.

But something to pay attention to is what the government would do after the August recess. Again, Trump is a very non ideological, meaning whatever makes him money, he’ll go with. Openai and anthropic are paying the most as of right now. so interesting to watch.

Also, one last relevant fact is that President Xi of China is coming to the US on a state visit for the first time in 10 years, so I suspect there will be some kind of deal as a gesture of goodwill from Trump, which might sounds crazy right now but the Chinese know ball and so does Trump.

A week ago we published the Q2 preview and said the market had changed what it was grading that growth was assumed and quality was on trial. We laid out four questions: what’s inside Cloud growth, does backlog convert, do margins hold, and is Search AI covering its own bill.

The print answered all four. Mostly with the right answers. And the stock sold off anyway because the market decided to grade a fifth question nobody can fully answer yet…how much is 2027 going to cost?

1. Cloud composition…answered, and better than feared. Cloud accelerated to +82% growth, and this time we got the decomposition we asked for. The majority of the $514B backlog is standard GCP contracts from a broad customer mix, not a handful of AI labs propping up the number. Customer acquisition velocity more than doubled year over year. Existing customers are exceeding their commitments by 50%+. And the concentration risk we flagged this whole time is mega contracts that could reprice or walk…looks smaller with demand described as broad based across products, geographies, and industries. Google recognized TPU hardware revenue for the first time, roughly $1B+, sold directly to customers and booked gross. That’s the lumpy, lower margin revenue stream we warned about in the preview, and it’s now real. But at ~$1B against a $119.8B quarter, it’s pocket change, not substance. Watch it ramp through 2026 and hit scale in 2027…that’s when the mix question returns.

2. Backlog conversion…answered. Backlog grew $52B sequentially to $514B, and just over half is expected to convert to revenue within 24 months. That’s the matching we demanded…backlog growing alongside accelerating recognized revenue, not instead of it. The tokens tell the same story, model APIs processing ~22B tokens per minute, up from ~16B last quarter. That’s 37% sequential growth in actual consumption. The warehouse of IOUs scenario is off the table for now. This is a flywheel now.

3. Margins… the one soft spot, and it’s explainable. Operating margin came in at 34.0%, modestly below expectations. The pressure sources are exactly what we listed in the preview, third party capacity bridging deals, TPU hardware mix, Wiz consolidation, accelerating depreciation, plus what looks like elevated legal charges, which the 10-Q should clarify Thursday. Note what this is a company supply constrained enough that it’s renting third party capacity to keep growing its customer base. That compresses margin and confirms demand simultaneously. There are worse problems.

4. Search AI…answered emphatically. Search grew +17% with every major vertical contributing. AI Mode crossed 1B monthly users and is driving incremental queries, the cannibalization fear inverted. Shopping ad relevance up 20% on better intent understanding. AI Max at ~500k advertisers unlocking billions of net new monetizable queries. Revenue per query versus cost per query, the ratio we called the most important in the AI trade is moving the right direction on both ends…costs down 30%+ from the efficiency work, monetizable volume up from the product work. The second engine is running.

So why did it pull back? Capex. The 2026 guide moved from $180–190B to $195–205B, a $15B raise. And management reiterated that 2027 steps up significantly from there. Our model now has 2027 capex at $378B, up 85% year over year.

Those are numbers that make generalist PMs reach for the sell button if you know what I mean, but its not that bad;

First, the raise reflects accelerated delivery of servers and data centers to serve contracted demand…pulling capacity forward, not paying more for inputs. In a supply constrained environment, delivering compute faster than your competitors is the entire game. Second, the demand indicators behind the 2027 step-up are long-term deals and renewals, the highestquality signal capex can follow. This is demand driven spend, and the distinction we drew in the Kimi K3 note applies here…disciplined and demand driven gets bought, and speculative and supply push gets sold. Google’s print sits firmly in the first category, even if the market’s first reaction didn’t.

The funding is the part to watch honestly ~$20B of debt and ~$85B of equity raised in Q2. The company says it won’t return to the equity markets, and we’d expect internal cash flow plus incremental debt from here against a balance sheet that remains resilient. But be clear eyed, the AI buildout has reached the scale where even Google finances it externally. That’s the era we’re in.

FX neutral revenue accelerated to +23%. ~90% of the Fortune 100 uses Gemini Enterprise. 9M+ developers on Google’s APIs monthly. YouTube grew +13% on both direct response and brand, with World Cup strength. On models, Gemini 3.5 Flash Cyber is out, 3.5 Pro is in testing, and Gemini 4.0 is in development on the most ambitious pre training run the company has ever attempted. The frontier position is funded and defended.

On The call

We reiterate our positioning…buyers on this pullback, December 2027 target of $420—25x our 2028 GAAP EPS estimate of $16.66.

The bear case this week is a single number: $378B. Although the models tend to be less competitive and now meta’s model is leading them which should not have happened and also three of the top talent was lost to competitors (mainly anthropic), one of whom was the noble prize winner. We watched Tae Kim on fox business news on Thursday and he covered the google bear case better than anybody we’ve seen, worth watching to understand the technological version of it.

But the bull case is everything that number is buying: 82% Cloud growth, a half trillion dollar backlog converting on schedule, a Search business whose AI layer is accretive at 1B+ users, TPUs opening a hardware revenue line, and full stack monetization across Cloud, Search, and subscriptions that no other company in the market can replicate.

In the preview we said the market would take one data point on Search AI economics over ten paragraphs of guidance. It got the data point. It got the Cloud decomposition. It got the backlog conversion. The quality questions, the ones that actually decide the multiple came back with the right answers.

The market sold the cost of the answers. That’s the opportunity…

• The headline numbers will almost certainly be fine.

• Cloud will grow fast.

• Search will grow.

• EPS will land somewhere near consensus.

None of that is the test anymore. This quarter, the market is going underneath the print into composition, conversion, and cost coverage. Four questions decide how this trades.

1) What’s actually inside Cloud growth?

Q1’s 63% Cloud growth was the number everyone quoted. It’s also a number that hides more than it reveals. The market now wants the decomposition:

How much is steady state enterprise GCP consumption…the recurring, high quality revenue that deserves a premium multiple?

How much is TPU hardware sales, which are lumpy, lower margin, and hostage to a handful of buyers?

How much is the Wiz consolidation flattering the comp?

And how much is concentrated in a few AI labs and mega-contracts that could reprice or walk?

63% growth built on consumption is a different company than 63% built on hardware shipments and acquisition math if that make sense, the analysts will catch up to this after the print if they’re smart enough.

2) Does backlog convert, or just accumulate?

Every hyperscaler now leads with backlog. The number goes up every quarter and the market has started treating it like deferred certainty. It isn’t backlog is a promise, and promises don’t pay for depreciation…

The test is matching now becomes—> does new backlog show up alongside synchronized growth in recognized revenue and operating cash flow?

Backlog that converts is a flywheel. Backlog that merely accumulates while current period revenue decelerates is a warehouse of IOUs and in a market that just spent two weeks repricing AI compute commitments, the difference is worth multiple turns of multiple.

3) Do Cloud margins hold the line?

Q1’s 32.9% Cloud operating margin bought Google enormous credibility…it proved the segment could scale profitably while capex exploded. Q2 stress tests it, Wiz consolidation costs, a rising depreciation wave from the buildout, and energy costs that keep climbing.

The bar is roughly 30%. Hold it, and the market will forgive almost anything below the operating line including free cash flow near zero. Break meaningfully below it, and the entire “AI capex pays for itself” thesis takes on water at the company best positioned to prove it.

4) Is Search AI covering its own inference bill?

This is the most important question of the four because it’s the one the entire AI trade is resting on. Google has said that after the AI Overviews and AI Mode upgrades, core AI response costs fell more than 30%. Good.

Now the other side…

Is query volume growing?

Are ad clicks and conversions improving?

Are AI answers cannibalizing traditional search ad inventory faster than they monetize?

And the one that synthesizes all of it…is revenue per query growing faster than cost per query?

If yes, Search AI is accretive and the largest advertising business in history just got a second engine.

If no, Google is running the world’s most expensive defensive remodel.

The market will take one data point on this over ten paragraphs of guidance.

The FCF number are so important and it will be misread

Free cash flow near zero…possibly slightly negative should surprise no one. The math has been visible for months.

Full year 2026 capex guidance sits at $180–190 billion. Q1 already consumed $35.7 billion, which leaves the remaining three quarters averaging $48–51 billion each. Stack Q2’s seasonal cash tax payments on top …recall Q2 2025 produced just $27.7 billion of operating cash flow and $5.3 billion of FCF even before this capex regime and a compressed or slightly negative Q2 FCF print is arithmetic, not news to dump the stock.

The distinction that matters here slightly negative is expected and survivable. Sharply negative is not. A modest dip says the buildout is on plan. A deep one says the operating engine isn’t keeping pace with the spend and in this market, that gap gets sold first and investigated later…ill follow up on this after the print!

Last week, Micron made the most credible case I’ve seen that it’s finally wrong and It made it with a number that shouldn’t be possible in a commodity business… an 84.9% gross margin, guided toward 86% next quarter, on a revenue line that beat the most optimistic estimate on Wall Street by 15%.

Then, days later, the market did what the market does. The KOSPI tripped circuit breakers, Samsung and SK Hynix cratered, $MU and $SNDK dropped double digits on Wall Street, and every trader with a memory of 2018 declared the top was in.

I want to address that selloff directly, because I think it’s the best gift long term investors have been handed all year but first, the quarter itself, because you can’t judge the selloff without understanding what it’s selling.

I went on Fox Business a day after the earrings and I tried to explain how impressive an 84.9% gross margin was and is higher than Nvidia’s most recent quarter. Higher than essentially any large US technology company has printed. And predicted on air that Micron will overtake Nvidia as the most valuable company on earth in the next 12-24 months.

This is a memory manufacturer, historically the most brutally cyclical, lowest margin corner of semis, out earning the company selling the most coveted chips on earth. And management guided it higher.

A number like that has only two explanations.

Either it’s the blow off top of a violent cycle and you should run, the traditional read, and the one this week’s sellers are trading.

Or the thing being measured is no longer the thing the old model assumes.

I’m going to argue it’s mostly the second, and the reason comes down to a single equation. Frankly this is the reason I changed the way I think about this…

In the AI inference era, token throughput, how much intelligence you can actually serve is governed by memory.

Specifically, throughput scales with HBM capacity multiplied by HBM bandwidth.

Which means that with every model generation, you don’t get to keep memory constant and add compute.

The memory itself, both its size and its speed has to roughly double.

so think about what that does to demand.

In the old world, memory demand grew with units shipped, more PCs, more phones, a mature and cyclical curve.

In the new world, memory demand grows with the square of model ambition, because capacity and bandwidth scale together, generation after generation. That’s not a cycle anymore imo, it’s an exponential, structural pull that doesn’t care whether the PC market is up or down this year.

This is why HBM behaves less like a commodity than anything memory has ever produced.

Think about what decyclicalizes a product…partial customization, structurally exponential demand, and rapid technological iteration that keeps resetting the competitive field. HBM satisfies all three.

It’s increasingly customized to specific accelerators.

Its demand is exponential by the equation above.

And it iterates so fast, Micron’s HBM4 12-high ramp is running twice as fast as HBM3E did that nobody gets to sit on a commoditized product long enough for prices to collapse. A thing with those three properties does not behave like DRAM did in 2018.

Also the cleverest part of the quarter wasn’t HBM, which everyone expected. It was that the traditional memory businesses beat too mobile, client, the supposedly mature stuff. That matters enormously, because the bear’s fallback was always “fine, AI memory is strong, but it’s narrow, the strength won’t spread.” It spread.

Why? Because agentic AI changes what ordinary DRAM is for.

Agents run long contexts, hold state across multistep tasks, run in isolated sandboxes all of which consume far more standard memory per CPU core, on top of an exploding number of CPU cores in the data center.

Even commodity DRAM is catching a structural new demand source.

NAND and SSDs get it too, KV-cache offload, AI video, agent storage are all new structural demand at the same tier.

When the strength is broad instead of narrow, “it’s just an AI niche” stops being a credible bear case.

The thing that actually kills the cycle…is the contracts

Micron signed 16 of them. These are not letters of intent. They’re five year, take or pay contracts running through 2030, with specified volumes and price floors and ceilings.

The financial commitments behind them are staggering…roughly $100 billion of contracted revenue floor, with around $22 billion in customer cash deposits and financial commitments expected about $18 billion of it as cash collateral. Customers are wiring Micron billions in advance to lock supply.

Now the line that, if it holds, ends the thirty year bear case.

Management said that even at the floor prices in these contracts, gross margins would remain far above the peak margins of any prior cycle.

The historical terror of memory was that peak margins evaporate when the cycle turns.

Micron is now contractually saying its floor is better than its old peaks and customers have prepaid to make it binding.

That is the mechanism that converts a cyclical into something with the earnings visibility of an infrastructure asset. It doesn’t eliminate cyclicality. It puts a contractual floor under it that the old model never had.

Which brings us to the selloff, the KOSPI circuit breakers, the double digit drops in memory names, the confident calls that the top is in…(sight)

A regional index that gained roughly 70% in a single quarter, where two chipmakers make up half the market cap, in a country whose currency sits near a 17 year low, unwound crowded positioning.

Korea has tripped more trading halts this year than it did in 2008.

That is a statement about leverage and concentration in Korea.

It is not a statement about HBM demand, and it is certainly not a statement about sixteen take or pay contracts with prepaid cash collateral.

Ask the only question that matters…did anything this week change the throughput equation?

Did a single hyperscaler cancel an order, walk back a deposit, or renegotiate a floor?

Did the HBM4 ramp slow?

No, no, and no.

What changed was price and price changing while fundamentals don’t is the definition of opportunity, not risk.

The selloff was positioning colliding with a headline vacuum.

Micron reported the single strongest quarter in the history of the memory industry, and within a week the stock was cheaper.

That is not how tops behave. Tops are when the fundamentals crack and the price refuses to…this was the reverse.

And on valuation, the math is almost embarrassing.

Take the contracted floor seriously… $100 billion of revenue with margins above prior cycle peaks and $MU at current prices trades at a multiple that assumes the old cycle reasserts itself almost immediately.

The market is pricing a 2018 style collapse into a company whose customers have prepaid billions specifically to prevent one.

Even after the run this stock has had, the earnings power implied by the floor not the bull case, the floor, isn’t in the price.

If the SCAs hold, this remains one of the most undervalued large caps in the market.

The week’s selloff didn’t dent that thesis. It improved the entry…

However, I’d be doing you a disservice if I made this sound free of danger…from here

First, ASP is still the engine, and ASP is the most reversible thing in this business.

Roughly 90% of the quarter’s revenue growth came from price, not volume, bit shipments grew only low to mid single digits.

Management itself flagged that next quarter’s margin reflects a clear slowdown in the pace of price increases. The direction is still up…the acceleration is easing. Price led booms are exhilarating and they are exactly what reverses hardest if the supply demand balance shifts.

Second, the whole thesis rests on supply discipline holding, and the test arrives in 2027-2028 when new capacity comes online. T

he reason margins are this high is that effective capacity is near full and new supply can’t arrive fast enough enormous fab build times, labor and energy constraints, HBM cannibalizing wafer capacity, limited cleanroom space.

Those constraints are real and they’re why this works. But every prior glut was built at exactly this moment at the top, with everyone certain demand was permanent.

The bull case is that the SCAs and structural demand make this time genuinely different.

The bear case is that “this time is different” is the most expensive sentence in markets, and it’s being spoken at 85% margins.

Third, the underappreciated one even Micron says future high value products will carry higher per bit costs. The transitions to next gen HBM, DDR6, LP6 add complexity that pushes blended cost per bit up.

The margin tailwind from pure pricing has a counterweight building underneath it.

Trade responsibly, ill follow up on this when micron overtakes Nvidia…again this is my personal thought and not an advise.

Start with where the stock is at, because the setup is everything. Memory is the most crowded long in technology right now.

And it’s crowded for good reasons, 1) the demand picture keeps improving in ways the models didn’t capture, 2) most recently the agentic CPU boom that nobody had penciled in six months ago, 3) and the ASP trajectory is rising with basically every channel check and company meeting.

Last quarter DRAM prices rose 60% QoQ and NAND rose 70%. Gross margins are sitting north of 80% a level that, in thirty years of memory history, has meant exactly one thing, THE TOP.

So the problem with today. A beat is already priced. The Street isn’t at company guidance of $33.5 billion anymore, it’s at $34.8 to $35.5 billion in revenue, EPS near $20, gross margin around 83%. Just clearing the company’s own guide does nothing.

To move this stock, Micron has to clear the most optimistic version of the Street, and then tell a story about why it lasts. The print is the easy part. The durability is the trade…let me break down what actually matters, in order;

1) the quality of the beat…not the size. A headline revenue number is meaningless until you decompose it. There are four sources of the upside, and they are not created equal.

ASP, price increases is the strongest part of any cycle and also the most fragile, because what tight supply gives, loosened supply takes away.

Bit shipments, actual capacity sold is more durable, because volume growth isn’t as easily reversed.

Product mix shifting toward HBM, high end DDR5, enterprise SSDs lifts margin structurally. And cost reduction is the hardest, highest quality profit of all, because it’s earned, not granted by the cycle, and competitors can’t replicate it overnight.

If today’s beat is mostly ASP, the bears have a point about fragility. If it’s mix and cost and bits, the story has legs. Listen for the decomposition on the call too.

2) bigger than this quarter, the guidance. This quarter is history the moment it prints. What rerates the stock is forward visibility. Goldman’s model for next quarter is aggressive, revenue near $48.8 billion, gross margin 86%, EPS approaching $30. Consensus is far more conservative, around $40 billion and $24 EPS. That gap is the whole battleground. If the guide comes in at $45 to $49 billion with EPS in the high 20s, the bullish sell side models get validated and the target price upgrade wave continues.

If it’s $40 to $42 with EPS at $23 to $24 still enormous numbers, it merely realizes the current stock price rather than catalyzing the next leg. And if it comes in soft while management dodges 2027, the pullback pressure is immediate and sharp, because at these margins, “in line” reads as disappointment.

3) the thing almost nobody outside our trading desk is watching closely enough, the Strategic Customer Agreements. This is the core of the entire bull thesis, and I can’t overstate it.

SCAs are Micron’s new multiyear contracts, 3 to 5 year DRAM and NAND deals with upfront payments and financial commitments, fixed volumes, and negotiated pricing that’s locked for the first part of the contract and variable for the rest. They had negotiated exactly one as of last quarter. The question today is whether there are more, and what terms get revealed.

I’ll expand on this because it’s so important…for three decades, the bear case on memory was never that it couldn’t make money, it’s that the market never believed the money would last.

Peak margins always meant a peak stock, because the next glut was always coming. SCAs are the mechanism that could finally break that reflex.

If Micron can demonstrate that a meaningful chunk of 2027 and 2028 demand is contracted fixed volume, committed price, real downside protection then memory stops being a cyclical trading at a peak discount and starts being valued like an AI infrastructure asset with visibility. That’s a multiple rerate imo.

So on the SCAs, listen past “we have robust multiyear agreements.” That’s last quarter’s language and it’s not enough.

Listen for five specific things, 1) coverage scope, does it span HBM, DDR5, LPDRAM, NAND and enterprise SSDs, or just AI data center products? 2) Commitment strength, is it take or pay, capacity reservations, prepayments, hard minimums, or just softer forecasts? 3) Pricing mechanism, are there floors, formula based pricing, ROIC linked terms? 4) Cancellation costs, if demand softens in 2027-28, do customers pay to walk, or do they just walk? 5) And capex linkage, is the expansion premised on these commitments, or is Micron levering up at peak prices and hoping?

If management can’t convince the market that downside margins are protected, the SCA help to the multiple gets discounted hard. If they make investors believe part of 2027 profit is already in the bank, the targets go higher.

On this, I’ll give you what I’m actually hearing from the buyside, because it’s the part you can’t get this anywhere else.

The desks are overwhelmingly positioned for more SCAs roughly three quarters expect additional agreements announced Wednesday, with only about one in ten thinking Micron holds off to preserve pricing leverage. On coverage, expectations split, about half expect SCAs confirmed but with no specific percentage disclosed, maybe just qualitative color on customer type like hyperscalers, about 40% are modeling a specific number, and that cluster lands around 30 to 50% of forward bit shipments.

On ASPs, the survey is basically in line with consensus, call it mid 40s percent for both DRAM and NAND. That’s the bar. Beat the SCA disclosure expectation and this works, merely meet it and the stock has to find a new catalyst.

Fourth, HBM4 with a counterintuitive twist. Micron’s HBM4 36GB 12 high enters high volume production in the second half, targeting Nvidia’s Vera Rubin, with bandwidth over 2.8 terabytes a second and 20%+ better efficiency than HBM3E, the 48GB 16 high samples are already out. HBM4E ramps in 2027.

But again, management said last quarter that nonHBM margins are currently higher than HBM. That 80%+ gross margin is being driven largely by extreme tightness in traditional DRAM, not by HBM alone. So when you hear the HBM roadmap, ask the questions that actually matter, do HBM4 yields mature faster than HBM3E? Does Micron hold or grow its HBM share against Samsung and SK Hynix, who are both already sampling HBM4E? And more importantly as the HBM mix rises, does overall gross margin keep climbing, or does advanced packaging cost dilute it? Is Micron a core supplier on Vera Rubin and Rubin Ultra, or a third tier filler? Share stability is the whole game, HBM4E is not a one sided story.

Fifth, the segment the market underrates, datacenter NAND and enterprise SSDs. This is what separates Micron from a pure HBM story.

AI inference, vector databases, and KV cache offload are driving data center NAND demand, Micron’s data center SSD share has risen four straight years, data center NAND revenue more than doubled QoQ to a new high. They’re shipping PCIe Gen6 SSDs now, the 9650, built for AI training and inference, 28 gigabytes a second, 5.5 million read IOPS.

This matters three ways, 1) it makes Micron an AI memory and storage provider, not a one trick HBM name 2) it upgrades the quality of the NAND cycle toward high value pools. 3) it explains why NAND pricing is strong on its own, not just dragged along by DRAM. If the call only talks HBM, the story feels incomplete. If they connect HBM, traditional server refresh, agentic AI CPU memory, LPDRAM, and enterprise SSD into one coherent demand picture, the valuation logic gets a lot more solid.

Now the three things that can break it…1) Capex. Last quarter the stock dipped after hours partly because a strong print was offset by a big capex hike over $25 billion for the year, with more coming in 2027. That’s the cyclicality ghost, spending at the top is how every prior glut got built. The framing today is everything. If the capex is presented as long term investment underwritten by customer SCAs and prepayments, with a clear ROIC target and the ability to delay if 2027 demand disappoints, it’s a positive. If it looks like a classic peak of cycle expansion, it’s a discount. Watch capex and SCAs together, they’re the same question asked twice.

Margin sustainability. At 80%+, the market won’t reward another two points. Management already cautioned that further price rises have diminishing margin impact from here. The real question, the one investors are circling, can 2027 margins sustainably hold above 70%? That threshold is the signal. The focus isn’t whether Q3 prints 82% or 83%, it’s whether management hints that even if prices pull back, the combination of SCAs, mix, cost curves, and supply discipline keeps margins at historical highs through the cycle.

Demand destruction. It’s real and you can’t wave it off. Prices rising this fast push PC, phone, and consumer customers to spec down memory or delay purchases. Management has acknowledged high prices are impacting some price sensitive demand even as the overall picture stays strong and they’ve said they can currently meet only 50 to 66% of customer need. Today, listen for whether PC and phone configs are dropping memory content, whether OEMs are delaying, whether auto and industrial are accepting the hikes, and whether customers are pulling demand forward (which borrows from 2027). Clear demand destruction is what would make the market fear 2026-27 is the peak.

So here’s where I land. The bull case is genuinely unprecedented, an elongated cycle, contracted demand through SCAs, and free cash flow generation so large that there’s a credible path to Micron producing something like a third of its current market cap in cumulative free cash by the end of 2028. That is not a cyclical stock outcome. That’s an infrastructure asset outcome. The bear case lives in one number and one reflex, gross margins this far north of 80% have only ever marked a top, and the entire SCA thesis is a bet that this time the supply discipline holds and the contracts make it stick. The valuation tension is real even among bulls, plenty of smart money doesn’t want to pay much more than 7 to 8 times earnings for a memory name no matter how good the story, because they’ve been burned believing “it’s different this time” before...

1. the United States and Iran signed a memorandum of understanding to reopen the Strait of Hormuz. The agreement pauses hostilities for 60 days, lifts the US naval blockade, allows ship traffic through the Strait, and specifies that Iran will not levy ship tolls for the duration. Shipping volume should rebound as freedom of passage is restored. The most acute risk to the global economy, a sustained closure drawing oil inventories below operational stress levels is now behind us.

2. Kevin Warsh chaired his first FOMC meeting, and it was hawkish well beyond what anyone expected.

This is the cross current that defines the moment. The energy shock that has hung over every forecast since the conflict began is collapsing. And at the same time, the Federal Reserve is signaling that the inflation generated by that shock combined with a labor market that keeps firming is pushing it toward rate hikes that markets were not positioned for. Oil down. Fed up.

Two forces pulling the second half outlook in opposite directions…

Warsh

We have anticipated for months that the Fed’s next move would be a hike. The magnitude of the hawkish tilt at last week’s meeting was still surprising.

Moving well beyond the removal of a 2026 rate cut, the June dot plot shows nine governors anticipating hikes this year with six looking for two or more moves.

This is a committee that has flipped its entire posture from cutting to hiking which is against the thesis that president trump hired Warsh for.

The motivation is clear in the revised projections, core inflation for 2026 moved up to 3.3% and for 2027 to 2.5%, while the committee shifted from a downside to a balanced assessment of growth and labor market risks. The message could not be more direct the focus of concern has shifted toward inflation, and rate hikes are under active consideration.

The market response was significant, and it underscores the power of Fed forward guidance in monetary policy transmission. Which is exactly what makes the Warsh communication strategy so consequential because he is trying to dismantle that tool while sitting on top of it.

Warsh has regularly criticized the use of forward guidance, and he received broad support this week in paring the FOMC statement down to a brief description of economic performance and two commitments, the 2% inflation goal and the ample reserve policy.

At the press conference he called for an even more substantial pullback, arguing that “financial markets work less efficiently when they ask a question, How will the Federal Reserve react to that incoming information?” He did not present his own dots. He deflected questions on his assessment of the outlook and the policy path required to deliver price stability. He downplayed the submitted dots themselves, suggesting they were “coming in with pencils, you know, those kind with the big erasers.”

This is the fog…Warsh’s recurring emphasis on returning inflation to 2% positions him as a hawk, our natural language scoring of the press conference confirms it.

But his refusal to elaborate on how inflation returns to target, combined with his remark that there is no trade off between achieving the Fed’s two mandates, makes it genuinely hard to assess where he stands on the policy path.

The uncertainty may not matter much if Warsh’s leadership does not magnify the influence of his single vote. The committee’s guidance is the most valuable guide available and the committee is already evenly divided on hiking this year.

We have not changed our base case. We still expect the first hike to be delivered next year, as a moderation in the pace of headline and core inflation arrives in the coming months. But the signal from the FOMC is unambiguous, the sustained rebound in job growth and elevated core inflation we forecast will deliver hikes, and the risk profile is now skewed toward the first move coming earlier than we anticipate. Hawkish surprises to the 2H26 forecasts embedded in the projections, stability in unemployment at 4.3% and a slide in core PCE to a 0.22% monthly pace, are exactly the outcomes that would tip the Fed into early tightening mode.

Resilience as Consumers Cool

The reopening of the Strait reduces the tail risk of a price spike and raises the possibility that oil settles well below our $100 per barrel Brent forecast for the second half. US gasoline has returned to $4 per gallon this week.

But there remains a near term drag from a three-month conflict during which the Strait was closed and global CPI inflation spiked. We expect that drag to reduce global GDP growth to a 2% annualized pace around midyear, led by a downshift in consumer spending.

The consumer cooling is already underway.

After a strong start to the year, global retail sales volumes rose at a 3.2% annualized pace in the three months through March spending fell in April.

The slowing is concentrated in China, where real spending fell in both April and May.

Elsewhere, gains slowed to a crawl, with the notable exceptions of the US and Japan. Last week’s firm US May reading points to a modest overall global spending gain despite a large China drag. US consumption is tracking a 2% annualized rise in the second quarter, impressive resilience in the face of an anticipated 6% annualized rise in consumer prices.

What cushions the consumer slowdown is the business sector.

This is the heart of our second half thesis, and the news remains encouraging. Global factory output rose 0.2% month-on-month in April, lifting the three month pace to a robust 5.0% annualized.

Against the backdrop of soft gains in China, that breadth reflects underlying business demand for both tech and nontech products and, if we are right, a turn in the inventory and hiring cycles that provides a bridge as the expansion absorbs the midyear moderation in consumer spending.

Next week’s June flash PMI is the early signal.

We expect Western Europe’s businesses experienced the largest drag from the conflict, with the sentiment shock weighing on both activity and future output expectations.

That is precisely where we look for a material step up as energy prices fall. A recovery in European business sentiment is the single highest upside surprise available in the second half data.

Central bank cross currents

The oil spike lifted headline consumer price inflation to nearly 6% annualized over the last three months. Its recent decline should generate a material reversal, if oil stays near current levels, the three month change in consumer prices could slide below 2% annualized next quarter and provide a boost to second half growth.

But the elevated level of energy prices will combine with input and transportation cost pressures to sustain roughly 3% core inflation this year. Declining energy pressure alongside a more hawkish Fed creates genuinely new cross currents for central banks to process.

In Europe, the Iran-US agreement removes a tail risk for the ECB and places the current level of prices below the staff’s baseline forecast. But if lower prices deliver the sentiment lift we expect, a sticky core and a hawkish Fed will likely keep the next ECB hike on track. The BoE retains a tightening bias with a second dissent but is signaling limited urgency until second round risks are clearer that signal plus lower oil pushes our hike call from July to November. Sweden’s Riksbank has shifted its rate path up to signal roughly even odds of a year end hike, we still expect one, but now in the fourth quarter. Norges remains the clearest hawk, with a September hike as the base case and rising odds of an earlier move.

In emerging markets, lower oil reduces pressure for credibility driven hikes in energy importing deficit economies but does not eliminate hawkish constraints where FX and financial stability dominate.

The Philippines delivered a smaller 25 basis point hike than we expected pre deal, we still look for further tightening but lower energy skews risks toward a shallower path. Indonesia’s hikes remain FX driven, we look for one more 25 basis point move in July, with a hawkish Fed tilt pushing risks higher. In EMEA, the CNB’s 25 basis point hike leaves us expecting another in August with rising odds of a one and done, Hungary’s NBH resumes cuts next week as lower oil lowers the bar for easing; and Turkey’s deescalation increases the odds of earlier operational normalization, with one week repo auctions potentially resuming as soon as the July 23 meeting and lowering effective funding by roughly 300 basis points to 37%.

In Latin America, lower oil helps greenlight another 25 basis point cut in Brazil, but stronger growth, above target inflation, and rising expectations raise the odds of a near term pause. Lower oil should buy more time for Chile and Peru to stay on hold even as growth improves we still expect the next move to be a hike in both but have pushed these to the first quarter of 2027 in Chile and the fourth quarter of 2026 in Peru.

China

China increasingly looks like the weakest link in the global economy. May activity data show a sharp slowing in domestic demand as we move into midyear. The housing slump shows no sign of bottoming. Retail sales are falling. Investment posted a sharp decline. The concentrated growth in exports and tech is capital intensive, with limited spillovers to employment and the rest of domestic demand.

We have cut our secondquarter GDP growth forecast to 3.3% annualized with risks skewed to the downside. We expect growth to stabilize below 4% in the second half even with a likely stepup in fiscal support.

The mechanism that prevents a worse outcome is policy reflexivity, the fiscal impulse waned after a first quarter push, and its return should stabilize growth. A sizable enough disappointment would itself spark new stimulus on top of current plans. That reflexive floor is real, but it is a floor, not an engine. China is no longer contributing the growth impulse the global economy grew accustomed to over the past two decades.

Japan

The BoJ raised its policy rate by 25 basis points to 1.0% this week. But the meeting carried two signals that complicate the path ahead a dissenting vote from the member appointed under the Takaichi administration, and the decision to halt reductions in JGB purchases from April 2027. Both highlight continued pressure from the government to maintain a very accommodative stance.

We expect the next hike in October, then a pace of one hike every six months thereafter. But the political friction undermines the BoJ’s guidance of hikes ahead, and it will likely show up as pressure on the yen and JGB yields. A central bank hiking while its government leans on it to stay loose cannot deliver clean forward guidance and markets know it.

The data, meanwhile, ran counter to our expectation of a midyear energy drag. Large firm business sentiment rose in June. April machinery orders posted strong growth, supported by a sharp increase in shipbuilding orders a sector designated as a priority area under the Japan-US tariff agreement. We now see upside risks to our second quarter forecast for a 1% annualized GDP gain.

Memory in ai is the most important bottleneck

There is an inflation story embedded in the AI buildout that deserves its own treatment, because it is starting to show up in the consumer price data and most macro investors are not tracking it.

AI is raising near term inflation through three channels.

First and most important, DRAM and NAND flash memory have surged in price on data center demand and a supply shift toward high bandwidth memory, with older storage technologies now seeing increases too.

Second, new AI features in software could raise measured prices.

Third, data center investment may raise retail electricity prices. The hardware channel is the dominant one. Each 10% increase in hardware costs raises core CPI and PCE inflation around 0.1%, and we estimate a 0.2 to 0.4% lift to inflation given the memory price shock with some of it already in the data.

The magnitude of the memory move is the part that has not been fully appreciated.

Based on our equity analysts’ estimates, the blended price of DRAM will have risen more than 400% between 2024 and the end of 2026, with NAND flash up close to 250%.

Prices are expected to remain elevated through at least 2028. Most of the increase is set to be complete by the current quarter, but between now and the end of 2027 our analysts still expect a further 21% rise in DRAM and 8% in NAND.

Costs have started rising for older non-SSD hard drives as well, as SSDs have become prohibitively expensive and prices are not expected to fall back to prior levels in the next couple of years.

The spike is already filtering through government price indexes. The CPI for software and accessories and the PPI for storage devices have both risen 23% since the end of 2024, with the import price index for computers, peripherals, and parts up 37%.

The pass through to finished products is where the consumer feels it. The evidence from manufacturers is broad, IDC reported computer makers warning of 15 to 20% price hikes late last year.

Gartner forecast PC prices rising 17% this year and smartphone prices 13%. TrendForce predicted mainstream notebooks could see 30 to 40% increases from memory and CPU pressure.

Dell hiked commercial laptop prices 10 to 30%. Some HP OmniBook models rose 24 to 50%. Microsoft raised Surface prices 25 to 35% on midrange models. IDC estimates average global phone prices rising from $450 in 2025 to $550 in 2026 just over 20%. Tim Cook confirmed in a Wall Street Journal interview that Apple price increases are coming, with TechInsights estimating the iPhone 18 Pro bill of materials could rise 25%.

A bottom up estimate frames the consumer impact. DRAM was previously 3 to 5% of the retail price of an average phone or laptop, NAND flash 1 to 4%. Assuming 4% for DRAM and 3% for NAND, combined with the 400% and 250% price forecasts, implies a 24% retail price increase if costs are fully passed through.

Applying a 20 to 40% price change to the relevant CPI and PCE weights yields a 20 to 40 basis point lift to both core measures a slight overestimate given some weight sits in memory insensitive peripherals, but the direction and rough magnitude are clear.

The memory supercycle is a measurable contributor to the core inflation that is pushing the Fed toward hikes.

The software channel bears watching but has not yet materialized.

The risk is real when Microsoft raised Office 365 personal pricing 43% in February 2025 and linked it to Copilot inclusion, that is exactly the kind of AI feature driven price increase that the matched model approach can mistake for pure inflation, especially as subscription models eliminate older versions from sale.

So far, software inflation in the CPI looks similar to earlier decades, with the exception of a brief spike to June 2025. Each 1% of extra software inflation produces essentially no extra CPI and about 1 basis point of core PCE. We are watching it. It is not yet a driver.

The ECB

Following last week’s 25 basis point hike and hawkish staff projections, the Iran-US agreement has pushed energy prices lower.

The ECB staff scenarios provide the framework for what comes next and the key point is that the initial market reaction does not remove the case for another hike.

The staff published a baseline and three scenarios last week, including a “mild” one with faster energy price declines. The latest spot and futures prices for oil and gas have moved significantly toward that mild scenario but are not there yet markets are pricing slower declines than the mild scenario assumes.

The announced agreement may create more balanced risks around our two-hikes-in-total call, with a final hike in September. Before the agreement, that risk had shifted modestly toward a third hike.

A lot depends on how energy markets evolve, but the crucial insight is that lower energy prices would also lift growth, while preconflict inflation was stickier than expected. Relief on one side is partly offset on the other. The case for another hike remains.

The UK

Andy Burnham’s victory at this week’s Makerfield by election is not assured, but the odds are high. A win would leave him on track to become the next Labour leader and Prime Minister by late summer. Burnham has indicated he would stick to current fiscal rules and the party manifesto a cautious approach on taxation and spending that would create significant tension at the next Budget given his hints at wealth taxes, nationalization of some industries, voting system changes, and devolution of power.

The barrier to the radical agenda is the lack of a fresh electoral mandate. The decision to call an early election lies with the Prime Minister, and it is fraught with risk. Labour won an unusually high number of seats in 2024, 411 seats on just 33.7% of the popular vote, made possible by first-past-the-post. That makes the party vote poor but seat rich, and vulnerable in any new contest.

We think Labour would need a polling boost of at least 10 points to around 28% for an early election to look plausible, and closer to 30% to be taken seriously. Historical patterns suggest a Burnham bounce of around eight points, short of what is required. An early election this year is not our base case, unless Labour can make clearer gains at the expense of the SNP and Reform.

This was a week of two consequential events pulling in opposite directions, and the interaction between them is the entire story.

Underneath both, the structure of the expansion is holding. Consumers are cooling, led by China, but the business sector lift is providing the bridge, global factory output at a 5.0% annualized pace, a turn in the inventory and hiring cycles, and the prospect of a Western European sentiment recovery as energy prices fall. The expansion is absorbing a midyear consumer moderation rather than breaking under it.

Next week’s June flash PMIs are the gut check. The manufacturing survey has been strong, the services survey weak, and the composite is pointing to 1% growth while the hard data suggest closer to 3%. A moderate pickup especially in Western Europe, would confirm that the business sector bridge is real. That is the number that tells you whether the second half looks like resilience or like the start of something softer.

The energy crisis is ending. The inflation fight is just beginning. And the Fed just told you which one it is more worried about…

World Trade Securities · Global Macro Intelligence · 21 June 2026

Production hell on the Model 3 was real, the company came within weeks of bankruptcy in 2017, burning approximately $3.5 billion in free cash flow as it attempted to fully automate an assembly line using robots to handle tasks including battery module assembly and wire harness routing at a time when the technology had not attained that level of sophistication.

The Cybertruck missed on sales expectations despite genuine engineering innovation a stainless steel exoskeleton, 48-volt electrical architecture, steer by wire technology, and towing capacity that outperformed most legacy ICE pickup trucks, because elevated pricing starting at approximately $70,000 and a differentiated design that did not resonate with the mainstream truck buyer slowed adoption.

FSD was promised on timelines that consistently slipped.

Elon Musk’s political involvement has had a quantifiable negative impact on brand equity, with Tesla’s brand value declining approximately 36% in 2026 per Brand Finance, US recommendation scores falling from 8.2 in 2023 to 4.0, European registrations declining for thirteen consecutive months, and a Yale study estimating 1 to 1.3 million fewer vehicles sold between October 2022 and April 2025 attributable to his political participation.

Every one of those criticisms is accurate. None of them change what the company has actually built, because the product at the end of the day is what really matters.

Tesla has constructed the most deeply vertically integrated technology and manufacturing platform in the world, one that spans raw lithium refining, battery cell production, semi fabrication, vehicle manufacturing, autonomous driving software, humanoid robotics, grid scale energy storage, and the AI training infrastructure that ties all of it together.

No other company on earth is doing all of this simultaneously from a single integrated stack. Not NVIDIA. Not Apple. Not Amazon at any comparable stage of development.

The investment thesis does not require believing that every program succeeds on schedule.

Tesla’s historical pattern is very consistent in my opinion, it delivers behind management’s ambitious timelines but ultimately produces materially superior solutions. Model 3 and Model Y. Gigacasting. FSD efficacy. 4680 cells. Megapacks.

In every case the delays were real and the ultimate product was superior to what anyone else had built.

The pattern repeats because the organizational culture that produces the delays, urgency driven, high agency engineering operating on stretched targets, is the same culture that produces the superior solutions.

The investment thesis requires believing that when robotaxi and Optimus arrive on delayed timelines, as they always do, they will prove materially superior to the rest of the field, and that the vertical integration and AI infrastructure advantages Tesla has built will compound faster than any competitor can replicate.

The Vertical Integration Flywheel

Every major technology platform in history has had a structural starting point advantage that competitors could not buy their way into quickly.

Intel had x86 and the two decade head start in processor architecture.

TSMC had advanced node process knowledge accumulated over decades of focused investment.

Amazon had the retail infrastructure that became AWS.

Apple had the iPhone installed base that became the App Store ecosystem.

Tesla’s starting point advantage is vertical integration across the full stack of physical AI, and it is deeper, more compounding, and more difficult to replicate than the market currently prices.

Start with the bill of materials, Tesla owns the only spodumene to lithium hydroxide refinery in North America, operational as of January 2026 in Corpus Christi. It has a cathode material processing facility in Austin. Its first LFP cell manufacturing facility is coming online imminently. The 4680 cells are ramping at Gigafactory Texas. In-house semiconductor fab is next via TeraFab, a joint venture designed to consolidate chip design, lithography, fabrication, memory, advanced packaging, and testing under one roof, targeting 1 terawatt of annual compute output at full scale, with a prototype R&D fab at Giga Texas capable of producing a few thousand wafers per month already committed at approximately $3 billion of capex.

The AI5 chip, tape out complete April 2026 is being fabricated on a 2nm gate all around process at Samsung’s Taylor, Texas facility and TSMC’s Arizona fab, with volume production ramping mid-2027.

This is not a company that buys components and assembles them. It is a company that is systematically eliminating every external dependency in its supply chain from the lithium in the ground to the silicon in the chip to the software running on it.

Every step reduces exposure to third party pricing cycles and supply chain disruptions that have repeatedly damaged every other automaker.

Every step also widens the cost advantage over competitors who cannot replicate this infrastructure without decades of investment and execution.

The manufacturing innovations compound the advantage further. Gigacasting replacing hundreds of individually stamped, welded, and bolted structural components with single large aluminum castings produced by 6,000 to 9,000 ton presses, eliminates more than 100 discrete parts per casting, reduces associated joining steps, compresses body shop cycle time, and lowers labor intensity per unit, all while improving structural rigidity.

Tesla operates one of the largest die casting machines in the automotive industry and has iterated on alloy composition and thermal management to minimize post casting machining.

Volvo, Hyundai, Ford, and Toyota are now pursuing similar approaches at earlier stages of maturity. The Juniper refresh in January 2025 a simultaneous transition of factories to a new specification within six weeks demonstrates the manufacturing agility and degree of automation achieved.

The Model Y today accounts for approximately 65% of Tesla’s overall deliveries and became the highest volume car sold globally in 2023, 2024, and 2025.

The unboxed manufacturing process planned for Cybercab takes the logic further, inverting traditional automotive assembly entirely, building large subassemblies independently and in parallel before combining them at the final stage, allowing open access from every angle for simultaneous installation.

Tesla has targeted long term cycle times of approximately 5 seconds versus 33 seconds at its Shanghai facility today.

The deepest part of the flywheel is the one most underappreciated if I was being totally frank, Tesla is using its own vehicle and cell factories as the test bed for Optimus.

The robots being deployed in Tesla’s manufacturing facilities are simultaneously lowering automotive COGS, generating the training data that improves Optimus capabilities, validating the product at industrial scale, and driving down unit costs for eventual enterprise commercialization.

This is structurally analogous to how Amazon built compute infrastructure to solve its own retail scaling problem and externalized it once proven and how Kiva Robotics, which Amazon acquired in 2012, was first deployed across Amazon’s own fulfillment centers before its technology became the foundation of Amazon Robotics. Tesla is running the same playbook at a larger scale across a more complex product.

The pricing power implications are significant and not yet fully modeled. Owning more of the bill of materials insulates Tesla from third party pricing cycles. The combined effect of the lithium refinery, cathode facility, in-house LFP cells, 4680 cells, and eventually TeraFab silicon creates a compounding COGS tailwind that provides headroom for further price cuts without the margin destruction that typically accompanies aggressive pricing in the automotive sector.

As R&D spend normalizes post development and input costs continue to compress, the resulting margin expansion creates a structural flywheel for pricing flexibility lower prices drive volume, volume feeds operating leverage, operating leverage funds additional cost reduction.

The flywheel is accelerating. The starting point advantage is absolutely the craziest set up i’ve ever seen. This is the foundation on which every other part of the thesis rests

The Pattern of Execution and delivered promises

The skeptical view of Tesla’s execution track record focuses on the delays which has became the talking point of Wall Street and frankly that also became what blinded them to see the actual value of this company. The accurate view focuses on the outcomes.

The Model 3 ramp went through production hell the company acknowledged it moved too fast into automation and course corrected with a manual labor driven tent assembly line. The vehicle ultimately became one of the best selling cars in the world. Learnings from Model 3 were directly transferable to Model Y, which shared the same platform and approximately 75% component commonality, shipped on schedule in March 2020, and went on to become the highest volume car sold globally in 2023, 2024, and 2025.

US market share expanded from approximately 0.9% in 2018 to approximately 4.2% in 2023. International share expanded from approximately 0.2% to approximately 2.1% over the same period.

The subsequent pullback from Musk’s political involvement and broader BEV demand softness has given back a portion of those gains, but the historical trajectory demonstrates the ability to scale manufacturing rapidly through vertical integration.

FSD progress has followed the same pattern. The vision only approach was widely dismissed when Tesla stripped radar from vehicles in 2021 and ultrasonic sensors in 2022.

The cumulative mile count now approaches 10 billion. The Texas robotaxi rollout features a live commercial service across Austin, Dallas, and Houston. The technology is working and improving iteratively.

The Cybertruck is the one clear miss, engineering innovation without commercial success at the expected price point. It is instructive rather than disqualifying in my view. The stainless steel exoskeleton is genuinely difficult to manufacture. The pricing at approximately $70,000 starting versus approximately $37,000 for the F-150 pickup truck priced out mainstream buyers. The lesson is that Tesla’s engineering ambition occasionally outruns its commercial judgment on pricing and market positioning. That is a risk to monitor on Cybercab and Optimus, not a reason to dismiss the platform.

The pattern across fifteen years of operation is consistent, ambitious timelines, real delays, superior outcomes. Again the investment thesis is built on the outcomes, not the timelines.

FSD

Tesla’s FSD program has accumulated approximately 10 billion supervised miles as of early 2026, approaching that threshold at roughly 1 billion miles every 50 days. The active FSD subscriber base stands at approximately 1.3 million as of first quarter 2026, each generating continuous training data on every drive. Even vehicles not subscribed to FSD contribute passive data via shadow mode telemetry, though the richness of that data is materially lower than what FSD active vehicles produce.

The comparison to Waymo’s 200 million fully autonomous miles is frequently cited as evidence that Tesla’s data advantage is overstated.

The comparison is not apples to apples. Tesla’s supervised miles are lower fidelity data because the system is not bearing full decision making responsibility, and corner cases where the human intervenes early may not fully stress test the model’s autonomous reasoning.

Waymo’s lidar based sensor stack captures denser three dimensional spatial data per mile, providing more geometric precision for training perception models at a hardware cost that limits fleet scalability.

The analytical question is whether volume of lower fidelity data or density of higher fidelity data ultimately wins at the frontier of model capability.

The industry has moved past a data quantity problem to a data sampling problem further increases in volume enable evaluation of edge case instances and tail risk events at sufficient frequency to meaningfully improve model robustness. Tesla’s argument is that it has sufficient volume at sufficient fidelity.

The Texas robotaxi deployment is the live test of that argument.

The vision only approach has been a perennial source of skepticism. The cost argument against lidar has largely evaporated, sensor pricing has collapsed to $200 to $1,000 from $75,000 a decade ago but the scalability thesis remains intact. T

esla’s thesis is that autonomy is fundamentally an AI problem, not a sensor problem, and that cameras provide sufficient information if the neural network is sophisticated enough to extract it. Tesla uses differentiated direct photon counting that bypasses the image signal processor, enabling cameras to operate without being blinded by glare, perceive in near total darkness, and perform in fog at least as well as the average human driver per company publications.

A reversal to lidar is not coming. The vision only approach is so deeply embedded in Tesla’s hardware architecture, customer commitments, training infrastructure, and corporate identity that pivoting would require redesigning the vehicles, retraining the neural networks on a fundamentally different data modality, revalidating safety across all operational domains, and rebuilding the existing installed base.

Industry participants confirmed that retrofitting existing Tesla units in operation with lidar is not feasible given the material hardware and software alterations required and Tesla had the opportunity to add lidar during the HW3 to HW4 compute swap when many owners were already learning their vehicles needed hardware upgrades, and chose not to. The more plausible course correction, if one becomes necessary, is regulatory rather than voluntary.

Counterintuitively, Waymo pivoting away from lidar may prove the harder pivot. The company’s entire perception, planning, and safety validation framework is built around multi sensor fusion.

Historical conversations with developers suggest lidar is involved in approximately 90% of Waymo’s driving decisions.

Removing it would require retraining the perception stack on a fundamentally different data modality, revalidating safety across all operational domains, and demonstrating to regulators and consumers that the degraded sensor suite maintains equivalent safety.

The hardware generation progression tells the story of where the technology has arrived. HW3, Tesla’s first custom designed FSD computer, deployed from second quarter 2019, featuring approximately 144 TOPS of neural network inference capability enabled FSD Beta at scale but cannot achieve unsupervised FSD due to insufficient memory bandwidth.

HW4, deployed from 2023, features improved camera resolution, wider field of view, and significantly greater compute at approximately three times HW3, and is the minimum hardware required for unsupervised FSD currently running on Tesla’s robotaxi fleet.

HW5, tape out complete April 2026, delivers approximately 5 times the compute and approximately 9 times the memory bandwidth of a dual HW4 setup. Volume production begins mid-2027. It is designed as the common silicon architecture for both FSD inference in vehicles and Optimus inference workloads on a shared platform.

The FSD software has progressed through architectural generations that matter. V11 and prior combined hand coded rules with neural network components for perception the neural network handled what the car could see, but the driving logic was largely programmed manually.

V12, introduced in early 2024, replaced the rules based driving stack entirely with an end to end neural network, the system learns driving behavior directly from real world data rather than relying on preprogrammed rules.

V13 in late 2024 represented iterative refinement.

V14 in early 2026 improved spatial judgment, driving behavior, and object avoidance.

V15, expected in late 2026 or early 2027, is characterized as the major architectural upgrade required before large scale unsupervised deployment ramps beyond the current limited robotaxi fleet.

The transition to FSD subscription only pricing at $99 per month, effective February 2026, signals increased management conviction.

Tesla is tightening control on the technology and positioning for pricing power as autonomous capabilities improve. Unsupervised FSD for general public launch is targeted later this year with delays possible, as always.

We forecast approximately 10.2 million FSD subscribers by 2035 and approximately 14.6 million by 2040 versus approximately 1.3 million today, driving approximately $20 billion and approximately $29 billion of subscription revenue respectively.

The Robotaxi As a Service

Tesla’s robotaxi service launched in Austin in June 2025 and has scaled to approximately 104 vehicles in operation as of late May 2026, with additional fleets in Dallas at approximately 7 vehicles covering approximately 31 square miles and Houston at approximately 8 vehicles covering approximately 25 square miles per Robotaxi Tracker crowdsourced data. Austin’s service area has expanded meaningfully from approximately 20 square miles at launch to approximately 244 square miles today.

The service operates through the Tesla Robotaxi app riders request a ride, select a pickup location, are notified when the vehicle arrives, tap a start ride button to begin, and use in-app controls for cabin climate and seat position.

At the end of the ride passengers gather belongings, exit, close all doors, and provide feedback.

The operating model includes dedicated depots where vehicles are staged, cleaned, inspected, maintained, and charged between shifts.

Teleoperations centers provide remote human monitoring and intervention capability.

Recent public zoning and permit documents indicate Tesla’s intention to build dedicated car wash facilities and robotaxi exclusive Supercharger depots in Las Vegas, Phoenix, and Texas.

Micro factories in major metro areas are planned for volume hardware upgrades for older vehicles seeking eligibility for higher autonomy capability.

Availability has been a consistent friction point, press reports indicate no vehicles available approximately 25 to 30% of the time and wait times exceeding 15 minutes in roughly half of all attempts. Initial accident data filed with NHTSA points to elevated incident rates versus human drivers at approximately 4 times higher depending on the baseline. The technology is viable and improving iteratively.

Scaling meaningfully requires further progress on safety incidents however.

Model Y deployments across the current robotaxi markets are serving as a deliberate bridge to Cybercab volume launch. Every mile driven, every edge case encountered, every intervention recorded feeds the neural network that Cybercab will run on. The fleet is already in the field doing the work that matters generating the training data and the regulatory track record that the scaled commercial service requires.

The regulatory path is the binding constraint on monetization timeline.

In the US, the absence of a unified federal standard forces operators to navigate a state by state patchwork. NHTSA caps exemptions at approximately 2,500 vehicles per manufacturer per year though Tesla is already using engineering vehicles with steering wheels and pedals to self certify to FMVSS standards, which waives the cap for Cybercabs specifically.

Texas is permissive.

California requires approximately 50,000 safety test miles and now allows law enforcement to ticket manufacturers for traffic violations effective July 2026.

The pending Motor Vehicle Modernization Act would raise the federal cap to 90,000 and critically preempt state rules passage would meaningfully accelerate the monetization timeline.

Internationally, China is the most permissive major market but geopolitical barriers likely preclude meaningful Tesla participation in Chinese robotaxi operations near-term.

The EU requires Whole Vehicle Type Approval, Tesla secured its first supervised FSD approval in the Netherlands in April 2026.

The UK’s Automated Vehicles Act 2024 remains in implementation phase.

The consumer adoption curve for autonomous mobility is gated by two variables price and wait times and neither alone is sufficient.

On pricing, as of today we’re still really early. Waymo is actually pricing above incumbent ride hail at approximately $20 per ride versus Uber at approximately $17 and Lyft at approximately $15 in San Francisco per Obi research, competing on experience and novelty rather than economics.

The first meaningful adoption phase is ride hail substitution, requiring not just a lower price point but sufficient fleet density to bring wait times below the approximately 5 minute threshold that Uber and Lyft have established as the consumer expectation.

The flywheel is circular fleet density improves wait times, driving adoption and utilization, which funds fleet expansion.

As cost per mile compresses through successive thresholds, the addressable market expands in phases from ride hail substitution to personal vehicle ownership displacement to public transit competition at sub-$0.25 per mile with each unlocking not just substitution but demand creation as lower costs induce incremental trips.

We size the global robotaxi TAM using a bottom up, vehicle miles traveled based framework across four geographies, US, Developed ex-US, China, and emerging markets, each with distinct regulatory access curves, urban VMT shares, and competitive dynamics.

The model layers five sequential cost phases from current ride hail pricing at approximately $2.00 to $3.00 per mile through to Tesla’s long term Cybercab target of $0.30 per mile, onto an accelerated VMT base that compounds faster as robotaxi availability expands.

We project combined Tesla robotaxi revenue reaching approximately $300 billion by 2035 and approximately $500 billion by 2040, with an implied Cybercab fleet of approximately 35 million vehicles by 2040 and a personal AV fleet of approximately 5.5 million in 2035 and approximately 23.8 million by 2040.

Cybercab

The Cybercab is a purpose built two seater autonomous vehicle without a steering wheel or pedals, engineered specifically for commercial ride hailing. The two seat configuration reflects the premise that approximately 90% of ride hail trips involve one or two passengers and do not require a full five seat vehicle.

It is optimized for 50 to 60 hours per week of duty cycle, with reduced top speed and gentler ride dynamics that extend tire and brake life and compress maintenance costs per mile.

Cybercab is certified at approximately 165 watt hours per mile, approximately 30% more efficient than Tesla’s most energy efficient Model 3 and Y variants, driven by the vehicle’s purpose built constraints, two seats, no steering column or pedal assembly, a teardrop shaped body optimized for aerodynamics, and a sub-50 kilowatt-hour battery pack expected to deliver approximately 300 miles of range. A smaller battery pack enables faster charging times and lower per vehicle battery costs versus the Model 3 and Y.

Cybercab entered pilot production in the first quarter of 2026 at Gigafactory Texas with an S-curve ramp running through the balance of 2026 and into 2027.

The sub-$30,000 target price at scale compares to Waymo vehicles estimated at $75,000 for the current generation Zeekr Ojai and more than $50,000 for the upcoming Hyundai Ioniq 5.

In China, Baidu, Pony.ai, and Xpeng have attained all-in manufacturing costs of approximately $30,000 to $45,000 per vehicle.

Tesla’s significant cost advantage in the Western hemisphere, driven by factory automation and vertical integration, creates a structural edge in fleet scaling capacity that no Western competitor can match.

The threat of customer owned tooling, cloud and hyperscaler customers attempting to bring the full autonomous vehicle design and manufacturing capability in-house follows the same pattern as the ASIC COT debate in semiconductors.

Recent evidence is unambiguous, Google’s COT attempt on the v8t lower complexity program experienced major delays. Google’s engagement with Tesla’s technology stack for the high complexity program is proceeding.

The pattern holds across industries, COT works for lower complexity programs where the customer has sufficient in-house expertise.

For complex, high tolerance, manufacturing intensive programs, the partner with the IP and the manufacturing expertise wins.

The primary unknowns are manufacturing costs, regulatory approvals, and fleet operations at commercial scale. These are the challenges of commercializing a demonstrated capability, not of inventing something new.

Cybercab is a two to four year execution story where the variables are pace and cost.

The feasibility question has been answered by the live service running in four cities today.

Optimus “THE MONEY GLITCH” product

Cybercab and Optimus are frequently bundled as Tesla optionality. They are fundamentally different investments in risk profile, timeline, competitive dynamics, and TAM construct, and understanding the distinction is essential to the thesis.

Cybercab extends a technology stack that is already functional and deployed on purpose built hardware, in a market where competitors already operate, where the revenue model is well understood, and where the primary gating factors are unit cost, regulatory approvals, and fleet infrastructure buildout.

Optimus is attempting to create an entirely new product category at commercial scale, addressing a market that does not yet exist in any structured form, with hardware tolerances tighter than automotive standards, supply chains never tested at automotive volumes, and a go-to-market motion that bears no resemblance to selling vehicles or miles.

The shared AI infrastructure and organizational capabilities create genuine synergies. The engineering challenges, customer bases, competitive edges, adoption curves, and probability weighted timelines to revenue are materially different.

With that framing established, here is why Optimus is still one of the most strategically important assets in the Tesla portfolio.

The TAM is an order of magnitude larger than robotaxi. I think the size of the global humanoid market using a bottom up labor substitution framework spanning four geographies and five primary labor categories, manufacturing, logistics, hospitality, eldercare and personal services, and consumer households.

The model starts with addressable workers and fully loaded wage pools, then runs through three sequential gates, a capability curve determining which task types become technically feasible over time, an economic eligibility test comparing amortized robot cost against the wage pool it can substitute, and vertical specific adoption curves.

Consumer humanoids are modeled separately through a diffusion style framework.

The model suggests global industry annual humanoid sales reaching approximately 12 million units by 2035 and approximately 31 million by 2040, with the global fleet scaling from approximately 31 million to approximately 109 million over the same period. Tesla’s share reaches approximately 1.9 million units in 2035 and 4.7 million in 2040, generating combined humanoid sale revenue of approximately $46 billion by 2035 and $94 billion by 2040, with another approximately $1 to $2 billion of annual licensing revenue estimated in 2040.

Current status and hardware evolution. The Optimus program has progressed through three visible stages, the 2021 concept reveal with no physical hardware, the 2022 first functioning prototype capable of basic walking but slow and fragile and Gen 2, demonstrated in late 2023, which featured Tesla designed actuators and sensors, approximately 30% faster walking speed, 10 kilograms lower weight, improved balance and full body control, an actuated neck, human like feet with toe articulation, and hands with 11 degrees of freedom and tactile sensing.

Gen 3 the first version engineered for mass production was previewed at the 2026 Appliance and Electronics World Expo in Shanghai in March and April 2026, with the formal public unveiling expected closer to the production start date in July or August.

Tesla is deliberately limiting disclosure on the latest hardware given competitive concerns.

Gen 3 features 22 degrees of freedom per hand versus 11 in Gen 2, incorporating a tendon and cable driven design with actuators housed in the forearm, multiple degrees of freedom per finger, and improved wrist articulation.

The hand is the most critical near term hardware focus and Elon has stated publicly that the arm is now complete and it was the hardest part. Useful factory work depends far more on manipulation, grip control, force feedback, and error recovery than on locomotion alone. Human hands have approximately 27 degrees of freedom.

Replicating that requires integrating dozens of small actuators, tendons, and sensors into a compact form factor while maintaining strength, precision, and durability.

The gap between Gen 2 and Gen 3 on hand capability is the most important hardware advancement in the program.

Manufacturing buildout. The Fremont factory, which ceased Model S and X production in early May 2026, is now being converted to Optimus production capacity with an anticipated conversion timeline of approximately four months an unprecedented speed for a factory of that complexity.

Nameplate capacity at Fremont is approximately 1 million units per year.

Gigafactory Texas is being prepared for a second generation line targeting approximately 10 million units per year at long term full capacity.

The early production ramp will mirror Model 3 and Cybertruck initial snags as the company resolves critical dependencies around humanoid hands, actuator supply, balance, batteries, onboard compute, thermal management, manipulation reliability, and safety protocols.

The S-curve follows, low volume internal pilot deployment, narrow factory rollout with productive hours as the governing KPI, scale manufacturing once task reliability and intervention rates are incorporated.

Software and AI architecture. Optimus operates on the same end-to-end neural network architecture that powers FSD in Tesla vehicles.

TSLA has shifted to exclusively using a vision only training model for Optimus robots as of August 2025.

The critical scaling constraint is compute (and we all know who has all the compute in 2026 (SpaceX)) training requirements for Optimus are estimated at approximately 10 times those for vehicle FSD, reflecting the much greater complexity of manipulation in unstructured environments compared to road driving.

The AI5 chip serves as the common inference architecture for both FSD vehicles and Optimus robots.

Cortex clusters at Giga Texas provide training compute, with Dojo 3 supercomputer and TeraFab augmenting capacity as Optimus scales.

Amazon built infrastructure to solve its own retail scaling problem and externalized it once proven.

Tesla is building labor infrastructure to solve its own manufacturing problem and will externalize it once validated internally.

The structural conditions align precisely, Tesla manufacturing is a hyperscale labor consumer that justifies the humanoid investment, just as Amazon retail was a hyperscale compute consumer that justified the AWS infrastructure build.

Optimus must demonstrably reduce factory COGS or the capex becomes unjustifiable just as AWS had to be cheaper than alternatives or Amazon’s retail margins suffered.

Both require extended precommercial maturation internally before external launch.

Both exhibit compounding scale economics, cloud through utilization, humanoids through training data and manufacturing learning curves.

The iPhone analogy applies to the consumer dimension hardware as the entry point, unit economics improving via manufacturing learning curves as the BOM compresses with volume, followed by the emergence of a services layer, task and skill marketplace, third party developer ecosystem, maintenance and software updates generating recurring revenue on the installed base.

TSLA captures approximately 20% of global smartphone units but approximately 80% of industry profits through platform economics a similar dynamic could emerge in humanoids if Tesla establishes the reference platform around which the broader ecosystem organizes.

Tesla’s structural advantage in humanoid robotics relative to every competitor is rooted in three things.

First, vertical integration across hardware, software, and manufacturing combined with being its own captive first customer eliminating the go-to-market friction, enterprise sales cycles, and external ROI validation requirements constraining every other player, and creating a self-reinforcing loop between deployment, data collection, and iteration that accelerates with scale.

Second, the shared AI architecture between FSD and Optimus enabling dual returns on infrastructure investments the same training compute and increasingly the same custom silicon serves both platforms.

Third, two decades of experience scaling complex electromechanical products at automotive volumes and quality standards directly applicable to the manufacturing challenges of humanoid production in ways that pure play robotics companies and technology companies cannot replicate from their respective starting points.

The humanoid robotics industry occupies an uncomfortable middle ground between late stage prototyping and early commercial deployment. Core technologies have matured enough to execute real tasks in constrained, repetitive settings, but no operator has demonstrated the durability, intervention rates, or multi task generalization that broad unstructured rollouts demand.

Figure AI carries the strongest Western execution proof point. Its Figure 02 robots completed an 11-month deployment at BMW’s Spartanburg plant loading more than 90,000 sheet metal parts, logging more than 1,250 hours of runtime, contributing to the production of more than 30,000 X3 vehicles, on a single pick and place use case.

The Figure 03 robot is now ramping, powered by Helix, a proprietary vision-language-action model. The dedicated BotQ factory operates at approximately 12,000 units per year with monthly deliveries doubling for three consecutive months through April.

BMW is expanding the pilot into additional European facilities.

Boston Dynamics brings unmatched institutional robotics expertise and is targeting heavy industrial use cases. The next generation all electric Atlas, unveiled at CES 2026, features 56 degrees of freedom and can lift up to 110 pounds.

Mass production is targeted at approximately 30,000 units per year by 2028, with initial deployment at Hyundai’s Metaplant America in Georgia.

A partnership with Google DeepMind for Gemini Robotics AI integration adds a sophisticated AI reasoning layer.

Notably, Boston Dynamics recently brought on the former head of Tesla’s Optimus project as a group adviser and outside director.

Agility is the most advanced in real world commercialization. Its Digit robot features a distinctive reverse knee bipedal design, capable of carrying up to 35 pounds over 8-hour shifts, with safety certifications permitting operation alongside human workers.

Digit has been deployed at Amazon, GXO Logistics, and under a RaaS agreement with Toyota Canada signed February 2026.

China accounted for more than 80% of all humanoid robot shipments in 2025, propelled by supportive domestic policy, a localized supply chain for actuators, sensors, and rare earth materials, and aggressive pricing.

Unitree’s average unit price collapsed from approximately $87,000 in 2023 to approximately $25,000 in 2025, with entry level units at approximately $5,600 and gross margins improving to approximately 60% through vertical integration.

Xpeng, UBTech, Fourier, and AgiBot are each targeting distinct segments industrial automation, assisted living, logistics at costs that reflect China’s supply chain depth in ways that Western competitors cannot yet match.

Tesla’s edge is not that it is furthest along on any single technical dimension. It is that no other company has the combination of manufacturing scale, AI infrastructure, vertical integration depth, and captive first customer status that compounds across every dimension of the problem simultaneously.

Same with FSD, Waymo never had a real chance against Tesla although it operates approximately 3,800 robotaxis across approximately 11 US cities as of early 2026, completing approximately 500,000 weekly rides and targeting more than 1 million by year end, with international launches planned for London and Tokyo in late 2026.

The fleet transitions from Jaguar I-Pace vehicles to the Zeekr Ojai and Hyundai Ioniq 5. The sixth-generation sensor suite reduces sensor count by approximately 40% versus the prior generation while improving capability 13-plus cameras including a 17-megapixel imager, 4 lidar units, and 6 radars.

Waymo operates a fully owned and operated model without personal vehicle integration and targets production capacity of tens of thousands of vehicles per year from its Phoenix facility.

Zoox operates a purpose built bidirectional robotaxi with four wheel steering, no traditional driver controls, and a rider centric cabin.

The company has logged approximately 2 million autonomous miles and carried over 350,000 riders across free rides in San Francisco and Las Vegas, expanding testing to 10 US cities.

Zoox is awaiting NHTSA approval to charge for rides and operate up to 2,500 vehicles commercially.

Chinese operators are furthest along on commercialization outside the US. Baidu Apollo Go has completed more than 20 million rides with more than 300 million autonomous kilometers driven, achieved unit economics breakeven in Wuhan, and launched fully driverless commercial service in Dubai.

Pony.ai expects more than 3,500 robotaxis across more than 20 cities by end of 2026 and targets vehicle cost below $33,000 by 2027.

WeRide operates across more than 40 cities in 12 countries with approximately 1,300 vehicles.

EV-native OEMs are moving a bit too. Lucid has partnered with Uber and Nuro to deploy Gravity vehicles as robotaxis, with Uber committing to at least 35,000 vehicles.

Rivian has secured up to approximately $1.25 billion from Uber to deploy 10,000 fully autonomous R2 SUVs starting in San Francisco and Miami in 2028, with Rivian’s third generation platform featuring a proprietary in-house chip, 11 cameras, 5 radars, and lidar.

Uber is positioning itself as the demand aggregation layer rather than building its own autonomy stack, partnering simultaneously with Waymo, Zoox, Lucid/Nuro, Rivian, WeRide, and Baidu. Uber is also investing in dedicated charging and depot infrastructure for autonomous fleets.

Tesla’s differentiation in this competitive context is unit cost, data scale, and the personal vehicle network. No competitor is building a product at Tesla’s target price point in the Western hemisphere. No competitor has 10 billion training miles. And no competitor has a path to a personal AV network of tens of millions of vehicles that contribute to the robotaxi fleet when idle.

Energy Storage

Tesla’s Energy Generation and Storage segment generated $12.8 billion in revenue in 2025, up 27% year on year, at a 30% gross margin, deploying approximately 47 gigawatt hours of storage. It is one of Tesla’s fastest growing and highest margin hardware businesses and it operates with structural tailwinds that are largely independent of the autonomous driving and robotics timelines.

Global battery energy storage capacity additions totaled approximately 307 gigawatt hours in 2025 and are forecast to grow at a 15% compound annual rate over the next decade to approximately 1,231 gigawatt hours by 2035, roughly four times current levels. The Americas grows at a 9% compound annual rate. EMEA grows at 16%. Asia-Pacific, driven primarily by China, grows at 14%.

The demand drivers are structural, not cyclical especially with ai needing a new grid essentially.

Grid infrastructure is aging out faster than it can be replaced, leaving reliability gaps that batteries can fill.

Overall electricity demand is rising as the economy electrifies vehicles, heating, industrial processes. Geopolitical instability has made energy security a priority for governments globally.

Battery cell prices have fallen more than 50% over two years, with LFP cells alone dropping more than 15% in 2025 per the IEA, making longer duration 4 to 6 hour systems viable for the first time at scale and pulling forward demand that would otherwise have materialized later in the decade.

The data center opportunity is the most immediate incremental driver. Data center and AI power consumption is projected to reach 9 to 17% of US electricity supply by 2030 from approximately 4% today per the Electric Power Research Institute.

AI workloads create extreme GPU power swings that gas generators physically cannot follow Megapack addresses this through both front-of-meter grid support and behind the meter load buffering. NERC has issued a rare Level 3 alert and the DOE declared a statutory emergency due to demand and supply imbalance. PJM now has authority to curtail data center power during severe grid emergencies. The strain is real and accelerating.

Tesla management has framed the opportunity at approximately $500 million of Megapack per gigawatt of data center load roughly 1% of data center capex implying more than $5 billion per year of incremental US demand at current buildout pace and potentially $80 to $90 billion cumulative globally over ten years.

The Imperial Valley AI data center in California integrates 220 Megapack 2 XLs for 862 megawatt hours, the largest BESS installation at any US data center to date. Meta’s Wyoming AI campus uses Megapack for dispatchable on-site storage.

Fluence’s data center pipeline is up more than 30% sequentially to 12 gigawatts of mostly hyperscaler master service agreements. ABB reported data center orders up triple digits in its first quarter 2026 earnings.

Siemens Energy booked 5 gigawatts of data center order intake in the second quarter of 2026 alone.

Tesla’s manufacturing footprint is scaling to meet this demand. Three Megafactories Lathrop California at 40 gigawatt hours, Shanghai scaling from 20 to 40 gigawatt hours, and Brookshire Texas at 50 gigawatt hours ramping in late 2026 bring total nameplate capacity to approximately 140 gigawatt-hours by 2027, roughly three times current deployment levels.

The product portfolio is advancing in parallel. Megapack 2 XL is the current flagship utility scale product at approximately 3.9 megawatt hours per unit using LFP chemistry.

Megapack 3, unveiled September 2025, features a simplified thermal design with approximately 75% fewer connections and a next-generation heat pump system, with production commencing at Brookshire in late 2026.

Megablock bundles four Megapack 3 units with an integrated medium voltage transformer and switchgear into a single 20 megawatt hour AC block, targeting approximately 25% faster installation and approximately 40% lower construction costs versus prior systems.

Megapack 4 with integrated substation functionality is in development.

The software stack,Autobidder for utility scale dispatch optimization and real time energy market bidding, Powerhub for distributed resource aggregation and virtual power plants, Opticaster for real time energy usage forecasting converts hardware into recurring, margin accretive revenue streams and creates a moat versus pure play integrators relying on third party SCADA systems.

Tesla’s Virtual Power Plant already aggregates thousands of customer owned Powerwalls into a grid scale energy resource dispatched during peak demand, proving the organizational competencies for fleet level orchestration of distributed hardware exist internally.

The supply chain risk is real but being resolved. Tesla currently sources Megapack LFP cells primarily from CATL and BYD, two of its direct competitors.

The $4.3 billion LG Energy Solution Lansing supply agreement running from August 2027 through July 2030, in-house LFP cell production coming online imminently, and late stage talks for a reported $2.1 billion Samsung SDI supply agreement from Indiana progressively address the dependency.

The FEOC compliance gap through mid-2027 is a near term headwind but solvable within the investment horizon.

However, eight of the top ten global BESS integrators are Chinese, with Fluence and Tesla as the only two Western players in the top tier.

In the US market Tesla competes primarily with Fluence, Powin, Wartsila, and GE Vernova. Legacy automakers are entering BESS in earnest, Ford launched Ford Energy as a wholly owned subsidiary in May 2026, repurposing its Glendale Kentucky battery plant targeting 20 gigawatt-hours of annual LFP-based stationary storage production with first deliveries in late 2027.

GM and LG Energy Solution’s Tennessee Ultium plant has shifted production lines from EV cells to ESS cells.

IRA domestic content provisions and the tariff regime provide a structural tailwind against Chinese imports for US-manufactured products, the 45X credit at $45 per kilowatt hour combined and the 30% ITC are critical economic enablers for domestically manufactured BESS.

Tesla’s LGES agreement progressively addresses the FEOC compliance requirement that is increasingly gating institutional market participation.

The Infrastructure, AI5, TeraFab, Cortex, and the Options Nobody Is Pricing

The base infrastructure powering Tesla’s autonomous driving, robotics, and energy software platforms creates optionality for capital light, high margin revenue streams that are not in any near term model and will not be until they arrive.

AI5 and the silicon roadmap. Tape out complete April 2026. Approximately 5 times the compute and approximately 9 times the memory bandwidth of a dual HW4 setup.

Volume production mid-2027. Being fabricated on 2nm gate all around process at Samsung’s Taylor Texas facility and TSMC’s Arizona fab. Designed as the common silicon architecture for both FSD inference in vehicles and Optimus inference workloads. The AI6 chip tape-out is targeted for December 2026 a custom silicon iteration cycle faster than Apple’s approximately 18-month cadence, which is typically considered the fastest in the industry, and materially faster than the industry wide standard. Tesla is two generations ahead in its own silicon roadmap and the cadence is accelerating.

TeraFab is the biggest product after Optimus without a doubt. A joint venture designed to consolidate chip design, lithography, fabrication, memory, advanced packaging, and testing under one roof.

The long term ambition is 1 terawatt of annual compute output.

The prototype R&D fab at Giga Texas committed at approximately $3 billion of capex will be capable of producing a few thousand wafers per month.

This is a long dated optionality but the strategic logic is straightforward bringing chip fabrication in-house eliminates the last major external supply chain dependency, reduces geopolitical risk, and amplifies the compounding benefits of vertical integration that have defined Tesla’s cost trajectory across every other part of the stack.

Cortex compute. Cortex 1 at more than 100,000 NVIDIA H100 GPU equivalents at Giga Texas was deployed in under a year.

Cortex 2, also in Texas, is ramping with more than 130,000 H100 GPU equivalents.

The compute is not gated by external cloud availability, pricing cycles, or competing customer workloads a constraint facing every competitor training on rented capacity.

Tesla is focused on ensuring compute is never a bottleneck on model improvement cadence, and dedicated infrastructure provides predictable uninterrupted access at a known cost per FLOP that should decline over time as in-house silicon ramps.

Distributed inference. Tesla’s vehicle fleet represents approximately 9 million geographically dispersed compute nodes sitting idle approximately 90 to 95% of the time. If activated for edge inference workloads, low latency AI tasks that benefit from proximity rather than centralized routing, the fleet becomes a decentralized infrastructure layer that scales automatically with vehicle sales, requires no incremental capex, and monetizes silicon the customer has already purchased.

Optimus units deployed at factory and warehouse scale represent an even more reliable distributed inference node given consistent power supply, persistent network connectivity, and predictable operating environments.

The longer term vision includes leveraging idle AI5 compute in parked vehicles and deployed robots as a distributed inference network that scales automatically with fleet size.

FSD licensing and compute licensing. Management has expressed openness to licensing autonomy software to third-party vehicles, including historical discussions with Ford.

At scale, per mile or per vehicle software revenue on third party fleets represents a capital-light revenue stream on infrastructure Tesla has already built.

Cortex capacity could be selectively leased to third parties during periods of lower internal utilization. None of these are near term revenue drivers.

As the infrastructure scales beyond what internal demand alone can utilize, they become natural extensions of the platform.

The market is not pricing them. It will eventually—sooner than we know or even modeled.

Near Term Catalysts, BEV Demand, Sales Momentum, and the Tesla Semi

The near term picture is more constructive than the headlines suggest.

Tesla’s sales momentum is accelerating into May, with registration data pointing to robust trends across European markets;

• UK up 44% YoY

• France up 655%

• Norway up 29%

• Denmark up 136%

• Spain up 113%

• Portugal up 349%

• Sweden up 71%

• and China up 39% YoY.

US sales tracked negative 17% YoY in May per Motor Intelligence, reflecting the ongoing brand impact from Musk’s political involvement, though the company is starting to see strong demand trends emerging from a shrinking competitive EV set as legacy OEMs dial down EV lineups, growing FSD awareness and consumer appetite for L2 and L3 plus solutions, and the surge in gasoline prices globally that is bolstering BEV cost advantages.gasoline price catalyst is quantifiable.

Assuming annual miles driven per driver at approximately 12,000, a 27 miles per gallon fuel efficiency and EV consumption of approximately 0.28 kilowatt hours per mile, and average US residential energy rates of approximately $0.19 per kilowatt hour, BEVs previously presented annual fuel savings of approximately $650 at $2.90 per gallon. With gasoline above $4.50 following the Hormuz closure, BEV fuel related savings have increased to approximately $1,295 approximately two times the pre conflict savings gap. A recent Cars.com study showed approximately 52% of in market shoppers expect to consider a BEV or PHEV in response to the gas price spike.

Tesla Semi volume production began April 29, 2026 at a dedicated 1.7 million square foot factory adjacent to Giga Nevada with 50,000 units per year nameplate at full ramp, nearly four years after the first pilot deliveries to PepsiCo in December 2022.

The truck is offered in Standard Range at 325 miles for approximately $260,000 and Long Range at 500 miles for $290,000 both at roughly half the Class 8 BEV industry average of $435,000 in 2024 and approximately $100,000 below Daimler’s eCascadia and Volvo’s VNR Electric, with 4680 cells produced on the adjacent Nevada line insulating the program from external cell supply disruption.

In California, stackable HVIP, LCFS, and local port and utility incentives cover up to approximately 90% of the purchase price for small fleet operators, Tesla captured 965 of 1,067 HVIP voucher applications between January 2025 and February 2026 versus fewer than 100 combined for Daimler, PACCAR, and Volvo.

Tesla’s January 2026 partnership with Pilot Travel Centers to build Semi-specific Megacharger hubs across California, Georgia, Nevada, New Mexico, and Texas addresses the charging infrastructure gap that has limited out of state adoption.

The Numbers

The near term earnings picture is intentionally under pressure and that is by design.

Tesla is in the middle of its third major R&D and capex cycle, the first covering Model 3 and Autopilot from 2015 to 2017, the second covering 4680, Shanghai, and neural networks from 2018 to 2020, and this one covering Cybercab, Optimus, TeraFab, and Cortex compute.

In every prior cycle the harvest period followed within two to three years and produced returns that justified the investment many times over.

We model overall revenue growing from $95 billion in 2025 to $203 billion by 2030, with Services and Other, Optimus, robotaxi, FSD licensing, and other adjacencies contributing approximately 50% of the total growth.

Automotive revenue grows at a more moderate pace reflecting the base EV business. Energy Storage continues its structural growth trajectory.

Company wide gross margin expands to approximately 28% by 2030 from 18% in 2025 as higher margin revenue streams displace automotive hardware in the mix.

EPS is $1.95 in 2026, growing modestly to $2.15 in 2027 before inflecting in 2028 and beyond to approximately $7.50 by 2030.

Capex is approximately $22 billion in 2026, slightly below guidance of $25 billion given the slow first quarter start, maintained at approximately $20 billion per year through 2030.

Free cash flow inflects positively in 2029 and beyond.

The current valuation, approximately 200 times forward two year earnings, approximately 12 times forward two year EV/sales, approximately 110 times forward two-year EV/EBITDA is lofty on near term numbers.

We are not arguing otherwise.

The argument is that valuing Tesla on near term earnings is the wrong frame for the same reason that valuing Amazon on retail margins in 2010 was the wrong frame.

The business being built is not the business being reported.

The sum of the parts framework values Tesla across five business lines reaching approximately $3.9 trillion by 2035, core automotive at approximately $400 billion at approximately 2.0 times EV/revenue, FSD subscriptions at approximately $300 billion at approximately 12.5 times EV/revenue, Energy Storage at approximately $80 billion at approximately 2.5 times EV/revenue, robotaxi at approximately $1.6 trillion at approximately 5.0 times EV/revenue, Optimus at approximately $700 billion at approximately 15.0 times EV/revenue, and infrastructure licensing at approximately $800 billion at approximately 10.0 times EV/revenue.

Discounted to year end 2026, the framework supports $800.

The Risks Honestly Stated

Execution across multiple frontier programs simultaneously.

Tesla must deliver on Optimus technology hands, manipulation, task reliability, productive uptime while executing the Cybercab production ramp and navigating unsupervised FSD regulatory approval across enough jurisdictions to build a meaningful revenue base.

A failure or material delay in any one area alters the earnings inflection expected later in the decade. No company has manufactured a dexterous humanoid at automotive scale. Actuator and harmonic reducer supply chains essentially do not exist at the targeted volumes.

Cybercab introduces new supply chains and manufacturing processes that historically create bottlenecks the Cybertruck experience is the cautionary template.

Organizational bandwidth is genuinely stretched across multiple key platforms simultaneously.

Elon Musk’s political involvement. Tesla’s brand value declined approximately 36% in 2026 per Brand Finance. US recommendation scores fell from 8.2 in 2023 to 4.0.

European registrations declined for 13 consecutive months. A Yale study estimated 1 to 1.3 million fewer vehicles sold between October 2022 and April 2025 attributable to political participation.

The thesis that product superiority eventually overrides brand damage is reasonable but requires the autonomous product to arrive on a timeline that does not allow the brand damage to compound indefinitely into structurally lost customers who have switched and are unlikely to return.

Key man risk and succession planning. There is no disclosed succession plan.

Tesla without Elon Musk would be a fundamentally different company. His personal involvement in the FSD approach, the Optimus program direction, the broader physical AI strategy, capital allocation toward building compute infrastructure at unprecedented velocity, and the organizational culture that attracts the high agency engineering talent the programs require these are not separable from the thesis.

A succession event before robotaxi and Optimus reach maturity and self sustaining economics would trigger meaningful multiple compression regardless of the successor’s capability.

A transition after unsupervised FSD is commercially deployed and Optimus is in volume production would be far better received.

Valuation on near term metrics. Approximately 200 times forward two year earnings leaves no margin for error on near term execution.

Index diversification risk, capital rotating from Tesla toward nearer term high growth stories could provide a better entry point.

We would view that as an opportunity rather than a reason to avoid the position, but it is a real near term risk to price performance.

Regulatory timeline. The technology can work, the fleet can be ready, and the unit economics can pencil, but unsupervised commercial robotaxi operation in enough jurisdictions to build a revenue base is not within Tesla’s control.

A single high profile incident can set an entire operator back by years.

The regulatory timeline for Optimus in industrial settings is even less defined safety certification frameworks for human robot coexistence in industrial environments are still being written.

Finally, Tesla has done something no other company has attempted, built a vertically integrated physical AI platform that spans the car, the robot, the energy system, and the AI infrastructure that ties all of it together.

The flywheel runs from lithium in the ground to autonomous miles on the road to humanoid labor in the factory to grid-scale energy storage at the data center with each business feeding cost, data, and capability back into the others.

The near term earnings are under pressure from the investments. The near term brand is under pressure from the politics.

The near term regulatory path is under pressure from the complexity of approving autonomous vehicles at scale in jurisdictions not designed for them.

None of that changes the destination. It only affects the timeline.

The sum of the parts reaches $3.9 trillion by 2035. The business that gets it there is already being built.

The market is pricing the car company and giving away the rest…which is perfect for people who are paying attention.

Disclosure: We initiated a long position in Tesla. This report is for informational purposes only and does not constitute investment advice. WTS and Nick’s research is produced independently. Past performance is not a guide to future returns.

Today, that changes hopefully forever…

Rising prospects of a midyear deal to reopen the Strait have reduced the tail of an oil price spike. August oil futures edged down to $87 per barrel, the lowest level since late April.

Retail gasoline prices have continued to drop, now around $4.10 per gallon, and there is a real chance energy makes a negative contribution to June CPI.

The latest developments reduce the tail risk of an additional negative price shock while increasing the likelihood that prices settle at a lower level, a development that could materially lower second half 2026 consumer price inflation.

This is not a story about the energy shock disappearing. Our baseline still has crude prices staying elevated around $100 per barrel through year end on a large inventory rebuild and a lingering geopolitical risk premium.

Our commodity team highlighted a slower global inventory drawdowns than expected, helped in part by an estimated rise in Persian Gulf exports to 9 million barrels per day in June.

If the Strait effectively stays closed, global oil inventories still hit operational stress levels by late June or early July pushed back from early June only on signs of weaker global petroleum demand.

What changed is the distribution of risk.

The downside tail is shrinking. And underneath it, the upside risks are building in a way the market has not yet priced.

The Business Sector Lift Is Broadening to Jobs

The most important macro development of 2026 is not the energy shock. It is the cyclical lift in business spending and hiring as last year’s sentiment shock fades and the evidence that it is now broadening from the goods sector to the labor market.

The latest Asia export data shows a tech spending boom sustained as we move toward midyear, alongside a nontech acceleration that looks set to be supported by a nascent turn in the inventory cycle. The global industry upturn has room to run, maintaining the rapid rate of expansion recorded so far this year.

The lift is now appearing in labor markets, where US and emerging market employment gains are firming.

This is the part of the story that matters most for the durability of the expansion because a business sector that is hiring provides the bridge that carries household purchasing power through the energy-driven inflation squeeze.

Our view breaks from the consensus on a crucial point. The widely held narrative links the expansion’s health to an AI-related surge in productivity. We do not.

We view the expansion’s health as far more closely tied to a recovery in business spending that strengthens employment supporting a more balanced and sustainable expansion than a productivity led story would imply.

The distinction is not academic at all.

A jobs led expansion provides the income growth that sustains consumer spending without a further drawdown in savings. A productivity led expansion does not. The data this year is pointing at the former.

Reinforcing the dynamic is a favorable financial conditions backdrop that has largely ignored the elevated tail risks, and a continued incremental shift toward easier fiscal stances.

Surging tech revenues are pushing up our estimates for EMAX second quarter GDP growth which in turn creates increased space for fiscal support across emerging Asia.

The risk rotation is straightforward from here, a reduction of the downside tails alongside increasing upside risks to our second half forecast of 2.1% annualized global GDP growth.

The immediate focus is whether a deal to reopen the Strait sustains the recent drop in oil prices, and how much of a positive impact that delivers to business sentiment particularly in Western Europe, where the conflict appears to have done the most damage.

A consumer that is bending but not breaking,US real consumption is tracking close to a 2% annualized increase in the second quarter despite a sharp rise in inflation is a positive message.

Next week’s May retail sales reports for the US and China will likely signal some softening as we move through midyear.

The question is whether that softening is the start of something or the noise inside a consumer that is absorbing the energy squeeze and continuing to spend.

The FED meet this week for the first time since Kevin took over

The inflation and financial stability threat from the energy shock has been a primary driver of the limited global central bank tightening so far.

As energy supply risks abate and oil settles into an $80 to $100 per barrel range, that pressure moderates.

But incoming news supports our view that core inflation was never going to follow an “immaculate” path toward target.

Service price inflation remains sticky at an elevated level. Goods price pressures are building from higher transportation costs, firming tech sector supply constraints, and firm demand. A broad but shallow central bank tightening cycle remains likely even as the relaxation of energy risk removes one source of near term pressure.

This week’s FOMC meeting is Jerome Powell’s last as a bystander and Kevin Warsh’s first as chair. We expect the committee to remove its easing bias and vote 12-0 to keep rates stable.

The statement and dot plot are likely to signal a consensus for stable policy rates through the end of the year.

This hawkish tilt should carry into next year, where some members will likely project hikes but the median dot still anticipates a 2027 ease.

We raise our second quarter GDP forecast to 3.0% as oil prices slip further. Next week’s retail sales should show a headline rise of 0.8% with control up 0.5%.

The wild card is whether the upside growth risks are realized and interact with weak labor supply to generate earlier and more substantial pressure on the Fed. That will not be resolved next week. But it is the question that defines the path into 2027.

The press conference will be the main event for the narrative.

Warsh has emphasized less Fed guidance, but the press conference is the chair’s best tool for controlling the policy narrative.

We see him likely to accurately reflect a committee discussion with no policy bias while delivering a more personalized perspective on the medium term forces he believes favor lower inflation.

He probably will not yet provide details on the “regime change” he has promised for Fed operations.

How he handles questions on the many areas where he has expressed a desire for change is the genuine uncertainty.

On core inflation, there is good news that may not last. Last year core goods prices rose 1.4% December over December. Over the past six months they have risen just 0.4% annualized, including a 0.1% monthly fall in May. Prices fell for home furnishings and recreation goods, apparel inflation is decelerating after an early year burst, and drug prices have been an important reason for the slowdown, Medicare negotiations authorized by the Inflation Reduction Act took effect at the start of 2026, GLP-1 prices are lower, and more generics are coming to market. The disinflation in goods is real. Whether it survives the building pressure from transportation costs and tech supply constraints is the open question.

Europe

The ECB delivered a 25 basis point hike this past week a move that likely heralds a broad move higher across Western Europe. The guidance emphasized data dependence, but the staff forecast emphasizing the need for a step adjustment was not challenged.

The baseline staff projection has three hikes, and even the “mild” scenario, which assumes a faster alignment of energy prices to pre-war levels — is consistent with further action.

We stick with our call of one more hike in September, with risks skewed toward a move in July. A more rapid reopening of the Strait could reduce inflationary pressure, but the direction of travel for the ECB is unambiguous.

When even the dovish scenario in your own staff forecast justifies more hikes, the bar to stop is high.

The pattern extends across the region. The Bank of England is expected to hold next week but shift incrementally more hawkish.

Weaker labor market signals and downside inflation surprises argue for patience, but inflation is still expected to rise above 3.5% later this year and surveys point to second round risks.

We look for additional dissents 7-2 with risks of 6-3 consistent with keeping a July hike on the table.

The Norges Bank should sound more confident that further tightening is needed and revise its rate path higher to show another hike in the third quarter.

Core inflation and wage growth are running above forecasts.

We see only a small chance of a hike next week, but the path is up. In Sweden, the Riksbank is likely to stay on hold, low headline core inflation is flattered by a VAT cut, but survey measures of pricing intentions have reaccelerated amid higher imported inflation.

We expect a modest upward revision to the rate path and a hike in September, by which time the ECB is projected to have hiked twice.

Western Europe is the region where the energy shock did the most damage to sentiment and where the inflation it generated is proving most stubborn.

Every major central bank in the region is moving in the same direction. The only debate is pace.

The BoJ

The Bank of Japan is likely to hike 25 basis points to 1.0% next week. The core issue is no longer whether the BoJ can hike, it is whether it can afford to keep waiting.

A delayed move risks intensifying upward pressure on long term yields and renewed yen weakness, both of which would weigh on the real economy.

Against that backdrop, we do not expect emphatic opposition either within the Policy Board or from the government.

There is an unusual wrinkle. Governor Ueda has been hospitalized for medical treatment and will be absent. Deputy Governor Himino serves as acting chair, and Deputy Governor Uchida holds the press conference on the Governor’s behalf.

The hospitalization is expected to be short, with Ueda scheduled to attend the July meeting. The decision itself should be unaffected, the direction toward a hike is broadly shared, Ueda will provide his views in writing, and if the vote tied, Himino would cast the deciding vote in line with Ueda’s intentions.

The press conference is the uncertainty. Uchida is viewed as one of the more dovish members of the board, but as a temporary stand-in he has little incentive to press personal preferences.

Any difference is likely more about style than substance, where Ueda, an academic, offers theoretical explanations, Uchida may respond from a more practical perspective.

If that comes through, the press conference could offer an unusually clear window into the BoJ’s internal deliberations.

The real market event is not the rate decision. It is the forward guidance on the hiking path and the conclusions of the JGB purchase review.

The BoJ’s core inflation measure, excluding fresh food and institutional factors, has risen to nearly 3%, and the market increasingly believes underlying inflation has already reached 2%.

Even after the hike, the policy rate at 1.0% would only reach the lower bound of the BoJ’s estimated neutral range, a range the Bank itself acknowledges may be biased downward by the legacy of past accommodation.

Here is the trap though…with additional hikes already priced, even a hike can be read as dovish if the balance sheet normalization message softens particularly if the BoJ halts reductions in JGB purchases after April 2027, as recent media reports suggest it may be considering.

We expect the existing reduction plan through March 2027 to be maintained, with the market’s focus shifting to the purchase plan from April 2027 onward.

The bar to be perceived as hawkish is high.

A genuine hawkish surprise would require signaling either a faster pace of tightening or a path above neutral beyond 2% which still looks like a marginal risk case rather than the base case.

The BoJ will try to sound hawkish. Markets have already priced the hawkishness.

That asymmetry is the whole meeting.

China

The outlook for China to expand at a trend like 4% annualized this quarter is underwater. April activity disappointed broadly.

This week’s May export data beat expectations, but the gain was narrowly based in tech and owed in large part to surging prices in memory chips and energy products rather than broad volume strength.

China’s May exports rose 3.1% MoM after an April jump, beating expectations by a wide margin and lifting annual growth to 19.4% YoY.

High tech shipments rose for the seventh consecutive month at 8.0% MoM, led by mobile phones, electronic ICs, and automatic data processing equipment with prices likely providing a meaningful boost.

The composition is the caveat.

The trade strength is narrower and increasingly price driven, concentrated in memory ICs and modules plus the three new products, EVs, solar, and batteries, which together accounted for around 60% of export growth.

That makes the read through to industrial production and the net export contribution far less clear cut than the headline suggests. Imports expanded for a sixth straight month, rising 2.9% MoM, but the strength was driven more by AI supply chain needs and commodity stockpiling than by a broad domestic demand recovery.

Crude oil imports continued to decline, partly reflecting last year’s reserve build and a preference to draw on inventories.

This week’s May activity data will challenge the quarterly outlook. We expect industrial production to rise 0.5% MoM, only partially reversing April’s decline. Retail sales should remain weak. YTD fixed asset investment likely stayed in contraction.

The silver lining is policy reflexivity. To the extent the current quarter disappointment owes to slower implementation of fiscal policies, there is upside for the second half.

A sizable enough disappointment could spark new fiscal stimulus on top of current plans. The second half export outlook should be supported by resilient global demand and a broadening of capex beyond tech, partly offset by renewed US tariff uncertainty as the administration pivots toward Sections 301 and 232 and the EU takes a tougher stance toward China.

China’s AI Edge

The most important structural story in China right now is its AI strategy and it is widely misunderstood.

China’s edge does not lie in winning the frontier model race.

Restricted access to advanced chips makes that race one it cannot currently win. China’s edge lies in scale and deployment.

The DeepSeek moment in early 2025 turned China’s AI story from policy aspiration into a market moving macro theme.

The R1 model challenged the view that US chip restrictions would cap China’s progress, reset assumptions on AI cost curves, and briefly shook the narrative around US hyperscaler dominance.

Since then, the performance gap between leading US and Chinese models has narrowed, and China’s strategy has looked less like a frontier race and more like an economy wide deployment agenda built on open source models, cost efficiency, and integration into manufacturing and services.

Following the 2025 AI+ action plan, the 15th Five Year Plan frames AI as a broad-based macro industrial strategy with the ambition to embed AI across 90% of the economy by 2030.

The “Six Networks” agenda proposed at the April Politburo meeting broadens the definition of infrastructure beyond roads and railways into integrated resources, energy, data, and connectivity.

The three most AI relevant pillars are;

1. Compute

2. Connect

3. Power

Together forming an integrated platform designed to lower unit compute costs, improve utilization, and ease the energy and connectivity bottlenecks that would otherwise constrain adoption.

Over time, AI competitiveness may depend as much on electricity generation capacity and cost, green power access, batteries, and grid deployment speed as on chips or models.

The power network will determine whether China’s compute buildout can translate into usable AI capacity.

While the US still leads in data center and hyperscaler infrastructure, China’s relative advantage lies in coordinating power supply, land availability, data center siting, and industrial policy.

The key bottleneck is grid delivery whether western power can be transmitted reliably to meet eastern demand.

China’s buildout is best viewed as a state guided macro industrial program, not a private tech capex cycle.

Where US AI capex is concentrated among hyperscalers and VC backed firms, China’s investment is driven by national and local agendas financed through government guidance funds and infrastructure programs.

Comparisons based only on private AI spending understate China’s true investment scale.

Government funds are estimated to have deployed roughly $912 billion across industries between 2000 and 2023, including about $184 billion toward AI companies.

On June 9, Bloomberg reported a potential 2 trillion yuan 1.4% of GDP five year state-led AI datacenter and computinghub buildout, with total investment possibly rising to around 5 trillion yuan, or 3.4% of GDP, if power grid integration is included.

The scale is directionally consistent with the AI+ push and the Six Networks agenda.

The key unknown is what the figure represents, state funding versus total investment mobilized, and incremental fiscal expansion versus budget reallocation.

If financed via special central government bonds, the macro outcome is likely K-shaped across AI beneficiaries versus other sectors.

The constraint is the chip layer.

US export controls have tightened China’s access to the most advanced GPUs and high end CPUs, especially since 2022.

This matters most for frontier model training.

Domestic substitution is advancing but the advanced chip ecosystem still lags.

Huawei anchors the AI accelerator push while SMIC remains the key advanced foundry but lags in process technology.

In memory, CXMT and YMTC are domestic champions in DRAM and 3D NAND, though gaps remain in HBM and process maturity.

Huawei’s proposed “Tau scaling law” shifts focus from transistor shrinkage to reducing system-wide latency, with ambitious claims including “1.4nm-equivalent” density by 2031 but the concept remains unproven without independent benchmarks or scaled production.

As Jensen Huang noted, much of the approach echoes existing trends in 3D integration and advanced packaging, while major hurdles remain in heat, power density, tooling, yield, and cost at scale.

The macro implication is that China’s strategy is less about the frontier and more about scaled deployment via cost efficient models, open source diffusion, and industrial adoption.

Chip restrictions slow frontier training but push China toward higher utilization, domestic substitution, and deployment led AI where its industrial base, policy coordination, and infrastructure buildout deliver the macro lift.

That edge could widen further if humanoid and industrial robotics scale off China’s manufacturing depth.

The export story is where this becomes visible.

As the global AI upcycle broadens from frontier models to datacenter buildout, China’s opportunity lies in the adjacent infrastructure layers servers, memory modules, optical interconnects, networking hardware, power equipment.

Memory chip and module exports appear largely price driven, reflecting global tightness, with IC export values in May roughly 3x the 2023 monthly average.

China’s more durable edge is in legacy chips and data center interconnect.

Optics is the main picks and shovels channel.

AI data centers require dense, high speed links and reportedly use 5 to 10 times or more fiber than traditional server rooms.

Some overseas order books now extend into 2027 and 2028, and China Customs data show fiber and cable export value rising to around 4 times the year ago level with average export prices up around 3 times.

Copackaged optics has become a bellwether for China’s onshore AI hardware ecosystem.

AI server upgrades are boosting demand for high end PCBs, while copper clad laminate is emerging as a critical upstream bottleneck and pricing lever. HGTECH optical interconnect exports rose 122% YoY in the first quarter, consistent with accelerating external demand.

The energy side adds another channel.

Batteries, solar products, and grid equipment are increasingly part of the AI infrastructure equation as datacenters face power, reliability, and decarbonization constraints.

But China remains a net importer in the frontier compute layers leading edge GPUs, high end AI memory, and other advanced semis.

The current cycle echoes China’s role in the Internet revolution.

China was not the original leader but became a pivotal force in accelerating global adoption by manufacturing at unprecedented scale, lowering the cost of the hardware backbone telecom equipment, servers, computers, consumer devices.

It then extended that cost and scale advantage into applications, ecommerce, digital payments, gaming, logistics.

The same playbook low cost infrastructure, mass adoption, high usage applications is now being repeated in the AI wave.

The macro impact so far is K-shaped.

The upside is most visible in AI infrastructure, where IT, industrial, and hardware firms are leading the first quarter Ashare capex cycle, and high tech production categories are outperforming.

The lower leg of the K is the limited near term spillover to the wider economy.

AI capex is concentrated in capital intensive, tech adjacent sectors, while traditional manufacturing, consumption, and services may benefit only gradually as adoption diffuses.

There is a labor market wrinkle worth flagging, AI raises underemployment risk for entry level workers while increasing demand for high skill workers a divergence that will shape China’s consumption story over the medium term.

Taiwan

Taiwan is the cleanest read on the tech cycle’s strength. May exports were strong across the spectrum, and we are upgrading our second quarter GDP growth forecast.

Next week’s Taiwan CBC meeting should reinforce a “hiking from strength” message amid firm CPI and very strong PPI the same dynamic pushing Korea and Singapore toward the same discussion.

CPI in China held steady while PPI extended its upturn, with AI related memory tightness feeding into communications tools CPI and broader IC and electronics PPI.

The tech cycle is now showing up in the inflation data across the region, not just the export data, a sign of how broad the AI driven supply tightness has become.

EM

Emerging market central banks are tilting more hawkish, but the driver matters.

There is an important distinction between financial stability pressures related to the energy shock and macroeconomic pressures related to strong growth and firming core inflation.

The clearest divide is in Asia.

Financial market vulnerability has driven recent hikes in Indonesia and the Philippines after this week’s off cycle 25 basis point BI hike, we still look for another next week, and we forecast a 50 basis point BSP hike with risks tilted toward a smaller move as energy pressure relaxes.

Tech led upside growth surprises are pushing Korea, Taiwan, and Singapore toward hiking from strength rather than from fear.

Elsewhere the picture is mixed.

In Turkey, the CBRT’s hold and dovish tone were in line with our base case of a reversion toward 37% effective funding as Middle East tensions recede.

Hungary remains an outlier, a stronger forint should enable a June cut.

Czech inflation and credit growth look too hot for the CNB to ignore, and we expect a 25 basis point hike next week.

In Latin America, a positive demand backdrop and stickier underlying inflation are shifting stances away from easing.

Mexico’s Banxico has entered an extended pause despite material slack.

Brazil’s COPOM faces higher imported price pressure and an increasingly expansionary domestic fiscal and credit mix, we expect another 25 basis point cut next week, with cautious guidance to bolster credibility.

USMCA

Canada and Mexico recently stated their intentions to renew the USMCA.

This week President Trump said he was not looking to renew the treaty.

An outright breakdown is unlikely, but the US position aligns with our view that negotiations drag on, with a resolution more favorable to US interests not reached until later in 2027.

The risk that negotiations fail and result in annual revisions is elevated, though not our baseline. Any agreement would likely not remove tariffs Mexico and Canada are focused on lowering them to levels that benefit North America relative to other US trading partners, in contrast to current rates such as those on autos.

The trade story is the counterweight to the risk rotation.

As the energy tail falls, the tariff tail rises.

The pivot toward Sections 301 and 232, the tougher EU stance toward China, and the USMCA uncertainty all point to trade policy becoming the dominant headline risk in the second half exactly as the energy shock recedes.

The defining macro development this week is a rotation in the distribution of risk.

The energy sector’s negative tail is falling as a deal to reopen the Strait of Hormuz comes back into view and oil drops to its lowest level since April.

At the same time, the positive business sector lift is broadening to labor markets, raising the upside risks to our second-half forecast of 2.1% global growth.

The market has not fully repriced either move.

The downside tail that has hung over forecasts for three months is shrinking. The upside risk from a jobs led expansion interacting with weak labor supply is building.

And the central bank response, a broad but shallow tightening cycle spanning the Fed’s removal of its easing bias, the ECB’s move higher, the BoJ’s hike to 1.0%, and a hawkish tilt across emerging markets is the policy reflection of an expansion that is proving more durable than the energy shock implied.

Two things to watch from here.

1. Where oil prices settle once a Strait deal is reached because that determines how much second half inflation relief actually materializes.

2. Western European sentiment because that is where the conflict did the most damage and where the recovery, if it comes, delivers the biggest upside surprise.

The expansion is not being derailed by the energy shock. It is being tempered by it.

And as that shock fades, what is left underneath is a business cycle with more momentum than the consensus has priced…

World Trade Securities, Global Macro Intelligence · 13 June 2026

Before I go any further, look at this latest earnings season and how saas stocks reacted, nearly all beat the top and bottom lines, and ask yourself one simple question, why are software companies beating earnings and selling off anyway?

ServiceNow beat. Sold off. Workday guided fine. Sold off. Adobe posted record revenue. Sold off. The old playbook says this is irrational, buy the dip, the fundamentals are intact, the market is overreacting.

The old playbook is wrong. The market is not reacting to the quarter. The market is pricing the terminal state. It has stopped paying for current cash flows because it no longer believes those cash flows survive the decade. That is what a structural repricing looks like in real time, price leading fundamentals by two years, the way it always does at the end of an era. The market knew the newspapers were dead in 2005 while the earnings were still fine. It knew Blockbuster was dead while the stores were still full. It knows now…

The age of superintelligence did not arrive gradually. It arrived in January. And every model release since has shortened the lives of all saas companies imo. When Anthropic shipped its Mythos model this past month, the capability jump was not incremental, entire categories of enterprise workflow that were “safe” twelve months ago became automatable overnight. That is the pattern that should terrify every software investor, the disruption curve is not linear, it is compounding, and every generation of frontier model permanently deletes another layer of the SaaS value proposition. The companies being disrupted have to out execute an opponent that doubles in capability every year and charges $20 a month.

They will not and I am certain of it…

Why this ends in bankruptcy and theres no recovery

The bull case for SaaS rested on three pillars,

1) recurring revenue

2) switching costs

3) and expanding TAMs

All three are not under pressure anymore. They are broken beyond repair.

Recurring revenue assumed customers keep paying. The renewal conversation has already changed from “should we switch vendors” to “do we need this vendor at all.” Every CFO running a Cowork class agent deployment is building the spreadsheet that compares $20 per month per agent against $150 per seat per month plus implementation, plus support, plus the army of admins. The renewals don’t collapse this quarter. They collapse at contract expiry, which means the revenue cliff arrives on a lag, concentrated in 2027 and 2028 as the multi year agreements signed in 2024 and 2025 roll off into a world where the alternative is two orders of magnitude cheaper.

Switching costs assumed migration is painful. AI deleted the pain. The integration work that justified seven figure consulting engagements is now an agent task. The data migration that locked customers in for a decade is now a weekend. The moat was never the software, it was the human cost of replacing the software, and the human cost just went to zero.

Expanding TAMs assumed the seats keep growing. The seats are the product being eliminated. Software priced per human, sold into companies using AI to need fewer humans, is a melting ice cube with a subscription attached.

Here is the part the saas bulls crowd refuses to model especially for my friends in investment banking in New York, who are still used to the old world of finance models and think we’re in a bubbl, which is the opposite in the West Coast. But these dynamics are self reinforcing on the way down. Revenue decelerates, so the multiple compresses. The multiple compresses, so stock based compensation stops retaining talent, so the real cash cost of the workforce rises exactly as revenue falls. Management defends the stock with buybacks, funded with debt as FCF erodes. Leverage rises into a declining revenue base, the one combination credit markets never forgive. Ratings fall, spreads widen, refinancing costs spike, and the companies that spent two decades being told they were annuities discover they are leveraged decliners. That is the death spiral, and it does not require a recession, a crisis, or a single bad quarter to start. It only requires renewals to keep doing what they have already started doing.

The timeline in my opinion is closer than Ive even modeled, but 2026 is the repricing. 2027 is the renewal cliff. 2028 is when the death spiral becomes undeniable, negative FCF at the weak names, going concern language in the audits, the first delistings among the subscale point solutions, the first restructurings among the leveraged laggards. The strongest balance sheets buy time, cash piles delay the reckoning, they do not prevent it. The destination is the same. Equity holders in the losers are holding claims on terminal decline cash flows financed by deteriorating balance sheets. Those claims converge to zero. The market started pricing that convergence in February. It will not stop and will only get worse…

How We Got Here

The catalyst was specific. Late January and early February disclosures from OpenAI and Anthropic demonstrated agentic systems executing multi step enterprise workflows not copilots, not demos, shipping products. Anthropic’s Claude Cowork put a $20-permonth agent in front of the exact workflows that ServiceNow, Workday, Palantir, and Salesforce charge tens of thousands of dollars annually to perform. The market watched it happen. So we’re still a few months into this, so we’re really really early to the deal outcome, so if you’re reading this, I beg you to rethink your saas investments…doesn’t mean you should sell, but understand where we’re going.

The disclosures landed while software management teams were admitting higher AI costs, uncertain pricing models, and longer payback periods with zero offsetting evidence of revenue protection. Equity repriced first. Credit followed. High Yield and Leveraged Loans set the tone, but this was not contagion, it was recognition moving up the capital structure.

And then the Mythos release confirmed the slope of the curve. Every quarter, the frontier models absorb another tier of work. Every quarter, the line between “what AI can do” and “what enterprise software charges for” moves in one direction. The incumbents are not competing against a product. They are competing against a capability curve and the curve is undefeated.

The sentiment is correct. The selloff is rational. The market is pricing the future before the present catches up, exactly as it should.

My team of geniuses and I were looking into actually buying into saas because we thought it was cheap, before we understood the real ai thesis, so this is coming from a place for a saas lover and a user of most of these products, wether on the enterprise side or personal side. Now we walk through the coverage universe and explain exactly why each name is going to go bust or, in the rare cases where data monopolies and regulatory entrenchment create genuine defensibility, why it survives the apocalypse

1. Accenture ($ACN) → Neutral. The labor arbitrage model that built this company is being eaten from the inside. When the work product of a thousand consultants is approximated by an agent running overnight, the billable hour dies. The balance sheet buys time. It does not buy a new business model.

2. Cisco ($CSCO) → Underweight. AI buildouts support high end switching demand, but the spending is concentrated among hyperscalers with brutal pricing leverage, and management is already flagging memory cost inflation crushing COGS from the other side. Spreads near Amazon and Alphabet levels are indefensible. Squeezed from both directions with no exit.

3. Expedia ($EXPE) → Underweight. When an AI agent plans the trip, compares the flights, and books everything, nobody opens Expedia again. The top of the funnel is being deleted, and the paid acquisition cost structure leaves no margin to fight back. Terminal.

4. FactSet ($FDS) → Underweight. A terminal business built on the scarcity of organized financial information, in a world where the scarcity is over. When a model ingests every filing, transcript, and report in real time and synthesizes it in seconds, the terminal subscription becomes unjustifiable at every renewal and that includes Bloomberg. The entire category is in runoff. Most of its participants do not know it yet.

5. Gartner ($IT) → Underweight. The product is aggregation and synthesis of industry information, the single thing AI does best. A Magic Quadrant equivalent now takes thirty seconds and costs nothing. The brand survives as nostalgia. The subscription does not survive the renewal cycle. This is one of the names we expect to be unrecognizable by 2028 and not in the good way.

6. HPE ($HPE) → Neutral. AI server upside offset by execution risk, competitive pricing, customer concentration, and weak cash conversion. There is a version of this company that wins from AI infrastructure. Management is not building it fast enough.

7. IBM ($IBM) → Underweight. Two decades of buying growth it could not build, now facing a disruption it cannot buy its way out of. A large debt financed acquisition to mask the decline, the move this management team has telegraphed its entire tenure costs the single A rating and accelerates the credit story. The slow decline is the base case. The fast one is one board meeting away.

8. Kyndryl ($KD) → Underweight. Spun out of IBM to die quietly. Management turnover, an SEC probe, collapsing confidence, and legacy infrastructure revenue that AI makes worthless faster every quarter. This is the first name on this list we expect to face genuine going concern questions. The 2028 delisting cohort starts here.

9. PayPal ($PYPL) → Underweight. A standalone intermediary in the era when intermediaries are being disintermediated. Agentic commerce embeds payments directly into the platforms and the models, the button is not a moat, the brand is not a moat, and the fraud business is losing to players with better data. Structural share loss with no support line in sight.

10. ServiceNow ($NOW) → Underweight. The best executing company in enterprise software for a decade, and it does not matter. The product is workflow orchestration, routing tickets, managing approvals, coordinating tasks. That is the first category agentic AI deletes, because an agent does not need a ticket routed to it. It executes the task. ServiceNow’s revenue is a direct function of enterprise workflow complexity, and AI compresses that complexity permanently. Priced for perfection, structurally exposed at the core, and the renewal cliff hits its install base hardest in 2027–2028. The most dangerous stock in software because it is the one everyone agrees is safe.

11. Workday ($WDAY) → Underweight. Software priced per employee, sold to companies using AI to need fewer employees. There is no version of the agentic future in which Workday’s pricing unit, the human seat is not in structural decline. Longer sales cycles and a soft FY27 outlook are not noise. They are the first data points of the terminal thesis showing up in the numbers. It will not be the last.

12. Palantir ($PLTR) → Underweight. The most dangerous valuation in all of software, and we are not hedging this one. The government business is real, long contracts, clearances, genuine barriers. That is worth something. It is not worth the richest multiple in the entire software universe, which is what you are paying, but even then, when we went into Venezuela and captured its leader, the model that supposedly “turned off” the Venezuelan army’s defensive capability was anthropic at a much much cheaper price. Same story in Ukraine and Iran. The pentagon is using anthropic models to executes  its mission and they have reportedly said they cannot work without anthropic models, so it’s really only a matter of time before its govt business deteriorates. The commercial business, the entire growth story is a labor arbitrage model dressed as a software platform. Forward deployed engineers integrating enterprise data and building operational applications that is precisely the task frontier models are absorbing fastest. Palantir charged millions because enterprise data integration required scarce human expertise. The Mythos models do the integration work as an agent task at $20 a month. The customers who bought Palantir because they could not build it themselves are watching the build it yourself cost collapse to zero in real time. Palantir is priced as the AI winner of enterprise software while its commercial moat is the single most exposed business model in our coverage. When the commercial growth decelerates and the renewal math says it does, the multiple does not compress gracefully. It collapses. From the richest multiple in software, the path down is the longest. The government floor catches the company. Nothing catches the equity at this price. By 2028 we expect this to be remembered as the defining valuation casualty of the agentic era.

11. Adobe ($ADBE) → Underweight. The most painful name on the list because the franchise is genuinely iconic and the disruption is genuinely existential. For thirty years the moat was that creative professionals had no alternative, Photoshop was the tool, the file formats were the standard, Creative Cloud was a tax on participating in the creative economy. That moat assumed content creation required skilled humans operating complex software. Generative AI broke that assumption at the root. The frontier models are not making Photoshop better. They are making Photoshop unnecessary, one use case per model release. The marketing team that needed a licensed designer now types a prompt. Firefly is competent and strategically defensive, Adobe using AI to protect a pricing structure that AI is destroying. You cannot charge $60 per seat per month for capabilities the model providers ship as features. The Figma failure removed the one repositioning move that mattered. Watch net revenue retention, it is the canary, and it is already coughing. The decline is gradual until the renewal base catches up to the capability curve. Then it is not gradual. Even for us, we used adobe’s products for the past 5 years or so across our companies and we recently canceled our subscription because it became too expensive compared to ai models that actually do a better and faster job than adobe’s and I suspect more companies will follow suite. Not to mention 40% of their recurring revenues are form overseas, when these so called countries learn about ai in the next year or so, maybe 80-90% of them will switch to better and cheaper models.

12. Salesforce ($CRM) → Underweight. The largest pure play SaaS company on earth is the largest pure play exposure to the terminal thesis. The pricing model is seats. The product is workflow software for humans. The growth algorithm for twenty years was more seats, more clouds, more price and every component is now in structural decline simultaneously. Agents shrink the seat count in sales and service organizations. Agentic tools replicate CRM workflows at the task level. And Salesforce is the largest line item on every software bill in the middle of the greatest vendor consolidation cycle in enterprise history. Agentforce is the admission, management is racing to cannibalize its own pricing model before the model providers do it for them. The transition math is fatal, every dollar of consumption based agent revenue replaces a multiple of committed seat revenue, and the competition for the agent layer is the model providers themselves, who own the capability and have no legacy revenue to protect. The cash pile delays the reckoning. It does not change the destination. Ive suggested to Marc instead of buying his stock, he should’ve invested that $27B into anthropic or openai inserted, so as his business get killed he wins on the side of the killer. They own a big stake of anthropic so I suppose that’ll help, and thorough their sales force venture in Japan, they’ve acquired a few startups in Asia that will turn out to be winners in the region.

    

    

The next ai winners that are still undervalued...that might be considered instead of saas or semis…

1. Dell ($DELL) → Overweight. When SaaS dies, someone still has to sell the servers. AI infrastructure demand, improving cash generation, falling leverage. The simplest thesis in the note.

2. Equifax ($EFX) → Overweight. You cannot train an AI to replace data you do not have. Fifty years of longitudinal consumer data, embedded in regulated decisioning workflows. The models need it. Full stop.

3. Moody’s ($MCO) → Overweight. The regulators will not let AI replace the rating agencies. Regulatory entrenchment, proprietary data, embedded compliance workflows. Structurally insulated.

4. MSCI ($MSCI) → Overweight. An index is not a feature an agent replicates. It is a coordination standard that trillions in mandates are contractually bound to. The widening is overdone. The mandate layer is the moat.

5. S&P Global ($SPGI) → Overweight. Regulation, market structure, and standardized data, the three places substitution risk goes to die. The AI era increases the value of what this company owns.

What to watch for in the next few Q’s

The terminal thesis confirms itself through five signals, in order.

1. Net revenue retention rolling below 100% at the seat based names, the renewals telling the truth before the income statement does.

2. FCF margins eroding as AI opex rises into decelerating revenue.

3. Debt funded buybacks defending stock prices, the move that converts an equity problem into a credit problem.

4. Desperation M&A, leveraged acquisitions to buy the growth organic execution cannot deliver.

5. And going-concern language at the weakest names as 2028 approaches.

Track them in that order. The first two are already visible. The third has started. The last two arrive on schedule…

Lastly, the market is selling these companies off regardless of earnings because the market knows what the future looks like. It is pricing the terminal state of an entire business model, software priced per human, in a world that needs fewer humans per task every single quarter, against a capability curve that compounds and an opponent that charges $20 a month.

2026 is the repricing. 2027 is the renewal cliff. 2028 is when the death spiral becomes undeniable, the first delistings, the first restructurings, going concern questions at the weak names, and terminal decline everywhere else that no balance sheet is large enough to reverse. The cash rich giants die last. They still die.

We are not hedging this. The conviction is total. Prefer Hardware over Software. Prefer data moats over feature moats. Prefer regulatory entrenchment over switching cost narratives. And above all, prefer the companies that control the data AI needs over the companies whose products AI replaces because the second group is going to zero, and the only open question is the order of the funerals.

The age of Software exceptionalism is over. Welcome to the age of AI exceptionalism. Choose your survivors carefully. There will not be many.

This report is for informational purposes only and does not constitute investment advice. WTS’s and Nick’s research is produced independently. Past performance is not a guide to future returns.

The stock fell anyway.

It fell because alongside that backlog, Oracle announced it would raise around $40 billion in debt and equity and spend $90-95 billion in gross capex next year to build against it. The market took one look at the cash going out the door and reached for the only word it knows in a moment like this, unsustainable. Free cash flow compresses. The valuation backstop weakens. Sell first, model later.

I think that reaction is the opportunity. Not because the capex isn’t real, it is enormous and it is the central risk in this entire thesis, and I will give it the full weight it deserves before I’m done. But because the market is making a category error. It is still grading Oracle as a mature saas company that has wandered into a capital intensive business it doesn’t belong in. The truth is the opposite. Oracle is now an AI infrastructure company with a software annuity bolted to its side, and infrastructure companies in the middle of a demand surge are not valued on this year’s free cash flow. They are valued on the contracted backlog and the credibility of conversion.

That reframing is worth, in my view, a path to $500 over the next eighteen to twenty four months.

What the market is misreading

The bear case on Oracle rests on a single chart, capex up, free cash flow down. It is a clean, intuitive, and incomplete story.

What it misses is who is paying for the buildout. In the fourth quarter Oracle signed $67 billion of AI infrastructure contracts, and the majority of them were what the company calls bring-your-own-hardware or prepaid structures deals where the customer either prepaid Oracle for the GPUs or bought the GPUs themselves and handed them over. Cumulative prepaid and BYOH contracts now sit at $75 billion. Of next year’s $90-95 billion gross capex, $20-25 billion is funded directly by customer prepayments, leaving net cash outlay closer to $70 billion.

Read that structure carefully, because it changes the risk profile entirely. When a customer prepays you to buy GPUs, two things happen at once. Your capex is derisked, you are not speculating on capacity, you are fulfilling a funded order. And the customer has posted collateral on its own demand. A company does not wire Oracle billions of dollars in advance for compute it is unsure it needs. The prepayment is the demand signal, and it is a far harder signal than any backlog footnote, because it is cash that has already moved.

The new CFO, Hilary Maxson, gave the Street the second half of the picture, steady state, project level returns on the infrastructure business run in the high 20s percent ROIC. That is a company deploying capital at returns most software businesses would envy, and the reason it depresses free cash flow today is simply timing, you spend the capital now and harvest the contracted revenue over the life of the deal. The free cash flow “problem” is the accounting shadow of a high return investment cycle. Confusing the two is exactly the error that creates mispricing.

The backlog is a conversion machine

Skeptics are right to ask the obvious question about any backlog, will it convert?

RPO is a promise, and promises in this market have lately been worth less than the paper they’re printed on. So look at what is actually being delivered.

Oracle handed over more than 1.2 gigawatts of capacity to customers in fiscal 2026. First quarter fiscal 2027 deliveries alone are approaching one gigawatt, nearly the entire prior four quarters combined, in a single quarter. Global GPU utilization sits at 97.5 percent. This is the company physically unable to build fast enough to satisfy contracted demand, lighting up data centers as quickly as power and hardware allow.

That distinction is the whole ballgame. A backlog that isn’t converting is a warning. A backlog this size that is converting at near full utilization is a flywheel.

And the conversion is already showing up in the reported numbers, total revenue grew 20 percent in constant currency, clearing that symbolic threshold for the first time since early 2023, with IaaS accelerating to 92 percent constant currency growth from 81 percent the quarter before. Acceleration, not deceleration, at scale.

The backlog is becoming revenue in real time.

The forward guide confirms the trajectory rather than walking it back. First-quarter total revenue is guided to 27-29 percent growth, cloud revenue to 58-64 percent. Oracle reiterated $90 billion in fiscal 2027 revenue and raised fiscal 2027 earnings to $8.05. Companies staring at a conversion problem do not raise next year’s earnings while guiding current quarter growth to accelerate by seven points.

The part nobody is paying for

Here is what gets lost entirely in the capex panic, Oracle is not a one engine company, and the other engines are quietly excellent.

Cloud database revenue grew 29 percent, but the number underneath it is the one that matters, multicloud database revenue grew 404 percent year over year, with bookings up 325 percent, as the Database@ partnerships with the hyperscalers scale.

Oracle spent years being mocked for missing the cloud transition. It has turned that supposed weakness into a toll booth, its database now runs inside Amazon, Microsoft, and Google’s clouds, and Oracle collects on every migration.

The most defensible, highest margin asset the company owns is inflecting at triple digit growth, and it is essentially free in the current valuation.

On the applications side, cloud apps grew 10 percent to $4.1 billion, a modest deceleration on the surface, but SaaS deferred revenue grew a faster 16 percent, which is the leading indicator, not the lagging one.

And Oracle is layering agentic monetization on top, token bundles that 33 customers prepurchased in a limited fourth quarter rollout, outcome based pricing, and over a thousand AI agents already embedded across the suites. While the rest of software debates whether AI cannibalizes the seat model, Oracle is quietly metering AI consumption on top of the seat. That is the move the entire SaaS industry is searching for, and Oracle shipped it without a press tour.

These higher margin, diversified vectors are the durability underneath the infrastructure ramp. They are why the margin story holds even as the capital cycle peaks. Proforma operating income came in at 44.8 percent margin, ahead of the Street, during the heaviest build in the company’s history. That is the tell. You do not hold 45 percent margins through a capex supercycle unless the underlying franchise is extraordinarily profitable.

Now the math, because this is where most bull cases get lazy and I refuse to.

At $500, Oracle is roughly a $1.4 trillion company against today’s $590 billion. On this year’s $8.05 in earnings, $500 is 62 times, and if that were the whole story, I’d be talking my book and insulting yours. It isn’t, and the reason is conversion.

Run the backlog forward. Fiscal 2027 revenue is guided to $90 billion, already roughly 34 percent growth.

With the first quarter guided to accelerate and a $638 billion backlog converting at near-full utilization, fiscal 2028 revenue in the $118-120 billion range is arithmetic. Fiscal 2029 around $150 billion follows the same conversion curve. Hold the proforma margin structure roughly steady, and earnings march from $8.05 this year toward $10.50-11 in fiscal 2028 and $13.50-14 in fiscal 2029.

Against those forward numbers, $500 is a different animal entirely, about 46 times fiscal 2028 earnings, and roughly 36 times fiscal 2029.

On revenue, $1.4 trillion is about 11.7 times fiscal 2028 sales and 9.3 times fiscal 2029 for a business compounding 30 percent with 45 percent operating margins and a triple digit growth database annuity attached.

Those are not stretched multiples for that growth and-margin profile. They are roughly what the market pays today for the hyperscalers Oracle is outgrowing.

So the entire thesis reduces to one sentence, $500 is what Oracle is worth if the backlog converts on schedule and the market keeps paying a premium growth multiple while OCI outgrows the hyperscalers. It is not a bet on a new story. It is a bet that the story Oracle is already executing keeps printing for another two years.

The rerating isn’t from “AI hype”, it’s the market moving Oracle from the “show me” bin to the “it’s converting” bin, the exact migration Google made last year and was rewarded for spectacularly.

BUT a bull case without a kill switch is a prayer, so here is mine, stated as plainly as I can.

The funding is the real risk, full stop. Oracle plans to raise around $40 billion next year and is leaning on debt to bridge a capex cycle that has run well ahead of free cash flow. If conversion slips, if power constraints, hardware delays, or a single large customer’s stumble pushes deliveries right then Oracle is carrying the financing cost of capacity that isn’t yet producing contracted revenue, and the FCF hole that bears are pointing at gets deeper before it gets better.

The valuation backstop the market normally leans on simply isn’t there during a build this large, which means sentiment, not cash flow, sets the price in the interim.

That cuts both ways and it is why this name will be volatile.

Concentration is the second risk. A meaningful share of that $638 billion traces to a handful of very large AI customers, and four of them each signed more than $8 billion in the quarter.

The prepaid structures derisk the capex, but they do not fully de-risk the backlog, if one of those customers’ own economics break, Oracle keeps the deposit and loses the multi year revenue stream the stock is now discounting. The diversification is improving, but it is early.

And the macro overhang is real, a genuine slowdown in AI infrastructure spend, a rate environment that punishes leveraged builders, or a compute pricing war that compresses the 30-40 percent infrastructure gross margins Oracle is underwriting any of these turns the flywheel into an anchor.

I weigh those risks and I still land net positive, because the prepaid structure, the near-full utilization, the accelerating reported growth, and the high 20s ROIC together describe demand that is funded, contracted, and converting — not speculative.

But if you take this position, size it for the volatility, because the path to $500 runs directly through quarters where the FCF line looks ugly and the bears get their headline.

Finally, Oracle spent a decade being the company that missed every transition missed cloud, missed mobile, missed the narrative. It is about to spend the next two being the company that quietly built the second largest AI infrastructure business on earth while everyone argued about its free cash flow.

The market is still pricing the software company it used to be. The backlog is describing the infrastructure company it has become. When those two numbers reconcile and a $638 billion book converting at 97.5 percent utilization will force them to reconcile the stock does not have a 210 problem.

But 500 is coming

I’ll wait. And I’ll be loud about it…

This report is for informational purposes only and does not constitute investment advice. WTS’s and Nick’s research is produced independently. Past performance is not a guide to future returns.

And every conversation I have with investors lately is the same conversation, when does the music stop?

I get it, but people made generational money in semis in eighteen months and now they’re sleeping with one eye open.

And my honest answer is nobody rings a bell. Stop waiting for one. There are exactly two things I watch, and only two.

1) the second derivative on AI lab revenue. Not a decline. The day that compounding curve so much as hesitates, every GPU order and vendor financed gigawatt on Earth reprices at the same time. Because all of it, is collateralized by the assumption the curve doesn’t bend.

2) the Blackwell-class models landing by year end. I genuinely have not seen what current silicon produces at the frontier yet. Nobody has. If those models are a step change, the trade extends and the bears wait another year. If they’re incremental... we just built the most expensive infrastructure in human history to find out about diminishing returns. They’d better be great.

But here’s what I actually want to talk about, because this is where the money is.

Eighteen months ago $GOOGL was an AI loser. Search was dead!!

You remember, everyone remembers, nobody admits it. That bucket flip was one of the great large cap repricings of our era, and the pattern repeats. So I’m not hunting winners. I’m hunting filing errors…and those errors are on our face

Two on my screen:

1) $META. The market has decided the superintelligence lab is Reality Labs all over again, money in, nothing out. Meanwhile the core business just grew 33% and the engine doing the growing is... AI. Think about that. AI is simultaneously why the business is great and why the stock is hated. Both can’t stay true. Zuckerberg has shipped his way out of every hole he’s ever dug. Connect is in September. I know which way I’m leaning.

2) $MSFT. The boring one, which is the point. Everyone’s still grading the Azure quarter. The tell was investors suddenly caring about M365 Commercial Cloud accelerating, on Copilot seats, while everyone on this website makes fun of Copilot. Let me say the quiet thing, mocked products with accelerating revenue is where money hides. It’s 20x CY27 earnings. Every PM alive has owned it before. They know the way home.

Semis made you rich. The rotation is what keeps you rich…

Ask which losers were never losers…

The consensus going into last week’s jobs report was that the US economy could grow at roughly 2% while adding fewer than 100,000 jobs per month. The argument was coherent, AI diffusion was reducing labor intensity, reduced immigration was compressing the labor supply, and the combination meant that trend-like growth would no longer deliver trend like hiring.

That argument is now DEAD!

The three month average increase in private employment has moved to 166,000 with revisions and a solid May reading. Not 80,000. Not 95,000. 166,000, the kind of number that does not get generated by an economy experiencing AI-driven labor displacement.

What it gets generated by is an economy where business sentiment is recovering from last year’s trade war shock, where consumer spending has provided a bridge through a hiring stall that firms are now exiting, and where the cyclical lift in goods sector activity is finally broadening to the labor market in the way the recoupling thesis always anticipated.

The AI productivity story was never the right frame for 2026. What happened last year was simpler, businesses pulled back on nontech hiring because sentiment collapsed around the trade war, not because robots replaced the workers.

Consumer spending held up despite weak labor income households drew down savings and perceived the income shock as transitory which boosted corporate profits and facilitated leaner inventories.

The recovery in business sentiment is now leveraging that support. The recoupling is not a prediction anymore, It is in the May jobs report ahead of next week’s Kevin' warash’s first FOMC.

The implication for the Fed is significant. A tightening labor market alongside sticky 3% core inflation is not a combination that justifies an implicit easing bias. The June FOMC meeting will remove it in my opinion.

The Global Picture

The global economy is absorbing a commodity shock of historic proportions with more resilience than the consensus expected entering the year. Global CPI is tracking a 6% annualized gain this quarter a full percentage point upward revision to our full year 2026 forecast, now at 3.7%.

The Middle East conflict has raised our Brent crude assumption to close to $100 per barrel through year end even assuming a Strait of Hormuz opening, roughly 65% above our forecast at the start of the year.

Our models point to a 0.8 percentage point drag on global GDP from this supply shock.

Our global GDP forecasts have been lowered by only 0.3 percentage points.

The gap between the shock and the forecast revision reflects the strength of the underlying expansion that absorbed it.

The May global manufacturing PMI edged up to a five year high, with details pointing to a positive turn in the inventory cycle.

Factory output increased at 3.7% annualized last quarter. Global composite PMI is tracking global GDP growth at an above potential 2.7% in May. These are not the numbers of an economy breaking under an energy shock.

But the vulnerabilities are real and they are concentrated. The services PMI has stepped down since the start of the Iran war, a cumulative 3.1 point decline concentrated in Western Europe, where the conflict appears to have throttled business and household sentiment simultaneously.

Despite upward revisions to Euro area retail sales and the flash PMI, regional indicators point to a near stall in GDP growth.

By contrast, US indicators point to solid service sector gains underpinning our 2.5% annualized second quarter GDP forecast.

Brent crude has fallen back from its peak on signs of progress in US-Iran negotiations.

The Strait of Hormuz is not open.

The risks of price spikes and disruptive physical supply constraints remain elevated. Next week’s US CPI report is expected to show a 0.5% MoM gain, reinforcing the message that the purchasing power squeeze is not yet fully absorbed.

What the Jobs Report Actually Means

Last year’s macroeconomic performance was imbalanced in a specific and unusual way.

Outside China, global capex rose a strong 5.2% but the pace of hiring slowed to a near stall. This decoupling in business behavior occurred alongside sustained above-trend GDP growth, a combination that historically does not persist because it requires consumer spending to carry the economy while businesses simultaneously expand capital spending and cut payrolls.

The mechanism that made it possible was consumer smoothing. Global real consumer spending rose 2% in 2025, providing a bridge for cautious firms that were pulling back on nontech spending and hiring while waiting for the trade war uncertainty to resolve.

That consumer bridge boosted corporate profits firms were delivering revenue growth without hiring costs and facilitated a move toward leaner inventories as firms ran down stock rather than restocking.

The juxtaposition of strong consumer spending and a hiring stall is what produced the decoupling. It was behavioral, not structural at all.

The behavioral catalyst for the recovery is twofold.

First, the fading of last year’s business sentiment slide as trade war fears have receded.

Second, the resilient consumer smoothing through perceived transitory income shocks that prevented the demand destruction that would have otherwise forced firms to cut capex as well.

The combination of these two dynamics reviving business confidence leveraging a lean inventory and strong profitability starting point is what the jobs report is now confirming.

The AI productivity story needs to be addressed directly because it has influenced consensus forecasts materially.

The thesis was that AI diffusion was reducing labor intensity at a pace fast enough to decouple GDP growth from employment growth on a sustained basis. The evidence for this thesis in the aggregate data has always been thin, the decoupling was concentrated in a period when business sentiment was depressed by trade war uncertainty, not when AI adoption was accelerating.

Productivity growth of approximately 3% over the past year is real but it is not eliminating hiring.

It is enabling firms to grow output per worker while still adding workers.

The alternative scenario, hiring recovers alongside productivity gains is more sustainable for the expansion because it provides the fuel for continued household spending growth without requiring a further drawdown in savings rates that are already low.

The Fed

The unemployment rate has been stuck around 4.3% this year. Average hourly earnings have moved down to 3.4% YoY. The May jobs report does not deliver the tight labor market signal that would force an immediate hike discussion at the June FOMC meeting.

What it does is remove the argument for patience.

The likely FOMC response at the June meeting is to remove its implicit easing bias in the statement and in the median dot projecting one 2026 ease.

The shift in the 2027 guidance will be more decisive. The number of members looking for material cumulative easing five had three cuts or more in March will dwindle significantly.

A meaningful contingent will likely forecast hikes.

Chair Warsh has expressed a distaste for forward guidance, and how faithfully he conveys this shift in the press conference will go a long way toward assessing his effectiveness in leading the committee.

The path to a hike runs through the labor market data over the next two quarters. If private sector job gains sustain above 150,000 per month and the unemployment rate drifts toward 4.1% by early next year as we forecast, the Fed will have the combination of tight labor markets and sticky 3% core inflation that makes holding rates restrictive difficult to justify.

The hike arrives in the first half of 2027 in our base case.

What prevents it is Western Europe’s deterioration spreading to the US.

That is the tail risk worth monitoring not as the base case but as the scenario that changes the committee’s calculus most rapidly.

A sharp drop in US services activity following the European template would flip the balance of risks back toward downside growth concerns.

The May data says that is not happening. The June data will tell us whether it is starting.

The ECB

The ECB hikes 25 basis points next week. Core inflation jumped to 2.5% YoY in May, the highest reading of this cycle.

Even with three hikes in its baseline, the staff forecast still shows core inflation remaining above 2% across the entire forecast horizon.

The Governing Council has no credible alternative to hiking when its own models show the inflation target being missed for a third consecutive year.

The discomfort is that the economy the ECB is hiking into is stalling.

The flash PMI points to near zero growth.

Germany is barely positive. Services confidence is collapsing. Consumer real incomes are being squeezed by energy prices that have not retreated to preconflict levels. Frankfurt will guide toward heightened uncertainty around the staff forecast and emphasize data dependence after the first move. The second hike in September is conditional on the inflation data holding but the inflation data is holding, and the ECB’s own models say it will continue to hold regardless of what happens in the Strait.

The ECB’s problem is not whether to hike. It is that hiking into a stalling economy creates the conditions for a policy reversal later in the year if the growth data deteriorates faster than the inflation data.

The committee is aware of this. It has no good options. It has a least bad one.

The BoJ

Governor Ueda struck a hawkish tone this week but with a June hike already fully priced, the yen failed to strengthen USD/JPY returned to approximately 160.

The market is telling the BoJ that one hike is not enough to move the needle on the currency, and another round of fiscal stimulus worth 0.5% of GDP from a new supplementary budget combined with media reports of a large food consumption tax cut next April worth 0.8% of GDP is offsetting the tightening signal with fiscal easing of comparable magnitude.

Japan’s inflation picture is more complicated than the headline CPI suggests.

Government subsidies are suppressing readings materially gasoline subsidies, tax cuts, temporary electricity and gas subsidies reduced core CPI excluding fresh food by approximately 1.0 percentage point in April.

Core CPI rose 1.4% YoY in April, but policy measures are estimated to have trimmed inflation by a further 1.4 percentage points.

Strip those out and underlying inflation is running above the BoJ’s target.

A second wave of energy price pressures arrives this summer as electricity tariffs which reflect oil price movements with a 3 to 6 month lag begin to pass through from mid year.

The government has announced plans for a supplementary budget to reintroduce electricity and gas subsidies over the summer, expected to lower core CPI by 0.4 to 0.5 percentage points in August through October. These subsidies cushion households from immediate energy cost impact.

They also increase the risk of a sharp rebound in energy inflation once they expire, pushing core CPI above 3% YoY by year end in our forecast.

The June hike proceeds. The path after June is the question the BoJ cannot answer cleanly because the fiscal policy it is hiking against keeps changing.

Trade Risk

While the Middle East conflict has dominated macro headlines, markets may soon refocus on trade.

The policy machinery has been moving in ways that are broader and faster moving than the 2017 to 2018 China measures.

After courts invalidated IEEPA and Section 122 tariffs earlier this year, the administration is shifting toward Sections 301 and 232.

Two Section 301 probes are underway one covering structural excess capacity across 16 economies with comments closed April 15 and no tariff level yet announced, and one covering forced labor enforcement across 60 economies where USTR has floated 10 to 12.5% tariffs with comments due July 6. Public hearings are scheduled for July 7.

Section 232 tariffs have been expanded to a wider set of downstream derivative products, and new 232 investigations are being considered for APIs, commercial aircraft, and drones.

India faces a proposed 12.5% tariff under the forced labor probe.

If implemented around the time the Section 122 tariffs expire on July 24, India would effectively see the 10% universal tariff replaced by a 12.5% tariff raising the effective tariff rate from 10.4% to 12.1% with additional upside risk from the parallel excess capacity investigation. New Delhi is pursuing a dual track response, contesting the proceedings on their merits while advancing the US-India Bilateral Trade Agreement framework announced in February.

USTR has also proposed a 25% base tariff on Brazil effective at approximately 19% with exemptions.

Mexico and Canada have supported a USMCA extension but new US tariffs on both suggest negotiations could slip past the July 1 deadline into next year.

The combination of new tariffs and heightened court scrutiny of the administration’s use of trade law authority points to another round of prolonged policy uncertainty headline risk broadening from geopolitics to tariffs, supply chains, and legal outcomes simultaneously.

Asia

The Asia picture this week is defined by three dynamics operating simultaneously, a strong tech cycle providing export revenue above preconflict baselines, fiscal policy cushioning the energy shock across multiple economies, and China navigating the oil disruption through substitution effects that are more creative than most models anticipated.

Korea is the clearest confirmation of the tech cycle’s durability. May customs exports set another record with the short-term trend remaining exceptionally strong. The May manufacturing PMI rose to its fastest pace since early 2021, with new orders and output hitting multi year highs, the composition of the improvement shifted from logistics related support to more fundamental demand and production strength. Semiconductors remain the standout. CPI firmed to 3.1% YoY in May with clearer first round spillover from energy into services, a signal the Bank of Korea will take seriously at the margin, reinforcing the July hike call.

Taiwan’s May manufacturing PMI accelerated to 56.1, the fastest pace since August 2021 and the sixth consecutive month of expansion. Output rose to its highest level in more than four and a half years. The AI semiconductor upcycle remains intact with US hyperscaler March results and robust committed orders pointing to multi year demand visibility rather than a one off surge. Energy conditions have stabilized through contingency measures and diversified imports. The drag from Middle East disruptions is being felt primarily in nontech sectors, especially petrochemicals, with downstream spillovers appearing manageable.

China’s oil substitution story is the most analytically interesting development of the week. China has built up oil reserves by importing approximately 10% higher than normal consumption since 2025, making the preconflict benchmark somewhat inflated by stockpiling needs. Refiners’ margins have turned negative due to state price controls on oil products, reducing incentive to maintain high utilization rates. But the downstream effects have been more limited than feared. In chemicals, the coal-to-chemical route has become more cost-competitive as higher oil prices improve its relative economics, coke oven operating rates have picked up, providing a partial alternative to oil based chemical feedstocks. In transport, the substitution is striking: while gasoline consumption declined, daily vehicles on road did not decline and was actually higher than preconflict levels. Electricity consumption at EV charging stations jumped 69% year-on-year in April versus 55% in January and February. Ride sharing orders, mostly run by electric vehicles, spiked 27% in April.

China’s economy is not immune to the oil shock. It is managing it with more structural flexibility than most observers anticipated, a consequence of the EV penetration and energy diversification investments made over the past decade being exactly the right infrastructure to have in place when a Strait of Hormuz closure tests the system.

The May PMI data from China offers mixed signals. The NBS manufacturing PMI edged down to 50.0 while RatingDog fell to 51.8. New orders slipped into contraction in the NBS survey. Export orders fell further into contractionary territory in both surveys. The downside risks to second quarter GDP that were present last week have not been resolved.

Our forecast for 3.2% annualized second half 2026 growth remains intact but the reflexive fiscal policy response that cushions unexpected weakness under implemented fiscal capacity being deployed as the economy softens is the mechanism that keeps the downside contained rather than the underlying data strength.

The EMAX complex emerging Asia excluding China sees positive pass through from the China cycle and from the global tech upcycle simultaneously.

Korea has delivered a large supplementary budget and retains scope for additional support.

India’s government is absorbing a meaningful share of the terms-of-trade shock, keeping retail price pass-through more modest.

Thailand has announced a 2.1% of GDP package to support second half 2026 growth.

The Philippines is expected to add a 0.7% of GDP supplementary budget to offset tighter monetary policy.

The Asian fiscal response is real, coordinated, and adding a meaningful cushion that is keeping regional growth above the level the energy shock alone would imply.

India

The RBI held rates this week as we expected, defying market pricing that had embedded roughly a 60% probability of a 25 basis point hike. The decision reflects the RBI’s inflation targeting separability framework, policy rates and liquidity manage domestic inflation and growth dynamics, while FX intervention and regulatory tools address external sector pressures.

With headline CPI forecast averaging approximately 5% in the coming fiscal year, well within the inflation targeting band and core inflation correctly measured running at just 2.1 to 2.2%, the threshold for hiking on inflation grounds has not been met.

Instead the RBI delivered a set of regulatory measures aimed at attracting capital inflows and shoring up the balance of payments amid rupee pressure.

The focus on capital inflows reflects the view that current rupee pressures are being driven more by the capital account than the current account, making capital inflow supporting measures a more targeted response than rate hikes that would tighten domestic financial conditions without addressing the source of the FX pressure.

The tariff risk is the more significant near term development for India’s macro outlook. The USTR has proposed a 12.5% tariff under the forced labor Section 301 probe, with public hearings scheduled for July 7 and a final determination to follow. I

f implemented around the time the Section 122 tariffs expire on July 24, India would see its effective tariff rate rise from 10.4% to 12.1% with additional upside risk from the parallel excess capacity investigation. New Delhi’s dual track response contesting the proceedings while advancing the US-India BTA framework, reflects the administration’s view that a negotiated outcome is available.

Whether that outcome arrives before the tariff determination is the key variable to watch over the next six weeks.

India’s first quarter GDP growth came in at a strong 7.8% YoY. April industrial production surprised to the upside at 4.9%. Next week’s May CPI is expected at approximately 4% YoY led by higher fuel and food prices with core momentum the key watchpoint for whether the RBI’s patience on rates remains justified.

EM Central Banks

Most EM central banks have shifted their focus to inflation and are responding with rate hikes. Two notable exceptions this week are India and Israel.

The Bank of Israel signaled greater sensitivity to further FX gains, with Governor Yaron expressing a preference to cut rates further before resorting to FX intervention.

We have pulled forward our next expected rate cut to July from September and added one additional cut to 3.0% by early next year from a previous forecast of 3.25% by mid 2027. Further shekel appreciation skews risks toward a deeper easing cycle.

Across EMEA more broadly the hawkish tilt is intact. The SARB delivered its well flagged 25 basis point hike and left the door open to July.

Turkey’s CBRT is on track to hike in June.

Hungary and Israel represent the dovish exceptions in a region that has otherwise moved decisively toward tighter policy.

Latin America

For the commodity rich Latin American economies, terms of trade are improving and current accounts are proving resilient.

Brazil’s oil exports are helping reduce the current account deficit despite Section 301 tariff risks, combined with fresh fiscal and credit measures amid a tense election race, activity is accelerating. This is adding to inflation risks and increasing the likelihood that the BCB will be forced to pause its easing cycle despite high real rates.

Rising metal prices are supporting Peru’s economy against the backdrop of the June 7 presidential runoff increasingly looking like a referendum on the post1993 framework.

Fujimori holds a narrow polling edge over Sánchez but a large undecided block points to a tight race.

Our base case remains continuity with BCRP policy tightening restarting as early as next week.

Colombia is the most consequential political story in the region right now. The first round election put right wing opposition candidate De La Espriella ahead of continuity candidate Cepeda, setting up a polarized June 21 runoff.

Our base case is regime change.

The macro implications, limited additional monetary policy tightening and a structural fiscal consolidation effort that reduces the sustainability concerns around Colombia’s consumption led growth model.

The market is still pricing limited political risk premium despite elevated uncertainty.

That gap is the trade.

Scandinavia, and Australia

Sweden has bought itself time with low inflation, core inflation running near 0% YoY in April, weighed down by a food VAT cut and SEK strength. But business surveys of activity and confidence have remained solid through May, and the May CPI showed a 0.6% MoM gain in services prices that Governor Thedeen described as creating “risks for inflation further ahead.” The Riksbank is not in a hurry. Assuming the ECB delivers two hikes by September, the Riksbank follows once in September.

Norway has been on the front foot, hiking in May when inflation was already running high. The next hike is likely September rather than June, the June meeting brings a pause as oil prices have fallen back, first quarter GDP came in below the Norges Bank’s expectations, and core inflation for May is expected to show a sideways move. The Bank could move earlier if greater pricing pressures emerge.

Australia’s first quarter GDP disappointed at 0.3% QoQ, below consensus and our expectations. Machinery and equipment investment surged 16.3% adding a solid 0.7 percentage points, largely offset by a 0.5 percentage point drag from higher imports. Non-dwelling construction made no net contribution. The economy grew 2.5% YoY over the last year with weaker recent consumption momentum.

The large minimum wage hike and ongoing Middle East risks keep a tightening bias baked into RBA commentary even as we forecast policy on hold. Sentiment surveys have reacted sharply to the conflict NAB business confidence fell 29 points in March.

Consumer inflation expectations have risen to their highest level since 2023.

More contemporary data on consumption and housing suggest further slowing into the middle of the year.

The UK

A new Survation poll this week showed Andy Burnham with a ten point lead over Reform ahead of the Makerfield by election on June 18 up from a three point lead last week. The momentum of Burnham’s campaign is building faster than the timeline implies.

Our working assumption of a new Labour leader by September remains intact but the pace of developments suggests the process may compress.

The PMI data was more constructive than the headline suggested the final composite output index was revised up 1.2 points to 49.7 in May.

The UK PMI is volatile, and the two month average of 51.2 is consistent with growth of approximately 0.9% annualized, close to our 2Q forecast of 1.0%.

The pricing data is the problem.

Composite output prices are at 61.6 and services output prices at 60.6, elevated readings that suggest relatively fast and broad indirect energy effects.

Business own price expectations in the BoE’s DMP survey fell from 4.4% to 4.0% in May.

A reading of 4% is still problematic for a central bank targeting 2% inflation. Wage growth in the DMP picked up from 4.1% to 4.3%.

Governor Bailey argued this week that the BoE was already responding to the shock simply by not delivering on expected rate cuts.

He acknowledged that expectations had been for 1 to 2 cuts and that a shock that proves large or persistent could require higher rates. Megan Greene looks ready to vote for a hike in June.

Our forecast assumes a majority develops for a hike by July.

The June decision hinges on whether the coalition of hawks, Greene and others who have been moved by the pricing survey data can bring enough votes to the table at a meeting where the governor himself is still emphasizing optionality rather than commitment.

Lastly, three things happened this past week that change the macro picture in ways that will be felt through the second half of 2026.

The US jobs report ended the debate about AI-driven labor market decoupling. The recoupling thesis is playing out. 166,000 private sector jobs over three months is not consistent with an economy where AI is displacing workers faster than growth can replace them.

It is consistent with an economy where business sentiment was depressed by trade war uncertainty last year, consumers bridged the gap, and firms are now hiring into a recovery that the macro data fully supports. The Fed removes its easing bias in June. The first hike comes in the first half of 2027.

The trade risk is broadening. Section 301 probes covering 60 economies on forced labor and 16 economies on excess capacity, Section 232 expansions across derivative products, proposed tariffs on India at 12.5%, USMCA negotiations at risk of slipping past the July 1 deadline, the headline risk is no longer just the Strait of Hormuz. It is a simultaneous energy shock, a tightening global monetary cycle, and a resurgent trade policy uncertainty that markets have not fully repriced.

Asia is managing the energy shock better than the models predicted, through oil reserve drawdowns, EV substitution, fiscal support, and a tech cycle that continues to deliver export revenue well above preconflict baselines. The second half risks to China growth are real but the reflexive fiscal response is the mechanism that prevents the downside from compounding into something structurally damaging.

The expansion holds. The margin of safety continues to narrow. June is when the data starts telling us whether the second half looks like a soft landing or something harder…

But it no longer applies in this market.

What is happening in the global memory market in 2026 is not a repeat of the 2021 upcycle or the 2017 upcycle or any prior version of this story. The demand structure has changed at the architectural level. The customer behavior has changed at the procurement level. The supply discipline has changed at the capital allocation level.

And the emergence of long term agreements as the dominant pricing mechanism has introduced a stability into memory economics that the industry has never had before.

DRAM pricing is up 220–250% YoY in 2026. The memory TAM reaches $1.3 trillion in 2027 and expands toward $1.7 trillion by 2028. HBM supply demand shortages persist through 2028. Enterprise SSD demand is compounding at 52% annually toward a market larger than HBM itself.

And CPU memory demand, the part of this story that almost nobody is talking about, is adding a demand layer on top of GPU demand that was not in any forecast twelve months ago.

The investors who are waiting for the cycle to turn are waiting for a script that has been retired.

The TAM That Nobody Modeled Even Wall Street

The memory TAM is expanding to levels that would have been dismissed as speculative two years ago. The path to $1.3 trillion in 2027 and $1.7 trillion by 2028 is not driven by a single product category running hot. It is driven by three simultaneous expansions that are largely independent of each other and are each in the early stages of their own multi year trajectory.

HBM demand compounds at 85% annually over the next three years as GPU and ASIC shipments scale and content per chip increases with each generation.

DRAM wafer capacity allocation to HBM rises from 24% in 2026 to 31% in 2028, capacity being pulled into the highest value product and away from conventional DRAM, creating supply tightness in conventional memory that is a direct consequence of the AI buildout.

Enterprise SSD is the underappreciated expansion. The market exceeds 500 exabytes in 2026, 43% of total NAND bit demand, and is forecast to exceed 1,100 exabytes within two years at a 52% compound annual growth rate.

The value TAM for enterprise SSD could exceed $300 billion in the next two years, larger than the HBM TAM.

Customers are consistently preferring higher density and higher speed configurations with meaningful content growth per server exceeding 40% gigabytes YoY.

SLC based high speed SSD is gaining traction alongside mainstream TLC SSD with multiple solutions hitting the market from the second half of 2026.

And then there is CPU memory, the demand layer that was not in any model twelve months ago and is now forcing upward revisions to the entire TAM framework.

Three Roads From Here Imo

The memory market has never had this degree of multi year demand visibility. But visibility is not certainty. The range of outcomes from here is wide enough to warrant a clear eyed view of what each path looks like and what drives the probability of each.

Bull Case, 25% Probability

Parameter expansion in foundation models continues at an accelerating pace with a sharp transition into physical and world AI ahead of consensus timelines. Software stack improvements remain limited, keeping hardware intensity elevated across every new computing architecture.

Token consumption follows an exponential curve with sharp AI agent monetization driving CSP capex meaningfully above current forecasts. Consumer demand proves resilient after a near-term unit correction and edge AI demand accelerates despite component cost inflation.

Suppliers and customers lock in multi year supply agreements with high investment and order visibility on both sides. China fails to scale up or out due to equipment export restrictions.

New form factors, HBF, LLW, and others, see active deployment with the wafer to die penalty ratio worsening sharply, creating additional demand pull. In this scenario the $1.7 trillion TAM estimate is conservative and HBM ASP surprises materially to the upside in 2027 and 2028.

Base Case, 60% Probability

Foundation model parameter expansion continues at a moderate pace with a gradual transition into physical and world AI. Memory intensity continues to grow in new architectures despite some software stack improvement.

Token consumption remains robust with moderate AI agent monetization enabling stable CSP capex growth. The mid and low end consumer segment continues to shrink gradually after a near-term correction while edge AI demand progresses slowly.

Suppliers and customers reach consensus on investment scale with reasonable order visibility and certain contractual bindings. China continues to face meaningful scale up and scale out challenges.

New form factors see selective adoption and the wafer to die penalty ratio gradually worsens. This is the scenario embedded in the $1.3 trillion 2027 TAM forecast and the HBM shortage-through-2028 base case. Memory stocks deliver 50% market cap upside on a 3.6x price to sales framework.

Bear Case, 15% Probability

Foundation model parameter expansion takes a pause. The transition into physical and world AI slows materially with rising funding concerns across the AI ecosystem.

A software stack breakthrough reduces hardware intensity and BOM cost pressure flattens memory demand in new architectures. Token consumption follows a linear rather than exponential curve and slow AI agent monetization translates into muted CSP capex spending.

Demand destruction hits all consumer segments and edge AI projects are delayed by multiple years. Suppliers and customers fail to comply with multi year agreements, limited order visibility emerges, and surplus capacity tips into oversupply.

China shows a meaningful pickup in bit and wafer productivity and product quality. Customers lose appetite for new form-factors and wafer to die penalty concerns dissipate. In this scenario the cycle turns faster than the supply demand model implies and the valuation rerating reverses before it completes.

The 15% bear case probability is not low enough to ignore. The key variables to monitor are the pace of AI agent monetization, the difference between exponential and linear token consumption is the single biggest swing factor in the demand model and China’s equipment constrained progress on advanced node capability.

If either variable moves materially in the wrong direction the probability distribution shifts toward the bear case faster than the LTA structure can absorb.

The CPU Demand Is worth being attention to here too…

The AI memory narrative for the past three years has been a GPU story. HBM for NVIDIA. HBM for Google TPUs. HBM for Amazon Trainium. GPU compute drives memory demand. That framework is correct and it remains intact.

But the biggest surprise in first-quarter 2026 earnings was not GPU demand. It was CPU demand.

NVIDIA guided to $20 billion in standalone Vera CPU sales in 2026.

AMD revised its server CPU TAM forecast upward to grow 35% annually to $120 billion by 2030 double its prior estimate.

Intel guided the CPU to GPU ratio moving from approximately 1:8 toward 1:4 and hinted at eventual parity as the compute requirements of agentic AI accelerate.

Three separate companies, in three separate earnings calls, delivering the same message, the CPU is becoming a first class citizen in AI infrastructure, not an afterthought.

The memory implications are substantial and have not been fully absorbed by the market.

The CPU to GPU ratio has been trending in one direction for three years, from 5.4:1 in 2023 to 3.2:1 in 2025, moving toward an estimated 2.4:1 by 2028.

Every step in that ratio is a step function increase in DRAM demand per server rack.

A Vera CPU comes in two memory configurations, 768GB based on four modules of 192GB SOCAMM2, or 1.5TB based on eight modules.

A typical two socket server carries approximately 1.5TB of memory content per CPU.

At 600,000 Vera CPU units in 2026 and 3 million in 2027, the incremental bit demand reaches 461 million gigabytes and 2.3 billion gigabytes respectively.

The CPU contribution looks modest as a percentage of total demand today. The trajectory is what matters.

As AGI infrastructure scales, the CPU orchestration layer, handling task coordination, state management, API execution, and logical control of AI workloads scales proportionally.

NVIDIA’s Vera CPU is the first purpose built chip for this role. ARM’s AGI CPU and Qualcomm’s ARM based server CPU are in development which goes back to my bullish thesis on ARM that I posted on X last week. The CPU memory demand story is in its first inning.

The bull case sensitivity analysis shows 6% and 20% server bit demand upside in 2026 and 2027 respectively under higher CPU content assumptions with the key bottleneck being memory supply itself, which will ironically limit the total DRAM TAM.

Memory makers running at near 80% margins are expected to slow ASP hikes to soften demand destruction and deepen the LTA relationships that provide multi year revenue visibility.

The DRAM Market: Supply, Demand, and the Shortage That Stays

DRAM blended pricing is up 220–250% YoY in 2026. The supply demand model shows shortage conditions persisting through 2028 even after accounting for the capex acceleration underway.

Three years of accumulated DRAM capex of $364 billion brings wafer starts per month to 2.8 million by end of 2028, up 880,000 from end of 2025 levels with 60% of the incremental allocation going to HBM.

That means 60% of new DRAM capacity is being built for the highest value product with the longest shortage runway, leaving conventional DRAM supply tighter than the headline capacity numbers suggest.

The critical supply constraint is not really capital. It is EUV. EUV procurement and infrastructure buildout capacity is the key bottleneck limiting how fast the industry can bring leading edge DRAM capacity online.

ASML produces a finite number of EUV machines per year and the semiconductor industry is competing for every one of them.

The DRAM capacity expansion timeline is therefore more constrained than the capex numbers alone suggest, a structural ceiling on supply growth that the demand trajectory shows no signs of respecting.

HBM Shortage will not end

High Bandwidth Memory is the most constrained product in the semiconductor supply chain and it will remain constrained through 2028. The supply demand model has been revised upward by 17–21% through 2026–2028 and still shows double digit percentage shortages persisting to the end of the forecast horizon,

The key changes driving the upward revision, ASIC demand acceleration the ASIC bit mix in HBM rises from 33% in 2026 to 39% in 2027 as Google TPU, Amazon Trainium, and other custom silicon programs scale volumes and upgrade to HBM3E.

NVIDIA Blackwell GPU shipments revised up to 8.9 million and 9.9 million units in 2026 and 2027 respectively. And conservative HBM4E content assumptions reflecting a longer 12-Hi product life cycle, the most advanced memory product staying in production longer because demand is absorbing every unit produced.

The ASP story is the one that changes the multi year earnings trajectory. HBM3E ASP, previously expected to decline 2% in 2026 and hold flat in 2027, is now expected to be flat in 2026 and rise 19% in 2027, a complete reversal driven by the resilience of demand from both GPU and ASIC customers.

Blended HBM ASP rises by mid teens percent in 2026, aided by a 29% HBM4 pricing premium over HBM3E.

In 2027, blended HBM ASP rises a record 32% YoY, 19% from HBM3E, 15% from HBM4, and HBM4E priced at a 61% premium to HBM3E due to severe supply chain tightness.

The wallet share dynamics are shifting. Samsung has begun shipping 12-Hi 48GB HBM4E samples for qualification driven by its 1cnm mono-die architecture and Samsung Foundry’s 4nm base die with improving back-end yield.

SK Hynix maintains leadership in HBM3E and the majority of NVIDIA wallet share.

Micron is gaining share aggressively through its US ID1 fab and Singapore packaging facility. The introduction of HBM4E triggers a new competitive wave competing on pin speed, logic die sourcing, and yield performance.

NAND: Under-Investment Creates the Next Shortage

NAND pricing grows approximately 150% in 2026 and approximately 10% in 2027. The supply discipline in NAND is even more pronounced than in DRAM, capital intensity is running below historical peak levels, greenfield capacity impact does not materialize until the second half of 2028, and evidence of 2029 greenfield projects is only now beginning to emerge. and as I am writing this, there was a report out of South Korea this morning, that Sk has confirmed “zero new NAND capacity”. Which will leave only YMTC and Kioxia adding supply into a market where AI SSD demand is accelerating and the inventory correction is behind us.

Kioxia is the direct beneficiary. No competition at all from here and greenfield site is operational. Frankly, I think there’s a reason why they’re doing this and it’s clear. SK Hynix don’t need to add new capacity at all because they’re printing at 75%+ margins on HBM.

Why dilute that with commodity NAND capex?

Less supply. More demand. Kioxia gaining share. NAND pricing just got yet another structural tailwind imo & the market hasn’t priced yet.

NAND is a lower priority than DRAM among the top three producers, but the eSSD demand trajectory creates a scenario where under investment today becomes a supply constraint problem in 2027 and 2028.

Three years of accumulated NAND capex of $86 billion brings wafer starts per month to 1.44 million by end of 2028, up 165,000 from current levels, with the majority of spending on technology migration rather than greenfield capacity.

Memory as The Procurement Revolution

The single most important structural change in the memory market is not the demand story. It is the procurement story as I call it.

Long term agreements and strategic customer agreements have become the dominant pricing mechanism for both DRAM and NAND.

Hyperscalers and cloud providers are locking in supply and pricing across multi year windows because memory has become a strategic asset in the words of Micron’s CEO and stable procurement is imperative for AI service operations.

You cannot run an AI inference cluster at scale if your memory supply is uncertain.

The hyperscalers have internalized this and they are paying for certainty with long term commitments.

Memory as a percentage of total cloud service provider hardware capex ranged between mid teens percent at the beginning of the AI buildout. It is now anticipated to exceed 50% this year.

That single statistic captures the architectural shift better than any other data point honestly and memory has moved from a commodity input in the computing stack to the primary value add component.

The implications for cycle dynamics are profound. Prior memory cycles were characterized by spot market pricing that moved violently in both directions.

LTAs remove the spot market from that role for the customers that matter most.

Memory makers are actively slowing ASP hikes as margins approach 80% not because demand is softening, but to soften the risk of demand destruction and to deepen the LTA relationships that provide multi year revenue visibility.

Supply Discipline

Capital intensity is running below historical averages across both DRAM and NAND.

This is supply discipline by design, not by accident.

The top three producers have internalized the lessons of the 2022/23/24 oversupply cycle and are managing capacity expansion to match demand visibility rather than to capture spot market share.

The capex acceleration underway, $450 billion combined over the next three years versus $300 billion in the December 2025 model is being driven by the multi year shortage outlook and the economics of a market where DRAM margins are approaching 80%.

But the EUV bottleneck means that capital deployed does not translate linearly into capacity delivered. The supply response is slower than the capex numbers suggest and the demand trajectory shows no signs of waiting.

China

China’s domestic memory producers are gaining volume share DRAM from 6% in 2025 to 8–11% through 2028, NAND from 12% in 2025 to 12–16% through 2028 with above industry capacity build out trends taking DRAM and NAND capacity shares each to approximately 18–19% by end of 2028.

The value share picture is different.

Inferior product mix primarily at mature node conventional DRAM and lower density NAND limits value share gains to approximately 10% DRAM and 12% NAND by 2028.

China’s domestic producers cannot qualify for HBM programs, cannot compete in enterprise SSD at the leading edge, and face ongoing equipment export restriction risk via the MATCH Act and related policy updates.

The two monitoring points that matter, first, how much DRAM wafer capacity China allocates to HBM equivalent products greater HBM wafer mix among global producers tightens conventional DRAM supply demand and partially offsets China’s conventional DRAM share gains.

Second, any updates to equipment export restriction policy that limits China’s ability to migrate to leading edge nodes.

The Valuation Framework Is Transitioning

Memory stocks continue to trade at a discount to earnings. The best memory market in history, and the stocks are not pricing it at least for now.

The reason is a valuation framework problem, not a fundamental problem. Investors applying traditional price to book multiples are getting answers that look stretched against historical ranges.

But those historical ranges were calibrated to a cyclical industry with volatile earnings and no pricing visibility.

The LTA structure, the strategic asset framing of memory by hyperscalers, and the multi year shortage outlook represent a permanent change in the earnings quality of memory producers one that warrants a transition from cycle based valuation to something closer to the cash flow multiples applied to structural growth businesses.

We estimate the memory TAM will expand to $1.3 trillion in 2027 with approximately 50% market cap upside throughout next year using a 3.6x price to sales multiple.

The upside is potentially greater using a market cap to operating profit approach, as memory stocks continue to trade at a discount to their earnings a discount that reflects the market’s reluctance to apply a new valuation framework to an industry whose earnings profile has structurally changed.

The path to a new valuation framework will be uneven. There will be quarters where spot pricing creates noise and the old framework reasserts itself.

But the direction is clear. Memory is no longer a commodity cyclical.

It is a strategic infrastructure input with multi year demand visibility, pricing power anchored by long term agreements, and a customer base that has publicly committed to treating it as a critical asset. The valuation will follow the fundamentals. It always does.

Where to Be Positioned

The global memory sector has outperformed the index and semiconductor ecosystem peers ytd in 2026 and the fundamental drivers supporting that outperformance are not weakening. They are broadening.

SK Hynix leads in the near term the strongest HBM3E position, the deepest NVIDIA relationship, and the most advanced HBM4 qualification timeline.

Samsung is closing the gap faster than consensus expected, with 12-Hi HBM4E samples already in customer qualification and Samsung Foundry’s 4nm base die providing a differentiated technology path.

Micron is the most aggressive capacity expander and the most levered to the US domestic supply chain buildoutvthe ID1 fab and Singapore packaging facility represent the most significant incremental HBM capacity coming online through 2027.

Kioxia is the enterprise SSD play the largest purevplay NAND producer with the most direct exposure to the eSSD TAM that reaches $300 billion in the next two years. The supply discipline in NAND combined with the SLC demand pickup creates a pricing environment that is favorable for the next several quarters at minimum.

Micron is the most accessible name for US investors and the most directly actionable. WTS maintains Overweight on MU. The CPU memory demand upside, the HBM share gain trajectory, and the eSSD market expansion all represent demand drivers that were not fully in the model at the start of the year and are still not fully in consensus.

The memory market has changed permanently. The cycle framework that has governed how investors think about this industry for thirty years no longer reflects the demand structure, the procurement behavior, or the supply economics of what exists today.

The investors who recognize that first will be the ones who are still positioned when the valuation framework finally catches up to the fundamentals…

This report is for informational purposes only and does not constitute investment advice. WTS and Nick’s research is produced independently. Past performance is not a guide to future returns.

Hyperscaler datacenter capex is tracking more than 80% YoY growth in 2026. CY27 is already tracking more than 50% YoY growth based on bookings and backlog at the AI compute and networking vendors. Order books are at record levels with visibility extending into CY27 and beyond. Thirteen semis and semicap companies presented at the TMC Conference this week. Every single management team delivered the same message which is essentially, bookings, backlogs, and customer escalations have continued to build in the weeks since first quarter earnings.

This is not a late cycle signal. This is a mid cycle (potentially the start) signal in a cycle that has structural tailwinds that prior semiconductor cycles did not have.

We maintained Overweight ratings on AVGO, MRVL, MU, NVDA, KLAC, ADI, AMAT, LRCX, CDNS, SNPS, ALAB, MTSI, and MKSI. The case for each is made below.

Where We Are in the Cycle

The semi industry has completed four full cycles over the past twelve years. In every one of them the down cycle was demand driven, not supply driven. The supply side of the value chain has become structurally more disciplined. The last four corrections 2010/11, 2015/16, 2018/19, and 2022/23/24 were all caused by demand falling away, not capacity overwhelming the market.

That structural shift in supply discipline changes the risk profile of the current upcycle fundamentally. Capital intensity is running below historical averages. The principal oversupply risk is not reckless capacity expansion, it is BOM inflation driven demand destruction in client and consumer markets if memory pricing rises fast enough to choke off end device purchases. That risk is real and worth monitoring in the second half of this year. But it is a consumer end risk, not a datacenter risk, and the datacenter is where the cycle actually lives.

Industry revenues accelerated to more than 20% growth in 2025 following 6–8% growth in CY24 excluding memory. The 2026 outlook, semi industry revenues up more than 30% YoY. Wafer fab equipment up more than 20%. Software up 10–15%. Approximately 80% of covered semiconductor and semicap companies delivered better than expected guidance over the past five quarters double the rate seen in 2024.

The cycle has two to four years of structural runway. The drivers are not cyclical. They are architectural, and the shift from CPU to accelerated compute, the buildout of AI inference infrastructure at hyperscale, the agentic AI transition that requires 10 to 100 times more compute per user than prior AI workloads, and the memory upcycle that is embedding pricing power into the supply chain for the first time in a decade.

The ASIC Market $70 Billion and Growing at 50% Annually

The custom silicon opportunity is the most important structural shift in semiconductors that the market is still underpricing.

Large OEMs, cloud providers, and hyperscalers are moving toward custom ASICs because off the shelf solutions GPUs included, cannot deliver the combination of differentiation, performance, power efficiency, and total cost of ownership that the most demanding AI workloads require. Google, Meta, Amazon, Microsoft, Anthropic, OpenAI, and SoftBank/ARM are all building custom silicon programs. The digital custom ASIC market is a $70 billion opportunity in CY26 growing at a 40–50% compound annual rate.

Two companies dominate this market and will continue to dominate it. Broadcom holds 80–85% share of the high end ASIC market. Marvell holds 10–12%. The reason neither share position is at risk is the same reason networking market share is not at risk, the IP portfolio, the SERDES capabilities, the advanced packaging expertise, and the chip design know how required to execute these programs at scale cannot be replicated quickly.

The COT, customer owned tooling threat that the market periodically revisits is not materializing at scale. The evidence is unambiguous. Google’s COT attempt with MediaTek on the v8t lower complexity chip has experienced major delays with no supply. Google’s engagement with Broadcom on the v8i high complexity chip has experienced no delays and is ramping now. Microsoft’s COT attempt with GUC on Maia 3nm was previously delayed. Microsoft’s engagement with Marvell on Maia 3nm is on track to ramp at the beginning of CY27. The pattern is pretty much consistent COT works for lower complexity programs where the hyperscaler has sufficient in-house expertise. For high complexity programs, the ones that actually move the needle, Broadcom and Marvell win every time.

Broadcom: The Infrastructure Layer the AI Economy Runs On

OW | Price Target $800 | Current $460

Broadcom’s AI networking silicon revenues grow more than 3x this year. In FY27 they reach $45 billion, more than double FY26, representing approximately 28% of total AI revenues. The company will drive $65 billion in total AI revenue in FY26, up from approximately $20 billion in FY25, as new programs ramp across Meta, Google, Anthropic, OpenAI, and SoftBank/ARM.

The Tomahawk 6, 3nm, 102.4 terabits per second, 1.6 terabit optical connectivity is ramping now in what is tracking as the fastest networking product ramp in company history. Broadcom is shipping T6 six to nine months ahead of NVIDIA and more than twelve months ahead of Marvell and Cisco. T6 is already almost sold out for CY27. The platform incorporates next generation Condor 3nm high speed 200 gigabit SERDES I/O, support for copackaged optics via the Davisson architecture, and early Ultra Ethernet Consortium specifications designed to support 1 million XPU clusters.

The ASIC pipeline is the other half of the story. Google TPU orders continue building the CY27 backlog currently around 6.5 million V7/V8 shipments worth approximately $100 billion in revenues. Meta is ramping its first 3nm program now with four additional programs through 2027, targeting 1 gigawatt in CY27. OpenAI’s 2nm/3nm first silicon is out of fab with functionality confirmed on track for CY27 ramp at 1 gigawatt. Google and Broadcom are assessing the v8i next generation design with four compute die and twelve HBM stacks. Either the v8i extension or the next generation v9/v10 2nm TPU strategy results in no revenue air pocket in CY28, the team is tracking 70%+ growth in CY28.

Broadcom now offers full rack scale capabilities and is ramping advanced substrate design and manufacturing capabilities in 2H26. Three of its five AI ASIC customers are already on next generation 3D SOIC chip stacking programs.

The 2nm Tomahawk 7, 204.8 terabits per second, 3.2 terabit optical begins sampling next year and ramps in 2028. At any given time Broadcom is designing its next two generations simultaneously. A competitor trying to match Tomahawk 6 today would be shipping into a market where Broadcom’s customers are already beginning to qualify Tomahawk 7.

The analysts modeling Broadcom as a chip company will keep getting the numbers wrong. This is an infrastructure company. The AI economy runs on its silicon and the upgrade path from that infrastructure runs through Broadcom’s next product generation by design.

Marvell, The Number Two That Is Executing Like a Number One

OW | Price Target $500 | Current $460

Jensen’s fav company, he said this week that Marvell will be a trillion dollar company and he’s not wrong, in fact, it's very possible very soon. Even after the stock shot up 23% today, we still 3x less than the trillion dollar market cap that he talked about but not because he said it, it means its going straight up, the numbers show it.

Marvell drives approximately $9.3 billion of data center revenue in CY26, up from $6.1 billion in CY25, and $14.6 billion in CY27. The ASIC pipeline has more than 70 cumulative designs across 12nm, 5nm, 3nm, and 2nm nodes with early engagement on sub-2nm already underway.

The near term pipeline is executing without delay. Amazon Trainium 3 at 3nm commenced production last quarter and is on track to ramp mid-CY26. Trainium 4 at 2nm is on track for 2H CY27. Microsoft Maia 3nm is taping out now with program ramp beginning CY27, full front end and back end design engagement. The Celestial CPO solution tied to Trainium 4 has seen upside to 2H27 orders. A multibillion dollar multi year Google SmartNIC/DPU ASIC win ramps in CY27.

Marvell’s HBM memory IP differentiation is underappreciated. The recently unveiled capability delivers 33% more HBM stacks per package, 70% lower interface power, and 25% more XPU silicon area. Marvell has won numerous HBM4 ASIC logic base die programs with DRAM suppliers, a position that compounds as HBM becomes the dominant memory architecture for AI inference.

The optical DSP business, strong 800 gigabit shipments now, 1.6 terabit ramp underway, coherent lite and initial CPO ramp beginning, adds a networking revenue stream that is growing alongside the ASIC pipeline rather than competing with it. Marvell is building toward the same kind of fullstack AI infrastructure position that Broadcom has, it is two to three years behind, but the pipeline execution suggests it is on track.

NVIDIA Keeps getting better

OW | Price Target $300 | Current $211

The debate about whether custom ASICs threaten NVIDIA misses the point. The hyperscalers building custom silicon are also continuing to buy NVIDIA GPUs for the workloads where off the shelf solutions remain the most cost effective option. These are not mutually exclusive strategies. The TAM is expanding faster than any single architecture can fill it.

Agentic AI is the demand accelerant that makes the NVIDIA bull case durable well beyond CY26. Agentic workflows require 10 to 100 times more compute per user than prior generative AI applications. The inference infrastructure required to run production agentic AI at scale is orders of magnitude larger than what was needed for the experimentation phase. NVIDIA’s Blackwell architecture is the product that inference at scale is being built on.

Datacenter capex is tracking more than 80% YoY growth in CY26 with CY27 already trending more than 50% growth based on order visibility. The AI accelerator TAM is growing fast enough that GPU and ASIC share gains are happening simultaneously. NVIDIA is not being displaced. The market is growing around it.

Memory The Upcycle With Pricing Power

Micron | OW | Price Target $1,600 | Current $980

I stared covering micron when it was trading at $80 and formally initiated a positioned around $111 for many reasons, I posted a full video on the thesis on X back then and said the target was $1000, now that we are here, I am revising my target higher but as simple as it was back then every datacenter, robot, phone, self driving cars and every model, to name a few, need memory.

DRAM blended pricing increases more than 160% in 2026 and 15–20% in 2027. NAND pricing grows approximately 150% in 2026 and approximately 10% in 2027. Long term and strategic customer agreements have emerged as the cycle shaping catalyst, locking in pricing and supply across both DRAM and NAND and providing the kind of multiyear revenue visibility the memory industry has never had in prior cycles.

DRAM bit demand grows more than 25% YoY in 2026 led by AI server demand and normalization of excess inventories. NAND bit demand grows more than 20% YoY led by enterprise SSD. Capital intensity is running below historical averages, limiting oversupply risk.

HBM tightness is extending well beyond CY26. Memory is increasingly framed by hyperscalers as a strategic asset for AI workloads the language of strategic customer agreements, not spot market procurement. That framing has structural pricing implications that the cycle models were not built to capture.

Wafer Fab Equipment The Longest Visibility in a Decade

The WFE cycle is elongating into CY27 and beyond. Customer buying behavior has structurally changed, rolling eight quarter forecasts and linear early order placement are providing the longest forward visibility in more than ten years. Near term WFE is not being driven by Greenfield pull-ins but by creative reuse of existing capacity combined with leading edge transitions that require new tooling regardless. The mix is shifting decisively toward advanced foundry logic, DRAM and HBM, and advanced packaging, the highest value segments for every large cap semicap name.

Tool and process intensity per wafer is compounding via leading edge transitions, gate all around, CFET, backside power delivery, HBM, and hybrid bonding. This drives SAM expansion and share gains for the large cap semicap names that outperform underlying WFE growth. Service intensity is rising structurally. CY27 WFE spending is expected to exceed CY26.

KLA Corporation | OW | Price Target $3,000 | Current $1940

KLA’s share of advanced packaging is rising to approximately 6% in 2025 and likely 7% or more in 2026 on bump shrink and hybrid bonding adoption. Visibility into the following year is the best seen in over a decade. The process control position in advanced logic and DRAM is structurally advantaged as leadingedge transitions multiply the number of inspection steps required per wafer.

Applied Materials | OW | Price Target $900 | Current $460

Approximately 80% of CY26 incremental WFE spend is concentrated in advanced foundry logic, DRAM, and advanced packaging, the three segments where Applied Materials has its strongest positions. The deposition and etch intensity increases at every leading edge node transition, and the move to gate all around and backside power delivery structures creates step function increases in Applied Materials’ addressable content per wafer.

Lam Research | OW | Price Target $740 | Current $460

Etch and deposition intensity per wafer compounds at every node and every new memory architecture. HBM production requires significantly more etch steps than conventional DRAM. The NAND recovery adds a cyclical overlay to a structural position that is already benefiting from AI-driven demand. Lam’s exposure to both leading edge logic and the memory upcycle provides dual engine growth that is rare in the semicap group.

EDA, The Software Layer Nobody Talks About Enough

The complexity of 2nm and 3nm designs, chiplet architectures, advanced packaging, and heterogeneous integration is driving structural demand growth in EDA that is not cyclical and does not decelerate when the memory cycle turns. Every new ASIC program at Broadcom, Marvell, Google, Meta, and Amazon runs through EDA software. Every leading edge node transition creates years of new design starts that require EDA tools throughout the lifecycle.

Synopsys | OW | Price Target $620 | Current $460

The pending Ansys acquisition, if completed, creates the most comprehensive design through simulation platform in the industry. Synopsys’ core EDA position is structurally protected by the switching costs and learning curves embedded in design teams that have used its tools for years.

Cadence Design Systems | OW | Price Target $600 | Current $460

Cadence’s AI-driven EDA tools are compounding the value proposition at exactly the moment when design complexity is creating the most demand for automation. The system design and verification portfolio is expanding into the hyperscaler custom silicon programs directly, a TAM expansion that was not visible two years ago.

Analog and MCU, The Cyclical Recovery Is Real and Early

The analog recovery is not priced correctly by the market because the market is still anchored to the 2022/23/24 down cycle narrative. That narrative is stale.

Texas Instruments and Analog Devices inflected positively YoY in first quarter 2025 and have continued to gradually improve. Gross margins are trending higher across the group on improved factory utilization, cost synergies, and premium mix shift toward data center. Pricing is holding with upward bias on emerging tightness in analog and mixed signal. Lead times are stretching. Backlog is building.

The MCU recovery is lagging analog by six to nine months, consistent with every prior cycle but the indicators are improving and revenue trends are stabilizing. AI at the edge and AI enablement are emerging as incremental growth vectors layered on top of the cyclical recovery, adding a structural dimension to what would otherwise be a purely cyclical story.

Analog Devices | OW | Price Target $600 | Current $460

ADI’s industrial and data center exposure provides the most balanced growth profile in the analog group. The data center power management and signal chain content is growing alongside the AI infrastructure buildout while the industrial recovery adds a cyclical overlay. Inventory restocking is in early stages with lead times stretching — the setup for sustained positive earnings revisions through 2026 and 2027.

SMID-Cap, The Higher-Beta Plays on the Same Structural Themes

Astera Labs | OW | Price Target $900 | Current $460

Astera Labs is the purest play on the connectivity infrastructure layer of the AI data center. PCIe and CXL retimer and smart cable module content per AI server rack is growing with every new generation of compute architecture. The company has no legacy business to defend and no exposure to the consumer or industrial end markets that are creating noise in the broader sector. Every dollar of AI data center capex is a dollar that eventually flows through the connectivity layer that Astera Labs provides.

MACOM Technology Solutions | OW | Price Target $780 | Current $460

MACOM’s compound semiconductor and photonics content is growing with the optical connectivity upgrade cycle, the same 800 gigabit to 1.6 terabit transition that is driving Broadcom’s Tomahawk 6 ramp. The defense and aerospace exposure adds a non-correlated revenue stream that is growing independently of the data center cycle. The combination of data center optical tailwinds and defense spending growth creates a multi year earnings revision story that the market has not fully priced.

MKS Instruments | OW | Price Target $520 | Current $460

MKS provides the power, gas, and plasma control infrastructure that semiconductor fabs cannot operate without. The WFE upcycle is a direct revenue driver. The leading edge transition to gate all around and advanced packaging structures increases process complexity and the number of control steps per wafer, driving content growth that outpaces the underlying WFE spend growth. The recovery from the 2022/23 down cycle is not complete and the path to earnings normalization provides multiple expansion alongside fundamental growth.

Conference Takeaways

The most power thirteen companies in the world and they all had the same exact message in my opinion.

The TMC 2026 Conference delivered the clearest demand confirmation of the year. Across every presenting company, without exception, management teams reported that bookings, backlogs, and customer escalations have continued to build in the weeks since first quarter earnings. This is real time demand confirmation from companies with direct visibility into what their customers are ordering and installing.

The memory tightness signal was the most important new data point. HBM, DRAM, and NAND tightness is expected to persist beyond CY26. Memory is being framed by hyperscalers and their suppliers not as a commodity input but as a strategic asset for AI workloads, language that has pricing implications that extend well beyond the current upcycle.

Microchip’s recovery is firmly on track with FY26 revenue up 35% YoYand April bookings at a near four year high. NXP echoed building demand momentum with conviction on double digit percentage growth for both CY26 and CY27. The industrial and automotive recovery is real, it is in early innings, and it is being layered with AI at the edge demand that was not part of the demand picture two years ago.

But I am not naive to ignore the bears and wall street’s arguments against my thesis, there are obviously risks involved like everything else but this is the single greatest asymmetrical risk i’ve ever seen, and the odds are with me I believe, they just dont see what I see and it will just be a matter of time before they jump onboard and it’ll be too late by then to generate alpha. With all due respect to them, these are the smartest and most capable people on the planet and its a zero sum game, somebody has to lose for the other to make money. They still live in the old economy, where their models don’t work anymore.

With that being said, the only three risks I warrant monitoring are as follow,

Demand destruction in consumer end markets. DRAM pricing up 160% in 2026 is already visible in low and mid end Android smartphone and consumer PC shipments. If memory pricing rises fast enough to meaningfully impair device economics at the consumer level, the client and consumer segments create a headwind to the broader industry revenue numbers. Data center is insulated. Consumer is not.

Geopolitical and export control escalation. Further restrictions on semis exports to China remain a risk that is difficult to model and easy to underestimate. The US-China technology competition is structural and the regulatory environment around semiconductor exports is directionally tightening regardless of near term diplomatic signals.

Agentic AI monetization timeline. The 10 to 100 times compute requirement of agentic AI is real. The timeline over which enterprise adoption converts that compute requirement into actual infrastructure spending is uncertain. A slower than expected enterprise adoption curve does not break the cycle but it affects the slope of the CY27 revenue numbers for the inference exposed names.

None of these risks are cycle ending in my opinion but they are all worth tracking.

The final point that I want to make here is that the semi industry is in the middle of the most structurally supported upcycle in its history. Four cycles over the past twelve years have all been demand driven corrections in a supply disciplined industry. The current upcycle has AI infrastructure demand at its core, a demand source that is not discretionary, not front loaded, and not reversing.

The custom ASIC market is a $70 billion opportunity growing at 50% annually with two dominant players who cannot be displaced on any realistic timeline. Memory pricing is up 160% with long term customer agreements locking in multiyear visibility. WFE is extending into CY27 with the longest order visibility in a decade. The analog and MCU recovery is in early innings with AI at the edge adding a structural layer to a cyclical story.

The stocks that benefit from all of this are not hard to identify. The question is whether you are positioned before the second half of 2026 confirms what the order books are already telling you.

I again reiterate and maintain Overweight on AVGO ($800), MRVL ($500), MU ($1,600), NVDA ($300), KLAC ($3,000), ADI ($600), AMAT ($900), LRCX ($740), CDNS ($600), SNPS ($620), ALAB ($900), MTSI ($780), and MKSI ($520).

The cycle has years left. The AI infrastructure buildout is the foundation. Everything else is downstream of that single fact…

This report is for informational purposes only and does not constitute investment advice. Nick and WTS’s research is produced independently. Past performance is not a guide to future returns.

Every two years, Broadcom doubles the switching throughput of its flagship Tomahawk chipset. It has done this without interruption since 2010. Not approximately every two years. Not roughly doubling. Every 24 months. Exactly doubling. While the rest of the semi industry was debating the death of Moore’s Law, Broadcom kept its own version of it running inside the one part of the stack that the AI infrastructure buildout cannot function without.

That cadence is the moat. Everything else, the $45 billion FY27 revenue number, the 70% market share, the fastest product ramp in company history is a consequence of it.

What Tomahawk 6 Actually Means

The current generation Tomahawk 5 5nm, 51.2 terabits per second of switching throughput, supporting 800 gigabit optical connectivity has been in volume production since 2023. It powers the AI clusters being built today at Google, Meta, Amazon, Microsoft, Alibaba, Oracle, ByteDance, and Tencent. Every major hyperscaler building at scale is running Broadcom’s silicon in the network layer.

Tomahawk 6 doubles it. 3nm process. 102.4 terabits per second. 1.6 terabit optical connectivity speeds. Next generation Condor 3nm high speed 200 gigabit SERDES I/O that reduces cost and power consumption and extends the viability of copper cables while positioning the platform for copackaged optics adoption. Support for AI specific features, cognitive routing, congestion management, load balancing, telemetry that the previous generation introduced and this generation deepens.

Tomahawk 6 started shipping in the second half of last year. Volume production is ramping now. The shipment crossover from T5 to T6 happens in the second half of this year and it is tracking as the fastest networking product ramp in Broadcom’s history. OpenAI and Anthropic join the existing hyperscaler base as T6 customers next year.

The product is not just ahead of where the market needs it. It is ahead of every competitor. Broadcom is shipping T6 six to nine months ahead of NVIDIA’s competing solution and more than twelve months ahead of Marvell and Cisco. In a market where the hyperscalers are building infrastructure on multiyear timelines and cannot afford to wait for supply, being twelve months ahead is not only a competitive advantage, its a market advantage.

The $45 Billion Number and Why It Is Probably Conservative

Broadcom’s AI networking revenues grow more than 3x this year. In FY27 they reach $45 billion, more than double FY26, representing approximately 28% of Broadcom’s total AI revenues. These are not aspirational targets. They are derived from a product roadmap that is already in volume production, customer relationships that have been in place for years, and a market that is growing at a 50%+ compound annual rate with no visible deceleration.

The $45 billion figure covers the full AI networking stack that Broadcom is ramping simultaneously. The scale out Tomahawk 6 switching platform is the headline, but running alongside it is the scale across Jericho 4 router platform 3nm, 200 gigabit SERDES and the scale up Tomahawk Ultra Ethernet networking platform. The company’s 1.6 terabit 3nm PAM4 DSPs complete the portfolio. These are not separate product lines competing for the same dollars. They address different layers of the same AI cluster architecture, and the hyperscalers buying them need all of it.

The reason the $45 billion estimate may be conservative is the same reason Broadcom’s numbers have consistently surprised to the upside, the AI cluster buildout is still in its early stages, and the move from 800 gigabit to 1.6 terabit optical speeds inside cloud datacenters is just beginning. Every hyperscaler currently running Tomahawk 5 infrastructure will need to upgrade. The installed base is enormous. The upgrade cycle is mandatory. Broadcom is the only company with the product ready to supply it at scale.

The Barriers Are Higher Than They Look

70% market share in a market growing at 50%+ annually sounds like an obvious target for competition. It has been an obvious target for competition for a decade or so now. The share has not moved even a bit.

The reason is not pricing or customer relationships, though both matter. The reason is that Broadcom is always designing its next two generations simultaneously. While Tomahawk 6 is ramping in volume production, the team has already started sampling and qualification of the 2nm Tomahawk 7 platform, 400 gigabit per lane SERDES, 204 terabit switching throughput, 3.2 terabit optical connectivity, with sampling expected next year. A competitor trying to match Tomahawk 6 today would be shipping into a market where Broadcom’s customers are already beginning to qualify Tomahawk 7.

The compounding nature of this cadence creates a lead that widens with each generation. Every chip integrates capabilities developed in the prior two generations. The AI specific features in T6, the cognitive routing, the congestion management architecture designed for 1 million XPU clusters, are built on top of the foundation laid in T5. A new entrant cannot buy its way into this position. The institutional knowledge, the customer qualification cycles, the SERDES IP, and the 16 years of unbroken execution are not replicable on any timeline that matters to the hyperscalers making infrastructure decisions today.

Very few competitors have the R&D scale and IP to match this cadence. None have demonstrated the ability to execute it consistently.

Also, the framing that misses what Broadcom has built is the framing that treats it as a semi company competing for design wins.

AVGO is not competing for design wins in the traditional sense. It is providing the infrastructure layer that makes AI compute scale. The largest AI clusters on earth at Google, Amazon, Meta, Alibaba, Oracle, ByteDance, Tencent, and now Anthropic and OpenAI, ALL run on Broadcom’s networking silicon. These are not clusters that can be rebuilt on a different networking architecture between product generations. The switching fabric is the nervous system of the cluster. Replacing it requires rebuilding the cluster.

That is what 70% market share actually means in this context. It does not mean Broadcom won 70% of a procurement competition. It means Broadcom’s architecture is embedded in the infrastructure that the AI economy runs on, and the upgrade path from that infrastructure runs through Broadcom’s next product generation by design.

The AI infrastructure supercycle running at 12% annualized business capex gains globally, with hyperscaler spending accelerating through every recent earnings report is not a headwind to Broadcom’s position. It is the engine of it. Every dollar spent building a new AI cluster is a dollar that flows, in part, through Broadcom’s networking silicon. Every cluster upgrade from 800 gigabit to 1.6 terabit optical speeds runs through Tomahawk 6. Every cluster built to support 1 million XPU scale runs through the Ultra Ethernet Consortium standards that Tomahawk 6 is already incorporating.

The company is not riding the AI wave. It is part of the substrate the wave runs on which is perhaps the biggest difference that Wall Street has not factored in yet frankly.

My view is that Broadcom is one of the few companies in the semi ecosystem where the competitive moat is only getting better and bigger as the AI infrastructure buildout accelerates. The 2-year cadence is intact. The customer base is expanding, OpenAI and Anthropic joining the existing hyperscaler roster with T6 next year. The product lead over competitors is measured in months that translate into years of installed base advantage. And the next generation is already in design.

AI networking revenues reaching $45 billion in FY27 more than double FY26, growing at 50%+ annually through the foreseeable future reflects a business that is structurally positioned at the center of the most important infrastructure buildout in a generation. The Tomahawk 7 roadmap means this is not a one cycle story.

Therefore, I think AVGO should be $800+ stock whereas of today it’s trading at $440, which is a complete steal. I covered it early this year when it was trading at $280 and posted the thesis on X, now this a revised version of the pervious thesis.

And one last thought for the street’s analysts, them modeling Broadcom as a chip company will keep getting the numbers wrong for the same reason they kept getting NVIDIA wrong, they are looking at the product when they should be looking at the position.

Disclosure: This report is for informational purposes only and does not constitute investment advice. Nick or WTS research is produced independently. Past performance is not a guide to future returns