It’s been four months since I published my deep dive on , and since then there have been quite a few developments in the photonics sector, from CPO delays and changes in technology to, in general, a lot of FUD.

For this reason, I think it’s a good idea to analyze AAOI’s Q2 2026 results to cut through the noise and focus on what the numbers tell us and, to a lesser extent, what management is saying.

With , the situation is the same as with many other companies: time is its best ally because we know that AI clusters require more optical bandwith and architectures as CPO could create an additional long term demand driver for high power lasers.

Before we focus on the results, I want to review everything that has happened over the last few months to provide some context for the sector and, in particular, for AAOI.

Table of Contents

1. Four months of Noise

2. What AAOI is becoming (Business model, vertical integration and ramp up)

3. CPO

4. Q2 earnings results analysis

5. Valuation

6. Opportunities

7. Risks

8. Conclusion

Language disclaimer: Spanish is my native language, and I write my articles in Spanish before translating them into English using tools such as DeepL or Claude. All research, analysis, investment views, charts, tables, and underlying ideas are my own. I may use AI tools to improve the translation and make the English read more naturally

1. Four months of Noise

When I first published the article on AAOI, the stock was trading at $102. The thesis was as follows: AI clusters are limited by the interconnect infrastructure between GPUs, not by the GPUs themselves, and AAOI makes its own lasers in Texas at a time when the industry is struggling to secure enough laser capacity.

Since then, the stock reached $233.67 on May 13, fell below $90 in late July, and stands at $138.80 today. A full cycle in fifteen weeks. Most of that volatility had nothing to do with AAOI business.

In June, SemiAnalysis said CPO would be delayed, and the stock dropped 17%. A few days later, NVIDIA said there was no delay and everything was still on track.

Later, InP supply became a much bigger issue for the industry. This is why we saw signing a long term deal with to secure InP capacity through 2031.

For AAOI, this is good and bad: Tighter InP supply increases supply risk, but it also makes AAOI’s vertically integrated laser manufacturing more valuable.

In the earnings call management says it has enough substrate supply secured through the end of 2027.

The latest: The FCC is considering restrictions on Chinese optical transceivers used in U.S. data centers. AAOI rose 18% that same day.

Before the earnings report, we knew that pluggables were performing well, lasers were in short supply, and U.S. policy could boost domestic suppliers.

After reporting earnings, AAOI fell more than 7% in after hours trading but then rose 10% the following day.

The volatility makes sense. Overall, the results were good, but there are still a few things we need to understand and others where we simply have to trust management. And as we already know from the past, management tends to be very optimistic.

So we’ll focus on the numbers and see if management’s story actually matches what the business is showing. But first, let’s quickly go over what AAOI does and why investors have mixed feelings about the company.

2. What AAOI is becoming

It’s true that AAOI generated only $456M in revenue in 2025. With that in mind, an $11B market cap can look very expensive.

The AAOI thesis is all about the ramp.

The company is moving from being a relatively small supplier of 100G/400G transceivers and CATV equipment into a high volume manufacturer of 800G and 1.6T products, and eventually external laser sources for co-packaged optics.

2.1 The business model

I know this piece is about AAOI Q2 2026 results, but for anyone who is new to the stock, and even for those who aren’t, it’s worth quickly reviewing the business model.

Basically, AAOI sells the hardware that moves data between servers, switches, and GPUs inside and around data centers.

Its main data center products are optical transceivers. These are modules that convert electrical signals into optical signals, or light.

Why? Because chips and GPUs work with electrical signals, but when you need to move huge amounts of data quickly between racks, switches, or clusters, it is much more efficient to send that data as light through fiber.

As speeds and distances increase, copper becomes much harder and less efficient to use.

The light travels through the fiber and, once it reaches the other end, is converted back into an electrical signal.

2.1.2 The Second Business: CATV

AAOI second major business is CATV, Cable Television. This business is not really about television, but about the network infrastructure used by cable operators to deliver broadband internet and other services to homes and businesses.

Cable networks use a combination of fiber and coaxial cable.

As data travels through the coaxial part of the network, the signal weakens over distance, so operators need equipment such as amplifiers to strengthen it and keep the network running at high speeds.

AAOI sells these amplifiers, optical equipment and software to cable operators.

The industry is now upgrading toward DOCSIS 4.0, which allows cable companies to deliver much higher broadband speeds using much of their existing infrastructure. That upgrade cycle is creating demand for AAOI newer 1.8 GHz amplifiers and related software.

CATV is still a meaningful business for AAOI. It generated $80.6M in Q2, or 42% of total revenue, and management continues to expect more than $325M of annual revenue from the segment.

2.2 Why AAOI matters for AI Infraestructure

The more GPUs you put into an AI cluster, the more data they need to exchange with each other. As clusters become larger, the network connecting those GPUs also needs to become faster.

That is why the industry is moving toward higher speed optical connections:

• 100G = 100 GB/s

• 400G = 400 Gb/s

• 800G = 800 Gb/s

• 1.6T = 1.6 Tb/s

Larger AI clusters are increasingly driving demand for 800G and 1.6T transceivers, which is where AAOI current capacity ramp is focused.

The opportunity is simple: more optical connections inside larger AI clusters, combined with faster and higher value transceivers. AAOI can potentially grow revenue through both higher volumes and a richer product mix.

So basically AAOI has both a volume opportunity and a higher value product opportunity.

2.3 AAOI Vertical Integration Advantage

An optical transceiver contains a lot of components, including lasers. Many optical companies buy those lasers from third party suppliers.

AAOI manufactures a large part of its own InP laser dies in Texas.

The process starts with an InP substrate, which AAOI buys from external suppliers. AAOI then grows the epitaxial layers on top of that substrate, manufactures the laser die, and integrates it into its optical products.

This allow the Company:

1. More control over supply because it is less dependent on external laser suppliers.

2. Less exposure to laser shortages, especially important when industry capacity is tight.

3. More control over costs and margins, AAOI captures more of the value chain internally.

4. More control over capacity expansion, it can invest directly in increasing its own laser production.

Management has already confirmed that demand exceed capacity, so the laser shortage is real.

If demand for 800G, 1.6T and eventually CPO continue to exceed supply, owning and therefore control part of the laser manufacturing process become very valuable, resulting in a multiple expansion.

2.4 The capacity Ramp and the Q4 test

AAOI is expanding production capacity at an extraordinary pace.

Combined 800G + 1.6T capacity increased from nearly 100,000 modules per month at the end of Q1 to approaching 200,000 modules per month by the time of the Q2 earnings call.

Management still expects capacity to exceed 650,000 modules per month by the end of 2026 and 930,000 per month by the end of 2027.

Subscribe now

The original roadmap called for 148K/month of combined 800G + 1.6T capacity in Q2. In this earnings call, management said capacity was already approaching 200K/month, versus nearly 100K at the end of Q1. The next targets remain +650K by end 2026 and +930K by end 2027.

Later in the article, we will look at whether those targets are realistic. For now, the revenue ramp already looks like this:

• Q1 2026: $151M

• Q2 2026: $192M

• Q3 guidance: $255M-$290M

But this is where things get interesting.

Management still expects around $1.1B of revenue in 2026.

With $343M already generated in H1, that means Q4 would need to deliver roughly:

• $502M if Q3 comes in at the low end of guidance.

• $484M at the Q3 midpoint.

• $467M if Q3 comes in at the high end.

The problem? Even after a very strong Q3, Q4 would still require another massive sequential growth.

This is the bear thesis of AAOI, EXECUTION: Can the company turn its rapidly expanding capacity into shipped product and revenue quickly enough?

Before we dive into the Q2 numbers and management comments, let me talk a moment on CPO, because it is one of the biggest long term questions around the AAOI thesis:

3. CPO: the next stage

As we have already seen, today AAOI benefits from selling pluggable optical transceivers.

In the current architecture, the full transceiver module sits at the front panel of the switch. Inside that module are the lasers and optics that convert electrical signals into light, send the data through fiber and convert it back at the other end.

With CPO (Co-Packaged Optics), that architecture changes.

Instead of keeping the optics at the front of the switch, they move much closer to the switch ASIC, the main chip responsible for processing and routing data.

So the difference is:

• Today: the optics sit inside a removable transceiver at the front of the switch.

• With CPO: much of the optics move next to the ASIC itself.

How this change impact in AAOI? This creates both a risk and an opportunity.

In the near term, its growth story is still centered on the ramp of 800G and 1.6T pluggable transceivers.

But as CPO adoption grows, AAOI could eventually sell fewer complete transceiver modules. At the same time, CPO creates a new opportunity for high-power lasers and external laser sources such as ELSFPs

CPO could reduce AAOI content in the traditional transceiver, while increasing the importance of the part of the value chain where AAOI is vertically integrated: lasers.

3.1 Why CPO Could Make the Laser Bottleneck Worse

Thompson point is that CPO could make the laser bottleneck much worse, not better.

An 800G transceiver today uses roughly a 70mW laser, while a 1.6T transceiver uses around 100mW. For CPO, the external laser source needs to be closer to 300mW.

The laser die required for CPO is roughly six times larger, meaning fewer dies can be produced from the same InP wafer.

But higher power is only part of the problem.

In simple terms, CPO requires more powerful lasers, larger laser dies, and tighter manufacturing tolerances.

That is why Thompson believes the industry may eventually need 8 - 10x more laser manufacturing capacity to support CPO demand.

So the transition may reduce demand for some traditional pluggable transceivers, but it could dramatically increase demand for the high power lasers and external light sources that AAOI can manufacture internally.

AAOI originally showed a very aggressive ELSFP capacity ramp reaching roughly 400K units per month by late 2027. On the Q2 call, management comments appeared to push that level of capacity closer to 2028.

So CPO should probably be viewed as a second growth curve, not the main driver of the current 800G/1.6T ramp.

4. Q2 2026 earnings results

4.1 Q2 2026 P&L

4.1.1 Revenue Growth

Revenue has increased from $103M in Q2 2025 to $192M in Q2 2026, while Q3 guidance of $255 - 290M implies another +33 - 51% sequential increase.

4.1.2 Revenue by Product

Read more

I’ve published deep dives on Velo, LPKF, Ouster, PDF Solutions, AAOI, Sandisk, IREN, PENG… and the bottleneck map around photonics and memory, which is the one I’m most proud of, covering everything from substrates and materials all the way to the lasers.

A lot of the companies I’ve covered have run up quite a bit, and others are about to once order volume starts making headlines week after week.

Since many of you discovered me through those deep dives, I think now is the right time to show the portfolio behind the research.

I decided to write this short piece so you can get to know me better and understand how I invest and how I split my capital.

Let me start with how long I’ve been investing and what I’ve achieved:

I started investing in November 2023 with less than 3,000€, and my average monthly contributions since then have been around 500€ to 600€:

For this article, I tried to calculate my annual return as precisely as possible and be fully transparent about it. Over the coming months, you will get to know my method, my mistakes, and how I think about investing. Artificially inflating my annualized return makes no sense.

So without further preamble: my annualized return has been 90%. Some will think that is low. Others will think it is too high. The truth is that luck has been a real differentiator.

My first two investments were Verve and AMD. AMD stayed basically flat, and was even negative at times, until early 2026, when it went 4x in less than 12 months.

A few months after I started investing, I found the company that changed my life: Palantir. It took me a while to understand the business model and what Palantir actually did. I have to admit it was Antonio Linares and Jack Prescott who helped me clear up my doubts, and whose videos were crucial for what I consider the single most important thing in investing: CONVICTION.

Palantir went 10x in 18 months and, together with a few other companies I will discuss below, became one of the main drivers behind my 90% CAGR.

I expect that figure to come down over time. I doubt I will do another 10x in 18 months, and if I do, it will probably be with a stock that is not a major position in my portfolio, meaning the impact on my overall returns would be much smaller.

Back then Palantir was a meaningful part of my portfolio, but I made the mistake of not adding to the position because of the bias against raising my average cost.

If you want, I can write another article walking through my entire journey as an investor: the companies I’ve bought, sold, and held, the mistakes and the lessons.

But for now, to keep this from running too long, I’m going to share my portfolio and explain why I invested in the companies I hold, as well as how I split my portfolio between long term and short term.

Four months ago I started my short term portfolio to ride stocks with momentum.

The idea is to allocate 1% to 3% of the total portfolio to each position. If I see that a company is executing well and could become a compounder, I move it into my long-term portfolio, adding to the position as it goes up (I’ll admit it’s hard to overcome the bias against raising your average cost on a stock).

The idea behind starting the short term portfolio was to maximize gains.

Let me explain: selling Palantir at $190 in December 2025 was mainly because I considered the valuation to be too high.

I’m not going to lie: I did not know that 2026 would turn into such a bad year for software companies ( SaaSpocalypse)

I still had other software names in my portfolio, such as TEM and LMND. I sold them for a small profit after seeing the weakness across the software sector and rotated the capital into the companies I mention below.

After those moves, my portfolio became much more concentrated around the themes where I see the biggest upside over the next few years.

This is my long term portfolio:

My goal is to reach $1,000,000 before 2031.

For paid subscribers, I’m opening the portfolio, showing every position I own, how much I have allocated to each one, and the thesis behind every bet.

Read more

1. Introduction

We all know Micron’s results were spectacular, but what I liked most were the hints it gave us about winners and losers. In other words, it gave us a map to make better investment decisions heading into 2027 and 2028.

The data center revenue is now the engine. Core Data Center revenue was $11.5 billion, while broader data center revenue topped $25 billion including Cloud Memory, an annualized run-rate above $100 billion. Data center SSD revenue alone passed $5 billion, more than doubling sequentially.

This no longer looks like the old cyclical memory story tied only to PCs, smartphones, and inventory resets. Micron is increasingly becoming an AI data center supplier: HBM, high-capacity server DRAM, SOCAMM, enterprise SSDs, NAND, and context storage.

But for me, the most important part of the quarter was not revenue, but supply.

DRAM and NAND demand still sits far above supply, and tight conditions are expected to persist beyond calendar 2027. Even as supply improves gradually in 2028, Micron still has no clear line of sight to when supply catches demand.

This “Issue” flows through the entire memory supply chain: wafer fab equipment, EUV, materials, substrates, advanced packaging, bonding, test, enterprise storage, CXL, KV-cache infrastructure, and memory pooling.

In this article, I will analyze who collects the memory tax and who pays it.

The free section covers the framework: what Micron’s results revealed, why supply stays constrained, why HBM4 raises the intensity of the entire chain, and why NAND is quietly becoming strategic too.

For paid subscribers, I go one layer deeper and map the full chain, separating the companies that benefit from memory scarcity from the ones that get squeezed by it.

Micron told us told us where pricing power is moving.

That is the map worth holding for 2027 and 2028.

2. Micron Results

Micron results have been truly exceptional, driven by brutal pricing power and supply that is unable to satisfy demand, as reflected in take-or-pay contracts and capex that will not fully address this demand until 2028, when new fabs come online.

Micron reported $41.46bn in revenue in Q3 2026, up 74% QoQ and 346% YoY. As if that were not enough, gross margin came in at 84.9%.

Guidance for Q4 2026 is $50bn in revenue and an 86% gross margin.

2.2 Data center is now the engine

Core Data Center alone was $11.5bn, but Micron’s broader data center revenue exceeded $25bn when including Cloud Memory, an annualized run rate above $100 billion.

Data center SSD revenue alone exceeded $5 billion, more than doubling sequentially.

This rewires the Micron narrative. The old Micron was a memory company levered to PCs, smartphones, and inventory cycles. This Micron is driven by AI data center: HBM, high-capacity DRAM, LPDRAM and SOCAMM for servers, enterprise SSDs, and storage for context memory.

2.3 The most important thing: The earnings call

SUPPLY.

Management said HBM demand, across HBM3E, HBM4, and even ahead of HBM4E qualifications, sits far above what Micron can supply, not just in 2027 but into 2028.

Micron often has to talk customers down, explaining that the volume on offer is all it can commit, less than customers want to sign for. When your customers are trying to commit to take-or-pay for more product than you will sell them, you are not in a normal market.

Then the CFO dropped the number that reframes the debate: Micron previously expected the HBM TAM to cross $100 billion in 2028.

It now sees that crossing happening in 2027. The HBM demand curve pulled forward a full year. That does not just affect HBM makers. It cascades to advanced packaging, bonding, test, substrates, WFE, CXL, and inference architecture.

2.4 How Micron Expect to solve the supply constraint

Micron expects DRAM and NAND to stay tight beyond calendar 2027. Even as supply improves gradually in 2028, the company says it has no line of sight to when supply catches demand. The causes are threefold:

1. Each new node is more complex and yields less bit growth per transition.

2. HBM consumes more wafers and more silicon per bit, and its rising trade ratio with each generation pressures non-HBM capacity.

3. The only real way to add incremental capacity is greenfield fabs, which take years: construction, permitting, power, skilled labor, tool installation, and yield ramp.

Greenfield capacity starts contributing meaningfully to bits in calendar 2028, and even with that improvement, management still does not see an intercept of supply with demand.

Micron is spending enormously to add supply, and that supply still arrives late.

2.5 HBM4 raises the intensity of the whole chain

Micron said that HBM4 12-high is ramping twice as fast as HBM3E 12-high, that it has already shipped over $1 billion in HBM4 revenue, and that it expects mature yields significantly faster than the prior generation.

HBM4 is not simply more memory. It means:

• More dies per stack

• More bonding

• More test

• More yield risk

• More packaging intensity

HBM4 raises the intensity of the entire supply chain. That benefits the memory makers, but also the suppliers that enable fabricating, packaging, inspecting, and testing HBM at scale.

2.6 NAND is quietly going strategic too

The market fixates on HBM, but the call made clear that NAND and enterprise SSDs are tight as well. Both DRAM and NAND are very constrained, though customer urgency is higher in DRAM.

AI does not only consume GPUs and HBM, it consumes context storage.

Agentic AI means more context, more working memory, more retrieval, more fast storage, more data movement between tiers.

Agentic AI is extending infrastructure beyond accelerator racks to include CPU racks for the control plane and storage racks for a rapidly expanding context store. That is bullish for enterprise SSDs, high-end NAND, controllers, storage infrastructure, and indirectly for memory-tiering, CXL, and KV-cache architectures.

3. What it means for the market

AI capex is not slowing, it is shifting toward new bottlenecks. For two years the debate was GPUs. The chokepoint is now moving toward memory, packaging, power, networking, and storage.

Micron validates three things:

1. AI infrastructure demand is still very strong, because hyperscalers signing take-or-pay agreements through 2030 with enormous cash deposits are not hyperscalers pausing capex.

2. Memory is moving from commodity cyclical toward strategic asset, which does not kill the cycle but makes the downside potentially less brutal than prior ones.

3. Memory inflation creates winners and losers inside the AI trade.

So far, the free section has covered the main framework: Micron’s results, what the earnings call revealed, and why memory is becoming one of the key bottlenecks of the AI infrastructure stack.

For paid subscribers, I will now go one layer deeper.

The goal is not just to say “memory is bullish.”

The goal is to map the full supply chain and separate the companies that collect the memory tax from the companies that pay it.

That means looking at the real winners and losers across HBM, NAND, WFE, advanced packaging, bonding, test, substrates, enterprise storage, CXL, KV-cache infrastructure, neoclouds and hyperscalers.

Who benefits from memory becoming scarce, strategic and expensive and who gets squeezed by it?

Layer 1: The cleanest Winner

Read more

Disclaimer: I have sold $SIVE after a +820% return.

Before attacking me or getting defensive, I only ask one thing: please read how I currently see SIVE as an early investor, both in terms of valuation and based on my interpretation of the latest results.

1. Introduction

The AI buildout is breaking the optical network inside the data center. The cleanest pinch point is not the switch or the GPU. It is the material the lasers are grown on, and the lasers themselves.

IQE and Sivers sit on opposite ends of that supply chain. One just got rescued and has real revenue. The other has a spectacular pipeline and a balance sheet that runs on fumes.

Why this sector, why now

Before going into the results of both companies, we first need to understand the sector they operate in.

This is important because management’s guidance is not as aggressive as it may look at first glance. Once you understand the market backdrop, it may actually be conservative.

LightCounting estimates that sales of Ethernet optical transceivers and co-packaged optics for AI clusters reached 16.5 billion dollars in 2025 and will hit 26 billion in 2026, a 60% growth rate in both years.

The supply side is where it gets interesting for these two names. The same report flags that shortages of VCSELs and indium phosphide lasers constrained shipments between 2023 and 2025. The bottleneck is not in chip design. It is in the compound semiconductor material that the lasers are built from, and in the lasers themselves.

Then come Rosenblatt and put hard numbers on the indium phosphide piece in a 28 May 2026 industry update:

Their proprietary InP supply-and-demand model has the companies making the lasers, modulators and photodiodes for optical transceivers and co-packaged optics scaling capacity roughly 12x between 2025 and 2030, from about 1.9 billion dollars of laser revenue capacity to 22.75 billion.

That is a 64% capacity CAGR. The kicker: even after that 12x build, Rosenblatt’s base case has datacom demand still running about 50% behind supply exiting 2030.

Yes, After a 12x build, the market is still short. That is what a structural, non-cyclical bottleneck looks like, and it is the opposite of the “soft landing” some forecasters pencil in for 2027. (Treat Rosenblatt as a sell-side model, not a filing-verified fact, but the direction is a strong tailwind for the whole InP food chain.)

One more detail from Rosenblatt that frames the whole thesis: NVIDIA reportedly asked the supply chain to grow InP laser capacity about 20x from 2025 to 2030. The vendors committed to roughly 12x. The gap between what the buyer wants and what the chain will deliver is the bottleneck.

2. Business Model of Both

To be honest, Co-packaged optics as a standalone market is still tiny and the estimates are all over the map, from under 100 million dollars to 2 billion in 2025 depending on who you ask. That tells you CPO is still mostly narrative, not revenue. The money today is in 800G and 1.6T pluggables.

Both IQE and Sivers know this, and both have positioned to sell into pluggables now and CPO later.

Where they differ is the layer of the stack:

• IQE sells the epitaxial wafer. The atomic layers of compound semiconductor grown on a substrate, on which everyone else builds the laser or the chip. IQE is upstream. If InP laser capacity has to go 12x, somebody has to grow 12x the InP wafers underneath.

  

  Thanks for reading! This post is public so feel free to share it.

  Share

• Sivers sells the finished laser and the wireless chip. Closer to the end product,

  higher potential gross margin, but further from having any real data center revenue today.

  

That single difference drives almost everything that follows. And there is a detail that makes the pairing even sharper: the two companies are on the same supply chain.

IQE supplies InP wafers to Sivers (and IQE grows them on Aixtron MOCVD tools). So this is not two separate ways to play one theme. IQE sits upstream of Sivers. IQE sells the wafer; Sivers grows the laser on top of it, in Glasgow and at its Taiwanese foundry partner WIN. When you own IQE you own a slice of Sivers’ input costs, plus MACOM’s, plus everyone else’s.

That is the picks-and-shovels argument in one sentence. (Source: third-party industry research; the wafer-supply relationship is not separately disclosed in IQE’s filings, so treat it as well-sourced but unconfirmed by the company.)

3. Who is better option now?

Sivers gives you direct exposure to external light sources, CPO and optical engines. But after the move it has had, investors are already paying for a lot of success before the data center revenue has actually arrived. IQE sits further upstream, trades like an unloved industrial turnaround, and could benefit from the same InP capacity buildout without having to win the final module race.

This is the part where I break down the uncomfortable question: after the run both stocks have had, is Sivers still the best way to play the optical bottleneck, or has the better risk/reward quietly shifted upstream to IQE?

Everything below the line is the answer: the rescue that changed IQE’s balance sheet, the 6-inch InP angle the market is misreading, the dilution math on both, and where I’d actually put money.

The question is where the best risk/reward now sits: downstream in the exciting laser story, or upstream in the boring wafer supplier.

Subscribe to unlock the full deep dive: both companies broken down to the filing level, the dilution math on each, the pipeline sized against real revenue, the valuation on forward multiples, and where I’d actually put my money.

Read more

1. Introduction

AI infrastructure, energy, and photonics are still the sectors I believe will generate the highest returns this year, and probably in 2027 as well. But I also think the market is not pricing one of the biggest black swans of the decade: a Chinese invasion of Taiwan, or at the very least, a blockade.

In this article, I explain why 10% of my portfolio (I will increase till 30%) is allocated to two defense pure-plays that I believe are extremely well positioned for the era ahead.

We will go through the macro context, TAM, government spending trends, and finally dive into the two companies I like the most and believe still have huge upside.

Without further ado, let’s begin.

2. The macro setup

US margin debt provided by broker-dealers hit $1.304 trillion in April 2026, the highest level ever recorded. The figure surged +$83 billion in a single month. Over the trailing twelve months, margin debt is up +$453 billion, or +53%. As a percentage of US GDP, margin debt now sits at 5.2%, approximately 3 percentage points above both the pre-2008 Financial Crisis level and well above the 2000 Dot-Com Bubble peak.

The leverage is the signal. When investors are this leveraged into equity exposure, the market is at maximum reflexive fragility. Any drawdown (or worse, a black swan) gets amplified by margin calls. Any volatility spike gets amplified by forced selling. The asymmetry of returns from this leverage level has historically been negative.

That backdrop is creating a structural capital rotation. We have already seen that capital is leaving virtual-economy assets (software-only business models, speculative growth narratives, AI applications that have not yet converted to cash flow) and rotating into tangible-asset cash-flow generators that have multi-year revenue visibility independent of macro volatility. The pattern is consistent with previous regime shifts where structural inflation expectations, supply chain fragmentation, and geopolitical realignment drive capital toward physical-asset cash flow rather than valuation multiples.

2.1 Defense and national security

Defense and national security sit at the convergence of three of those drivers:

1. State-mandated demand

2. Multi-year fiscal commitment

3. Binding geopolitical necessity

Unlike historical event-driven spikes in defense valuations (Gulf War, post-9/11), the current rotation reflects a structural re-engineering of national fiscal priorities. The US Fiscal Year 2027 defense budget request is $1.5 trillion in total national defense budgetary resources, a 44% increase over the FY2026 baseline of approximately $1.06 trillion.

The dual-bill framework pairs a $1.15 trillion base discretionary budget with $350 billion in defense-specific reconciliation resourcing authorized under Public Law 119-21. Approximately $750 billion of the FY27 request is earmarked for capability development and weapon systems procurement, guaranteeing a multi-year revenue pipeline for hardware manufacturers.

The European side matches the US. In 2025, for the first time, all 32 NATO allies met or exceeded the 2% of GDP defense spending target first set in 2014. Combined European and Canadian defense spending rose approximately 20% in real terms to $574 billion. At the 2025 Hague Summit, allies committed to a new target of 5% of GDP annually to core defense and industrial security by 2035, with a mandatory floor of 3.5% of GDP for core military capability.

This is not a 12-month trade. This is a multi-decade (yes, until 2035/2040) structural reallocation, and it has not yet finished re-pricing inside public equities.

3. Where defense spending is concentrating

Defense spending in 2026 is concentrating in five major areas:

1. Autonomous and uncrewed systems
Military drones and counter-drone systems are becoming a massive spending priority. The US alone plans to spend more than $70B in this category, while global military drone spending is expected to more than double over the next decade.

2. Directed energy and high-energy laser weapons
The economics of warfare are changing. Destroying a $1,000 drone with a $1M missile is unsustainable. High-energy lasers can neutralize drones for roughly $10-20 per shot, which is why governments are aggressively accelerating procurement.

3. Subsea and undersea security
Submarine cables and underwater infrastructure have become strategic assets. Around 95% of global internet traffic flows through subsea cables, and recent incidents in the Baltic Sea and Red Sea have turned underwater surveillance, autonomous vehicles, and sonar systems into major NATO priorities.

4. Solid rocket motors and energetics
Missile production capacity and supply chains are now a bottleneck. Governments are investing heavily in suppliers and critical components to secure domestic production capabilities. (This is bullish for )

5. Software-defined autonomous combat platforms
This is the Anduril & Palantir thesis: AI-driven defense systems, autonomous warfare platforms, and integrated command-and-control software. The company doubled its valuation to $61B in less than a year and secured a $20B US Army contract, showing how aggressively defense is shifting toward software-defined warfare.

3.1 The TAM

We have seen how defense spending is set to keep increasing over the next decade.

But I did not want to stop there. I wanted to go one level deeper and identify which areas within defense are likely to benefit the most, so I could better select the companies best positioned for this cycle.

1. Undersea security and autonomous underwater systems: The UUV market is expected to grow from roughly $4-5B in 2025 to $7-9B by 2030. The highest-growth segment is extra-large underwater autonomous vehicles, driven by US, UK, Australian and European naval programs.

Seabed warfare is becoming a real procurement priority. Submarine cables, pipelines, ports, and underwater infrastructure are no longer passive assets. They are strategic targets.

2. Counter-drone and directed energy. The economics of warfare have changed. Using a $1M missile to destroy a $1,000 drone does not work at scale.

That is why governments are accelerating spending on counter-drone systems, remote weapon stations and high-energy laser weapons. The directed energy market alone is expected to grow from roughly $5B in 2026 to $15-20B by 2030.

3. Platform protection. One of the companies covered in the full report has already been selected for a major US Army vehicle protection program. Management believes the long-term opportunity could reach up to $3B over 15 years as the system is rolled out across domestic and export fleets.

Defense procurement moves slower, but once a supplier gets qualified, revenue can compound for years.

That is why I think these two small-cap defense pure-plays are worth a deep dive.

4. COMPANY 1

Read more

1. Introduction

I am writing this piece because of recent tweets from Serenity about $PENG.

Before anything else, let me say it directly: Serenity is right about the photonics hype. There has been a wave of investors framing Penguin as a silicon photonics play, and that framing is wrong. Penguin does not fabricate optical chips, does not make lasers, does not produce transceivers. If you want a pure-play photonics exposure, the answer is Marvell (post-Celestial acquisition), Coherent, or Lumentum. Not Penguin.

But here is what I think Serenity missed, and what I want to address in this piece: 90% of the investors who are long Penguin (myself included) did not initiate the position because of photonics. We initiated because of CXL, KV cache, and the memory wall. The photonics piece is a side story for us. The main story is something else entirely.

So this article is a deep dive on what Penguin Solutions actually is, what it actually does, why I bought it last week, and where the photonics question fits inside the larger picture. I will be as objective as I can given that I hold a position.

Let’s go.

2. The macro context: why inference broke the GPU story

For two years, the AI infrastructure conversation was a GPU conversation. NVIDIA earnings, Blackwell ramps, hyperscaler capex.

To train an AI model (LLM), you need a lot of processing power. Training is compute-bound. The GPU is the binding constraint. That part of the story has not changed.

But to use a model day-to-day, you run inference. And inference is memory-bound, not compute-bound.

Every time an LLM generates a response, the system processes text token by token. To keep the thread of the conversation and avoid recalculating the entire sequence on every interaction, the AI architecture temporarily stores the historical context data in an intermediate memory tier called the KV cache (Key-Value cache).

The KV cache is the model’s working memory. It grows with context length and with the number of concurrent users. In a macro scenario of exponential AI adoption, where context windows are expanding (people are now asking models to process entire documents, full financial statements, complex regulatory analyses) and where thousands of users query simultaneously, the KV cache grows proportionally.

2.1 Where the “memory wall” begins

GPU memory (HBM) is extraordinarily fast but extraordinarily limited. It is also extraordinarily expensive to scale. When the KV cache grows beyond what HBM can hold, the system either has to recompute tokens redundantly (which destroys latency and Time-to-First-Token), or has to offload to slower storage like NVMe (which destroys latency even harder), or has to buy more GPUs (which destroys the unit economics of inference).

The third option, “buy more GPUs,” is the default answer the market has been pricing for two years. But here is what the data is saying right now:

A paper by Gholami and others (2024, arXiv) quantified the divergence. AI model compute requirements grew about 410x every two years. Memory bandwidth scaled only 1.6x per two years. Interconnect bandwidth grew 1.4x per two years.

GPUs became much faster than the memory systems feeding them.

That creates a structural problem in inference. In inference, especially when the model responds token by token, the situation changes.

To generate each new word, the model needs to constantly access memory: the model’s weights, the previous context, the KV cache, etc. The problem is no longer that the GPU cannot calculate, but that the GPU does not receive the data fast enough.

Analogy:

You have a Ferrari capable of going 300 km/h, but it is stuck on a single-lane road full of traffic. The engine is extremely powerful, but it cannot use all that power because the flow of cars ahead is too slow.

That is what happens to the GPU during inference.

This is the core problem of inference. Modern GPUs have enormous compute capacity, but during token-by-token generation they often spend a lot of time waiting for data to arrive from memory. The bottleneck is no longer the ability to calculate. The bottleneck is the ability to feed the GPU fast enough.

This is the memory wall.

And here is the independent confirmation from the IDM side. Jeremy Werner, Micron’s Senior Vice President and General Manager of the Data Center Business Unit, said the following in a May 2026 interview on The Circuit Podcast:

Memory has become a key strategic asset for breaking bottlenecks in data center inference and is also the core support for training the world’s most advanced models. I don’t think this trend will slow down.

When the memory IDM (Micron) and the system integrator (Penguin) describe the same constraint from opposite ends of the value chain, the framing is not marketing positioning.

In addition, there is a popular thesis going around that cheaper tokens will reduce total AI infrastructure spend. The argument is that as models get more efficient and compute gets cheaper, total spending compresses.

That thesis is wrong.

Cheaper tokens do not reduce total compute spend. They unlock products that consume more tokens, longer context windows, and larger KV caches. AI demand is no longer limited only by FLOPs. It is limited by memory architecture. Lower per-token cost expands the use case set, which expands aggregate token consumption. The math goes the other way.

So the new question becomes:

How do you keep those GPUs fed with memory, context, and data without breaking the economics of inference?

That is the question Penguin Solutions is answering.

3. What is Penguin Solutions, actually

Penguin Solutions is not a new company. It is the combination of two heritage businesses that the market is just now starting to value correctly together:

1. SMART Modular Technologies is the memory engineering heritage. Nearly three decades of building precision memory modules for industrial, defense, networking, and now AI infrastructure customers. The DRAM module manufacturing know-how, the conformal coating IP for liquid immersion, the relationships with Samsung, Micron, SK Hynix as primary suppliers. This is the “memory engineering” half of the company.

2. Penguin Computing is the high-performance computing (HPC) heritage. Acquired in 2018. Penguin was one of the early commercializers of the Beowulf architecture (cluster-based HPC). They built supercomputers for governments and Tier 1 enterprises before “AI factory” was a marketing term. They have integrated more than 3.3 billion GPU runtime hours of operational data, which becomes important when we get to the software side.

In February 2025, the company rebranded the broader platform around what it now calls the AI factory architecture. New CEO Kash Shaikh (since early February 2026) reorganized the platform around six elements:

1. ICE ClusterWare: the cluster orchestration software (more on this later)

2. MemoryAI: the CXL KV cache server family (this is the new product line that drives the equity thesis)

3. Advanced Computing: the HPC and AI hardware systems (the heritage Penguin Computing business)

4. OriginAI: reference architectures and AI factory designs

5. End-to-end services: deployment, integration, ongoing operations

6. Partner ecosystem: NVIDIA Elite Partner, SK Telecom strategic anchor, Dell distribution, CDW channel

   

The SK Telecom anchor is critical context. On December 13, 2024, Penguin closed a $200 million strategic investment from SK Telecom via Astra AI Infra LLC. The investment took the form of 200,000 convertible preferred shares at $1,000 per share, with SKT receiving board representation rights, pro-rata rights, and registration rights. SK Telecom is a sister company to SK Hynix within the SK Group, which means Penguin has direct relationship depth into the most aggressive memory IDM in the world. At a moment when memory supply is the binding constraint on AI infrastructure, having that relationship matters more than the market has yet priced.

3.1 MemoryAI KV Cache Server: what Penguin actually launched at GTC

On March 16, 2026, at NVIDIA GTC in San Jose, Penguin announced the MemoryAI™ KV Cache Server. Per the company’s own press release:

The industry’s first production-ready KV cache server that utilizes CXL memory technology to address the critical ‘memory wall’ challenge in AI inferencing.

Let me explain what this actually means in plain terms.

Inside a modern AI server, you have a memory hierarchy:

• HBM (High Bandwidth Memory, attached to the GPU): extremely fast, extremely limited (80-192 GB per GPU), extremely expensive

• System DRAM (server motherboard memory): fast, more capacity (1-2 TB), more affordable

• NVMe SSD (server storage): slow, very high capacity, very cheap

The problem with current AI inference deployments is that the KV cache mostly has to live in HBM because everything else is too slow. Once HBM fills up, the system has bad options.

MemoryAI introduces a new tier: CXL-attached memory.

CXL (Compute Express Link) is a high-speed interconnect that lets a server attach and share additional memory outside the GPU. The MemoryAI KV Cache Server is a 4U appliance with up to 11 terabytes of memory (3 TB of DDR5 main memory + up to 8x 1 TB CXL Add-in Cards).

The performance numbers from Penguin’s launch materials:

• 10x faster than NVMe-based caching for KV cache workloads

• 3.8x faster than RDMA approaches. The practical alternative to CXL local memory is pulling KV cache from another server via InfiniBand or similar, which adds network latency that CXL eliminates.

• Native compatibility with NVIDIA Dynamo (NVIDIA’s software architecture for KV cache memory offloading, released alongside GTC)

The framing from Penguin’s own press release is the most important data point in the entire equity thesis:

Penguin Solutions itself is telling investors that inference workloads are 70% memory-driven and 30% compute-driven. The GPU is no longer the constraint. The memory architecture is.

CEO Kash Shaikh said it directly on the Q2 FY2026 earnings call:

Model training was largely compute bound. Inference, however, agentic AI is memory bound and latency sensitive. We believe this is driving a re-architecture of the data center across compute, memory, interconnect, and software.

CTO Phil Pokorny added on the launch:

CXL-enabled KV cache technology delivers faster time-to-first-token, reduced time per output token, and increased overall end-to-end token throughput. These critical performance improvements enable enterprise-scale inference across many users that expect low latency and timely access to AI-generated insights.

And here is what makes MemoryAI different from every other CXL play in the market:

It is the only production-ready CXL-based KV cache server actually shipping at scale:

• XConn Technologies and MemVerge demonstrated a joint CXL pool at SC25 in November 2025, but as a demonstration, not a shipping product.

• SK Hynix has showcased memory-centric AI architectures with CXL pooling but has not productized them into a turnkey server.

• Astera Labs ships Leo CXL Smart Memory Controllers, but those are silicon components, not an integrated system appliance.

• Marvell (post-Celestial acquisition) has photonic memory technology in development, but commercial product is 2027+.

Penguin shipped a 4U production server with 11 TB of CXL-attached memory, NVIDIA Dynamo compatibility, and customers already in deployment as of Q2 FY2026.

Two specific customer wins disclosed on the Q2 call:

1. A tier-one financial institution purchased CXL-powered KV cache servers for an on-premise AI factory deployment.

   

2. A generative AI company building inference workloads placed a “substantial order for CXL cards” for KV cache offload.

   

   I know you read “photonic memory appliance”. We are going to talk about it later.

Dollar values were not disclosed, but the customer categories are exactly the buyer profiles that justify the product. Financial services has regulatory pressure for on-premise inference. Generative AI companies have the technical sophistication to identify the bottleneck and the budget to solve it.

This is the product. Now the financials.

4. Financials

31.2% non-GAAP gross margin is structurally higher than any commodity AI hardware integrator. Supermicro, for comparison, reported gross margin of approximately 9.9% in its most recent results.

The gap is roughly 3x. That gap is what tells you Penguin is not a pure integrator. We will get to the moat that creates that gap in a moment.

One important one-time item in Q2: Penguin disclosed a $27.5 million GAAP gain (approximately $32 million cash) from the disposition of its equity stake in Celestial AI, the photonic memory company that was acquired by Marvell Technology in February 2026 for approximately $3.25 billion. Penguin was an early Series B and Series C investor in Celestial. The $27.5M gain is excluded from the non-GAAP numbers above, but the cash hit the balance sheet and contributed to the $489M cash position.

The headline -6% revenue decline obscures the segment-level story:

The Integrated Memory segment grew 63% YoY in a single quarter. That is the AI-driven memory demand thesis converting into actual revenue. CFO Nate Olmstead on the call:

On the memory outlook, we’re really pleased with the demand that we’re seeing as well as the favorability that we see in the pricing environment. For the increase that we’re seeing in the second half, that’s majority pricing, but demand is also very strong. AI-driven demand is just very strong. In fact, to get to the high end of that outlook really just refers to our ability to secure materials, which is really the only inhibitor.

That is the same allocation language Samsung, SK Hynix, and Micron used on their Q1 calendar 2026 earnings calls. The memory cycle is in allocation mode, not commodity mode.

The Advanced Computing decline of -42% YoY needs honest framing. It is driven by two specific items:

1. The ongoing Penguin Edge wind-down (legacy hardware business being deliberately exited)

2. The non-recurrence of a large hyperscale hardware sale from Q2 FY2025

Strip out those two effects and the underlying non-hyperscale AI/HPC business grew 50% YoY in the first half of FY2026. Seven new AI/HPC logos closed in H1 FY26 versus three in H1 FY25.

4.1 Guidance

The doubling of the revenue growth guide in a single quarter is the data point that signals management has visibility on the underlying memory demand strength and the MemoryAI customer pipeline conversion. Companies do not double growth guidance on speculation.

5. The photonics question.

What Serenity got right, and what changed in Q2

This is the section that directly addresses what Serenity raised. Before I started this piece, I would have framed it as a clean binary: PENG is not a photonics company, the Celestial position was sold, that exposure is monetized and done. End of story.

BUT then I went back and read the Q2 FY2026 earnings call transcript again. And the framing changed.

Let me walk through what the call actually said, because this is the part where retail investors who only read headlines are getting the story wrong.

What Serenity got right

The photonics-hype framing is wrong. PENG does not fabricate optical chips, does not make lasers, does not produce transceivers. Marvell (post-Celestial), Coherent, and Lumentum are the photonics pure-plays. If your thesis on PENG was “this is a silicon photonics manufacturer,” your thesis was wrong, and Serenity was right to push back on it (again, this was 0.1% of investors).

What changed in the Q2 call

CEO Kash Shaikh disclosed on the April 1, 2026 earnings call something that the market has not yet priced. Direct quote from the prepared remarks:

And on the Q&A, when asked directly how CXL and photonics interact:

Translate that out of management-speak:

1. PENG has an active product in development called PMA (Photonic Memory Appliance, formerly OMA). This is a real product, not a marketing concept. The naming change from OMA to PMA happened between the Q1 and Q2 calls.

2. The Celestial AI partnership continues with Marvell. PENG sold its equity stake in Celestial when Marvell acquired it ($27.5M gain in Q2). But the co-development partnership did not end. Marvell inherited it. PENG is now working with Marvell on the PMA.

3. The PMA is the “next level” version of the MemoryAI KV Cache Server. Same architecture (external memory appliance for AI workloads), different physical layer (photonic interconnect instead of CXL electrical). The product roadmap is sequential: CXL MemoryAI shipping today, PMA in active development for the next architecture wave.

4. The exposure is product-level, not just equity-level. The equity stake was monetized and is gone. The product partnership is active and ongoing.

Why CXL and photonics are sequential, not competitive

The instinct is to read “photonics is the next level” as “photonics is faster than CXL” and assume PMA replaces MemoryAI when it ships. That framing misses what the two technologies actually do.

The difference between CXL and photonics is not speed. It is distance.

CXL electrical works within a server. Copper interconnects, maximum physical reach of approximately 1.5 meters. That is enough to connect additional DRAM memory inside the same server box as the GPU. The MemoryAI KV Cache Server with 11 TB of CXL-attached memory operates inside this constraint: it sits next to the GPU servers in the same rack.

That solves the first bottleneck: memory capacity per server. The GPU runs out of HBM, the CXL appliance gives it 11 more terabytes within reach.

But AI data centers are evolving toward a different architecture. Multiple GPUs across multiple servers across multiple racks all sharing a unified memory pool. At 10+ meters of physical distance between racks, copper interconnects degrade signal and inject latency that destroys the benefit. Electrical CXL cannot reach that far.

Photonics can. Fiber optic and silicon photonics transceivers move data over tens to hundreds of meters without signal degradation. That solves a different bottleneck: memory pooling across the entire data center.

So the two products serve sequential problems:

• CXL MemoryAI today: solves “the GPU ran out of memory inside its own server”

• PMA tomorrow: solves “the entire data center needs to share one gigantic memory pool”

When CEO Kash Shaikh says photonics provides “increased capacity to share the memory,” the keyword is share, across distances that electrical interconnects cannot cover. He is not saying photonics is faster. He is saying it scales to distances and pool sizes that CXL physically cannot.

Both bottlenecks exist. CXL is being deployed against the first one today. Photonics will be deployed against the second one when AI architectures need it (industry consensus: 2027-2028 for early deployments, 2029+ for at-scale adoption).

PENG is positioned across both. That is the actual roadmap.

How to think about it now

PENG is not a photonics pure-play. Marvell is. If you want a pure photonics investment, buy Marvell, Coherent, or Lumentum.

But PENG is also not a “photonics is irrelevant to this company” play, which is what most retail commentary has settled on after the Celestial sale was reported. The Q2 call disclosed that PMA is in active development in partnership with Marvell. That partnership is the operational continuation of the Celestial relationship, not its termination.

The honest frame:

• Today’s revenue thesis is CXL. MemoryAI KV Cache Server, shipping to Tier 1 financial services and generative AI customers. This is what generates revenue in FY2026 and FY2027.

• The next-architecture optionality is photonics. PMA in active development with Marvell. If CXL adoption scales as expected through 2026-2027 and the memory wall persists into 2028+, photonic interconnects become the obvious next step for AI memory architectures. PENG is positioned to ship the appliance layer when that transition happens.

If you are long PENG because of CXL and the memory wall today, you are in the right thesis. The PMA is upside optionality, not the primary reason to buy.

If you are long PENG because you think it is a photonics manufacturer today, your thesis is wrong, and you are eventually going to be disappointed.

Serenity was right to push back on the photonics-pure-play framing. The framing creates expectations that the business cannot deliver in the short term. But the underlying photonics roadmap exists, is disclosed primary, and is more advanced than most investors realize. The PMA disclosure on the Q2 call is the part that the market has not yet priced.

That is the photonics question, honestly answered.

6. The moat

Why PENG is not a commoditized integrator

Now we get to the question that determines whether the 31.2% gross margin is sustainable or whether it compresses to integrator-level margins (10-15%) over time.

This is the most important section of the piece.

For comparison, a pure server integrator like Supermicro reports gross margins of approximately 9.9% in its most recent results. That is what commodity AI hardware integration looks like. If Penguin’s gross margin compresses to that level, the equity is dramatically overvalued.

If Penguin’s gross margin holds in the 28-32% range (where it sits today), the equity is correctly priced or undervalued.

At $2.35B market cap and approximately $2.3B enterprise value, PENG trades at roughly 1.7x forward EV/Sales on the raised FY2026 guide. For context, Supermicro trades at approximately 1.5-1.8x forward EV/Sales.

The market is currently pricing PENG at integrator multiples despite a 31.2% non-GAAP gross margin (vs Supermicro’s 9.9%), 24% adjusted EBITDA margin, +63% Integrated Memory growth, and the only production-ready CXL KV cache server in the market.

The question is why the market is already pricing it as an integrator despite the margin profile. The answer depends on what justifies the gross margin structurally. If the margins are real and sustainable, the equity has room to re-rate. If they compress to integrator levels over time, the current valuation is approximately correct.

Five concrete moat elements determine which scenario plays out:

Moat 1: MemoryAI patent stack and the Conformal Coating IP

Penguin does not just integrate cooling. Through SMART Modular, they design and manufacture memory modules with a patented Conformal Coating that allows DDR5 RDIMMs to operate in dielectric fluid immersion without corrosion.

When AI rack density exceeds 45 kW per rack (which is what Blackwell and Vera Rubin require), passive air cooling stops working. The industry transitions to direct-to-chip liquid cooling and increasingly to single-phase and two-phase immersion systems.

The challenge: memory modules immersed in dielectric fluid degrade and corrode over time due to chemical interaction with the fluid.

SMART Modular solves this by applying a patented Conformal Coating: an ultra-thin polymer layer sprayed with precision over the memory modules. The layer hermetically seals the electrical contacts and silicon against corrosion and oxidation without insulating thermal transfer from the chips. The modules can sit in dielectric fluid tanks for years without degradation.

No integrator can replicate this. It is materials science IP backed by manufacturing infrastructure.

Moat 2: ICE ClusterWare and AIM software

A large AI cluster is not a collection of independent computers. It functions as one giant machine. Without extremely advanced orchestration software, the GPUs in a data center hit constant bottlenecks and operate at maybe 50% of their theoretical value.

Penguin provides its own proprietary software: ICE ClusterWare (formerly Scyld ClusterWare) and the AIM module (Autonomous Infrastructure Management).

ICE ClusterWare automates the configuration of bare-metal AI and HPC systems, allowing thousands of servers to be set up, updated, monitored, and managed as one coordinated cluster.

AIM is the differentiator. It leverages 3.3 billion GPU runtime hours of telemetry data accumulated by Penguin to apply predictive maintenance algorithms. The software monitors thermal, electrical, and performance signals on each GPU node in real time. Using the runtime hours dataset, it detects early signs of failure and automatically isolates or remediates problematic components before they cause an interruption to training or inference workloads.

In plain terms: Penguin doesn’t just help customers build AI supercomputers. It helps keep them online. That is recurring software revenue that has no analog in pure integrator businesses.

Moat 3: SMART Modular memory engineering heritage

Nearly three decades of precision memory packaging. Long-standing relationships with Samsung, Micron, SK Hynix as primary suppliers. The CXL NV-CMM E3.S 2T module passed CXL Consortium certification in January 2026 and was added to the official CXL Integrators List, making Penguin one of very few interoperable silicon providers ready for mass deployments of in-memory databases.

This is preferred allocation access during a memory shortage. The same dynamic that gives Apple priority allocation from TSMC during a foundry shortage applies here. When the memory IDMs are choosing who to allocate, they prioritize the customers with longest history and deepest engineering relationships. Penguin sits in that tier.

Moat 4: NVIDIA Elite Partner + DGX-Ready Managed Services

Penguin is recognized as an NVIDIA Elite Partner and a DGX-Ready Managed Services Provider. The company has direct hardware-level integration relationships with NVIDIA for next-generation architectures.

Reference deployment: Penguin worked with Meta’s operations team on hardware integration and infrastructure setup for accelerated H100/H200 GPU clusters, coordinating Pure Storage storage system connectivity over low-latency InfiniBand networks.

You don’t get that level of co-engineering access by being a generic integrator.

Moat 5: SK Telecom alliance and the Haein supercomputer

The most important strategic alliance Penguin has formed is with SK Telecom and SK Hynix.

In December 2024, SK Telecom completed the $200M strategic investment in Penguin. At CES 2025, the partnership expanded into a three-way agreement with SK Hynix for research, development, and commercialization of next-generation AI Data Center (AIDC) infrastructure.

The first major demonstration of this synergy is the Haein supercomputer, one of South Korea’s most advanced sovereign AI infrastructures. Put into production in August 2025 in a record installation time of just two months. Key specifications:

• Over 1,000 NVIDIA Blackwell B200 GPUs integrated into the SK broadband Gasan data center

• 50+ miles of precision optical fiber cabling orchestrated

• Combined software stack: Penguin’s ICE ClusterWare + SK Telecom’s Petasus AI Cloud Kubernetes virtualization platform, enabling GPU-as-a-Service (GPUaaS) at market scale

The strategic cooperation with SK Hynix guarantees SMART Modular (Penguin) joint development and preferred allocation of advanced CXL memory and HBM silicon wafers, mitigating supply chain bottleneck risks at a moment when memory supply is the binding constraint on the entire AI infrastructure industry.

That is the moat. Five concrete elements, each defensible. The market is starting to price this. Penguin still has room to re-rate.

One more proof point: the strategic divestment

In March 2026, Penguin completed the sale of Zilia Technologies (its Brazilian commodity memory module subsidiary) to Lexar for $46.08 million.

This matters editorially. The new CEO Kash Shaikh’s leadership team is explicitly exiting the commodity hardware manufacturing business and concentrating capital into high-margin, high-value-added AI infrastructure solutions. The Zilia sale is the operational proof that the strategic repositioning is real, not just messaging.

A company that is becoming a pure integrator does not divest its commodity manufacturing arms. A company that is becoming a value-added platform does.

7. Management

This part is genuinely insane

I have to stop here for a moment because the board composition at Penguin Solutions is the part that almost made me triple my position size after I dug into it.

Most retail investors do not look at boards. They should, especially for AI infrastructure plays where the technical signal is buried inside who is willing to associate their reputation with the company.

On May 18, 2026, Penguin Solutions filed an 8-K announcing it had expanded its board to eight members and appointed:

1. David Heard, President of Network Infrastructure at Nokia ($NOK), as a Class I director and member of the Compensation Committee, with a term running to the 2029 annual meeting.

2. Penny Herscher - Chair of the Board (since 2022). Penny is concurrently Chair of the Board at Lumentum Holdings ($LITE), one of the largest silicon photonics and optical component manufacturers in the world. She has been a director at Penguin since September 2021. Prior board service includes Rambus (2006-2018), Cadence Design Systems (EVP and CMO), and CEO roles at FirstRain and Simplex Solutions. Cambridge mathematics. Two decades at the intersection of EDA, semiconductors, and photonics governance.

3. Mark Papermaster - Class I director (since August 2022). Concurrently the sitting Chief Technology Officer and EVP of Technology and Engineering at AMD ($AMD) since 2011. Before AMD: Silicon Engineering Group at Cisco, SVP of Devices Hardware Engineering for iPod and iPhone at Apple, and multiple roles in technology and server development at IBM. He is responsible at AMD for technical direction and product development including microprocessor design, I/O and memory, system-on-chip methodology, and advanced research. Member of the Global Semiconductor Alliance Board. Engineering background in microprocessor and memory design at the highest tier of the industry.

4. David Heard - Class I director (just appointed, May 18, 2026). Concurrently President of Network Infrastructure at Nokia ($NOK) since July 2025, where he sits on Nokia’s Group Leadership Team and reports directly to CEO Justin Hotard. He joined Nokia in February 2025 via Nokia’s $2.3 billion acquisition of Infinera, where he had served as CEO since June 2017. Before Infinera: JDS Uniphase, BigBand Networks, Somera Communications, Lucent Technologies, and AT&T. Four decades of senior leadership inside optical networking and photonics.

That Infinera deal is the part most retail investors miss. Infinera was the photonic chip company that vertically integrated Nokia from a reseller of someone else’s photonics into a company with its own optical chips. Nokia Optical Networks grew 20% year-over-year in Q1 2026 and management raised Network Infrastructure full-year guidance from 6-8% to 12-14% on the back of that integration. The man who built that integrated optical business at Nokia just joined the Penguin Solutions board.

5. Min Yong Ha - Class III director (since December 2024). SK Telecom’s Investor Designee on the board, appointed under the terms of the $200M SKT strategic investment. SK Telecom is the sister company of SK Hynix within the SK Group, direct relationship pipeline into the most aggressive memory IDM in the world.

6. Mary G. Puma - Class II director (since July 2023). Former President and CEO of Axcelis Technologies (semiconductor capital equipment, ion implant systems for chip manufacturing) from 2002 to 2023. 25+ years inside the semiconductor industry.

7. Sandeep Nayyar . Class II director (since 2014), Audit Committee Chair. Long-tenured finance leadership at the board level.

8. Bryan Ingram and Maximiliane Straub round out the eight independent directors with deep technology operating and finance backgrounds.

And on the operating side, Ian Colle (Senior Vice President and Chief Product Officer since March 2026, not a board member) came over from Amazon Web Services, where he was General Manager of Advanced Computing and Simulation, the team that built and scaled AWS’s HPC and AI infrastructure portfolio into a multi-billion-dollar business from the ground up. Before AWS: Red Hat and Intel.

Now look at the network this composition creates.

Penguin Solutions has on its board the sitting Chair of one of the largest public photonics companies (Lumentum), the sitting CTO of one of the two largest public chip designers (AMD), the President of Nokia’s optical networking division (just joined), and a direct SK Telecom representative (which means SK Hynix ecosystem access). On the operating side, the Chief Product Officer came from AWS’s HPC and AI infrastructure team.

This is not the board composition of a commoditized integrator. This is the board composition of a company that the technical leadership of the AI infrastructure industry believes will become structurally important to where the industry is going.

What board composition tells you is who is willing to put their reputation behind the company. When the people who run AMD’s silicon engineering, Nokia’s optical networking, and Lumentum’s governance are all willing to associate themselves with Penguin, that is a signal that the technical thesis is real to the people who know the industry best.

8. Risks

1. Penguin Edge wind-down headwind. Roughly 14 percentage points of FY2026 net sales growth is being absorbed by deliberate Edge business exit. The optics of declining Advanced Computing revenue can obscure the underlying AI/HPC business inflection through FY2026. There is a real risk that the market views Penguin as a slowing-growth company through the back half of FY2026 even as the underlying business is accelerating.

2. Bookings-to-revenue lag. Penguin’s AI/HPC project deployments run 12-18 months from prospecting to revenue recognition. Shaikh acknowledged on the Q2 call:

Most of the bookings that we are expecting may not materialize into revenue for the second half of this fiscal, but we believe that it will have a positive impact obviously going into the first half of the next fiscal.

That means FY2026 numbers reflect prior-period bookings, not the current inflection. The reward comes in FY2027 if execution holds.

3. Gross margin pressure. Memory is a lower-margin segment than Advanced Computing services. The mix shift toward memory and higher memory input costs both compress consolidated gross margin. The full-year FY2026 28% non-GAAP GM outlook is below prior-year trajectory. Whether MemoryAI margins are accretive over time depends on attach rates of higher-margin services and ClusterWare software.

4. Customer concentration. Top ten end customers accounted for 66% of net sales in 2025 (vs 58% in 2024 and 60% in 2023). Two individual customers exceeded 10% of net sales individually in 2025. The concentration is increasing, which is consistent with the AI infrastructure thesis but adds single-customer dependency risk.

5. Photonics adoption ahead of CXL. If silicon photonics commercializes faster than expected (Marvell ships Celestial-based photonic memory products in 2026 rather than 2027+), CXL-based memory tiering could be displaced before Penguin’s MemoryAI line achieves full revenue ramp. I assign low probability to this, but it is real and asymmetric.

9. What you are actually buying with

Let me close with the framework that I am operating from, since you have come this far.

You are not buying a photonics company. You are buying a CXL memory wall play with a strategic partnership anchor (SK Telecom + SK Hynix) and the only production-ready KV cache appliance shipping at scale in the market.

You are buying a company with:

• 31.2% non-GAAP gross margin (vs 9.9% commodity integrator average) sustained by five concrete moat elements

• +63% YoY memory segment growth with FY2026 guidance raised to +65-75%

• Net cash position ($489M cash vs $443M debt, no maturities until 2029)

• NVIDIA Elite Partner status and DGX-Ready Managed Services Provider designation

• SK Telecom $200M strategic investment with board representation, anchored into the most aggressive memory IDM ecosystem globally

• The Celestial AI $27.5M one-time windfall that strengthened the balance sheet

• A production-ready CXL KV Cache Server (MemoryAI) compatible with NVIDIA Dynamo, shipping to tier-1 financial services and generative AI customers

• An active Photonic Memory Appliance (PMA) in development with Marvell, as the next-architecture extension of MemoryAI when photonics matures

• Six core platform elements (ClusterWare, MemoryAI, Advanced Computing, OriginAI, services, partner ecosystem) under a focused new CEO

• A board composition that includes the Chair of Lumentum, the sitting CTO of AMD, the President of Nokia Network Infrastructure (former CEO of Infinera), and a direct SK Telecom representative — the technical leadership of the AI infrastructure industry has put its reputation behind this company

You are absorbing:

• A FY2026 headline growth optics drag from Penguin Edge wind-down (14pp)

• A bookings-to-revenue lag that pushes the visible inflection into FY2027

• Customer concentration (66% in top ten)

• Mix-shift gross margin pressure through FY2026

• Memory cycle correlation

• An asymmetric but low-probability risk that photonics scales faster than CXL

The framing question is whether you can hold a position through 6-12 months of headline optics that do not reflect the underlying business trajectory, in exchange for a position in the AI factory platform that I think wins the CXL memory tiering category for the next 3-5 years.

I made that bet last week.

That is what I am long $PENG.

If you want a photonics pure-play, Marvell or Coherent are the answer, not this. If you want the CXL memory wall infrastructure play with the only production-ready KV cache appliance in the market, a tier-1 strategic anchor with SK Hynix access, and a Photonic Memory Appliance in active co-development with Marvell as the next-architecture optionality, $PENG is the position.

Serenity was right that the photonics-pure-play framing is hype. PENG does not make optical chips, and pretending otherwise sets up the thesis to fail. But the underlying CXL memory wall thesis remains, the PMA roadmap is real and disclosed primary in the Q2 call, and that is the combined thesis I am invested in.

Share Nicolas

If you want the full 4-player Memory Wall map, I covered it in depth in Part 4 of The Bottleneck Map series:

Not investment advice. Disclosure: I own shares in $PENG, opened the position last week.

All claims trace to primary sources: Penguin Solutions 10-K, 10-Q, Q2 FY2026 earnings call transcript (April 1, 2026), Q2 FY2026 earnings release, NVIDIA GTC 2026 MemoryAI product launch announcement (March 16, 2026), Celestial AI / Marvell acquisition disclosure (February 2026), SK Telecom strategic investment disclosure (December 2024), Penguin Solutions 8-K filing for David Heard board appointment (May 18, 2026), Penguin Solutions DEF 14A Proxy Statement (FY2026 board composition), Nokia Group Leadership Team announcement (David Heard promotion July 1, 2025), and SEC filings from Vanguard, Invesco, FMR LLC, and William Blair.

The Bottleneck Map series, Part 4 of an ongoing analysis of AI infrastructure capacity constraints. Part 1 covered advanced packaging. Part 2 covered HBM and process equipment. Part 3 covered materials and substrates.

This piece moves the analysis inside the AI system itself, into the memory wall that AI inference is hitting.

1. Introduction

The market spent two years pricing AI infrastructure as a GPU story. The hyperscaler capex, the HBM ramp, the foundry orders… reduced to one question: how many accelerators can you buy? That question is increasingly incomplete.

A reasoning model does not just generate an answer. It stores context, maintains a KV cache, calls tools, retrieves documents, loops through multiple steps, and synthesizes outputs across long sessions. Every one of those actions increases the amount of data that must be stored, moved, and accessed without blowing up latency, cost, or power.

The old AI question was: how many GPUs can you buy? The new AI question is:

how do you keep those GPUs fed with memory, context, and data without breaking the economics of inference?

2. Thesis

AI model compute requirements grew about 410x every two years, while memory bandwidth scaled only 1.6x per two years and interconnect bandwidth 1.4x.

GPUs became much faster than the memory systems feeding them.

That creates a problem in inference. LLMs generate answers token by token. For each new token, the model must pull weights, context, and KV cache from memory. The GPU has plenty of compute available, but the amount of useful work per byte moved is very low.

In plain English: the GPU is ready to calculate, but it is waiting for data.

That is why inference latency is increasingly dominated by memory bandwidth, not raw FLOPs(*). Once memory becomes the binding constraint, “buy more GPUs” stops being the answer. You need to feed the GPUs better.

* Raw FLOPs are the theoretical compute power of a GPU: how many mathematical operations it can perform per second. But theoretical compute only matters if the memory system can feed the GPU with data fast enough.

The distinction between training and inference matters because they hit memory differently:

Training is compute-bound at scale: large batch sizes, weight gradients flowing in parallel, FLOPs utilization regularly above 50%.

Inference splits into two phases with very different profiles:

• Prefill: processing the input prompt in parallel, compute-bound, looks like training.

• Decode: generating output tokens one at a time, memory-bandwidth-bound, where the GPU spends most of its life waiting. Agentic AI, reasoning models, long-context workflows, and RAG systems all extend the decode phase relative to prefill. The longer the chain of tokens being generated, the more memory-bound the workload becomes.

  

2.1 The memory cycle is ratifying the thesis

The first ten days of May 2026 produced the most violent inflection in memory pricing in over a decade. Korea is the leading indicator of global memory pricing because Samsung and SK Hynix collectively export the majority of the world’s DRAM and NAND, which means Korea’s monthly export price data reads the cycle before US earnings catch up to it.

Per Korea Customs Service export data, in the first ten days of May 2026 versus the April average:

• DRAM (including modules) export prices: up 18% by weight

• DRAM (excluding modules): up 35%

• NAND flash memory: up 47%

• SSD exports: up 140%

The four charts look the same shape: a decade of flat-line pricing, then a vertical exponential move in the first week of May. That is the price data ratifying what Samsung, SK Hynix, SanDisk, and Micron said on their Q1 2026 earnings calls (24-30 April 2026): demand exceeds supply structurally, long-term agreements with binding financial commitments are locking up multi-year supply at fixed prices, and the memory market is exiting commodity behavior.

SK Hynix DRAM head Park Joon-deok told investors:

Demand for the next three years far exceeds supply and that the company was managing for balanced growth rather than maximizing revenue.

Samsung EVP Jaejune Kim said:

Demand fulfillment rate is at a record low” and that “customers are bringing 2027 demand into 2026.

Memory IDMs (Micron, SK Hynix, SanDisk, Samsung) have already re-rated on this thesis. The harder question is who captures the layer above the IDMs where the same scarcity translates into pricing power further down the AI data path.

That is what this piece is about.

Four names that sit one layer above the memory IDM, each capturing a different point in the chain of feeding the GPU.

3. Why inference changes the rules

Training was bursty. You build a cluster, run a training job for weeks, then stop. Inference is continuous. It runs twenty-four hours a day, scales with usage, and as AI products move from chatbots to agents, coding systems, deep research workflows, and enterprise automation, each user session consumes more context, more memory, and more data movement.

Important take: That is why cheaper tokens do not reduce total compute spend. They unlock products that consume more tokens, longer context windows, and larger KV caches. AI demand is no longer limited only by FLOPs. It is limited by memory architecture.

The confirmation came from the memory IDM side a month later. Jeremy Werner, Micron’s Senior Vice President:

Memory has become a key strategic asset for breaking bottlenecks in data center inference and is also the core support for training the world’s most advanced models. I don’t think this trend will slow down.

KV cache is one of the hidden bottlenecks in long-context inference:

When an LLM generates text, it does not start from zero every time it produces a new token. It stores the previous attention information in memory so the model can reuse it instead of recalculating the entire context again. That stored state is called the KV cache.

The problem is that the KV cache grows with context length and user concurrency. A short chatbot session may not break the system. But long-context agents, enterprise RAG, coding assistants, deep research workflows, and multi-step reasoning models can create very large KV caches that sit in expensive GPU memory.

That is where CXL comes in:

CXL, or Compute Express Link, is a high-speed interconnect that lets a server attach and share additional memory outside the GPU. Instead of forcing all inference state to live inside scarce HBM or falling back to slow NVMe storage, CXL creates an intermediate memory tier: larger than GPU memory, faster than storage, and more flexible than traditional server DRAM.

In plain English:

KV cache is the model’s working memory. CXL is a way to give that working memory more room without buying more GPUs.

4. Where the memory wall splits

The memory wall is not one bottleneck. It is four constraints stacked together, and each one has a different supplier, a different ramp curve, and a different time horizon:

1. Industrial AI memory modules and CXL-enabled expansion. DDR4 is not disappearing overnight, especially in industrial, defense, medical, and embedded systems where qualification cycles last years. At the same time, new AI designs are moving toward DDR5, higher-capacity modules, and CXL-enabled memory expansion. Legacy memory can stay tight for longer than expected, while next-generation memory becomes a new design-in opportunity. The companies that serve industrial, edge AI, and high-value vertical markets capture this layer.

2. KV cache and CXL memory expansion for inference. Long-context models and agents create large KV caches. The more steps an agent takes, the more memory state the system needs to hold. CXL changes the architecture by allowing memory to be expanded, pooled, or attached more flexibly across servers. Inference memory is becoming a system-level bottleneck. Whoever ships the integrated memory systems that solve this, captures the buildout.

3. Data movement and AI fabric silicon. Once memory becomes distributed across GPUs, CPUs, accelerators, SSDs, and CXL devices, the problem becomes moving data between those points quickly enough. PCIe Gen 6 retimers, CXL controllers, smart fabric switches, and high-radix scale-up fabric silicon determine whether the rack-scale AI system can scale efficiently. The AI cluster is no longer a box of GPUs. It is a distributed memory system, and the silicon that connects it is the constraint.

4. Enterprise storage and the AI data layer. AI systems do not run only on model weights and GPU memory. They need access to enterprise data: documents, logs, images, codebases, transaction histories, customer records, and retrieval databases. As companies move from pilots to production AI workflows, the storage layer becomes more strategic. Agents need fast access to clean data. RAG systems need retrieval. Long-running workflows need persistence. Multimodal models need more storage intensity. Enterprise all-flash storage that can keep up with hyperscaler-class workloads becomes a category-defining product.

   

5. Four companies, four layers of the same wall

Company 1 is exposed to the strange combination of legacy memory scarcity and next-generation CXL design-ins.

Company 2 is building around the KV cache problem, where long-context inference turns memory capacity into a system-level constraint.

Company 3 sits inside the rack-scale fabric layer, where GPUs, CPUs, memory, storage, and accelerators have to behave like one distributed system.

Company 4 owns part of the enterprise data layer, where production AI systems need fast access to the information agents actually use.

None of these companies are pure memory producers.

If the first-order winner is the memory manufacturer, the second-order winner is the company that makes scarce memory usable.

That is where this article goes next.

AI inference is becoming always-on, memory-heavy, and data-hungry. The companies that solve the memory wall capture the next layer of the AI infrastructure trade, the one above the IDMs.

Read more

The first piece in this series covered packaging, the bonders and substrates that turn GPUs and HBM stacks into working modules. That bottleneck gets all the attention because TSMC’s CoWoS capacity is the public story:

1. Introduction

The next bottleneck is invisible by design:

The materials that go into every step of every chip that gets made.

If packaging is the constraint on how many chips ship, materials are the constraint on whether those chips can be made at all:

Photoresists, ABF substrates, CMP slurries, specialty gases, filters, epitaxial wafers, none of these are sexy. None of them have brand recognition outside the semiconductor industry.

But every advanced logic and memory chip that comes off a TSMC, Samsung, or SK Hynix line passes through hundreds of process steps, and at every step a specialty material company is being paid.

The most under-priced moats in semiconductors live here:

The chemistries are protected by trade secrets and decades of customer qualification cycles. New product introductions take 24-36 months from sample to volume revenue. The customer concentration is brutal but on the supplier side, not on the customer side. Three or four global suppliers control most of the high-purity materials volume. When AI workloads accelerate node transitions and HBM stack heights, content per wafer goes up, and most of that incremental dollar flows through the materials supply chain.

2. Sector

There is a second macro driver running underneath all four companies that did not exist twelve months ago at this scale:

The migration of AI data center optical interconnect from pluggable transceivers toward Co-Packaged Optics (CPO).

Three of the four players in this piece sit upstream of the CPO ramp. CPO is not a 2030 story anymore.

Lumentum’s CEO described it on the most recent earnings call as “the largest single growth driver” with “a massive supply demand imbalance” and “very much in its infancy.” Himax management guided “mass-production readiness” in 2026 with volume ramp from 2027. Samsung Foundry committed to silicon photonics commercial ramp in H2 2026 (covered in the Memory Makers Cross-Call piece). The materials supply chain that feeds the CPO ramp is what this piece is about.

The Four Companies cover four different layers of the materials stack:

ABF substrates (Company 1), CMP slurries and process consumables (Company 2), compound semiconductor epitaxy for photonics (Company 3), and Photonics substrates for silicon photonics and Co-Packaged Optics (Company 4).

Each is exposed to the same macro driver: the structural shortage of advanced packaging and high-purity materials capacity but at different points in the value chain.

The TAM

The materials and chemistry layer of the semiconductor stack is harder to size than packaging because it spans hundreds of distinct sub-markets.

Three sub-segments matter for AI infrastructure investing:

• ABF substrates: per Mordor Intelligence, the advanced IC substrates market reached approximately $5.3B in 2025 and is forecast at $9.0-9.5B by 2030-2032 (CAGR 5.6-10.6% depending on source). The narrower “ABF substrates for server and HPC” sub-segment is forecast at $311M (2025) growing to $525M by 2032 at 7.71% CAGR per 360iResearch. The Mordor Intelligence advanced IC substrates 2026-2031 outlook reported lead times for 14-layer ABF substrates stretching to 28 weeks in early 2025 as AI demand outran supply by 18 percentage points. Average selling prices climbed from $65 in 2024 to $82 in 2025. Per the same source, hyperscale operators installed more than 1.2 million AI accelerator units in 2025, each integrating four to six flip-chip BGA substrates.

• CMP slurries and semiconductor process materials: a fragmented pool with multiple sub-segments. CMP slurries alone are roughly $2-2.5B globally per industry estimates, growing at high single digits. The broader “specialty chemicals and consumables for semiconductor manufacturing” pool, covering CMP, photoresists, specialty gases, filtration, and chemical handling, is multiples of that. Company 2 alone reported $3.2B in 2025 revenue across these segments.

• Compound semiconductor epitaxial wafers: per Yole Group’s 2024-2025 forecasts, the InP datacom die market reaches $1.85B by 2029. The GaAs VCSEL mobile and consumer die market is forecast at $996M by 2029. The MicroLED total panel revenue is forecast at $100M by 2029. The InP consumer wearable sensing die market reaches $1B by 2029. The total compound semiconductor opportunity sits in the multi-billion range, with AI photonics and 3D sensing as the fastest-growing applications.

The connecting thread across all three segments is node transitions and AI.

As advanced logic moves to 2nm, NAND moves from 250 to 300 layers, HBM moves from 12-high to 16-high, and chiplet architectures slice monolithic dies into smaller tiles, the content per wafer of materials goes up. More process steps. More chemistries. More substrates. More CMP planarization steps. More high-purity gases. More qualified suppliers.

The materials suppliers capture this as accretive content per wafer, they sell more material per unit of GPU produced, on top of unit growth.

3. Where the bottleneck splits

Four sub-bottlenecks make up the materials layer, each with a different leader and a different timing:

• Substrates, present generation. ABF (Ajinomoto Build-up Film) substrates have been the standard for FC-BGA packages used in CPUs, GPUs, and AI accelerators since the 1990s. AI server demand accelerated faster than substrate suppliers expected, creating real shortages. AMD entered into a $400 million venture with Siliconware Precision Industries in 2025 to lock up 30% of a new Taiwanese facility’s output, fabless designers are now co-investing in substrate plants. The bottleneck inside ABF is at the high-layer-count end: 14-layer and 16-layer substrates for advanced AI accelerators face the longest lead times.

• Process consumables, every node. CMP slurries, specialty gases, filtration, and contamination control are the “tax” the materials suppliers collect on every wafer that moves through a fab. Each new node transition (2nm logic, sub-300-layer NAND, HBM4) adds process steps and accretive content per wafer. The customer concentration is high: TSMC, Samsung, SK Hynix, Intel, Micron account for the majority of demand, but the supplier base is also concentrated, which protects pricing power.

• Compound semiconductor epitaxy, the InP frontier. As 800G and 1.6T optical transceivers become standard in AI data center buildouts, the demand for indium phosphide (InP) and gallium arsenide (GaAs) epitaxial wafers is inflecting. InP is the preferred material for high-speed lasers used in optical transceivers and, increasingly, Co-Packaged Optics (CPO). The customer base: Coherent, Lumentum, MACOM, Broadcom, and emerging silicon photonics startups is consolidating, but the supplier base for compound semiconductor epitaxy is small. Two or three companies globally have the technology and capacity to serve the AI photonics ramp.

• Photonics-SOI engineered substrates, the silicon photonics base. This is a sub-bottleneck that did not warrant a dedicated section twelve months ago. Silicon photonics chips, the photonic integrated circuits that route light inside CPO modules and high-speed optical transceivers, are not made on standard silicon wafers. They require specialized engineered substrates with ultra-precise top-silicon thickness uniformity (220nm with sub-nanometer tolerance) and buried oxide layers tuned to optical wavelengths. The dominant technology is Photonics-SOI (Silicon-on-Insulator engineered for photonic applications). Without these substrates, silicon photonics chips cannot be manufactured at the precision required for data center optical engines.

4. Inside the 2026 CPO Inflection

The first two articles were about making AI chips and HBM manufacturable. This section is about making AI data center connectivity manufacturable. CPO is the bridge between the semiconductor materials layer and the optical interconnect bottleneck.

4.1 CPO TAM

Goldman Sachs published the most aggressive scenario in April 2026, a chart projecting CPO contribution to AI optical interconnect TAM growing from $169 million in 2026 to $91 billion by 2028 under a 29% penetration rate in scale-up networks. This is not a typo.

Now you understand why everyone is so bullish on CPO. The base case is probably much lower but the direction is what matters:

CPO is moving from science project to real infrastructure bottleneck.

The Goldman scenario is far above consensus and should be read as a bull-case illustration, not as a base-case TAM. What all the forecasts agree on is the inflection:

2026 is the year CPO moves from sampling to early revenue, with volume ramp expected from 2027 onward. The InP and Photonics-SOI suppliers sit upstream of every dollar that flows through that ramp.

The CPO ramp map: where the supply chain stands in 2026. Reading the public commentary from the photonics ecosystem, 2026 is the year of supply chain validation, not mass production. SemiAnalysis estimates initial CPO unit shipments at approximately 10-15 thousand units in 2026, meaningful but a long way from the hundreds of thousands per year that hyperscaler buildouts will require by 2028.

Pluggable transceivers at 800G and 1.6T continue to dominate the volume. What is changing is that CPO orders are now placed, capacity is being expanded, and the supply chain is moving from demonstrations to first commercial deliveries.

The interesting part about the CPO ramp is not the optical module itself but the invisible materials layer sitting upstream of every module that gets built.

One company supplies the InP epitaxy used for the high-speed lasers.

Another supplies the Photonics-SOI engineered substrates that silicon photonics chips are built on.

Another supplies the CMP slurries, photoresists and contamination-control consumables flowing through every wafer at the leading foundries.

The 2026 ramp is still small in absolute units.But that is missing the point:

• 2026 is the validation year.

• 2027-2028 is when the materials revenue starts compounding.

The optical bottleneck is slowly becoming a materials bottleneck.

The investment logic mirrors packaging: own the right name at the right point in the curve.

• Company 1: for the substrate shortage that's binding now.

• Company 2 for the diversified consumables tax that compounds across nodes.

• Company 3 for the compound semiconductor optionality with strategic validation just landed.

• Company 4 for the base layer that every silicon photonics CPO chip is built on, currently trading through a cyclical trough that masks AI photonics segment growth.

They are picks-and-shovels businesses sitting underneath the physical scaling limits of AI.

And in many cases, the market is still valuing them like cyclical industrial suppliers instead of strategic infrastructure providers.

Let’s break them down one by one:

Read more

Part 1 of this series focused on the advanced packaging bottleneck.

The thesis was simple: AI capex is no longer just a GPU supply story because the constraint has moved downstream into the physical packaging stack: CoWoS capacity, HBM stacking, TC bonding, hybrid bonding and eventually glass substrates.

In Part 1, I broke down that bottleneck and analyzed 3 companies with significant upside potential, each positioned at a different point in the technology curve:

TC bonding today: Company 1
Hybrid bonding next: Company 2
Glass substrates at the frontier: Company 3

1. Introduction

Once you accept that HBM and advanced packaging are the bottleneck, the next question becomes more interesting:

What has to happen around HBM for the memory makers to actually scale production?

This is where Part 2 comes in.

This article is about the process-equipment layer that makes HBM and advanced memory manufacturable at scale.

And the latest calls from SK Hynix, Samsung, Micron and SanDisk all point in the same direction:

AI customers are no longer just buying components. They are reserving capacity. But reserved capacity is useless unless memory makers can turn it into high-yield supply.

2. Bottleneck Thesis

The hyperscaler earnings round of April 29, 2026 was a forward-looking validation of the entire HBM thesis.

Every major buyer of HBM-driven AI infrastructure raised 2026 capex guidance, and every one of them gave the same reason: components are scarce, and the bottleneck is not closing in 2026.

Combined, the five hyperscalers , , , and are now expected to spend hundreds of billions on AI infrastructure in 2026 alone.

The prior estimate already pointed to roughly $695-725B of 2026 capex. Morgan Stanley now raises the stakes even further, estimating hyperscaler capex approaching $805B in 2026 and $1.1T in 2027.

That is the first layer of the thesis.

The money is going into AI infrastructure at a scale the market still struggles to process.

But the second layer is more important:

Where does that capex actually land inside the supply chain?

That is where the memory calls give us information:

Samsung, SK Hynix, Micron and SanDisk all gave the same answer from different parts of the stack.

In other words, the hyperscalers tell us how much money is being spent.

The memory makers tell us where the bottlenecks are forming.

And the answer is clear: the next layer of upside is not just in the memory makers themselves. It is in the equipment suppliers that help turn capex into usable, high-yield capacity.

That is why this article moves one layer deeper than Part 1.

Part 1 covered the direct packaging bottleneck: TC bonding, hybrid bonding and glass substrates.

Part 2 focuses on the process-equipment layer behind the AI memory ramp: molding, CMP, cleaning, metrology and front-end process quality.

2.1 Memory Companies Hint:

The most important line came from SK Hynix.

Management said that HBM competitiveness is not only about DRAM process technology. It is also about TSV, packaging, yield, quality and supply stability. That is the whole reason this article exists. HBM is not just a memory die. It is a manufacturing stack:

SK Hynix also said that customers are prioritizing securing volume over pricing because supply remains constrained and memory has become increasingly critical in AI computing. In other words, customers are behaving differently. They are not trying to time a normal memory cycle. They are trying to make sure they get allocation.

Micron said the same thing from another angle. It increased FY2026 capex guidance to over $25 billion, with the increase still dominated by DRAM and HBM additions. Management also said meaningful new capacity from major cleanroom projects will not arrive until FY2028.

That is the key point:

The demand is immediate but the supply response will take years.

Samsung also confirmed that memory equipment spending should rise through the year as the company deploys cleanroom capacity already secured. It expects 2026 capex to increase substantially because of AI demand.

SanDisk does not validate HBM equipment directly, but it validates the broader regime. Customers are signing multi-year supply agreements, posting financial guarantees and trying to secure future capacity. SanDisk disclosed five multi-year partnerships, more than $11 billion of financial guarantees and about $42 billion of minimum contractual revenue from the first three agreements signed during the quarter.

The message is clear:

Capacity has become strategic.

3. What this means for the equipment layer

Memory makers who have just locked in multi-year demand commitments need to build production capacity that can actually deliver.

As we already saw in Samsung's Earnings call transcript, equipment orders are about to inflect. The memory makers have the demand certainty they needed to commit.

This is what makes the HBM equipment layer the trade:

A meaningful portion of incremental hyperscaler capex flows into the AI memory and advanced packaging stack. Every additional HBM stack requires the equipment to compress-mold it, planarize it, measure it, and anneal it.

Four companies sit at the structural choke points of that production.

Their FY2025 numbers reflect cyclical weakness from a 2024-2025 trough that is now visibly inflecting in their orderbooks (this is one of my best indicators).

The hyperscaler signals are the leading-edge confirmation, and the memory makers' own confirmations from last week are the supply-side proof point.

This piece covers those four names, where their FY2025 filings actually leave them, and why their forward setup looks fundamentally different from where the headline numbers suggest.

4. Why this layer is important?

Part 1 of this bottleneck covered the three sub-bottlenecks at the front of the HBM packaging stack:

- TC bonding (today)

- Hybrid bonding (next)

- Glass substrates (frontier).

Those are the technologies that join the dies together.

The HBM equipment layer covered here is the other half, the steps that prepare each layer of memory before stacking, measure alignment between layers, and protect the finished module after.

Disclaimer: The full HBM manufacturing process is much more complex and includes many more steps. For simplicity, this article focuses only on the selected equipment layers that are most relevant to the investment thesis: molding, CMP/cleaning, metrology, and high-pressure annealing.

Full HBM Manufacturing process:

5. What our 4 Companies do

• High-pressure hydrogen annealing (HPA). Before the packaging steps even begin, the wafer has to be made reliable enough to perform. Advanced chips have tiny defects at the transistor level. That is where high-pressure hydrogen annealing comes in. It uses hydrogen under pressure to repair these microscopic defects and improve transistor performance.If those defects are not repaired, performance, power efficiency and yield suffer. Below 16nm logic, which is essentially all advanced AI silicon, HPA is required, not optional.

• Chemical Mechanical Planarization (CMP). Before bonding can happen, the wafer surface has to be extremely flat and clean. Even tiny imperfections can hurt yield. That is where CMP comes in. CMP is basically the process of polishing the wafer with both mechanical force and chemical slurry until the surface is flat enough for advanced processing.

  As HBM stacks get taller and logic moves toward 2nm, wafers need more CMP steps and tighter surface control. Company 2 is interesting because it sells both CMP equipment and the slurry used in the process, giving it exposure to both tools and consumables.

• Overlay metrology. Every new layer of a wafer has to line up with the layer below it. If the alignment is off by even a few nanometers, yield and performance can suffer. That is where overlay metrology comes in. It measures whether each layer is placed exactly where it should be.

  As nodes move toward 2nm, HBM moves toward taller stacks, and hybrid bonding requires tighter alignment, overlay control becomes more critical.

  Company 3 is interesting because it is an alternative in a market historically dominated by global metrology leaders. It does not own the bonding step, but it helps make advanced bonding and high-density memory manufacturing possible at yield. Spoiler: the company supplies SK Hynix and Samsung.

• Compression molding. After memory dies are stacked, the assembly is encapsulated in resin to protect the chip from physical and thermal stress. As HBM stacks move from 8-high (HBM3) to 12-high (HBM3E) to 16-high (HBM4) to 20-high (HBM4E), the encapsulation process becomes more demanding. Company 1 dominates this step.

  

Each of these four steps has a different leader, different timing, and different exposure to the HBM ramp. Together, they capture the part of the equipment supply chain that doesn’t show up in mainstream “AI infrastructure” coverage.

So, in simple terms:

Compression molding: Company 1

CMP / cleaning / slurry: Company 2

Overlay metrology: Company 3

HPA: Company 4

Their upside has not fully played out yet. And as we saw at the beginning of the article, the HBM equipment layer is one of the clearest beneficiaries of rising AI memory capex.

The memory makers are already re-rated.

But the tools that allow them to turn capex into real, high-yield supply may be where the more asymmetric opportunities sit.

Read more

This is the first piece in a five-part series on where AI infrastructure capex is actually hitting walls and the under-the-radar names positioned to capture each one.

The premise of the series is that "AI infrastructure" is not one trade. It is at least five distinct bottlenecks (for now), each with its own physics, its own timing, and its own set of winners.

Most retail-focused commentary lumps these together and ends up either chasing the obvious large-cap names or buying generic "AI infrastructure" baskets. This series breaks the stack apart and identifies three to four under-the-radar players per bottleneck, names that genuinely capture the structural shortage rather than just trade on the narrative.

The five bottlenecks covered:

1. Packaging: TC bonding, hybrid bonding, glass substrates

2. Materials & Substrates: ABF, CMP, compound semiconductor epitaxy

3. The Memory Wall: CXL, disaggregated memory

4. Power & Thermal: liquid cooling, >100kW racks

5. Power Generation: behind-the-meter, gas turbines, nuclear, SMRs

Each piece stands alone and can be read independently, but the full picture is sequential: the further down the stack you go, the more under-priced the moats become and the further the names sit from mainstream coverage.

With that, let's get into the first one.

1. The Packaging Bottleneck

The first phase of AI capex was framed as a GPU supply story but that is no longer the full picture.

The constraint has moved one step downstream: advanced packaging. Making the GPU is no longer enough. Each Blackwell, Blackwell Ultra and Rubin module has to be packaged together with HBM memory at extremely high bandwidth.

The technology that does that is TSMC’s CoWoS (Chip-on-Wafer-on-Substrate). It’s the hard ceiling on how many modules actually ship.

Even with enough wafers, enough HBM and enough demand, shipments are limited by how much CoWoS capacity TSMC can provide.

UBS estimates Nvidia alone could demand 678,000 CoWoS wafers in 2026, around 40% above 2025. Morgan Stanley expects TSMC to reach 120,000 - 130,000 CoWoS wafers per month by end of 2026, up from 13,000 - 16,000 wafers per month in 2023. That is a massive capacity ramp and it still isn’t enough.

Apple, Qualcomm and Broadcom are reportedly struggling to secure slots. Apple has started recruiting engineers familiar with Intel’s EMIB technology (Embedded Multi-die Interconnect Bridge, Intel’s alternative to CoWoS that connects chiplets via small embedded silicon bridges instead of relying on a large interposer) as a hedge.

The packaging bottleneck has three sub-layers:

• HBM stacking* (TC bonding today, hybrid bonding next),

• The GPU-to-HBM interconnect (CoWoS-S, CoWoS-L, future glass interposers),

• The substrate underneath.

*HBM stacking is the process of building the memory tower next to the GPU. Today that tower is mostly assembled using thermocompression (TC) bonding. In the future, the industry may move toward hybrid bonding to make the connections denser, faster and more efficient.

This piece looks at where the value sits inside that bottleneck and identifies three under-the-radar names positioned to capture different parts of it.

2. The TAM

Per Yole Group’s Status of the Advanced Packaging Industry 2025 report, the advanced packaging market grew 19% year-over-year in 2024 to $46B and is projected to exceed $79.4B by 2030, a 9.4% CAGR through 2030.

The fastest-growing segment is telecom and infrastructure (AI accelerators, GPUs, data center), forecast at 14.9% CAGR and that segment is the one drawing essentially all the incremental capex.

Inside advanced packaging, Yole tracks the back-end equipment market that supplies the bonding tools. Total back-end equipment revenue is estimated at $6.9B in 2025, growing to $9.2B by 2030 (5.8% CAGR).

Within that, the two pieces that matter for HBM and AI accelerators have very different trajectories:

• TCB bonders: $542M (2025) → $936M (2030), 11.6% CAGR

• Hybrid bonders: $152M (2025) → $397M (2030), 21.1% CAGR

Yole’s framing matters: TCB and hybrid bonding are among the fastest-growing back-end equipment segments through 2030.

Most retail-focused commentary lumps these together as if they were the same trade. They aren’t. They have different leaders, different timing, and different customer adoption curves.

The HBM supply backdrop is structurally tight. Per industry coverage of the most recent earnings calls, HBM supply is fully allocated through 2026, including HBM3E. Samsung has signaled high-teens to low-20% HBM contract pricing increases for 2026. CoWoS is oversubscribed through at least 2027.

The April 29, 2026 hyperscaler earnings round confirmed this dynamic in the most direct way possible. Meta raised 2026 capex guidance from $115-135B to $125-145B, a $10B increase at both ends of the range. The reason Meta gave in its 8-K filing is the data point that matters most for this piece:

“This reflects our expectations for higher component pricing this year and, to a lesser extent, additional data center costs to support future year capacity.”

Read the framing carefully. Meta is not buying meaningfully more units. It is paying meaningfully more per unit. The same dynamic showed up across the hyperscaler cohort:

Total 2026 capex guidance for the top five hyperscalers is now tracking toward $700-750B combined, up roughly 60-67% year-over-year, with components HBM, GPUs, advanced packaging, driving most of the increase.

This is the bullish translation of the bottleneck. When supply is constrained and demand is inelastic, prices do the work that volume can’t.

The customers buying HBM are passing the cost through. The companies making the equipment that produces that HBM, TC bonders, hybrid bonders, advanced packaging tools, are capturing the operating leverage of fixed-cost businesses selling into rising-prices end markets.

That is the structural setup behind the names covered in this piece.

3. Where the bottleneck splits

Three sub-bottlenecks make up the packaging stack, each binding at a different point in the technology curve:

HBM stacking, present generation. Memory dies bonded together using TC bonders (Thermal Compression Bonders). HBM3E uses 8-high or 12-high stacks; HBM4 introduces 16-high. Each layer makes the process more thermally fragile and yield-sensitive. TC bonders use heat and pressure to solder the dies through micro-bumps.

HBM stacking, next generation. Above 16 layers, and especially above 20, TC bonding hits physical limits. Heat dissipation becomes a wall, micro-bump pitch becomes too tight, and signal integrity degrades.

Hybrid bonding replaces the bumps entirely with direct copper-to-copper bonds at sub-100nm precision. Per Player 2 2025 Annual Report, hybrid bonding versus the most advanced TCB delivers 15x interconnect density, 11.9x speed, 191x bandwidth density, more than 100x energy-efficient performance, and reduces HBM stack temperature by 20%. Critically, it also delivers 10x lower cost per interconnect once volumes scale.

Glass substrates, frontier. As organic ABF substrates run into physical limits and silicon interposers become expensive at larger package sizes, the industry is exploring glass interposers and glass-core substrates.

But glass requires a new equipment layer. The key process is TGV formation, or through-glass-via drilling, which creates the vertical connections through the glass.

This is where LPKF becomes relevant. Its LIDE technology is designed to process glass with high precision and without microcracks, making it one of the key equipment plays on the glass substrate transition.

Per Yole's 2025 advanced IC substrate report, the total advanced substrate market reaches $31B by 2030, and within that, glass-core substrates are moving from nascent to commercial through 2026-2030.

The TGV equipment market that enables this transition is much smaller (under $1B) but growing at 21% CAGR through 2031, which is where Player 3 competes.

The timing of the transition from TC bonding to hybrid bonding is not fixed.

TC bonding is not disappearing tomorrow. In fact, recent industry updates suggest it may have a longer runway than expected.

According to TrendForce and Korean media reports, JEDEC reportedly relaxed the maximum HBM package height requirement from 775μm to around 900μm in early 2026. That matters because it gives memory makers more physical room to keep using existing TC bonding processes for taller HBM stacks.

SK Hynix is a good example. The company evaluated fluxless bonding for 16-high HBM4 in Q4 2025, but ultimately decided to stay with MR-MUF, suggesting that hybrid bonding is still too early for mainstream 16-high HBM production.

Samsung, however, is already pushing the next step. At Nvidia GTC 2026, Samsung demonstrated hybrid bonding and claimed more than 20% improvement in thermal resistance versus TC bonding.

The takeaway is simple:

TC bonding still owns the current HBM ramp, but hybrid bonding is becoming the next-generation technology to watch.

The investment logic is not “pick a packaging stock.” It is “understand which technology is binding now and who captures that flow.”

The three companies covered here capture three time horizons:

1. Present (TC bonding): Player 1

2. Near future (hybrid bonding): Player 2

3. Frontier (glass / TGV): Player 3

Each is exposed to the same macro driver, the structural advanced packaging shortage, but at different points in the curve.

4. Player 1

Read more

Revenue down 32%. Adjusted EBIT loss widened. Cash position deteriorated. And yet, after reading the results and the call, I became more bullish.

Here’s why the call mattered more than the numbers. For the $LPKF full Deep Dive:

The headline

LPKF reported Q1 2026 this morning and the headline numbers are bad. Revenue of €17.1M, down 32.4% year-over-year. Adjusted EBIT of -€5.7M, widening from -€3.4M a year earlier. Free cash flow of -€7.6M versus zero a year ago. EPS diluted at -€0.30. Cash dropped from €7.0M at year-end to €5.8M. The current credit line drawdown of €4.2M moved the company from a +€7M net cash position to roughly -€1.5M net debt.

If you stop reading here, the print looks like a fresh disappointment in a stock that has just rallied 100% in a few weeks. That is the lazy read. It is also wrong.

The actual print, and especially the call, told a very different story.

Order intake +18% and book-to-bill 1.4

Order intake came in at €24.1M, up 18% versus Q1 2025’s €20.5M, with a book-to-bill ratio of x1.4. Backlog at €34.1M is down year-over-year (was €46.1M in Q1 2025), but that comparison is contaminated by the Solar collapse. Sequentially, backlog grew from €27M at year-end to €34.1M, which is the trajectory that matters for the FY26 guidance.

To hit the FY26 revenue midpoint of €112.5M, $LPKF needs roughly €100M of additional order conversion during the year. Q1 alone delivered roughly a quarter of that. If Q2-Q4 maintain even half the Q1 intensity, the FY guidance becomes very achievable on revenue.

The driver mix matters. Order intake came mostly from Development and Electronics, with strong rapid prototyping demand and SMT recovery. The CEO described “a very strong uptick in order entry in the SMT field in Q1”

Translation: the tariff-driven investment paralysis that hurt 2025 is fading. Customers who postponed projects from 2025 are now releasing CapEx and placing the orders.

Welding contributed a bulk order from a smart robotics customer, exactly the order Klaus Fiedler hinted at on the Q4 call when he described “humongous plans” from a frontrunner in the field. That order has now landed. It is not a one-off; it is the foundation for the Welding turnaround thesis.

The Solar segment intake stayed weak as guided. Klaus reiterated that 2026 is the weakest Solar investment year of the decade as the industry transitions from cadmium telluride to perovskites. He expects to sell first-of-a-kind perovskite factory equipment in 2027 and volume ramp in 2028.

The positioning phase is over

This is the single most important sentence Klaus said on the call, and it crystallizes the change in management posture.

For two years, Klaus has been describing LPKF’s role in glass substrate manufacturing as “qualifying with customers”, “being selected for prototyping”, “in positioning phase”. That language has now ended. The positioning phase is checked off. The next phase, winning production capacity orders, has officially started.

And on this call, for the first time, he gave a concrete data point to back it up:

This was not in the Q1 report. It was disclosed only on the call. A first capacity expansion order has been booked from a Korean customer in Q1 2026. This is the milestone the LIDE thesis was waiting for. It is one order from one customer at this stage, Klaus was careful to say “single-digit machines per customer” for these initial production orders rather than 50-100 machine orders that come later.

But it is the first concrete confirmation that customers have moved from “qualified, evaluating” to “buying production capacity”.

The most important part of the Q&A

The scale question:

Johannes Ries asked whether mass production would mean only a couple of machines per customer or something closer to 50-100 machines per customer. Klaus Fiedler’s answer was short: “In the high volume, yes.”

That matters. The first ramp-up orders are not supposed to be huge. Management is talking about single-digit machines per customer at the beginning. But if customers move into true high-volume production, the equipment requirement can scale to dozens of machines per customer. That is the operating leverage embedded in the LIDE thesis.

Klaus described where this lands on the timeline:

That language is meaningfully different from the Q4 2025 call, where the same point was framed as “I see a handful at the moment” of customers “ready for that now”. The shift from “ready to start” to “have started” happened in one quarter.

Segment breakdown: Solar is the villain, ex-Solar the story is positive

Here is the segment table for Q1 2026 from the report:

The full €8.2M revenue decline year-over-year is fully attributable to Solar (-€9.3M). Stripping out Solar entirely, the rest of the business actually grew slightly in Q1.

The full €2.3M Adjusted EBIT decline is also dominated by Solar (-€2.7M swing) partially offset by the Development turnaround (+€0.8M).

This is the same playbook as 2025: Solar is the structural drag the company is willing to accept while the perovskite transition matures. Klaus stated his target on the call: sell perovskite factory equipment in 2027, volume ramp in 2028. The €1.3M of Solar revenue in Q1 is not the new normal forever; it is the lowest point of the cycle before the perovskite restart.

Management says customers are no longer investing in cadmium telluride capacity, while perovskites are not yet ready for high-volume factory ramps. That creates an air pocket in 2026.

But it also creates optionality. LPKF is already participating in perovskite prototyping equipment and trial lines. If perovskite moves from pilot lines to factory-scale production in 2027-2028, Solar could shift from today’s drag to a second bottleneck-driven growth driver.

If that transition works, LPKF may have another picks-and-shovels angle.

Solar penetration keeps rising, but the next efficiency/cost step may require better technology. If perovskites move from pilot lines to factory-scale production, companies selling the equipment for that transition could benefit.

Higher solar penetration → more need for capacity, efficiency and lower-cost production → higher relevance of next-generation solar technologies like perovskite and tandem cells → potential demand for the equipment needed to manufacture them at scale.

The Electronics segment EBIT loss of -€2.8M (vs -€2.4M prior year) reflects continued investment in LIDE-related capability and absorbing the costs of the NeXaR product expansion. As LIDE volume orders convert to revenue in 2027, this segment becomes the operating leverage story. The CFO explicitly said this segment is “stable” with continued investment in the leading technology.

The Development segment turnaround to +€0.2M EBIT (from -€0.6M) is a real improvement, driven by deferred US government shutdown projects landing in Q1 2026 plus growing defense R&D demand. This segment is structurally underrated.

The Welding segment EBIT was flat at -€1.5M despite revenue growing 21%. Why no operating leverage? Because the smart robotics bulk order required upfront investment to secure the delivery and stabilize the production process. This is a one-quarter timing issue. As Welding revenue scales through 2026, this should swing positive.

The implicit TAM upgrade

In the Q&A, Johannes Ries from APUS Capital pushed Klaus on market sizing. Klaus’s answer is the second most important thing said on the call:

Translated: the existing €500M LIDE addressable market estimate is likely being reassessed, with an upward bias, because the underlying HPC wafer-start assumptions are moving higher.

Management did not give the new number yet, but the direction of travel sounds positive. This is not formal guidance, but it is a clear signal that the prior TAM framework may be too conservative.

And then Klaus added a separate, specific TAM figure for the new product lines:

So the Nexa-Ablate process step alone (RDL singulation) has a €100M TAM that is on top of the €500M LIDE TAM. Bonding is a separate market that Klaus refused to size yet (”we are really currently with the customers and their architectures in deep discussions”). CPO is “huge” but too early to size bottom-up.

The math here is that the 2030 LIDE-related TAM, including ablation, is already over €600M. With wafer starts moving up, the headline number is likely to be revised toward €700-800M in the next investor update. That is the TAM in which LPKF claims to have over 80% qualification share and a stated 70% market share target in production.

North Star is delivering

The cost-saving narrative held up in Q1. Personnel costs dropped from €14.8M to €14.4M (-€0.4M) reflecting the 35-person headcount reduction. Other operating expenses fell from €6.1M to €5.0M (-€1.1M), with R&D down €0.5M, opex down €0.4M, and temporary labor down €0.3M. Material cost ratio dropped from 34% to 26% on product mix.

The Fürth plant closure is “finalized”, with all personnel out by end of April. The footprint is now three production sites (Garbsen, Suhl, Naklo) instead of four. Welding production has moved to Suhl with the first systems already shipping from there.

Restructuring costs in Q1 were €1.2M (vs €0.5M in Q1 2025). This is consistent with the FY26 guidance of 3-4% of revenue in restructuring costs (which would be €3.5-4.8M for the year). The company is on track on cost reductions.

But the cash burn was the genuine soft spot. Operating cash flow was -€6.4M (vs -€2.8M Q1 2025). Free cash flow was -€7.6M. Cash dropped from €7.0M to €5.8M, with the company drawing €4.2M of additional credit lines. The net cash position swung from +€7M to roughly -€1.5M net debt.

The CFO’s framing: this includes one-time restructuring outflows, plus inventory build (+€3.5M to €23.2M) “in preparation of revenue ramp-up in Q2”. So part of the cash drain is timing, building production for orders that will ship in Q2 and Q3.

The bigger backstop: the syndicated loan facility was renegotiated in February to extend through December 2028, with €25M in cash facility (of which €12.5M still undrawn) and €7.5M in guarantee facility. The company has approximately €18M of total liquidity headroom (cash + undrawn credit) before any real stress.

This is comfortable for the trajectory implied by the guidance.

The CEO acknowledged the social media rally

In the CEO letter (page 4), Klaus did something unusual:

Management is explicitly telling shareholders that the rally is driven by social media attention, not by a change in fundamentals beyond what was already disclosed in March. This is a CEO trying to manage expectations downward not because the business is worse than expected, but because he does not want the stock to run too far ahead of the operational reality.

This is the opposite of what a hype-chasing management team does. It is a meaningful signal of conservative IR culture, exactly the kind of culture an activist like AOC would push for after two years of guidance misses. Klaus is no longer overselling. He is explicitly underselling.

So what have change?

Three things shifted in this quarter that did not exist before:

1. The first concrete capacity expansion order from a glass customer. Previous quarters had qualification orders, prototype orders, single-machine evaluation orders. Q1 2026 had the first commercial production capacity order from a Korean customer. This is the operational milestone that converts the LIDE bull thesis from “future opportunity” to “current revenue stream”.

2. A management posture that has shifted from “we hope” to “we now know”. Compare these two excerpts. Q4 2025 call (March 2026): “I see a handful of customers being ready for that now... We’re going to see several customers.” Q1 2026 call (today): “Four or five players are at the technical and yield level now to make their first ramp-up investments. That’s happening here and now.”

3. Explicit TAM upgrade signaling. Klaus pre-announced that the €500M LIDE TAM is going to be revised upward in coming weeks because wafer-start projections have moved up. He also gave the first concrete TAM number for ablation (€100M). The TAM keeps growing.

What did not change

The 2026 guidance is unchanged. Revenue €105-120M. Adjusted EBIT margin -3.0% to +4.5%. Restructuring costs 3-4% of revenue. Solar weak.

Klaus reiterated that “potential volume orders from advanced packaging are not included” in this guidance, meaning any production capacity orders that ramp through 2026 (including the Korean order) are pure upside. This is the same sandbagging posture from the Q4 print.

My read

I came into this print already long the thesis. I came out of it more confident, not less.

The numbers are bad on a year-over-year basis because Solar collapsed and because Q1 is seasonally the weakest quarter. The numbers are also bad on cash because of restructuring outflows and inventory build for Q2 deliveries. None of that is news.

What is news is the order book. €24.1M of intake, book-to-bill 1.4, an actual capacity expansion order in glass from a Korean customer that was not in the guidance. That is the first concrete data point that the LIDE thesis is converting from positioning to commercial.

What is also news is the management posture. The CEO acknowledged the social media rally and asked shareholders to stay grounded in the operational reality. He pre-signaled a TAM upgrade. He confirmed that LPKF will not raise equity to bridge cash. He stated that the positioning phase is over.

If you understand that this is a multi-year transition story where the first commercial production orders matter more than quarterly revenue, today was a good day.

The catalysts ahead remain visible. The next H1 report on July 23 will give the first full read on H1 order intake including the Korean capacity expansion order. The TAM update Klaus pre-signaled is likely in Q2. The shareholder meeting on June 4 is a venue for additional narrative. And the production-ramp purchase orders from “several customers in advanced talks” should land through the year.

I am bullish.

1. Introduction

Organic substrates have worked for a long time, but they have physical limits. Intel has already said that by the end of the decade the industry will likely reach limits in scaling transistors on silicon packages using organic materials, partly because of power, shrinkage and warping issues. Glass, according to Intel, offers better thermal and mechanical stability, higher interconnect density and better dimensional stability, making it a serious candidate for next-generation AI, data center and graphics packages

Intel just told the market that glass substrates are a gross margin headwind.

Translation: demand is outrunning supply across the chain. Here's what's happening in the glass equipment market, and a side-by-side of the two pure-play tooling names: LPKF and SCHMID Group.

One has a clean balance sheet, an activist accelerating the operational turnaround, and a CEO confirming "low triple-digit" revenue ambitions for the glass business. The other has a going concern qualification, a highly dilutive convertible financing, and insiders lending to their own company at 15% interest. Both are riding the same industry transition.

In this article we will compare the two pure-play names that I think matter most: LPKF Laser & Electronics SE (XETRA: LPK / OTC: LPKFF) and SCHMID Group N.V. (NASDAQ: SHMD).

They are not the same business. They are not even direct competitors in the same step. But they are both how a public-market investor gets exposure to this specific transition.

Two Pure-Play names

Most of the conversation about glass equipment focuses on a small number of public companies. $LPKF and are the two that come up most often in serious sector analysis, because they are the only two pure-plays where glass exposure is a meaningful part of the thesis. Onto Innovation, Coherent, IPG Photonics, ASMPT, Besi all have some adjacent exposure, but glass is a small fraction of their revenue and the read-through is diluted.

LPKF and SCHMID, by contrast, are both German micro-to-small caps where glass is a strategic priority and where the share price is sensitive to glass adoption signals. They are both worth understanding, but they are very different businesses with very different financial profiles. Investors who treat them as substitutes are making a mistake.

2. What each company actually does

The most common confusion in retail conversations about these two names is that they compete head-to-head. They do not. They sit at different points in the glass substrate manufacturing flow.

LPKF sells laser equipment. Specifically, their flagship technology is LIDE (Laser Induced Deep Etching). LIDE is a two-step process: a precisely-tuned pulsed laser modifies a column of glass without ablating material, creating a chemically-reactive damage line. A subsequent selective wet etch removes only the modified zone, leaving smooth-walled through-glass vias with high aspect ratios. The laser does the modification; the wet etch does the removal; LPKF sells the integrated tool.

This is the drilling step in the flow, the part where you create the holes that will eventually carry electrical signals through the glass. Full deep dive of $LPKF:

LPKF is a focused tooling company. Beyond LIDE, they have legacy businesses in PCB stencil cutting, depaneling, plastic welding, and laser scribing for thin-film solar cells. The non-glass businesses contributed roughly 87% of 2025 revenue. But strategically, LIDE is the asset that defines the bull case.

CEO Klaus Fiedler stated on the Q4 2025 earnings call that “over 80% of customers preparing for ramp-ups for glass and advanced packaging are now qualifying with LPKF equipment”.

His personal target is 70% market share in production-volume orders. And on the same call, when an analyst asked whether his prior expectation of “low triple-digit sales” for the LIDE business still held, Fiedler answered: “absolutely my expectation. First customers have now shared their volume demand profiles for the coming years.” Translation: LPKF management has visibility on €100M+ of LIDE revenue at the destination. They are not promising it for 2026 or 2027, but they are now anchoring it as their target run-rate based on actual customer-disclosed forecasts.

Important Update!

https://www.thelec.kr/news/articleView.html?idxno=55792

Fresh reporting from The Elec shows that LPKF is building a broader equipment portfolio around glass-based advanced packaging. LPKF Korea said the company has already supplied 3D optical waveguide formation equipment to a customer for co-packaged optics, after working for four years with a specific semiconductor company on commercializing its Direct Light technology.

That matters because Direct Light is not just another drilling tool. It uses a laser to form 3D optical waveguides inside glass without physically cutting the material and without the traditional etching step. LPKF also said it has secured equipment and technology for multilayer 3D waveguides on large 510×515mm glass panels, at a stage suitable for mass-production lines rather than just R&D.

The company is now grouping its glass packaging equipment under the NeXaR brand: Nexa-LIDE, Nexa-Direct Light, Nexa-Bond and Nexa-Ablate. That changes the debate. LPKF is still not a full turnkey line provider like SCHMID, but it is no longer just a single-tool LIDE story.

SCHMID Group plays a different role in the glass substrate value chain. While LPKF focuses on the critical drilling/structuring step, SCHMID is positioned more broadly across the production line.

Its equipment is used for wet processing, plasma processing and metallization, the steps needed after the glass has been structured. In simple terms, SCHMID helps turn a piece of processed glass into an actual substrate with conductive copper traces.

Its key technology is called ET, or Embedded Traces. Instead of using the traditional drill-and-plate approach, ET embeds copper traces directly into the insulating material. The goal is a cleaner, more precise and more scalable process for advanced PCBs and substrates.

SCHMID’s glass substrate exposure is not as direct as LPKF’s.

LPKF owns its own laser-based process for glass structuring. SCHMID’s role is different. Its glass strategy is partly linked to its ET process, which can be applied to glass core substrates, and partly linked to its partnership with TRUMPF.

TRUMPF brings the laser drilling capability. SCHMID brings the downstream process know-how: etching, wet processing and the steps needed after the holes have been created. Together, the two companies are trying to offer a combined process for glass interposers, where TRUMPF creates the through-glass vias and SCHMID supports the etching/process steps required to make the substrate usable. SCHMID says this approach can shorten process times by a factor of ten.

This is an important difference versus LPKF. LPKF is more directly exposed to the critical glass drilling/structuring step. SCHMID is more exposed to the broader process flow around substrate manufacturing.

That also explains the difference in tone. LPKF says more than 80% of customers preparing glass and advanced packaging ramp-ups are already qualifying with LPKF equipment, although this is not yet volume revenue.

SCHMID is more cautious: it describes glass substrate manufacturing as still focused on R&D, pilot lines and early OEM engagements, with volume adoption expected only in the mid- to long-term.

So the layering becomes clearer:

What this layout tells you: LPKF started with one critical step in the glass flow, but the story is now broader. LIDE remains the entry point, while Direct Light, glass bonding and RDL ablation expand LPKF into a multi-step glass advanced packaging platform.

SCHMID is a process equipment company with a broader portfolio across glass, PCB, PV and energy storage, where their position in glass depends on a partnership for the laser drilling step.

3. The Business Model Debate: Complete Line vs. Critical Step (The gap is narrowing)

A lot of investors frame the debate as simple: SCHMID sells the full production line, while LPKF only sells one machine. Therefore, SCHMID has the better model.

I think that is too simplistic.

SCHMID’s turnkey model (selling a “ready-to-run” solution) can be powerful. A customer buying a more complete line has lower integration risk, one point of accountability and potentially higher switching costs. If SCHMID’s ET process becomes a standard, the customer may stay locked into SCHMID’s ecosystem through service, upgrades, consumables and future expansions.

But LPKF’s more focused model also has advantages. In semiconductor manufacturing, the most valuable companies are often not the ones selling the full line. They are the ones that control one mission-critical step. ASML does not sell the whole fab. KLA does not sell the whole fab. Lam and Applied Materials do not sell the whole fab. They win because they are extremely strong in specific process steps.

That is the point with LPKF. If LIDE becomes the preferred solution for high-precision glass processing, LPKF does not need to own the full line. It only needs to be indispensable in the part of the process where failure is expensive.

So the debate is not “turnkey is better” or “specialty is better.” It depends on the customer:

Large semiconductor players usually prefer multi-vendor, best-of-breed flows. They do not want to be hostage to one supplier. Smaller OSATs or PCB manufacturers may prefer a more integrated solution because it reduces complexity.

This means both LPKF and SCHMID can win, but they are different bets.

SCHMID has the broader model if ET becomes a standard for glass core substrates. But that is still not proven. ET was originally developed for high-end PCB and organic substrate applications, and SCHMID itself frames glass substrate adoption as still early-stage.

LPKF is narrower, but further along in the actual glass qualification cycle.

In simple terms:

SCHMID has the broader line-level model.

LPKF has the more validated glass position and is expanding its share of wallet around that bottleneck.

We have covered why glass substrates could become one of the next real bottlenecks in AI packaging, what LPKF and SCHMID actually do, and why their business models are not the same.

Now comes the important part:

Which company is better positioned to capture the economics if this market scales?

I will go deeper into their customer signals, partnerships, competitive positioning, upside potential, risk/reward and balance sheet quality

I will also share my personal view on which company looks better positioned today not just technologically, but as an investment setup.

Read more

In this article, we will look at how a company that has been around for more than three decades, and was once valued like a low-multiple (3-4x sales) niche software business, could be turning into a much more important player: the data and orchestration layer connecting a huge and increasingly complex semiconductor supply chain.

Founded in 1991, PDF Solutions first grew as an analytics platform provider, but the company is evolving toward AI-driven collaboration and orchestration across the fabless (design house), the foundry (the wafer manufacturer), OSAT (assembly operation) and the test operation generate compatible data and share it in real time. That coordination didn't exist. Today it's the industry's bottleneck.

Table of contents.

1. The Setup

2. The Sector and the TAM

3. What the Company actually does

4. What the Market is Missing

5. Moat

6. The Financials

7. Valuations

8. Risks

9. Catalysts

10. Conclusion

1. The Setup

The semiconductor industry is on its way to becoming a $1 trillion market by 2030, but making chips is getting much harder. As Moore’s Law slows, the industry is moving toward more complex 3D designs and chiplet-based architectures.

In the past, yield analytics was mostly a tool used to study problems after they happened. It sat on top of the manufacturing process rather than being part of it.

Today, PDF Solutions is trying to become something more important: a core data and coordination layer for the semiconductor industry. As chip manufacturing becomes more global and more fragmented across different companies and production steps, PDF Solutions is positioning itself as the neutral platform that helps connect the whole ecosystem:

PDFS does not design or manufacture chips; it provides the software and connectivity that allows the four primary tiers of the industry to collaborate without compromising intellectual property:

1. Fabless (e.g., NVIDIA, AMD): Needs to know if their complex 3D designs are manufacturable.

2. Foundry (e.g., TSMC, Samsung): Needs to ramp up new nodes (2nm) faster and reduce waste.

3. OSAT (e.g., ASE, Amkor): Needs full traceability from the final package back to the original wafer.

4. Equipment Vendors (e.g., ASML, Applied Materials): Needs to monitor their multi-million dollar tools inside remote fabs.

The Paradigm Shift: PDFS is moving from being a tool used by one of these players to being the orchestration platform that links all of them.

That’s what explains the sharp recent move this week. And it’s also why the valuation exercise in this deep dive is harder than for a typical software company.

We’re valuing a business-model transition happening in real time.

The companies above are illustrative examples of each tier, not confirmed PDF customers.

2. The sector and the TAM

The reason PDF matters now is that semiconductor manufacturing is getting harder faster than most people appreciate. Management frames the shift around three forces: more 3D structures, a more complex global supply chain, and the need to apply AI across design, manufacturing, test, and packaging.

The industry is not just getting bigger. It is getting harder to coordinate. That is the part that matters for PDF. Management’s investor deck says the semiconductor market could reach roughly $1.1 trillion by 2030, with AI representing about 67% of that market and growing about 3x faster than the broader semiconductor market.

The same deck sizes the sub-2nm foundry market at about $100 billion by 2030 and advanced packaging at about $80 billion.

Even if you haircut those numbers, the message is the same: as AI systems get larger, more heterogeneous, and more packaging intensive, the value of data infrastructure, process control, and cross-company coordination rises.

The addressable market history in the deck is more useful than the headline TAM. Management says PDF’s addressable market expanded from about $1.3 billion in 2019 to about $2.6 billion in 2023 and could reach about $5.3 billion by 2030 as the company moves from yield ramp into manufacturing analytics, high-volume manufacturing, equipment and supply-chain orchestration, and AI-driven workflows.

There is also an adjacent market angle: The company website now highlights Exensio Battery, which applies the same analytics stack to battery cell manufacturing. It is not part of the core thesis today, but it is a credible adjacent option because it reuses the same core capability: messy industrial data, process variability, and yield improvement.

3. What the company actually does

The product stack is easier to understand if you stop thinking about PDF Solutions as a single software product and instead view it as a semiconductor data stack. Each layer solves a different part of the manufacturing problem.

• Exensio Platform is the analytics core. It ingests large volumes of manufacturing data, standardizes it, and helps engineers identify yield loss, quality issues, and process variation across fabs, test, and assembly. The simplest way to think about it is as a semiconductor-specific data and analytics platform.

• DirectScan and eProbe are the differentiated data-generation layer and the main hardware piece of the story. Instead of scanning full images, eProbe takes targeted electrical measurements on the wafer, allowing customers to detect certain functional defects inline before they create larger downstream problems. Strategically, this matters because PDF is generating valuable proprietary data, not just analyzing third-party data, and management is increasingly shifting the model toward subscriptions and leases.

• SecureWISE is the secure connectivity layer. It allows equipment vendors and semiconductor manufacturers to exchange data through a zero-trust architecture. Its importance is that it moves PDF beyond analytics and into governing how data flows across the semiconductor ecosystem.

• Cimetrix is the equipment connectivity layer. It lets manufacturing machines communicate with the factory’s central systems. It is not flashy, but it is deeply embedded in day-to-day fab operations.

• Sapience Manufacturing Hub is the orchestration layer. It connects manufacturing data with enterprise systems like ERP, linking engineering, operations, and finance. The key value is that it helps customers move from isolated analytics to coordinated decision-making across the business.

  

Taken together, PDF Solutions is no longer just a yield analytics company. It is building a broader semiconductor data stack that combines analytics, differentiated hardware-generated data, secure connectivity, equipment integration, and workflow orchestration. The strategic appeal is that this stack is vendor-agnostic. PDF does not need to own the fab, the tool, or the ERP system. It just needs to become the neutral layer that helps all of them work together more effectively.

So far, we have covered the setup, the market, and the business model.
From here, we get into the part that really matters: what the market is still missing, where the moat is real or overstated, the key catalysts and risks, the numbers, and the valuation models.

Read more

Table of Contents

1. The Setup

2. The Business

3. Moat

4. Numbers

5. Valuation (Multiples, reverse DFC, peer comparison and scenario analysis)

6. Risks

7. Catalysts & Timeline

8. Conclusion

1. The setup

LPKF Laser & Electronics SE ($LPKF / $LPK / $LPKFF) is a German laser equipment maker that has spent the last decade quietly building the dominant industrial process for drilling micro-holes in glass, a step every major foundry now needs as the semiconductor industry pivots from organic substrates to glass core substrates for AI accelerators.

Intel reached high-volume manufacturing of glass substrates at its Chandler facility in Q1 2026. Samsung is targeting full mass production by H2 2026 at Sejong. TSMC is building a CoPoS pilot at AP7 Chiayi to serve NVIDIA's Rubin architecture. Rapidus targets 2028. Absolics is already producing prototypes in Georgia.

Over 80% of these customers are qualifying with LPKF's LIDE equipment. Yet LPKF trades at €14/share with €350M market cap, 3x trailing EV/Sales, after a brutal 2025 in which Solar revenue collapsed 31% and the company missed its original guidance by 13% on revenue and 96% on EBIT margin.

The market is treating LPKF as a struggling industrial restructuring story. The filings show something different: a quasi-monopolist on a $20B TAM ramp-up, hiding behind two non-core segments that are stabilizing, an activist shareholder who entered in September 2025, and a CEO who hinted at "tangible numbers" on LIDE order entry coming in the Q1 2026 report on April 30. This is the analytical setup.

Macro context: why glass and why now

The semiconductor packaging industry is at an inflection point that has nothing to do with the front-end node race and everything to do with physics. As AI accelerators push past 1,000 watts of power per package, integrate more chiplets, and demand higher I/O counts, the organic substrates that have served the industry for two decades are hitting hard physical limits. Warpage above 55mm package size, dielectric loss at high frequencies, thermal expansion mismatch with silicon, and signal integrity degradation are all converging into what packaging engineers call the “warpage wall.”

Glass solves this. Glass core substrates offer 5-10x better dimensional stability than organic alternatives, ultra-low coefficient of thermal expansion (sub-5 ppm/°C), exceptional flatness (sub-20μm warpage across 100mm packages), and 40% higher interconnect density. They enable 60-80mm packages that integrate 8-16 chiplets with HBM stacks, architectures that simply cannot be built reliably on organic substrates. This is why Intel started developing glass substrates over a decade ago, and why every major player has now committed to glass roadmaps in the 2026-2030 window.

Every leading foundry has a glass roadmap, NVIDIA has dictated glass for Rubin, and Broadcom/Marvell have integrated CPO platforms. If glass captures 10-15% of the advanced packaging substrate market by 2030 (the IDTechEx and Future Markets Inc projections suggest this is the conservative scenario, with 20-30% by 2036), the laser equipment TAM scales 4-6x faster than the numbers implies.

LPKF is not the entire story of glass packaging. It is one specific tool in a longer process chain. But it is the dominant tool for the most patent-protected, hardest-to-replicate step: drilling thousands of high-aspect-ratio micro-vias through glass without cracks, with sub-micron precision, at industrial throughput. This is what LIDE does.

2. The business: what LPKF actually sells

LPKF operates in four reporting segments:

1) The Electronics segment is where LIDE lives. CEO Klaus Fiedler said on the Q4 2025 earnings call that the advanced packaging semicon portion of Electronics was “still in the low 8 figures” in 2025, meaning roughly €10-15M of revenue.

The rest of Electronics is the legacy PCB stencil cutting, depaneling, and flexible PCB processing business. The negative segment EBIT in 2025 reflects a combination of LIDE qualification line costs, R&D for new LIDE generations, and weak utilization in core PCB tools because of US tariff disruption to SMT customer ramp plans.

2) Solar produces laser scribers for thin-film solar cell structuring. The 2025 revenue collapse to €28.3M (from €41.2M) reflects two things: the loss of a planned major project in China, and customer investment paralysis as the industry transitions from established silicon thin-film to perovskites, where volume production is not yet ready. Management expects 2026 to be similarly weak. Importantly, Solar is still profitable at the segment level, just structurally smaller.

3) Development supplies in-house PCB prototyping equipment to research institutions, universities, and corporate R&D labs. This is a steady, defensive business with North America as the dominant market. The US government shutdown from early October to mid-November 2025 dampened Q4 order entry but did not change the structural trajectory. With defense R&D budgets growing, this segment is positioned for low-to-mid single digit growth.

4) Welding produces laser systems for plastic welding, used in consumer electronics, medical devices, and automotive supply. The 2025 turnaround in this segmen, from -€4.8M EBIT to +€0.5M, on +30% revenue growth, was driven by a single large consumer order out of China. Importantly, LPKF also won a substantial smart robotics customer in 2025 with continuing orders into 2026, which Fiedler described in the call as having “8 figures” of opportunity if the customer scales.

What the market is missing: insights from the filings

1. The sharp decline in backlog is the main bearish signal, but it also helps explain why 2026 guidance appears conservative. The midpoint of guidance can be reached with a normal level of 2026 book-and-bill orders even without large LIDE volume orders. The company also stated that potential Advanced Packaging volume orders are not included in the forecast, which implies that the current guidance does not capture upside if those orders arrive.

2. The Welding segment improved materially in 2025, moving from a loss to positive EBIT. In addition to the turnaround, management pointed to a potentially meaningful opportunity in smart robotics, suggesting that Welding may have more growth potential than a typical legacy industrial segment.

3. LIDE’s economics are not disclosed separately because they are embedded inside the Electronics segment, which reported a loss in 2025. That makes the segment look weaker than it may actually be, because LIDE is still absorbing qualification, R&D, and ramp-related costs. If LIDE moves from qualification into volume production, incremental revenue could carry much stronger margins than the current segment figures imply.

4. LPKF also has about €50 million of unrecognized tax loss carryforwards and temporary differences. If profitability improves, these could reduce cash taxes for several years and potentially increase future free cash flow and equity value.

3. MOAT

LIDE (Laser Induced Deep Etching) is a two-step process. A pulsed laser modifies a precise zone within the glass without ablating material, creating a chemically reactive damage line. A subsequent selective wet etch removes only the modified zone, leaving micrometer-scale through-glass vias with smooth sidewalls and high aspect ratios (>15:1 demonstrated). This is fundamentally different from direct laser ablation (which leaves rough sidewalls and microcracks) or photosensitive glass methods (which require special glass formulations).

The IP position is real. Fiedler stated in the Letter that LPKF has “successfully confirmed our protective rights in Europe and Korea, [and] are currently actively pursuing patent infringement in China and have initiated proceedings there against a patent infringer. Our goal is to consistently prevent the unauthorized use of our technology and ensure that products manufactured with such imitation technologies cannot access markets outside China.” This is a defensive posture but it is being executed actively.

The competitive landscape Fiedler discussed on the Q4 call:

• Schmid Group (Germany): traditional wet processing player, complementary to LIDE in some flows but competitive in others. Mentioned as a credible alternative.

• Philoptics (Korea): laser equipment supplier, named by an analyst on the call as a competition concern. Fiedler’s response: “If they go in with their own technology they developed, fair, that’s good. It cannot be a single source market. If it’s competitors copying us, there, we will be very active in avoiding it.”

• Chinese copycats: explicitly mentioned, with active patent litigation. Fiedler’s framing: “We need to make sure that products manufactured with such imitation technologies cannot access markets outside China”, implicit acceptance that China will have local-for-local equivalents, but defense of the rest of the world.

• Other Western players: “I don’t see any viable competitor in the Western countries who is close to our offering.”

Fiedler’s stated personal target is 70% market share in production ramp-up orders, against a current >80% qualification share. The 10-percentage-point compression assumed in his target is the realistic erosion the market should price in. Anything better than 70% is upside.

The non-IP elements of the moat are also meaningful:

• +5 year customer relationships in development stage (Fiedler: “we sometimes work with them for more than 5 years”)

• Customer-pull product expansions (ABF singulation was customer-requested, not LPKF-pushed)

• Vertical integration: LPKF also produces glass components in-house using its own LIDE equipment, generating qualification data and process knowledge that pure equipment vendors lack

• Adjacent process expansion: glass bonding, ABF singulation, CPO research already in motion

The big risk to the moat: yield curves on customer LIDE machines. If Schmid or Philoptics ships a competitive tool with similar throughput at 20-30% lower price, customers might split purchase orders to maintain dual sourcing. This is the realistic erosion vector, not technology displacement.

4. The numbers

This table shows a business that looks much weaker on the surface than it likely is underneath. Revenue grew sharply in 2022, stayed broadly flat in 2023 and 2024, and then fell 6.2% in 2025 to €115.3M. That 2025 decline matters, but it was not a collapse across the entire company. The biggest drag came from Solar, while Development held up well and Welding rebounded strongly. In other words, the weak top line reflects a mix issue and delayed customer spending more than a breakdown of the whole franchise.

What to watch in the Q1 2026 report (April 30)

This is the single highest-information-density event for LPKF in 2026. Klaus Fiedler signaled explicitly on the Q4 call that he will share “tangible numbers” on LIDE order entry. The market will be looking for:

1. LIDE order intake quantification: did the “handful of customers” Klaus mentioned actually issue POs in Q1? Magnitude matters, €5-10M of Q1 LIDE orders signals a potential €20-40M run-rate; €1-2M signals continued slow ramp.

2. Number of LIDE customers ordering for production: Klaus said “more than one or two.” Confirmation of 3-5 named customers (or quantified range) materially derisks the bull case.

3. Total order intake: if Q1 order intake is €25-30M, the full-year guidance becomes very achievable. Below €20M is a yellow flag.

4. Solar order pipeline commentary: any update on perovskite project timing or large customer recovery

5. Welding order entry from smart robotics customer: any quantification of run-rate

6. North Star cost savings realization: how much of the structural cost reduction has flowed through

7. Q1 cash position and any drawdown of credit lines: confirms whether the financing runway is comfortable

A guidance raise at the H1 2026 report (likely late August) is plausible if Q1 trends positive. The AGM on 4 June 2026 in Hannover is also a venue for narrative updates.

Up to this point, the qualitative case is clear: LPKF is not just a weak Solar story or a messy restructuring case. It is a company with a strong position in a critical glass-packaging step, a cleaner cost base, and meaningful upside if LIDE converts from qualification into real production orders.

The next step is where the thesis becomes investable: valuation, the specific risks that can break the story, and the concrete catalysts that will tell us whether the ramp is actually starting. That is also where the asymmetry becomes much easier to see.

Read more

The goal here is simple: separate the noise from what the company is actually saying.

In this article, I will break down why is no longer just a lidar company, why the latest results need to be normalized, where the business is heading, and what the filings really tell us about its customers, competitors, catalysts, valuation, and overall setup.

If you have only followed the headlines, you are probably missing the real story.

Table of Contents

1. The Business

2. Competitors

3. Guidance

4. Moat

5. The Financials

6. Valuation (Multiples, reverse DFC and peer comparison)

7. Risks

8. Catalysts & Timeline

9. Conclusion

1. The Business

Ouster designs and sells digital lidar sensors, perception software, stereo cameras and edge AI compute through the Stereolabs acquisition. The company positions itself as “Physical AI’s first unified sensing and perception platform.”

Translated from marketing, that means: Ouster wants to be the one-stop-shop for the sensor stack and low-level perception software that runs on autonomous machines in four verticals.

To understand what Ouster does, you have to understand what makes "digital lidar" different from what most of the industry still builds. Every lidar sensor does essentially the same job: emit light pulses, measure how long they take to bounce back off objects, and convert those time-of-flight measurements into a 3D point cloud the machine can use to perceive its surroundings. The differences are in how the emitting and detecting hardware is built.

Traditional analog lidar is built from many separate components, so improving resolution usually means adding more parts, more complexity, and more cost.

Ouster’s approach is different.
Instead of improving the sensor by adding more separate hardware, Ouster tries to improve it by putting more of the system into the chip itself. The lasers, the detectors, and part of the signal processing are more tightly integrated into silicon. The idea is simple: rather than making the product better by adding more pieces, Ouster wants to make it better by building a better chip.

That matters because each chip generation can improve range, resolution, and reliability without requiring a completely new architecture. L3 is the current generation, L4 is now in validation, and Chronos is the next step for flash lidar. So the real advantage is not just better sensors today. It is the possibility of improving performance over time on a more efficient cost curve.

That distinction matters because, if Ouster is right, the advantage is not just a better product today. It is a better architecture that can keep improving with each chip cycle. And if the advantage really sits at the architecture level, then the IP should matter too.

That is where the Hesai example becomes important. Hesai is one of the biggest lidar manufacturers in the world by volume, especially in automotive. Instead of designing around Ouster’s digital lidar patents, it chose to license the technology. That does not prove Ouster’s advantage will last forever, but it does suggest the architecture has real commercial value

The new piece is Stereolabs. Ouster closed that acquisition on February 4, 2026 for $35.4 million in cash plus 1.85 million newly issued shares. Management says Stereolabs is a high-growth, high-margin business that should be accretive to consolidated results and expands Ouster into camera vision, AI compute, sensor fusion, and foundational AI model training.

Product lines

Ouster does not sell just one product. It has several product lines, and each one plays a different role in the business.

OS product line (the core business): Four models, all built on the same digital architecture but with different optical configurations for different use cases.

• OSDome: for indoor or ceiling-mounted applications, like buildings and infrastructure

• OS0: short range, useful for robots, warehouses, and small autonomous vehicles

• OS1: the most versatile model, used across many industrial applications

• OS2: long range, for mining, highways, and long-distance mapping

DF series. This is Ouster’s automotive product for ADAS and autonomous driving. It is the part of the story that sounds exciting because investors always like large automotive opportunities. But the reality is that it is still early. The product is not yet a meaningful commercial business. Ouster has prototypes, and the Chronos chip is still being tested. So for now, DF is best understood as future upside, not as something driving results today.

The software layer, which is where the re-rating story really starts:

This is the part of the business that makes Ouster more interesting than a pure hardware company.

Gemini is Ouster’s perception software for smart infrastructure. It handles tasks like detection, classification, tracking, and monitoring. That matters because customers are not really paying for raw point clouds. They are paying for outcomes. They want to know what is moving, where it is moving, and whether something needs attention. Gemini helps turn Ouster’s hardware into a usable system rather than just a sensor.

It matters for another reason too. Gemini is built specifically for Ouster lidar. So every Gemini deployment makes the customer more tied to Ouster hardware. That is what makes the software layer strategically important. It adds value, but it also makes the relationship stickier.

Then there is BlueCity, which is basically Gemini adapted for traffic operations and road safety. It is the same broader idea, but focused on a very specific use case. Think intersections, traffic flow, pedestrian safety, signal timing, and near-miss detection. Management has already talked about deployments in Tennessee, Utah, and New Jersey, with additional pilots planned for Europe and the Middle East in 2026. BlueCity is important because it shows how Ouster can package the same underlying technology into a clearer, more valuable end-market solution.

The platform extension through Stereolabs

The February 2026 Stereolabs acquisition adds something Ouster simply did not have before.

With ZED cameras and AI Compute, Ouster now has a first-party camera and compute offering alongside lidar. That means stereo vision, neural depth perception, and edge inference hardware all become part of the portfolio. Before this deal, Ouster could talk about perception. After the deal, it can offer more of the perception stack directly.

That is a meaningful shift. Cameras are a core sensor in many robotics and physical AI applications, and compute is what makes sensor fusion useful in the real world. So Stereolabs does more than just add another product line. It broadens what Ouster can actually sell to customers.

In addition, the TAM for lidar and the broader sensing and perception market is enormous, especially now that Ouster is no longer just a lidar company:

What Stereolabs actually changes for Ouster:

1. Adds the second major sensor modality (cameras) to complement lidar, enabling sensor-fusion-based solutions that neither company could sell alone. Multi-sensor AI training (lidar + cameras fused) is a technical lever Pacala flagged as “the obvious next step” on the Q4 call.

2. Brings a ready developer community. Stereolabs has had a developer-first distribution model since founding, thousands of software developers are actively using ZED cameras in production systems. Ouster couldn’t build this organically.

3. Adds edge AI compute hardware. The Stereolabs AI Compute platform (likely Jetson-based) extends Ouster’s offering from sensors-only to sensor-and-compute, which matters because customers increasingly want integrated solutions.

4. Creates narrative coherence. “Unified sensing and perception platform for Physical AI” is a stronger marketing position than “lidar sensors with some software.” The narrative coherence matters for multiple expansion even if underlying unit economics evolve more slowly.

If Ouster is a lidar company, why did it buy Stereolabs?

Lidar is very good at measuring distance, shape, and position in 3D, especially in low light or difficult conditions. Cameras add something different: color, texture, and visual context. In simple terms, lidar tells you where something is and what shape it has. Cameras help you understand what it is.

That is why the strategy makes sense. Ouster did not buy Stereolabs because lidar is not enough. It bought Stereolabs to move up the stack. Instead of selling only a lidar sensor, Ouster now wants to sell lidar, cameras, compute, software, and sensor fusion as one integrated perception system.

The customer benefit is simple: less integration work. Many customers do not want to buy lidar from one company, cameras from another, compute from another, and then spend months putting everything together. Ouster wants to make that easier by offering more of the stack itself.

Strategically, this helps most in markets like industrial robotics, automation, humanoids, visual inspection, and smart infrastructure, where using multiple sensors together is often better than relying on just one.

So the real goal is not to replace lidar with cameras. The goal is to make Ouster more than a lidar company. Lidar is still the core, but cameras and compute let Ouster sell more per customer, make the product harder to replace, and strengthen the shift from sensor vendor to platform supplier.

How the business makes money

This is where the deep dive has to be careful, because Ouster reports revenue in a way that obscures the real operating picture. I'll break out the actual decomposition.

Royalty revenue was unusually high because Ouster was finally allowed to recognize revenue that had been sitting on the balance sheet from earlier periods. The 10-K says Ouster recognized $22.8 million of royalty revenue in 2025, and $16.1 million of that was old deferred royalty revenue that could only be recognized once the uncertainty around the license contract was resolved. Then on the earnings call, Ken Gianella said 2026 royalty revenue should be less than $5 million.

Ouster does not break out product revenue into “hardware revenue” and “Gemini/BlueCity software revenue” in the 10-K. So we cannot know the exact software revenue number from the financial statements alone. What we do know from the call is that software-attached bookings more than doubled in 2025 and represented over 15% of sensors shipped. Ouster also shipped 25,000+ sensors in 2025.

So the rough interpretation is:

• software is clearly growing

• more customers are buying software together with the sensor

• but software is still not large enough for Ouster to report it as a separate revenue line

The business flow:

1) Hardware sales. Customers buy lidar sensors, and Ouster records that revenue when the unit ships. This is still the main business today.

2) Software through Gemini and BlueCity. Customers buy the hardware, but they can also pay for software licenses on top. This part matters because software is recurring, higher margin, and makes customers more tied to Ouster’s system.

3) Royalties. This comes from other companies licensing Ouster or legacy Velodyne technology. It helped revenue in 2025, but management has already said this should be a much smaller line in 2026.

4) Services, like support or custom work. This is a very small part of the business and not strategically important.

So the simple picture is:

The software is what turns lidar data into something useful. On its own, the sensor just produces a 3D point cloud. That is raw data. Most customers do not want raw data. They want answers. They want to know if a person entered a restricted area, if traffic is backing up at an intersection, or if two objects almost collided.

That is what Ouster’s software does. Products like Gemini and BlueCity take the lidar data, detect what is happening, classify objects, track movement, and turn all of that into alerts, analytics, and real-world actions.

The big shift is that Ouster no longer wants to be just a sensor company. It wants hardware to bring customers in, and software to make the business more recurring and more valuable over time.

2. Competitors

Before getting into the explicit moat analysis, it’s worth mapping where Ouster sits in the broader lidar/sensing industry. The competitive field has been reshuffled significantly over the past 18 months.

• Innoviz (INVZ): $55M revenue 2025, auto-focused, recent Daimler Truck L4 win, struggling balance sheet.

• Aeva (AEVA): $18M revenue (est), newly announced Nvidia partnership, FMCW technology approach (different from Ouster).

• MicroVision (MVIS): Sub-scale, recently bought Luminar’s assets, remains in transition.

• AEye (LIDR): Very small scale, structurally marginal.

The private and Chinese:

• Hesai Technology: By volume the world’s largest lidar company, public on NASDAQ (HSAI). Approximately $350M revenue 2025. Now confirmed Ouster IP licensee.

• RoboSense, Seyond: Large Chinese players.

• Koito Manufacturing: Japanese Tier 1 with lidar business.

Ouster’s position in this field is unique: it has real scale (25,000+ sensors/year, top 3-5 by US-based public company volume), a legitimate balance sheet, a broad multi-vertical product portfolio (most competitors are 1-2 verticals), and now with Stereolabs a multi-modal sensing offering. Not the largest by volume (Hesai dominates that), but arguably the highest-quality public pure-play left standing.

3. GUIDANCE

The Q4 call gave a much clearer picture of 2026 than the headline numbers suggest.

Ouster guided Q1 2026 revenue to $45 million to $48 million, and that already includes about seven weeks of Stereolabs revenue after the February 4 close. Since Stereolabs did about $16 million of revenue in 2025 and management said roughly 60% of that business is weighted to the second half, the Q1 contribution from Stereolabs is probably only around $1.5 million to $2 million. That implies the core Ouster business is guiding to roughly $43 million to $46 million of Q1 revenue, compared with $41 million of product revenue in Q4 2025.

In other words, the underlying lidar business still looks like it is growing at roughly a mid-30s rate, which is consistent with what Gianella later suggested on the call.

Management also reiterated its broader 2026 framework. The company is still targeting 30% to 50% annual revenue growth and 35% to 40% GAAP gross margin, but importantly, that framework is being measured against a 2025 pro forma baseline excluding royalties. Using Ouster’s $146 million of core product revenue plus roughly $16 million from Stereolabs, the relevant starting point is about $162 million, which implies a 2026 revenue range of roughly $211 million to $243 million. That is the number investors should anchor to, not the inflated reported 2025 revenue that included one-time royalty catch-up.

That normalization matters because management was very explicit that 2026 royalties should be less than $5 million, with most of that landing in the second half. So one of the biggest differences between 2025 and 2026 is that the royalty line should shrink sharply. At the same time, GAAP operating expenses are expected to grow 5% to 8% in 2026, slightly above the original long-term framework because of Stereolabs integration costs. Management also disclosed something important on liquidity: even after the acquisition, Ouster believes it still has roughly four to five years of runway until it reaches operating cash flow breakeven.

What the call really tells us

The simplest way to read the call is this: the core lidar business still looks healthy, Stereolabs adds growth on top, but reported 2026 numbers will look noisier because the royalty benefit largely disappears and because Stereolabs is seasonally back-half weighted, which means Q1 and Q2 may look softer than the full-year story.

The most useful signal from management was not the formal framework, but the tone around the underlying business. Gianella effectively pointed investors toward mid-30s growth in the core product line, while Pacala kept reinforcing that the bigger story is no longer just lidar units. Software-attached bookings more than doubled in 2025, software was attached to more than 15% of sensors shipped, and management said customers are already asking to buy the full hardware and software stack from one supplier. That is the strategic message underneath the numbers.

The other important part of the call was what management did not hype. Pacala was fairly cautious on defense and humanoids, suggesting both are still more of a medium-term opportunity than a 2026 revenue driver. By contrast, management sounded much more confident on smart infrastructure, industrial automation, and the combined lidar plus cameras plus compute platform after Stereolabs.

So if you strip the call down to its essentials, the message was simple:

Normalize out the royalties, assume the core business is still growing around the mid-30s, expect Stereolabs to help more in the back half, and recognize that the platform story is becoming more important than the pure hardware story.

4. MOAT

The Business section above covers what Ouster is and the competitive landscape it sits in. This section asks the harder question: how durable is the competitive advantage, and which specific sources of moat will compound versus erode over the next 5-10 years?

The long-term bull case depends far more on software and integration than on hardware alone.

1) Underlying technology and IP. Management clearly sees the digital lidar architecture and custom silicon roadmap as core differentiators. But patents mainly protect against direct copying. They do not stop competitors from reaching the same market through different technical approaches. So the IP is valuable, but it is not a permanent shield.

2) Manufacturing and scale. Ouster believes its digital design helps on cost, yields, and production efficiency, and the company already manufactures the majority of its products through Benchmark and Fabrinet in Thailand. That matters because customers buying into multi-year programs want proof that the product can actually be built and delivered at scale.

Still, this is better understood as table stakes than as a deep moat. Manufacturing credibility matters, but it is not the part of the business that should command a premium multiple on its own.

3) The real MOAT: Software. Gemini is not just an add-on. The 10-K says it is optimized exclusively for Ouster’s digital lidar sensors, which means every Gemini deployment strengthens the bond between the customer and Ouster hardware. BlueCity extends that idea into traffic and road safety, turning the same underlying technology into a more specialized solution. And the company is already seeing real traction here: software-attached bookings more than doubled in 2025, software was attached to more than 15% of sensors shipped, and Ouster’s AI models are now running at more than 1,200 Gemini and BlueCity sites.

This is the closest thing Ouster has to a compounding moat, because it creates switching costs, improves the product with more real-world data, and adds recurring revenue on top of hardware.

That is also why Stereolabs matters strategically. Ouster is no longer trying to sell only a lidar sensor. Management is now clearly presenting the company as a unified sensing and perception platform that combines lidar, cameras, AI compute, sensor fusion, and perception software. The more of that stack Ouster can sell, the more valuable the customer relationship becomes. A company that sells one sensor can be replaced. A company that sells the sensor, the software, the camera layer, and the compute layer is much harder to dislodge.

4) The balance sheet, which in a sector like this does function as an advantage. At year-end, Ouster had $211.2 million of cash, cash equivalents, restricted cash, and short-term investments, with no debt outstanding. Management went further on the call and said the company still has roughly four to five years of runway to operating cash flow positivity even after the Stereolabs acquisition. The balance sheet is not a classic moat in the way software can be, but it is still a real commercial asset today.

So the clean conclusion is this: Ouster’s moat is not primarily about having the best sensor forever. The stronger version of the thesis is that the company is using hardware as the entry point, software as the lock-in layer, and platform integration as the long-term value driver. The IP helps. Manufacturing helps. The cash balance helps.

But the part that can truly compound over time is the software and perception stack. If that keeps scaling, Ouster can become harder to replace and more valuable than a pure hardware vendor. If it does not, then the company risks remaining a solid but ultimately more replaceable sensor supplier.

Ouster’s real moat is not the sensor by itself. It is the combination of software, integration, and balance-sheet credibility built around that sensor.

Once you understand the business, the guidance, and the moat, the story starts to sound compelling. The next question is the only one that matters: what is all of that actually worth?

The business is interesting. The moat is real. The guidance is better than the headlines suggest. Now comes the part that separates narrative from investment: the financials, the valuation, and the risk-reward at today’s price.

By this point, the story should be clear. What is not obvious, and what really matters, is whether the market is still mispricing it. That is where the financials and valuation do the heavy lifting.

The qualitative story can get your attention. The next section is where the thesis either holds up or falls apart.

We will look at the financials, the valuation through the lens of multiples, peer comparison, and reverse DCF, as well as the key risks, catalysts, and the final conclusion.

Read more

Not exactly the profile that screams “buy.”

Fourteen months later, the stock trades above $930. That’s a 30x return. SanDisk just entered the Nasdaq-100, replacing Atlassian. Analysts at Bernstein have a $1,250 target. Jefferies raised to $1,000.

The question isn’t whether this happened because the chart is right there. The question is why, and what it reveals about how technological shifts can completely re-price assets that the market had written off.

The Setup: A Business Nobody Wanted

SanDisk’s backstory matters. The original SanDisk Corporation was a pioneer in flash memory, founded in 1988. Western Digital acquired it for $19 billion in 2016. For nearly a decade, the SanDisk brand lived inside WD as a product label, not a company.

In October 2023, Western Digital announced it would split its NAND (flash memory) business from its legacy hard drive division. The logic was straightforward: the NAND cycle was in a brutal downturn, and the combined entity’s stock was being penalized for housing two very different businesses under one roof. The flash business was the “problem child”: cyclical, capital-intensive, and hemorrhaging money.

The numbers from the 10-K tell the story. In the fiscal year ending June 2023, SanDisk’s attributed operations generated $6.1 billion in revenue on gross margins of just 7.1%. That’s not a typo. Revenue was overwhelmingly tilted toward Client ($3.6B) and Consumer ($1.9B), with Cloud representing a mere $500 million. The business was essentially a high-volume, low-margin consumer electronics supplier.

When the spin-off completed in February 2025, WDC retained 19.9% of shares (later disposing of 14.6% through a debt exchange in June 2025). SanDisk took on $2 billion in Term Loan B debt, paid WDC a $1.5 billion distribution, and started life as a public company with negative net income, negative operating cash flow history, and a market that fundamentally did not want to own pure-play NAND.

The stock opened around $28.

What Changed: The NAND Supply-Demand Inversion

Two things happened simultaneously, and their intersection created the conditions for a 30x re-rating.

1. AI fundamentally altered the demand profile for NAND. Data center storage went from afterthought to bottleneck. The AI inference cycle generates enormous amounts of data that must be stored and read at high speed and NAND flash is the only technology that scales to meet that demand. Jensen Huang’s CES comments highlighted a key point: AI systems need far more NAND than before, something SanDisk’s CEO had already been emphasizing for quarters:

   “NAND is no longer a commodity input, it’s a strategic component of AI infrastructure.”

On the Q2 FY26 earnings call, Goeckeler quantified it: data center exabyte growth forecasts for 2026 jumped from mid-20% two quarters ago, to mid-40% last quarter, to high-60% now. That’s three consecutive upward revisions in a single fiscal year. And that excludes the KV Cache demand (the extra memory AI needs to store context during inference), which management estimates at 75-100 additional exabytes in 2027, doubling the year after.

For the first time, data center is expected to become the largest end market for NAND in calendar 2026, displacing smartphones and PCs, which had dominated for decades.

2. Supply did not respond. This is the critical piece. NAND fabrication requires extraordinarily capital-intensive fabs with multi-year build timelines. SanDisk operates through a joint venture with Kioxia across two sites in Japan. The industry had gone through a brutal 2022-2024 downturn that decimated margins and led to production cuts, underutilization charges, and deferred capacity investments across the entire supply chain.

When demand inflected, there was no spare capacity to absorb it. SanDisk’s management has been explicit: they’re targeting mid-to-high-teens bit growth annually, delivered through technology transitions (BiCS8 node) rather than greenfield expansion. New fabs take years. The supply response to the current demand surge is structurally limited.

The result: a pricing environment unlike anything the NAND industry has experienced. Prices across segments are rising simultaneously, not just data center, but edge and consumer too, because NAND is fungible and flows to the most attractive market. When data center pricing pulls supply, every other segment tightens.

The Numbers: From 7% Gross Margin to 51% in Two Years

The financial transformation is staggering. Here’s the quarterly trajectory, straight from SanDisk’s filings and earnings deck:

Q2 FY26 revenue of $3.025 billion beat guidance of $2,550-$2,650 million by over $350 million. The overdelivery came entirely from pricing strength that accelerated during the quarter. But the Q3 FY26 guidance is where it gets extraordinary: $4.4-$4.8 billion in revenue at 65-67% gross margin, implying $12-$14 in non-GAAP EPS in a single quarter. That’s the midpoint guiding to $13 EPS, more than double the prior quarter.

At the annual revenue run rate, Q3 guidance implies over $18 billion annualized, at gross margins that would have been inconceivable 18 months ago.

On the balance sheet, SanDisk has gone from $2 billion in debt at separation to $603 million by Q2 FY26, generating enough cash to deleverage rapidly while funding capex and R&D. CFO Luis Visoso stated on the call that the company’s priority is continuing to reduce debt, building prudent cash reserves, and fully funding innovation, all from operating cash flow.

The Cloud Pivot: $325M → $960M

The revenue mix shift tells the real story. SanDisk’s Cloud segment, which is essentially enterprise SSDs for data centers, generated $325 million in FY24, $960 million in FY25 (up 195%), and is now accelerating further. In Q2 FY26, data center revenue was $440 million, up 64% sequentially, and management expects another substantial step-up in Q3.

This matters because data center carries structurally higher margins than consumer flash. Enterprise SSDs require qualification at each hyperscaler (a multi-quarter process), specific performance configurations, and ongoing technical support. This is not a commodity market, it’s closer to a design-win model where switching costs are real.

SanDisk has already secured approval from major hyperscalers for its high-performance SSDs, and its next-generation, lower-cost storage product is expected to start generating revenue soon.

The Structural Thesis: Business Practices Are Changing

Perhaps the most underappreciated aspect of SanDisk’s story is the shift in how NAND is being bought and sold.

Goeckeler’s framing on the call was direct: for decades, NAND operated on a quarterly auction model: spot pricing, short-term negotiations, and brutal cyclicality. When supply exceeded demand, prices collapsed and so did margins. When demand exceeded supply, prices spiked but producers had already committed capacity at lower rates.

That model is breaking down. Data center customers, hyperscalers spending tens of billions on AI infrastructure, need supply assurance measured in years, not quarters. They’re planning 2027, 2028, even 2029-2030 capacity needs. SanDisk disclosed on the call that it has signed one long-term agreement (with a prepayment component) and has “several in the queue.”

CFO Luis Visoso framed it clearly: customers across end markets and geographies are reaching out. They want to assure supply and are willing to commit to multi-year agreements with firmer pricing and, crucially, prepayments. This isn’t incremental, it’s a potential structural shift from cyclical commodity to contracted infrastructure, similar to what happened in semiconductor foundry (TSMC’s model) over the past decade.

If long-term contracts become common, SanDisk’s earnings would be much more stable and the market would likely value the business higher.

What This Teaches About Market Regime Changes

SanDisk’s 30x move isn’t primarily a story about a great management team executing a turnaround (though the execution has been strong). It’s a story about what happens when an exogenous technological shift, the AI infrastructure buildout, collides with a supply-constrained asset that was priced for permanent mediocrity.

Every part of the investment case was visible in the filings:

• The demand inflection was visible in the data center exabyte forecasts, which management updated transparently each quarter.

• The supply constraints were visible in the capital expenditure data and the multi-year timelines for new fab construction.

• The margin expansion potential was visible in the pricing dynamics described in the earnings calls, where selling prices were rising simultaneously across segments.

• The balance sheet deleveraging was visible in the cash flow statements, which swung from negative to nearly $1 billion in adjusted FCF per quarter.

The market at $28 was pricing SanDisk as a structurally impaired business exiting a downturn. It was not pricing in the possibility that NAND demand could undergo a regime change driven by AI, that data center could overtake smartphones as the primary end market, that business practices could shift from spot to contracted, and that margins could reset at structurally higher levels.

What Comes Next

SanDisk reports Q3 FY26 earnings on April 30, 2026. At guideline midpoints ($4.6B revenue, 66% gross margin, $13 EPS), the stock at $930 trades at roughly 17x the Q3 annualized earnings run rate. That’s not cheap for a historically cyclical business, but it may not reflect the full earnings power if long term agreements reduce cyclicality and the data center buildout extends through 2028-2030 as management suggests.

Recent catalysts reinforce the trajectory: the $1 billion equity investment in Nanya Technology with a multi-year DRAM supply agreement (announced March 25, 2026) signals SanDisk is expanding from pure-play NAND into adjacent memory technologies to serve the full AI storage stack. And the Kioxia JV extension through 2034 locks in supply infrastructure for the next decade.

The bear case is straightforward: NAND has always been cyclical, and the current pricing environment is historically abnormal. If AI capex slows, if hyperscalers find storage alternatives, or if supply responds faster than expected, the margin structure collapses. Bernstein’s Mark Newman himself noted the stock could go to $3,000 if ASPs jump 75% sequentially for two consecutive quarters but the inverse scenario is equally dramatic.

The Takeaway

Sometimes the best investments aren’t found in what’s new, but in what’s existing and suddenly relevant. SanDisk was making the same product before and after the AI revolution. What changed was the world around it. The technology that was a commodity became a bottleneck. The business that was an orphan became essential infrastructure.

The lesson for retail investors: structural demand shifts don’t announce themselves with a press release. They show up in the filings, in the exabyte forecasts that keep getting revised upward, in the management commentary about customers calling to secure supply, in the margin expansion that comes when demand persistently outstrips supply.

The data was there at $28. Most people weren’t reading it.

That’s the edge: reading the filings, doing the math, and recognizing when the market is pricing yesterday’s reality while the business is living in tomorrow’s.

This content is for informational purposes only and should not be considered financial advice. I am not a financial advisor. Always do your own research before making any investment decisions.

But none of that changes the actual fundamentals shown in the 10-K, earnings call, and Q1 2026 preliminary results.

This piece cuts through the noise and focuses on what really matters: the business, the numbers, and the key catalysts.

If you own InmunityBIO or are thinking about buying it, this is what you need to know.

1. The Business

Most cancer therapies shrink tumors but leave patients with lymphopenia, a depleted immune system. Checkpoint inhibitors wake up T cells, yet if there aren’t enough functional immune cells left, the effect is limited.

ImmunityBio’s thesis is different: restore and re-train the full immune system (innate + adaptive + memory) so the patient’s own body fights the cancer long-term and prevents relapse.

ANKTIVA is designed as the backbone that can combine with BCG, checkpoint inhibitors, CAR-NK cells, low-dose chemo, and vaccines. This “Cancer BioShield™ platform” approach sits at the leading edge of immuno-oncology:

• NK cells (Natural Killers): The immune system’s rapid-response force. ANKTIVA massively expands and activates them to attack tumors immediately.

• CD8+ T cells (Tumor Killers): Precision assassins that recognize and destroy cancer cells. ANKTIVA multiplies and empowers them.

• Memory T cells: The long-term defense system. These cells remember the cancer and prevent relapse, the reason ANKTIVA generates durable complete responses lasting 12–24 months.

The mechanism turns immunologically “cold” tumors into “hot” ones. This is not a narrow bladder cancer drug — it’s a platform backbone designed to combine with BCG, checkpoint inhibitors, CAR-NK cells, low-dose chemotherapy, and vaccines across multiple tumor types. ImmunityBio calls this the Cancer BioShield platform.

The current revenue model is simple: ANKTIVA is administered intravesically (directly into the bladder) in combination with BCG in urology clinics. It received a permanent J-code (J9028) effective January 2025, which unlocked standardized Medicare reimbursement and removed a major commercial friction point. More than 240 million U.S. lives are now covered.

To put it simple:

ANKTIVA takes a different approach than traditional medicine: instead of just attacking the cancer, it focuses on rebuilding and retraining the body’s immune system so it can fight on its own.

Think of it this way: traditional therapies remove the enemy, but leave the army exhausted. ANKTIVA aims to rebuild that army, strengthen it, and train it to recognize the threat in the future.

That’s why the goal isn’t just a temporary response, but a durable, long-term defense. In simple terms IBRX isn’t just treating cancer, it’s trying to turn the patient’s immune system into a lasting anti-cancer shield.

2. The Numbers

Revenue

FY25 was the inflection point, and Q1’2026 preliminary results show that growth is continuing:

IBRX reported $44.2M in Q1 2026 revenue, up +168% YoY and +15% QoQ, with growth every quarter since ANKTIVA launched.

Annualizing Q1 gives $177M for FY2026, likely conservative. With continued U.S. adoption and international expansion (now Dr. Pat is in Europe), a $180–200M range is a reasonable base case.

Importantly, unit sales also grew +168% YoY, confirming that growth is driven by real adoption, not pricing.

Cash

On March 31, IBRX strengthened its balance sheet with two moves:

• $75M non-dilutive financing (revenue-based, no equity dilution)

• $25M debt conversion into equity, reducing debt with minimal dilution

As a result, cash rose to $380.9M (from $242.8M), a +57% increase in one quarter. This cash will be deployed to GLOBAL EXPANSION.

3. Pipeline & 2026 Catalysts

NCCN Guideline Update

In March 2026, the NCCN updated its guidelines to include ANKTIVA + BCG for additional patients with papillary-only disease. In simple terms, this means more patients are now eligible for treatment, and doctors have official guideline support to use ANKTIVA.

NCCN guidelines are essentially the playbook that U.S. oncologists and urologists follow, so this update directly supports adoption. It expands the patient pool, reduces hesitation among physicians, and should lead to more prescriptions over time.

This development received little market attention because it coincided with the FDA warning letter. However, from a commercial perspective, this is the more relevant catalyst.

Near-Term Revenue Drivers:

1. The BCG-naïve NMIBC trial is already fully enrolled with 366 patients and is on track for results, with a potential filing in late 2026 to early 2027. This is the most important catalyst, as BCG-naïve patients represent a much larger population than the current indication. If successful, it could significantly expand the market, potentially doubling or tripling the addressable patient base.

2. The company is expanding within its existing market by targeting papillary-only patients. A supplemental BLA was resubmitted in March 2026, which would allow more patients within the same setting to receive treatment. This means growth can come from the same doctors and infrastructure, simply by increasing the number of eligible patients.

3. ImmunityBio is moving into new cancer types, such as lung cancer (NSCLC). The therapy is already conditionally approved in Saudi Arabia, and a Phase 3 trial is ongoing. If successful, this would open an entirely new market beyond bladder cancer.

In simple terms, the company is not relying on a single path: it is expanding the number of patients, broadening its current use, and entering new indications, all of which support future revenue growth.

But the pipeline goes much deeper:

2026 Watchlist: What Confirms or Breaks the Thesis

4. Conclusion

IBRX has crossed the line from speculative platform biotech to commercial-stage company with real revenue, a global regulatory footprint, and funded expansion. The $400 million raised in 2025, the $100 million March 2026 financing package, and the $381 million cash position give management the capital to execute through the key 2026 catalysts.

The FDA warning letter is a self-inflicted wound, a promotional compliance issue that should never have happened and reflects poorly on executive judgment. But it does not change ANKTIVA’s efficacy, its label, its reimbursement, or its clinical pipeline. The class action lawsuits are noise.

The short interest, 135 million shares, 35% of the float creates both risk and opportunity. If the BCG-naïve data reads out positive and revenue continues its sequential acceleration, the short thesis faces a fundamental problem.

The central question for 2026 is not whether ANKTIVA works, as the data clearly shows it does, but whether ImmunityBio can scale the franchise across indications, geographies, and new verticals (see pipeline), expanding the TAM exponentially.

The next three quarters will answer that question decisively.

Track revenue, international traction, BCG-naïve data, and balance sheet management.

Those four metrics tell the real story.

This article is for informational purposes only and does not constitute investment advice, a recommendation, or an offer to buy or sell any securities. The views expressed are my own and may change without notice. Always conduct your own due diligence.

Aehr Test Systems is up almost 10x in less than a year, driven by a simple narrative: AI is creating a new bottleneck in chip reliability, and Aehr sits right in the middle of it.

But behind that narrative, there’s a much more complex reality, a business transitioning between end markets, a revenue model that is not yet visible in the financials, and a valuation that already assumes a large part of the future.

This deep dive is built to separate signal from noise.

We’re going to break down the business from first principles, reconstruct the numbers using filings and management commentary, and then build the valuation framework needed to understand what is actually priced in.

Table of Contents

1. Introduction

2. What Aehr Actually Does (And Why It Matters Now)

3. What the Market Is Missing: Three Non-Obvious Insights

4. What Is the Moat?

5. The Numbers

6. Bookings, Backlog, and the FY2027 Setup

7. Valuation (DFC, reverse DFC, Multiples and Scenario analysis)

8. Risks

9. The perfect entry point (Chart analysis)

10. Conclusion

1. Introduction

Aehr Test Systems is a 48 year old company that makes semiconductor burn-in and test equipment. For most of its history, it was a niche, sub $50M revenue microcap that barely registered on institutional radars. Between FY2022 and FY2024, silicon carbide demand for EVs briefly pushed revenue to $65-66M and the stock to $54. Then EV demand softened, revenue fell, and the stock cratered to $6.27 by Q4 FY2025.

What happened next is why we’re here.

In the span of six months, AEHR began pivoting from a single market EV play toward multi-vertical AI infrastructure, a transition that is underway but far from consolidated. Q3 FY2026 bookings hit $37.2M more than the company booked in the entire first half of the year combined. The effective backlog reached a record $50.9M. The stock has gone from $7.71 to $69, a near-10x move, on the thesis that FY2027 and beyond represent a step-function change in revenue.

One thing to keep in mind: backlog in semiconductor capital equipment is not the same as revenue. Orders are cancellable with limited penalties. Delivery schedules shift. The 10-K says it plainly:

“the Company’s backlog as of a particular date may not be indicative of net revenues for any succeeding period.”

The $50.9M is the strongest demand signal has ever had, but it is a signal, not a guarantee.

The question for us: is this the real inflection point, or are we paying FY2028 prices for FY2026 fundamentals?

This deep dive will answer that by doing what nobody else is doing: reading every page of the 10-K, the 10-Q, two investor presentations, and the most recent earnings call, then building the valuation models that let you decide for yourself.

2. What Aehr Actually Does (And Why It Matters Now)

Why Burn-In Is Becoming Critical

The semiconductor industry has historically treated burn-in as optional for many device types. Standard logic chips, GPUs for gaming, mobile processors… most of these shipped without production burn-in. The failure rates were low enough, and the devices cheap enough, that it was more economical to deal with field failures than to burn-in every chip.

That’s changing fast, and the earnings call gives us the clearest picture of why and how early we are in that change.

1. AI processors are too expensive to fail. An NVIDIA H100 or B200 GPU costs thousands of dollars. When you package it with HBM memory stacks on a CoWoS substrate that costs more than the silicon itself (per Gayn Erickson, Q3 FY2026 call), a single early-life failure destroys a module worth tens of thousands of dollars. The economics of burn-in flip when the cost of failure exceeds the cost of testing. And critically, each new generation breaks the tools that existed for the previous one. As Erickson put it:

   “The processors are getting higher power from generation to generation and breaking all the tools that are out there. Even the tools that were out there... their power is substantially more, which would require new tools. The ones that they’re working on in a year and out break the current tools.”

   This means the TAM isn’t just growing from new customers. Existing customers need new equipment with every generation cycle.

2. Advanced packaging creates a yield multiplier. Modern AI processors use multi-chip packages: multiple compute dies, HBM DRAM stacks, sometimes photonic transceivers, all assembled on a single substrate. If one die fails after packaging, you lose everything. Catching bad die at the wafer level, before they enter the package, saves the cost of the substrate and every other good die in the module. Erickson quantifies this on the call:

“Weeding out bad devices before they’re packed together with these other devices is significantly cheaper than the yield loss if these are burned in at package level and the entire multi-chip package is thrown away.”

And the value equation gets more extreme with every generation: “The CoWoS substrate is more expensive than the silicon itself of the processor, which sounds crazy. They would be very interested in doing the wafer level to screen out the die before they have to throw away everything else.”

3. The installed base of burn-in equipment barely exists and Gayn quantifies it. This is the most surprising data from the Q3 call. When asked what percentage of AI chips are burned in today, Erickson gave specific estimates: only about 5-20% of ASICs (custom chips) undergo production burn-in, and roughly half of AI accelerators do. “I’ve been a little bit surprised at how many devices are not yet doing production burn-in,” he said. He explained the historical context: foundries and OSATs (the contract manufacturers building these chips) historically spent nothing on burn-in. “There were no burn-in systems in the foundry OSAT models. People weren’t spending on it. They spend enormous amounts of money on test, and it’s growing, and they’re going to be spending a significant portion of their test budget on burn-in going forward.” Before Aehr’s Sonoma and FOX solutions existed as viable production options, many companies either burned in at the system/rack level (expensive, slow) or simply shipped devices into their data centers and “dealt with the fallout.” Now there’s a cost-effective alternative, and adoption is accelerating.

4. The TAM is expanding but needs careful framing. This is where the earnings call provides critical context, but also where we need to apply our own discipline. Erickson told Craig-Hallum’s Christian Schwab:

   “The TAMs are multi-hundreds of millions of dollars for package level burn-in. Wafer level burn-in, if you say it displaces package level, is even higher.” His aggregate estimate: “The spend in burn-in probably total spend measured in multiple billions of dollars per year in the next couple of years on an annual basis.”

5. Hyperscale data center capacity is expected to nearly triple by 2030. Erickson cited industry forecasts on the call: this expansion is driven by both new builds and upgrades to existing infrastructure. Every new data center needs AI processors, every AI processor increasingly needs burn-in, and every next-generation processor needs new burn-in tools because the power levels keep rising. The demand driver is structural and compounding.

The Business

AEHR makes machines that stress-test semiconductor chips before they ship. The process is called “burn-in”: you subject chips to elevated voltages, currents, and temperatures for hours or days to weed out the ones that would fail in the field. Think of it as a stress test for chips the same way a bank stress test works: find the weak ones before they cause real damage.

The Two Product Platforms

AEHR operates with two complementary product lines:

1. Wafer-Level Burn-In (FOX Platform): AEHR tests entire wafers at once, before chips are cut and packaged. Instead of testing chips individually at the end of the manufacturing process, the FOX platform stresses all chips on a wafer simultaneously under extreme conditions. Defects are caught before expensive packaging, which is especially critical for AI chips going into multi-die architectures where a single bad chip destroys a module worth thousands of dollars.

Wafer-level burn-in filters out weak chips early, improving yield and reducing total system cost. Originally proven in silicon carbide, this platform is now expanding into AI processors, silicon photonics, GaN power devices, and next-gen memory.

2. Package-Level Burn-In (Sonoma/Tahoe/Echo Platform): AEHR tests chips after they’ve already been packaged, which is the current industry standard. The Sonoma platform is designed for very high-power devices: AI accelerators, GPUs, custom hyperscaler chips. The key advantage is that customers can run burn-in without changing their chip design, no redesign required, easier qualification, immediate deployment. This makes it the fastest path to adoption, and it’s the platform that won AEHR its first hyperscaler production customer.

Why having both matters (this is the edge). Most competitors offer one or the other. AEHR offers both, and this creates a natural customer progression:

1. Customer starts with package-level (Sonoma): low friction, no chip redesign needed.

2. Ramps into production at scale: more systems, more consumables.

3. Designs next-generation chips for wafer-level (FOX): Chips designed with built-in test features can be tested more effectively, making advanced burn-in both possible and more valuable.

The result: higher revenue per customer, longer customer lifecycle, and multiple consumable streams from each relationship.

The hyperscaler customer is the concrete proof. Erickson walked through it step by step on the Q3 call:

“The first device they ran with us... is on Sonoma. Their second device, they just awarded us with production for that one and are planning the ramp of that with us right now. They’re already on the roadmap talking about the third device, and they’ve asked us about the DFT to specifically put into the third device because they would like to consider that for wafer level on our FOX systems.”

Half of the companies currently engaged for package-level qualification “have also expressed interest in wafer-level burn-in in addition to our package-level burn-in solutions.” Erickson described the customer dynamic on the call: “The last two customers that were in the last two weeks... the customer will come in and say, ‘I want to talk about package level.’ About halfway through the tour, they’re like, ‘What is that?’ We talk about wafer level, and they’re like, ‘Whoa, how do I do that?’”

Why this matters for the model: AEHR’s revenue per customer should increase over time as relationships deepen:

A typical customer might start with just a handful of Sonoma systems to qualify burn-in. If the process works, that small initial deployment can quickly scale into dozens of systems as production ramps. And as chip designs evolve (especially in AI), that same customer may begin adopting wafer-level burn-in on FOX, adding another layer of systems and consumables on top.

In other words, what starts as a relatively modest deal can evolve into a multi-year, multi-phase relationship. The key insight is that revenue per customer doesn’t stay flat, it grows over time, often significantly, as the relationship deepens and expands across both platforms.

Revenue Composition

Revenue breaks into three categories, with a meaningful shift underway:

AEHR reports revenue across three categories: systems, contactors (consumables), and services. But to understand what’s actually changing in the business, it’s important to look one layer deeper.

The most significant shift in FY2025 came from the packaged parts business (Incal / Sonoma), which contributed approximately $19.8M, or 34% of total revenue. This sits largely within the systems category (which represented 37% of revenue), highlighting how quickly Sonoma has become a core driver of system sales following the acquisition.

At the same time, the contactor line (WaferPaks, BIMs, BIBs) remains a critical part of the model. These are consumables, they wear out over time and need to be replaced every 2–7 years depending on usage and device lifecycle. As a result, they represent the foundation of AEHR’s recurring revenue.

Historically, consumables have made up a significant portion of revenue, 52% in FY2025 and 57% in FY2024, reflecting a strong installed base and ongoing replacement cycles.

In Q3 FY2026, this mix dropped to 29%, but management explicitly framed this as temporary, driven by customers who “bought systems ahead of the need and have grown into capacity.”

Implication: AEHR is not just selling systems, it is building an installed base that drives recurring revenue over time. As more systems are deployed (particularly through Sonoma), the base of consumables expands. Over time, this should translate into:

• More stable revenue

• Higher margins

• Increasing revenue per customer

In that sense, what looks like a traditional semicap business begins to resemble something closer to a hybrid model: part equipment / part recurring revenue platform.

3. What the Market Is Missing: Three Non-Obvious Insights

There are three things the market is missing. Two we’ve already covered, and one we’ll explore next:

1. The Incal acquisition is generating an incredible ROI

AEHR paid $22.2M for Incal in July 2024. In its first year, Incal contributed $18.6M in revenue and $3.8M in net income (from the 10-K, Note 4). That’s a 17% net income return on purchase price in Year 1. But the real value is what Incal unlocked: the hyperscaler production win, which alone could drive $40-70M in FY2027 revenue. If that materializes, AEHR will have generated 2-3x the acquisition price in revenue from Incal-derived customers within two years of closing. For context, the median M&A deal in semi equipment takes 3-5 years to generate a positive return. AEHR did it in under 12 months.

Important: this is a one-year snapshot during the initial honeymoon period. Integration risk is ongoing (they just consolidated the Incal office and took restructuring charges). The $3.8M net income benefits from Incal’s existing customer base, sustaining and growing that requires continued execution. And the cyclicality of semi equipment means Year 1 returns are not indicative of lifetime returns. Still, the strategic value is durable regardless of near-term cyclicality.

2. The consumable revenue drought is masking the true margin structure of the business

Gross margins have collapsed from 50% to 31%. Most analysts attribute this to the SiC downcycle and Sonoma’s lower margin profile. That’s partially right but there’s a deeper structural story buried in the revenue mix data.

In FY2023, contactors (WaferPaks + BIMs) were 34% of revenue at 50% gross margin. In FY2024, contactors surged to 57% of revenue as the SiC installed base bought replacement WaferPaks and gross margin held at 49%. Now in FY2026, contactors have crashed to 29% of revenue because SiC customers “bought systems ahead of the need and have grown into capacity” (CEO, Q3 call). The margin decline isn’t just about Sonoma’s lower margin, it’s about the highest-margin revenue stream temporarily drying up.

Here’s what matters: the CEO explicitly said this is “an outlier” and targets consumables at “30% or more of total revenue” with “margins will increase as sales of these value-add consumables grow.” As the installed base (which is now much larger thanks to Sonoma) begins generating consumable pull-through, the margin trajectory should inflect upward independent of product mix. Wall Street models gross margin as a function of revenue mix between wafer and package. The real driver is the consumable percentage within each category.

At the midpoint ($20M), consumable revenue would represent ~14% of my FY2027E revenue of $142M, well below the 30% target. For consumables to reach 30% ($43M), either the installed base needs to roughly double from today’s levels (plausible given the backlog) or per-system consumable spending needs to increase as customers add new device types to existing systems (also plausible as AI chip generations turn over every 12-18 months). Both dynamics are likely by FY2028.

3. Half the pipeline is requesting both platforms and nobody is modeling the revenue-per-customer expansion this implies

This is the most underappreciated data point from the entire earnings call. Erickson said that of the companies currently engaged for package-level qualification, “about half of these have also expressed interest in wafer-level burn-in in addition to our package-level burn-in solutions.”

The key insight is simple: customers are not adopting a single platform, they’re moving across both. What starts as a small Sonoma deployment can scale into a multi-platform relationship, with FOX layered on top. The result is a step-change in revenue per customer that is not yet reflected in current financials.

4. What Is the Moat?

Aehr’s advantage is real but it’s not where most people think.

This is not a story about selling machines. It’s about controlling the interface between the chip and the test system.

Before we go further, two definitions:

WaferPak: a custom interface that connects an entire wafer to Aehr’s system, allowing thousands of chips to be tested simultaneously. Think of it as a full-wafer probe card with up to 50,000 contact points, designed to operate at 150°C and deliver thousands of amps.

DiePak / BIM: custom carriers that hold and connect individual packaged chips or die for burn-in testing. Each one is designed specifically for a customer’s device.

Moat

Proprietary wafer-level capability. Aehr’s core differentiation is its ability to test entire wafers at once. This requires extremely complex contact technology, reliably connecting thousands of chips under high heat and power, maintaining uniform temperature across a 300mm wafer, and doing it across 18 wafers in parallel in a single system. No competitor has demonstrated this at scale for AI processors. Aehr holds 131 active patents across the US, Singapore, China, Japan, and Korea (10-K, p. 9-10), with expirations ranging from 2025 to 2045. This is where the IP matters.

Consumables-driven lock-in. Every chip type requires a custom interface, a WaferPak for wafer-level, a BIM/BIB for package-level. These are designed specifically for each device. Once qualified, they become embedded in the customer’s production process. Switching vendors would require redesigning the interface and re-qualifying the entire workflow, a process that takes months and costs hundreds of thousands of dollars. That’s slow, expensive, and risky, which creates real stickiness.

At the same time, these components wear out and need replacement every 2-7 years depending on device lifecycle (10-K, p. 8). This turns the model into a razor-and-blade business: sell the system once, sell consumables for years. The CEO’s target: consumables at 30%+ of revenue, at higher margins than systems. His exact words on the Q3 call: “our margins will increase as sales of these value-add consumables grow.”

Dual-platform positioning. AEHR is the only company offering both package-level (Sonoma) and wafer-level (FOX) burn-in at scale. This is not just product breadth, it shapes how customers adopt the technology. Customers often start with package-level and expand into wafer-level over time. What looks like two products is actually a progression path and AEHR captures value at every stage. (we already cover this)

Manufacturing capacity. Aehr’s Fremont facility (50,000+ sq ft) supports up to 400 wafer-level blades and 20 packaged-part systems per month (investor presentation, p. 3). Add the contract manufacturer chain now qualified for Sonoma production (20+ systems/month), and total Sonoma capacity reaches 40 systems/month. Combined capacity annualizes to $240-336M in Sonoma revenue alone, far above current demand. A competitor would need 12-18 months minimum to replicate this supply chain even with a competitive product.

The moat is not the system, it’s the installed base + consumables + qualification lock-in. Once a customer adopts AEHR, switching is difficult, revenue becomes recurring, and the relationship expands over time. Aehr is not defending a product. It’s defending a position inside the manufacturing process. And once it’s in, it’s very hard to remove.

5. The Numbers

Revenue

The decline in revenue from FY2024 to FY2026E is largely explained by one factor: the downturn in silicon carbide (SiC) tied to the EV cycle.

During the upcycle, SiC was a major driver of demand for Aehr’s wafer-level burn-in systems and, more importantly, its consumables (WaferPaks). As EV expectations reset and customer demand slowed, this segment pulled back sharply.

Contactor revenue declined from $37.6M in FY2024 to $9.1M in the first nine months of FY2026, reflecting a sharp drop in SiC-related demand. This implies an annualized run-rate of roughly $12M, although this likely understates normalized levels given current underutilization.

Systems revenue has been more resilient, partially supported by the contribution from Sonoma (Incal). However, this has not been enough to fully offset the decline in high-margin consumables.

The critical data point is buried in the 10-Q:

Wafer-level contactors (WaferPaks) declined by $21.2M in the nine months ended February 2026 versus the prior year period, reflecting the sharp slowdown in SiC demand.

At the same time, the newer AI-driven segments are growing: package-level consumables (BIMs/BIBs) increased by $3.6M, and package-level systems by $2.5M.

The key point is that this shift is not visible in the headline numbers. Within the contactor line, the legacy SiC business is contracting rapidly, while the AI/Sonoma business is scaling, just not yet at a level sufficient to offset the decline.

What looks like a broad-based slowdown is, in reality, a transition between end markets.

Margins and cost structure

Gross margins have declined materially, driven by lower utilization, unfavorable mix (package-level carries lower margins than wafer-level), and higher costs, including tariffs on imported components.

Management expects margins to improve as volumes recover and fixed costs are better absorbed.

A key structural point: wafer-level products carry higher margins than package-level, while consumables carry higher margins than systems. As a result, margins will depend on two factors: recovery in wafer-level demand and re-acceleration of consumables tied to installed base utilization.

On the cost side, OpEx has increased meaningfully following the Incal acquisition, higher stock-based compensation ($5.2M in FY2025 vs. $2.5M in FY2024), and continued investment in R&D.

At the current cost base ($30M annually), the company requires a meaningful step-up in revenue to return to profitability. Until volumes recover, the model remains under-absorbed.

Balance sheet

The balance sheet is clean, but the cash trajectory matters.

Cash declined from $49.2M to $24.5M between May 2024 and May 2025, driven by the Incal acquisition, capex, and operating cash burn. It then recovered to $36.9M as of February 2026, primarily due to $19.6M raised through ATM equity. Without the ATM, cash would be meaningfully lower.

The company has been free-cash-flow negative over the past two years, with $5.1M of operating cash burn in the first nine months of FY2026

6. Bookings, Backlog, and the FY2027 Setup

Record Backlog

This is where the bull case lives.

The Bookings Inflection

The $37.2M in Q3 bookings is the largest quarterly booking number in company history. It was driven by: the $14M follow-on order from the lead AI wafer-level customer, Sonoma orders from the hyperscaler for their next-gen processor, SiC orders from a new Taiwan customer, and the initial silicon photonics customer order.

Effective backlog at the time of the earnings call (April 7, 2026): $50.9M, consisting of:

• $38.7M quarter-end backlog (Feb 27, 2026)

• Plus $12.2M in bookings received in the first five weeks of Q4

This is a company record and represents more than a full year of trailing revenue.

What Management is Guiding

FY2026 (ending May 29, 2026): Revenue at “the high side of the $45-50M range.” Non-GAAP EPS of -$0.13 to -$0.09. Expect return to non-GAAP profitability in Q4.

FY2027: No specific guidance, but management stated they are “well positioned to exit the fiscal year ending May 29th with a strong backlog and deliver significant revenue growth in fiscal 2027.” The backlog alone ($50.9M) exceeds FY2026 full-year revenue guidance.

Share Nicolas

Up to this point, we’ve broken down the business from first principles: what AEHR does, why burn-in is becoming critical in AI, where the moat actually sits, and how the current numbers reflect a transition rather than a decline.

But none of that answers the only question that ultimately matters:

What is this business actually worth today?

In the next section, we move into the most important part of the analysis valuation.

We’ll build it from multiple angles: a detailed DCF, a reverse DCF to understand what the market is pricing in, relative multiples, and a scenario framework to map outcomes.

From there, we’ll cover the key risks, the catalysts that could drive the next move, and where the optimal entry point may be from both a fundamental and technical perspective, before closing with a clear conclusion.

7. Valuation

At $70 per share ($2.2B market cap), Aehr is no longer a hidden story.

The stock is up 800% from its 52-week low of $7.71. The market is not paying for what Aehr is today, a $49M revenue company losing money. It’s paying for what Aehr could become: the dominant supplier of burn-in equipment across AI, photonics, power semiconductors, and memory.

The question isn’t whether the story is real. It is. The question is whether the price already reflects the outcome.

Read more

What the market hasn’t internalized is simpler and more uncomfortable: a large chunk of the supply that’s supposed to meet that demand won’t show up on time.

We now have data to back that up and the numbers have gotten worse since most people last checked.

The pipeline is wildly overstated

Sightline Climate tracks 777 large data center projects (>50 MW) announced since 2024, roughly 190 GW of headline capacity. For 2026 alone, the industry is pointing to 16 GW across 140 projects.

Here’s the problem: only 5 GW of that is actually under construction right now. The other 11 GW is still at the “announced” stage, with no visible construction progress, despite typical build timelines of 12–18 months.

Announcements don’t build data centers.

And this isn’t just a theoretical risk. In 2025, 26% of expected capacity got delayed, and another 10% pushed back timelines without any public notice. If that track record holds, 30–50% of the 2026 pipeline probably doesn’t hit its dates. Sightline puts it plainly in their February 2026 outlook: half of the 2026 pipeline may not materialize.

Subscribe now

Nearly half of all projects in the database haven’t disclosed a power strategy at all. No confirmed source, no timeline. That’s not a delayed project, that’s a spec sheet dressed up as a pipeline entry.

Goldman keeps revising the demand number up

Meanwhile, demand estimates keep moving in the other direction.

Goldman Sachs has now updated their data center power demand forecast three times in under two years.

The original estimate: +165% by 2030 vs. 2023 levels. Then their GS Sustain team revised it to +175% in November 2025. In March 2026, they revised again, now projecting +220% by 2030, driven by a $300B+ increase in combined hyperscaler capex and R&D budgets for 2026–2027 alone.

To put that in physical terms: Goldman’s base case is the equivalent of adding a new top-six power-consuming country to the global grid by the end of the decade. By 2030, AI workloads alone are expected to represent roughly 39% of total data center power demand, up from 14% today.

The Goldman framework worth understanding: the 6 Ps

Goldman’s GS Sustain team built their analysis around six factors, the “6 Ps”, that either accelerate or constrain data center power demand. It’s a useful map for thinking about where the real risks sit.

Pervasiveness is how widely AI actually gets deployed across industries. If adoption disappoints, demand growth slows.

Productivity is the efficiency of compute per watt, and this is where the “DeepSeek fixes everything” argument falls apart. Goldman’s base case assumes hyperscaler budgets, not physics, are the real binding constraint. If budgets are unconstrained, upside dominates; efficiency improvements get absorbed by more deployment, not less power consumption. The GB200 and GB300 architectures are significantly more power-dense than what preceded them, and net total demand keeps rising regardless.

Prices and Policy set the cost and regulatory environment for new megawatts. Hyperscalers are already paying what Goldman calls a “green reliability premium”, they’re willing to pay more just to get power faster. The potential phase-out of IRA incentives post-2028 is a policy risk, but it’s not changing near-term procurement behavior. Parts, transformers, switchgear, turbines, cooling plants… have 2+ year lead times and the entire industry is competing for the same constrained pool of equipment.

The P Goldman is most concerned about: People. They estimate the US alone needs approximately 500,000 new jobs to build out the generation capacity and grid infrastructure that AI data centers require, with a shortage of 78,000 electricians and transmission/distribution technicians, roles that take up to four years to train. You can order transformers; you can’t rush a licensed electrician.

Share Nicolas

The bottleneck is physical, not financial

This is the part most financial models quietly skip. The constraints aren’t on the demand side and they’re not a function of capital availability, hyperscalers have committed $736B in capex for 2025–2026 alone. The constraints are physical and logistical: grid interconnection queues, transformer lead times, skilled labor, transmission capacity.

The hyperscalers have already reached that conclusion. On-site and hybrid power projects represent less than 10% of total project count but nearly half of total announced capacity. Google acquired Intersect Power’s 10.8 GW pipeline. Amazon moved directly into solar and storage at the project level.

They’re not waiting for the grid to catch up, they’re building around it.

The market is still pricing GPU supply and capex cycles. The real variable is time-to-power: who has it secured, when, and how much. That’s what separates executable capacity from press release capacity.

Where IREN fits and what the numbers actually say

This is the thesis for IREN. They’re not in the announcement category.

IREN has >4.5 GW of grid-secured power across multiple sites. Sweetwater 1 (1.4 GW) is targeting energization in April 2026. Childress (750 MW total campus) is under active construction, with Horizon 1–4 delivering 200 MW of liquid-cooled critical IT load in phases through 2026 for Microsoft. Their British Columbia air-cooled sites are already live and generating revenue, which is how they were in a position to place a purchase order for 50,000 B300 GPUs for phased deployment through H2 2026. A new 1.6 GW campus in Oklahoma brings the total secured power portfolio above 4.5 GW.

The Microsoft contract is the anchor: $9.7B over five years, with Microsoft providing a 20% prepayment per tranche. As of Q2 FY26, IREN also has $0.4B ARR under contract at Prince George (BC), with ongoing negotiations the company says could support an additional >$0.5B. Combined, that’s roughly $2.3B of ARR contracted, though the filing is explicit that “contracted” doesn’t mean revenue-generating until the relevant GPUs are delivered, commissioned, and in service. Execution against the Childress deployment schedule is the key variable to watch.

The stated target was $3.4B ARR by end of CY2026 across 140k GPUs, Microsoft Childress ($1.9B) plus the remaining $1.5B coming from BC air-cooled deployments and smaller contracts with Together AI, FluidStack, and Fireworks AI. With the subsequent 50k B300 order for BC bringing the total fleet to 150k GPUs, the revised target moves above $3.7B. The gap between today’s $2.3B contracted and either target is almost entirely a construction and delivery execution question, not a demand question.

Subscribe now

What the Market Is Pricing and What Would Change It

At $35 per share and $12B EV, the market is pricing roughly 3x forward ARR and ~4.6x CY2027E EBITDA. That's a meaningful discount to CoreWeave (13-15x CY2026E Adj. EBITDA, compressing to 5-6x on its exit run-rate) and Nebius (6x CY2026E revenue). Both peers carry more debt (CoreWeave $21B) or less proven revenue (Nebius $228M quarterly), but trade at materially higher EV multiples, reflecting their larger backlog and further-along deployment timelines. The discount is warranted today, IREN has $17M/quarter in AI Cloud revenue, heavy customer concentration, and a $6B ATM overhang, but it creates asymmetry if execution confirms the thesis.

The catalysts:

1. Q3 FY26 earnings (May 13, 2026). First quarter with meaningful Microsoft revenue. If AI Cloud revenue jumps from $17M to $50-100M, the ramp is real.

2. Sweetwater energization. Confirmation that the 1.4GW substation is live removes a binary risk. An anchor tenant announcement could add $10-20B in EV.

3. ATM usage disclosure. Every quarterly filing shows share issuance. Lower-than-feared dilution is bullish.

4. CY26 exit run-rate. The $3.7B ARR number needs to be approached credibly by Q4.

The current $11.7B market cap is pricing almost none of the Sweetwater or Oklahoma optionality, which is either an opportunity or a reflection of how much can still go wrong between here and there.

The Revenue Build

IREN’s revenue is splitting into two streams moving in opposite directions.

Bitcoin mining is the dying business. Q2 FY26: $167.4M, down from $233.0M in Q1. IREN is literally ripping ASICs out of BC data centers and replacing them with GPUs. This line goes to near-zero within 12-18 months.

AI Cloud is the entire thesis. $17.3M in Q2, up from $7.3M in Q1. Early-stage ramp at Prince George. The real inflection starts when Horizons 1-4 begin delivering to Microsoft (Q2 CY26) and Mackenzie/Canal Flats deploy the 40,000 additional GPUs ($1.0B ARR). Revenue recognition will be lumpy through CY26, these are not SaaS contracts. GPUs need to be physically delivered, installed, commissioned, and accepted. The Microsoft Agreement includes a delivery acceptance process and termination rights for missed delivery dates.

Valuation (2026 - 2030)

Here’s the model, with explicit assumptions at each stage:

Read more

There is a structural problem in how the United States builds mission-critical metal parts, and it has been hiding in plain sight for decades.

A jet engine turbine blade. A rocket combustion chamber. A ground combat vehicle bracket that fails under stress. These components share something in common: they are manufactured using casting, forging, and machining processes that were optimized for a world of stable supply chains, predictable lead times, and high-volume production runs. That world no longer exists.

Traditional metal manufacturing operates on a simple logic: pour metal into a mold, cut it to shape, repeat thousands of times to amortize the tooling cost. It works when you need 50,000 identical brackets for a commercial vehicle. It breaks when you need 200 highly complex propulsion components for a classified defense program, and you need them in six months instead of eighteen, and the only foundry qualified to make them is operating at full capacity with a two-year backlog.

This is not a hypothetical. The U.S. Department of Defense has publicly acknowledged that fragile, geographically concentrated supply chains are a national security vulnerability. Traditional casting and forging for defense and aerospace applications routinely involve lead times of 12 to 18 months. A single casting defect can force a full tooling redesign, adding cost and delay that low-volume defense programs, unlike automotive, cannot absorb through scale. The Defense Industrial Base is struggling with exactly this kind of bottleneck.

Metal additive manufacturing, 3D printing metals, layer by layer, from a digital file; offers a fundamentally different approach. No tooling. No molds. Design-to-part in weeks instead of months. Parts with internal geometries that casting physically cannot produce. And critically, a digital file that can be sent anywhere and printed on any qualified machine, decoupling manufacturing from geography.

The macro tailwinds here are converging from multiple directions simultaneously.

Re-shoring and supply chain resilience. The U.S. government is actively investing in domestic manufacturing capability. The Department of Defense awarded $23.4 million under the Defense Production Act specifically to expand domestic production of advanced metals for aerospace, defense, and space applications. Programs like Project FORGE (Fabrication of Operationally Responsive and General-purpose Elements) are explicitly designed to eliminate manufacturing bottlenecks by qualifying additive manufacturing as a production pathway for critical weapon systems.

The defense budget is flowing toward AM. The aerospace and defense additive manufacturing market was valued at approximately $5.2 billion in 2025 and is projected to reach $11.5 billion by 2030, growing at a 17.1% CAGR (The Business Research Company, 2026). The broader metal AM market follows a similar trajectory: $6.0 billion in 2025 to $13.1 billion by 2030 at 16.9% CAGR. North America accounts for roughly 40% of the A&D AM market. Direct Metal Laser Sintering (DMLS), the technology category Velo3D operates in, holds the largest share at approximately 32%.

Subscribe now

Within this broader market, the serviceable addressable market for large-format, mission-critical metal parts, the specific niche where Velo3D’s SupportFree technology is most differentiated, is estimated at $5-10 billion by 2030. At $65M in guided 2026 revenue, Velo3D is penetrating less than 1% of its addressable market. The runway is enormous. The question is whether the company survives long enough to capture it.

The space economy is pulling demand forward. Reusable launch vehicles require propulsion hardware, combustion chambers, injectors, heat exchangers… with geometries that push the limits of what traditional manufacturing can achieve. Companies building rockets need parts that are lighter, stronger, and producible in weeks, not quarters. Additive manufacturing is not an option for these applications. It is increasingly the only viable path.

The physics of the problem. A rocket engine combustion chamber has hundreds of internal cooling channels. A traditional approach requires manufacturing each channel separately and brazing them together, a multi-step process with failure points at every joint. Additive manufacturing can produce the entire structure as a single part, eliminating joints, reducing weight, and improving thermal performance. This is not incremental improvement. It is a category shift in what is physically manufacturable.

But here is the nuance that most macro narratives miss: the transition from prototyping to production-scale additive manufacturing is not automatic. Legacy AM systems have been printing metal parts for nearly 30 years. Most of them remain stuck in prototyping because they cannot deliver the repeatability, quality control, and qualification rigor that production programs demand. A defense program will not accept a part unless the manufacturer can prove it will come out the same way, on the same machine, on a different machine, in a different facility, every single time.

This is the gap. The macro demand is real. The TAM is large and growing. The policy tailwinds are strong. But the bottleneck has shifted from “can we print metal?” to “can we print metal at production scale, repeatably, with mission-critical reliability?”

That is the question Velo3D is trying to answer. Whether they can with $39 million in cash, a going concern flag, and 15 printers on their floor is the subject of the next seven sections.

2. The Business: What Velo3D Does and How It Makes Money

Velo3D is a metal additive manufacturing company headquartered in Fremont, California. It designs, assembles, and sells large-format metal 3D printers, the Sapphire family, along with proprietary software and production services. The company was founded in 2014, went public via SPAC merger in September 2021, nearly went bankrupt in 2024, and is now attempting what may be the most ambitious business model pivot in the AM industry.

To understand the investment thesis, you need to understand the four layers of the Velo3D platform and why the company believes they create something greater than the sum of their parts.

The Integrated Platform: Hardware + Software + Quality + Process

Layer 1: Flow - Print Preparation Software. Before a metal part can be printed, someone has to translate a 3D design file into precise laser instructions. Flow is Velo3D’s proprietary software that does this. It generates custom laser tool paths, the exact patterns the laser will follow, layer by layer, optimized for the specific geometry of each part. This is where Velo3D’s SupportFree capability originates. Traditional laser powder bed fusion (LPBF) systems require internal support structures to hold parts in place during printing. These supports must be manually removed after printing, which adds cost, limits geometric complexity, and introduces failure points. Flow’s algorithms can prescribe tool paths that eliminate or drastically reduce the need for supports, enabling geometries, internal channels, overhanging features, thin walls, that other LPBF systems simply cannot produce.

Flow Developer, the latest version, gives customers controlled editor-level access to print parameters within validated process boundaries. This is a significant detail: it allows users to customize settings for specific applications without breaking the “Golden Print File” methodology that ensures repeatability.

Layer 2: Sapphire - The Printer Family. Velo3D sells four printer configurations, all based on the same LPBF technology. The Sapphire XC (”extra capacity”) is the second-generation workhorse, 600mm diameter by 550mm tall, up to 155 liters, roughly 400% larger build volume than the original. The 1MZ variants of both extend the build height to one full meter, enabling the production of tall components like rocket engine chambers and large structural parts.

The Sapphire XC is the system at the center of both the RPS strategy and the defense pipeline. All recipes developed for the original Sapphire are fully compatible with the XC line, a design choice that enables the distributed manufacturing model the company is building toward.

The machines print from 16 qualified metal alloys: titanium, nickel superalloys (Inconel 718, Hastelloy X), aluminum, copper alloys, stainless steel, and others. Assembly takes 8 to 15 weeks depending on configuration, with final assembly, calibration, and testing performed in-house at Fremont.

As of December 31, 2025, Velo3D has shipped over 126 Sapphire family systems. There are approximately 140 printers in the field and 15 on Velo3D’s own production floor.

Layer 3: Assure - Quality Validation Software. Assure monitors the print process in real time using data from approximately 1,000 sensors embedded in each Sapphire system. It tracks layer-by-layer build quality and flags deviations. The platform is designed to enable “Closed Loop Control” using sensor feedback to calibrate laser systems within prescribed tolerances during the build itself.

Layer 4: Intelligent Fusion - The Binding Layer. Intelligent Fusion integrates Flow, Sapphire, and Assure into a single closed-loop system. It manages information flow, sensor data, and laser control across the entire print cycle. This integration is the foundation for the “Golden Print File” concept, a validated set of build instructions that can theoretically reproduce the same part across any Sapphire system, in any location, with consistent geometry, material properties, and performance.

How Velo3D Makes Money: Two Revenue Engines

Engine 1: Equipment Sales (the legacy model). Velo3D designs, manufactures, and sells Sapphire systems to customers who operate them in-house. A Sapphire XC system sells for roughly $1 million or more. The typical order-to-site-acceptance cycle is three to nine months. The company also historically offered operating leases but this model has been wound down to near zero (0.2% of 2025 revenue).

For FY2025, 3D Printer and Parts revenue was $39.2 million, representing 85% of total revenue. This line item now includes both system sales and RPS printed parts revenue.

Engine 2: Rapid Production Solutions (RPS, the new model). This is the strategic pivot. Instead of selling a $1M+ printer to a customer, Velo3D operates its own fleet of Sapphire XC systems and manufactures parts for customers directly. The customer gets the part without the capital investment, the qualification timeline, or the operational complexity of running their own AM capability.

The economics are meaningfully different. On the earnings call, CEO Arun Jeldi stated that RPS margins are 40% to 60% depending on the project, compared to approximately 35% gross margin on a newly built system sold directly. RPS was 10-15% of 2025 revenue, and management expects that percentage to roughly double each year, targeting up to 40% of 2026 revenue.

The supplementary slides add critical color: RPS backlog increased 230% from Q3 to Q4 2025. Average deal size tripled from approximately $140,000 in Q1 to $450,000 in Q4. Ten customers who placed their first order in 2025 already made significant repeat orders within the same year. And 79% of bookings came from existing customers.

Revenue Composition: FY2025

Total revenue: $46.0 million (up 12% YoY headline, or 26-28% excluding the one-time $5M SpaceX patent license from 2024).

By segment: 3D Printer and Parts (including RPS): $39.2M (85.2%). Support Services: $6.2M (13.5%). Recurring Payment (leases): $0.1M (0.2%). Other: $0.5M (1.1%).

By end market: Defense ~56%, aerospace/aviation ~24%, energy/semiconductor/other ~20%.

By geography: Americas $39.9M (87%), Europe $4.5M (10%), other $1.6M (3%).

Bookings for FY2025 were $59 million, up 90% from $31 million in 2024. Backlog at year-end was $31 million (up 94% from $16 million), all within 12 months duration. Q4 2025 bookings were the largest quarterly bookings in company history.

3. The Customer Map: Who Is Actually Buying

Velo3D does not name its top customers in the 10-K financial tables, they appear as “Customer 1, 2, 3...” But by cross-referencing the 10-K, earnings calls, investor conference transcripts, press releases, and SEC filings, a detailed picture of the customer base emerges. This matters because the quality, diversity, and stickiness of the customer base is what determines whether the RPS pivot has a real foundation or whether it is a handful of contracts dressed up as a business model.

Tier 1: Named Strategic Customers

SpaceX. The foundational customer. Over 126 Sapphire systems shipped since 2018, the largest fleet of any single customer by far. Five major parts in current Raptor engines are qualified specifically for the Velo3D platform. The company is already printing components for the next-generation Raptor 4 engine (confirmed by the CEO at the iAccess Alpha conference, December 2025). SpaceX was an early strategic investor.

The SpaceX relationship remains strategically important, it validates the technology in the most demanding application imaginable, but its direct revenue contribution is declining as the customer base diversifies.

Anduril. Confirmed as an active customer by the CEO at the iAccess Alpha conference in December 2025: “Anduril is currently the customer of Velo. We are combining our forces to get more contracts together.” Velo3D is working on multiple Anduril programs and actively expanding both the number of parts and programs with the defense technology company. Given that Anduril is widely expected to IPO in the coming years, being embedded as a key manufacturing supplier has significant strategic value. Contract values have not been disclosed, but if the relationship is growing rapidly, Anduril could be one of the top 3 customers for 2026.

RTX Corporation (Raytheon). Named in the 10-K as a current customer relationship. Publicly collaborating with Raytheon Technologies Research Center on the DEVCOM AvMC program ($4M) to develop aluminum CP1 LPBF processes for air and missile defense. RTX is the most likely candidate for the unnamed “key U.S. defense prime contractor” behind the $11.5M full-rate production RPS contract, though this is not confirmed.

Lockheed Martin. Named in the 10-K and by the CEO at the December 2025 conference as an active commercial relationship. Collaboration on ramjet/scramjet components through the LIFT program has been publicly documented.

General Dynamics. Named by the CEO at the December 2025 conference alongside Lockheed Martin and Kratos Defense as active commercial relationships.

Kratos Defense. Named by the CEO at the December 2025 conference. Kratos manufactures drones and defense systems at scale, a natural end market for rapid AM production of metal components.

Honeywell International. Named in the 10-K as a historical relationship and in earlier Velo3D materials as a recurring customer. Unclear whether the relationship generates significant current revenue.

L3Harris (formerly Aerojet Rocketdyne). Identified by the CEO as a relationship in the context of defense propulsion. Historical partnership on titanium thrusters.

Tier 2: Customers and Partners with Known Contract Values

Department of War / Defense Logistics Agency (DLA). $9.8M IDIQ contract, 5 years, under the JAMA Pilot Parts Program. Announced late March 2026. This is direct production of spare parts for the defense supply chain: Army, Navy, Air Force, Marines, and Space Force. An IDIQ (indefinite delivery/indefinite quantity) is a vehicle that allows the government to place direct orders against the contract ceiling. Not a guarantee of $9.8M revenue, but establishes Velo3D as a qualified supplier with a direct procurement pathway.

Defense Innovation Unit (DIU) / U.S. Navy (Project FORGE). $32.6M contract. Supports rapid prototyping and qualification of AM components for a major weapon system program. Collaboration with DIU, the U.S. Navy, and a key industry partner. Includes the potential to develop what could be the largest format LPBF printing capability in the U.S. Milestone-based — revenue recognition depends on meeting qualification and production milestones.

Unnamed U.S. Defense Prime. $11.5M multi-year RPS contract. Delivers essential system components for a sensitive national security program entering full-rate production. This is not prototyping — it is sustained production, which validates RPS in a defense context at scale.

The AR Concentration Warning

While revenue concentration improved (top 3 = 33.2%, down from 47%), accounts receivable tells a different story. Customer 2 holds 34% of AR ($2.1M of $6.3M total). Customer 4 holds 16.2%. That is 50% of receivables in two customers. The company also booked $1.9M in credit loss provisions on trade receivables and $1.4M on contract assets in 2025, someone is not paying on time. If the concentrated AR belongs to an early-stage space company or a financially stressed counterparty, the collection risk is material.

4. The Competitive Moat: SupportFree, Intelligent Fusion, and the Data Advantage

Every metal AM company claims differentiated technology. The question is whether Velo3D’s differentiation is real, defensible, and economically meaningful.

After reading the entire 10-K, the earnings call, and the supplementary slides, the honest answer is: the moat is real today, narrow but deep, and its durability depends on whether Velo3D can scale RPS fast enough to convert a technology lead into a data lead before the window closes.

What SupportFree Actually Means

In LPBF, traditional systems require internal support structures that hold parts in place during printing and must be manually removed afterward. These supports restrict geometry (anything overhanging beyond ~45 degrees needs support underneath), add 30-50% to post-processing time, and introduce failure points at every contact point.

Velo3D’s proprietary process eliminates or drastically reduces supports through a non-contact recoater system (the blade spreading powder doesn’t touch the part) and Flow’s algorithms that manage thermal stress without relying on supports. The result: geometries that no other commercial LPBF system can produce without redesign.

The IP Position

As of December 31, 2025: 57 issued patents (45 U.S., 12 foreign), 34 pending applications. The earliest patent expires in 2035, the latest in 2047. The portfolio covers systems, methods, software, and even the printed parts themselves, the 10-K notes that parts produced using Velo3D technology “have a signature that is readily recognizable and traceable.”

The SpaceX license from September 2024: SpaceX obtained a non-exclusive license to certain Velo3D patents for internal use for $5 million. The license is non-exclusive, Velo3D retains all rights with other customers. A one-time monetization event, not a sign of customer defection, but a data point to monitor.

The Real Moat: Process Knowledge and Data Accumulation

Velo3D’s hardware and patents create initial differentiation. But the durable moat is in the process data that accumulates as the platform scales. Every build generates ~1,000 sensor readings, captured by Assure, integrated through Intelligent Fusion, linked to specific process parameters from Flow. Over time, this maps the relationship between inputs and outputs with increasing granularity.

A contract manufacturer using EOS or SLM machines develops process parameters independently for each new application. They don’t have access to a centralized, continuously updating dataset across hundreds of builds. Velo3D, through RPS, will (if the fleet scales as planned).

The Competition

Direct LPBF competitors: EOS (Germany, private, ~$500M+ revenue, market leader), SLM Solutions (acquired by Nikon), Trumpf (multi-billion dollar conglomerate). Larger installed bases, more materials, bigger R&D budgets. But none offer SupportFree.

Publicly traded AM peers: 3D Systems (~$450M revenue), Stratasys (~$570M revenue), Markforged (~$100M revenue), Nano Dimension (~$60M revenue). Larger but less focused on high-performance mission-critical metal AM.

Service bureaus: Xometry, Protolabs, specialized AM service bureaus. Offer metal 3D printing as a service but operate multi-vendor fleets without Velo3D’s vertical integration.

My assessment: the moat is real and meaningful in the specific niche of complex, support-free, mission-critical metal parts. But the moat is narrow, it exists only in a specific segment. And it is time-limited: larger competitors will eventually close the gap. The question is whether Velo3D can convert the technology lead into a scaled production and data advantage before that happens.

5. The Numbers: Revenue, Margins, Cash, and the Path to Profitability

This is where the narrative meets reality.

Revenue: Growth Is Real, but Context Matters

FY2025 total revenue: $46.0 million, up 12.1% from $41.0 million in FY2024. But the headline understates underlying growth. In 2024, Velo3D recognized a one-time $5.0 million licensing fee from SpaceX. Stripping that out, core revenue grew approximately 26%.

Quarterly trajectory: Q1 $10.3M → Q2 $13.1M → Q3 $13.1M → Q4 $9.4M. The Q4 decline was attributed to product mix, timing, and a Q4 government shutdown that delayed defense-related shipments. The backlog data ($31M exiting Q4, 100% within 12 months) suggests this was timing, not demand.

Bookings tell a stronger story: $59 million in FY2025, nearly double 2024 and well above revenue, meaning the company is booking faster than it can recognize. Q4 bookings were the highest in company history.

Gross Margin: Negative, But the Trajectory Matters

FY2025 gross margin: negative 16.1% (gross loss of $7.4M). The largest factor was a $7.0 million non-routine inventory obsolescence write-off in Q4, the second consecutive year of major write-downs ($7.2M in 2024). Excluding the write-off, adjusted gross margin was approximately negative 1%. In Q2 and Q3 2025, the company achieved low single-digit positive gross margins on a normalized basis.

The margin bridge to 30%+ in 2H 2026 rests on: newly built systems carrying lower overhead, RPS at scale carrying 40-60% gross margins, and operational efficiencies. Aggressive but not impossible if RPS scales as bookings suggest.

Operating Expenses: Dramatic Cost Reduction

Total GAAP operating expenses: $47.5M in 2025, down 38% from $76.7M in 2024. R&D fell 31% to $10.7M. Sales and marketing fell 47% to $6.8M. G&A fell 38% to $30.1M. Non-GAAP operating expenses: $40.0M, down 40%.

The Cash Position and Burn Rate

Cash at December 31, 2025: $39.0 million. This is entirely the product of capital raises: ~$68M raised in 2025 through the August offering ($17.8M net), December PIPE ($28.0M net), convertible notes ($15M), and equipment sale-leaseback ($10M).

Operating cash burn in FY2025: $27.3M. The 2026 guidance adds $40-50M in capex. Total cash requirement for 2026 likely exceeds $60M. Against $39M in cash: Velo3D needs to raise additional capital in 2026 (see below, shelf offering)

The Balance Sheet: Clean-Up in Progress

Post-balance sheet, both the January Note ($5M + interest) and February Note ($10M + interest) were converted to equity in March 2026. The Secured Notes ($3.2M) were repaid in full. Total debt as of the filing date: approximately $9.7M (just the Varilease equipment loan). The balance sheet is cleaner than it has been in years but this came at the cost of issuing 1.54 million additional shares.

Post-Earnings Market Reaction

The stock fell 21.6% the day after earnings (March 25, 2026). Despite record bookings and a growing defense pipeline, the Q4 revenue miss ($9.4M, down 25% YoY) and the $40-50M capex guidance spooked a shareholder base that had ridden a 330% rally over the prior twelve months. As of this writing, the stock trades around $8.59, down approximately 30% from pre-earnings levels. The market is telling you something: it believed in the narrative at $12-14 but wants to see execution before paying up again.

Now let’s get to what really matters: valuation, scenario analysis, risks, and what ultimately drives the thesis:

Read more