Schulz Duggan

Semiconductor & AI supply-chain research.
Deep dives on memory, optics, and the chokepoints
nobody prices in.
Single-stock theses · Not investment advice.

[Schulz Duggan @Schulz_Research]

Three optics companies reported this week. The one with the smallest revenue had the biggest margin expansion. That is not a coincidence — it is the whole thesis.

$LITE $COHR $AXTI

Start with the number that should reorganize how you think about this supply chain.

GAAP gross margin, one year apart:

$AXTI (InP substrates, most upstream): 8.0% → 44.9%. +3,690bps
$LITE (laser chips, midstream): 33.3% → 47.4%. +1,410bps
$COHR (vertically integrated): 35.7% → 38.5%. +280bps

Revenue growth, same period:

$AXTI +164% · $LITE +109% · $COHR +34%

Both series are monotonic, and they run in the same direction: the closer you sit to the chokepoint, the more violent the expansion.

Most people are trading this sector by picking the best optics company. The data says the more important variable is where in the stack you stand.

1. The surface story: Lumentum is out-earning a company twice its size.

$COHR Q4 FY26: revenue $2.05B, non-GAAP gross margin 40.2%, non-GAAP EPS $1.74. Beat consensus of $1.65. Guided Q1 FY27 to $2.2–2.4B and EPS $1.85–2.05.

$LITE Q4 FY26: revenue $1.01B, non-GAAP gross margin 50.4%, non-GAAP EPS $3.23. Guided Q1 FY27 to $1.225–1.275B, midpoint +130% YoY, with EPS $4.05–4.35.

Coherent has 2.0x the revenue. Lumentum has 1.9x the EPS.

Then there is this pair of guidance lines, which I had to read twice:

Lumentum guided next quarter's OPERATING margin to 39.5–40.5%.
Coherent guided next quarter's GROSS margin to 39.5–41.5%.

What Lumentum keeps after paying for all R&D, sales and administration is what Coherent keeps before paying for any of it.

2. Part of the gap is a business Coherent can't grow.

From tonight's investor deck, five quarters of segment revenue:

Datacenter & Communications: 1,018 → 1,090 → 1,208 → 1,362 → 1,615 (+58.6% YoY)
Industrial: 511 → 491 → 478 → 444 → 431 (−15.7% YoY)

21% of Coherent's revenue has now shrunk five quarters in a row. It drags group growth from 58.6% down to 33.8%. Lumentum carries no such passenger.

But that only explains half. Coherent's datacenter-only growth of 58.6% is still barely half of Lumentum's 109%. The rest is structural, and it is the more interesting half.

3. Lumentum's pricing power does not come from being faster. It comes from yield.

CEO Michael Hurlston, asked directly on the call about Chinese companies bringing new indium phosphide fabs to market:

"We haven't seen any impact on our numbers as yet. I don't expect there to be any impact... Even in CW, we've been surprised at our ability to price up... customers are seeing far better yields given our ability to deliver consistent performance on our lasers. They simply don't deviate. The width of the spec is very narrow, and that results in far better transceiver yields... As a result, we are able to command a nice price premium that we expect to sustain, if these Chinese guys come online. I caution people also, some of these Chinese laser suppliers are not delivering in the market today, so there is no recourse when they throw out these big numbers."

Read what that actually says. Customers pay Lumentum a premium not because its lasers are faster, but because using them makes the customer's own transceiver yields higher. That is a switching cost embedded in someone else's factory. It is far harder to copy than a spec sheet.

4. Now the reversal. Why is Coherent sacrificing margin?

Coherent is spending to pull indium phosphide in-house: internal InP output doubling by year-end, more than doubling again by 2027, a 6-inch platform producing EMLs, CW lasers and photodiodes at higher yields than its 3-inch lines, backed by a $50M CHIPS Act letter of intent on top of $20M already in from Texas.

The obvious read is "Coherent is squeezing out its suppliers, bad for $AXTI." I built that framework earlier today and the data destroyed it.

Here is what AXT actually reported for the same quarter:

Revenue $47.6M, up 77% sequentially and 164% YoY. Record InP revenue of $30.7M. GAAP gross margin 44.9%, from 8.0% a year ago. GAAP net income $11.1M versus a loss. A completed $632.5M raise funding Tongmei's InP expansion and 6-inch R&D. Quarterly InP capacity targeted at roughly $60M by end-2026 and $130M exiting 2027.

A supplier being disintermediated does not expand gross margin by 37 percentage points in four quarters.

So invert the question. Coherent isn't building InP capacity to punish a supplier. It's building it because substrate margins went from 8% to 45% and it wants that margin back.

$COHR's capex is the most credible third-party validation of $AXTI's position available. It is a competitor voting with a balance sheet.

5. And the two stories cross-check each other.

If Chinese InP supply were genuinely arriving at scale, substrate gross margins could not run to 45%. Hurlston says those suppliers "are not delivering in the market today." AXT's P&L is the corroboration. Two independent sources, one conclusion: the substrate bottleneck is still tight.

Which also means "Chinese InP expansion" means opposite things to each name. To $AXTI it is future competition. To $LITE it is cheaper input, with its own premium sitting downstream in customer yield, untouched. To $COHR it is a risk already hedged by building its own.

6. On CPO, the two rivals accidentally agreed.

$LITE: ultra-high-power laser chip ramp in H2 calendar 2027, ahead of customer scale-up deployments in 2028. First external light source module order already booked for H2 2027 delivery. NPO is described as "completely additive."

$COHR investor deck: CPO/NPO listed as an H2 2027 growth engine. Thermal solutions H1 2027.

Two competitors, independently, put the same window on the calendar. Anyone selling pluggable optics today on CPO fear is early by roughly eighteen months.

7. What I think is actually happening.

This is a textbook contest between specialization and vertical integration, and the market just priced round one.

Lumentum specializes: buy the substrate, spend everything on laser consistency, extract premium from the customer's yield. Result: 50.4% gross margin, 40% operating margin guided.

Coherent integrates: own the substrate, control the cost curve, accept lower margin now. Result: 40.2% gross margin, capex still climbing.

Round one goes to specialization, decisively. But Coherent's bet is priced in years, not quarters. If 6-inch InP yields hold, its cost structure gets redrawn — and the deck already says 6-inch is running better than 3-inch.

AXT sits between them, and both directions are live simultaneously: it is the beneficiary of the specialization model and the casualty of the integration model. Its largest-scale customers are, at the same time, its most credible future competitors. That is an uncomfortable position, and it is also why it is mispriced in both directions depending on which quarter you look at.

Framework note: the vertical-integration lens here is from Rui Mingjie's Industrial Economics (Ch. on 纵向一体化, p.433) — integration is a choice about where in the chain the scarce rent sits, not a general virtue. Tonight the rent sits in substrates. That is why Coherent is moving.

Falsifiable, so you can score me next quarter:
1. $AXTI quarterly InP revenue toward ~$60M exiting 2026. It printed $30.7M. That requires near-doubling in two quarters.
2. $COHR non-GAAP gross margin above 41.5% — the top of its own guide. Below it, the integration spend is not yet paying and round two goes to specialization again.
3. $LITE sustaining 50%+ gross margin into the Q1 print. Hurlston claims the premium holds even if Chinese supply arrives. That is a testable claim with a date on it.
4. $COHR Industrial revenue. Five straight declines. A sixth means the 21% drag is structural, not cyclical.
5. First real CPO revenue. Both say H2 2027. If either pulls it into H1, the whole pluggable-versus-CPO timeline resets.

Source discipline: All financials from the 8-Ks and investor decks filed this week and AXT's Q2 release, not press summaries. Hurlston's quote is from the Lumentum call transcript. Coherent's call transcript was not yet published at the time of writing, so no COHR management verbal claim appears here that isn't in a filed document — one media summary put Coherent's capex commentary at "increasing sequentially from Q3's $290M," but I could not verify whether that figure is from tonight's call or May's, so it is excluded. Management at all three is talking its own book; the cross-checks above rely on rivals corroborating each other, which is the only kind of corroboration worth much here.

Prices, after hours Aug 12: $LITE $926.99 (−0.6%) · $COHR $340.17 (−4.4%) · $AXTI $77.02 (−1.9%).

[Schulz Duggan @Schulz_Research]

Very insightful analysis, especially comparing $POET to a 'reverse AAOI'—it hits the nail on the head regarding the current mismatch between capacity and demand. Regarding the goal of 12 million units of capacity by 2027, what do you think is the most critical catalyst? Aside from massive contracts like Lumilens, what specific signals are you watching to validate the demand?

[Schulz Duggan @Schulz_Research]

Brilliant thesis validation. The transition from distressed to critical AI DC infrastructure is a textbook turnaround. I completely agree on the unspoken national security implications—it changes the valuation multiple entirely when a company becomes a strategic asset rather than just a supplier. Looking forward to how the £126.5m guidance plays out.

[Schulz Duggan @Schulz_Research]

Completely agree with the 'risk mitigation' angle here. Given the increasingly strict power and thermal requirements for optical comms in AI data centers, this 3.4M program acts like a dirt-cheap 'call option' forSIVE. If SemiNex's high-power InP and SOA tech can actually solve the CPO bottlenecks, the risk-reward on this is massive. Do you think this puts early pressure on current players in the silicon photonics space?

[Schulz Duggan @Schulz_Research]

@amitisinvesting The density of information makes me wonder if I need a brain-computer interface to help me digest and process it. Hahaha

[Schulz Duggan @Schulz_Research]

Jensen is right that NVIDIA compute is financeable. The credit market agrees — and it has already put a price on it.

Two facilities, same month, same collateral class, both backed by NVIDIA GPUs:

$CRWV closed its DDTL 5.5 at SOFR + 5.50%. Ba2 / BB+.
$NBIS closed its first secured facility at SOFR + 2.50%.

300 basis points apart, on the same asset.

Which is the part worth sitting with. Fungibility, utilization and durability are properties of the chip. Financeability is not. It's a property of the chip plus the balance sheet borrowing against it — and lenders are pricing that second term very differently.

The sharper detail is in how CoreWeave's facility is built. It's a five-year loan against customer contracts averaging roughly three years. CoreWeave says so in its own release, and correctly calls it an unlock.

But read what lenders are actually underwriting there: not contract risk. Renewal risk.

Which is to say — "how long does a GPU keep earning?" is no longer only an accounting debate. It has been securitized. SOFR+550 versus SOFR+250 is the market's answer, quoted daily, with real money behind it.

Jensen argues A100s stay mission-capable through 2029. Meanwhile $NBIS just extended server useful life from four years to five, and Michael Burry is short the opposite view.

The credit market didn't wait for that argument to resolve. It just priced the uncertainty and moved on.

$NVDA $CRWV $NBIS

[Jensen Huang @JensenHuang]

The mighty A100 fleet are mission-capable from 2020 through 2029. NVIDIA computing is more than chips. CUDA gives developers and NVIDIA engineers a common platform to continually upgrade Ampere, Hopper and Blackwell throughout their useful lives.
CUDA makes NVIDIA computing versatile. Versatility makes it fungible. Fungibility drives utilization and extends durability, making NVIDIA compute a productive asset: rentable, durable and financeable.

[Schulz Duggan @Schulz_Research]

Jensen hits the exact core of GPU financialization: CUDA isn't just a developer layer; it’s a balance sheet protection mechanism.
​By ensuring backward software compatibility across hardware generations, CUDA flattens the depreciation curve. It turns GPUs from fast-depreciating CapEx into collateralizable, yield-generating infrastructure assets. Compute is officially a debt-financeable asset class.

[Schulz Duggan @Schulz_Research]

@kenchai96563783 the latter

[Schulz Duggan @Schulz_Research]

Cerebras told you its $25.4B backlog contains no AWS and no hyperscaler business. Do the subtraction.

$CBRS

The stock is down 16% after hours despite beating its own guidance on revenue, gross margin and operating margin, and despite raising full-year guidance across every metric. That combination is unusual enough to be worth explaining properly.

1. The concentration didn't go away. It changed address.

Before the IPO, the story was Abu Dhabi: G42 at 87% of revenue in H1 2024, then G42 down to 24% of 2025 while MBZUAI rose to 62% — two Abu Dhabi entities, 86% of 2025 revenue. One analyst called it exactly right at the time: "Concentration rotated, it didn't go away."

Tonight CEO Andrew Feldman said this on the call:

"Our $25 billion in RPO does not reflect any backlog of business from AWS or any other hyperscaler at this time."

And CFO Bob Komin, describing how the year started:

"We entered the year having won one of the largest technology deals ever, creating RPO of more than $25 billion."

Put those two sentences next to each other. The $25.4B backlog is, in substance, one contract. OpenAI's — over $20B across three years for 750MW.

That is the second rotation. Abu Dhabi → OpenAI. The denominator changed again. The structure didn't.

2. And this time the cost of the rotation is visible on the income statement.

GAAP revenue $180.1M ← LOWER than core
Core revenue $209.9M
GAAP gross margin 14%
Core gross margin 41%
GAAP operating margin −265%
Core operating margin −16%

Sequentially it is sharper still:

Core revenue: $191.3M → $209.9M. Up 9.7%.
GAAP revenue: $193.4M → $180.1M. Down 6.9%.

Same company, same quarter, opposite directions.

Cerebras states the reason plainly in its own reconciliation: amortization of customer warrants is equity granted to customers and recorded as contra-revenue, booked as a reduction against both hardware and cloud revenue lines depending on what that customer is buying.

Those warrants are OpenAI's ~10–11% stake. OpenAI also advanced roughly $1B in working capital to build the data centers it will then pay to rent.

So the anchor customer is simultaneously the anchor shareholder, the lender, and essentially the entire backlog — and the equity it received is being amortized against the revenue it generates. The more that relationship delivers, the more GAAP revenue is clipped on the way in.

Gross profit tells the same story without any adjustment: $32.1M a year ago, $25.6M today — down 20% — on revenue up 74%.

3. Now mark my own work. I published a preview before this print with five tests. I got one wrong.

T2 — Gross margin. Wrong. I said it would print at or below the 36–38% guide, reading "capacity constrained with falling margin" as cost escalation. Core gross margin came in at 41%, above the top of guidance, and full-year margin guidance was raised from 38–41% to 41–43%. My error was treating a newly-public company's conservative guide as an honest description of reality rather than as a number set to be beaten.

T1 — Shape right, timing wrong. I argued the year was entirely back-half loaded and put the test on Q3. Q3 guided to $214–216M. But FY26 was raised to $880–890M. H1 core is $401.2M, so H2 needs $479–489M. With Q3 at ~$215M, Q4 is implied at $264–274M — a 23–27% sequential jump immediately after a quarter that grew under 10%. The year didn't get de-risked tonight. It got concentrated into Q4.

T3 — Answered, above. Rotation, not resolution.

T5 — Gulf/NVIDIA availability. Not addressed. Worth noting what was said: the data center footprint is Alabama, Dallas, Denver, Minneapolis, Santa Clara, Stockton, plus France, Finland, Manitoba, Montreal, Norway, Saskatchewan and Toronto. No Middle East site was named. The revenue concentration and the physical footprint now point in different directions.

4. The part I underrated, and it's the real moat.

I wrote in the preview that Cerebras "routed around HBM." It's more deliberate than that. From the release:

"Through our wafer-scale architecture, we avoid many of the components that are currently in short supply. We do not use HBM memory, CoWoS packaging or 3nm fabrication technology, all of which are currently supply limited."

And Feldman on the call:

"Our ability to get wafer supply also benefits from the fact that we were able to deliver industry-leading performance while running on TSMC's five nanometer node, where wafers are less expensive and supply is less constrained."

Read the strategy in full: everyone else is competing for the three scarcest resources in semiconductors — HBM, CoWoS, and leading-edge nodes. Cerebras deliberately uses none of them. It trades silicon area, which is abundant, for supply chain position, which is not. Building the largest chip in the industry on a mature node is not a compromise. It is a procurement strategy disguised as an architecture.

5. And the physics argument is stronger than the benchmark argument.

The most substantive thing said on this call had nothing to do with tokens per second:

"GPU solutions can support high throughput, but only at low speeds. When configured to support even moderate speeds, GPU throughput drops precipitously. This is true not just for GPUs, but also for ASICs and all solutions that use HBM. The HBM memory architecture forces a trade-off between throughput and speed. SRAM-based architectures like Cerebras' are the exact opposite."

If that holds, it is a constraint no amount of engineering removes, because it is a property of where the weights live. Speed determines user experience. Throughput determines inference economics. HBM lets you optimize one at the expense of the other. That is a structural claim, and it is testable by anyone with access to both platforms — which makes it far more valuable than a vendor benchmark chart.

6. Which is why the strategy quietly changed tonight.

Cerebras is no longer positioning as the company that replaces GPUs. It is positioning as the company GPUs need.

Disaggregation splits inference into prefill (processing input — parallelizable, suits GPUs and HBM) and decode (generating output tokens — sequential, memory-bandwidth-bound, the bulk of the compute, suits wafer-scale SRAM). Cerebras now has disaggregated solutions with $AMD (Helios racks, production Q4 2026) and $AMZN (Trainium; Bedrock availability Q1 2027).

The combined solution keeps Cerebras speed and lifts throughput 5x.

That is a materially better business than being an $NVDA alternative. It converts every existing GPU deployment into a potential attach point rather than a competitor, and Feldman said the economics work "with any GPU." It also, in his words, "allows us to serve more of the demand that we have in RPO" — which matters because of the next point.

7. The bottleneck moved. It is no longer chips.

"Data center space continues to be the bottleneck for the entire industry, and we are no exception."

Wafer supply: secured with TSMC through next year. Manufacturing: already 4x H1 2025, scaling more than 10x in 2026 across Flex, Sanmina and Rocket EMS. Over 600MW live or under contract for delivery by end-2027, with a pipeline "measured in gigawatts."

Cerebras also argues it needs no gigawatt-scale single sites because inference doesn't require training-cluster footprints — which is a genuine flexibility advantage in a permitting-constrained world.

So the chokepoint I flagged in the preview — TSMC capacity — has been relieved. The one that replaced it is real estate and power.

8. The customer base is diversifying, and one win is genuinely new.

Six deals above $30M signed in Q2. Coding: Figma, Cognition, Lovable. Agentic: Block, AlphaSense, GSK. And $CRWD — inline LLM security screening enterprise traffic in real time, which Feldman described precisely: "an application that only exists if AI is fast... the speed creates the invisibility."

That is not a customer win. That is a category that could not exist on GPU latency. If enterprises come to expect large fractions of their traffic inspected this way, it is a TAM that belongs structurally to SRAM architectures.

Hyperscaler revenue starts mid-2027, ramping 2028. None of it is in the RPO yet.

9. Two numbers to sit with.

R&D was $320.2M against $180.1M of GAAP revenue — 1.78x revenue, up 427% YoY. Some of that is IPO-related equity recognition, but the run-rate is what a company spends when it is buying 2028, not harvesting 2026.

GAAP operating loss was roughly $477M in a single quarter, against $34M on a core basis — a 14x gap. Liquidity is not the issue: $8.6B cash plus $850M of debt capacity after a $6.4B IPO. The issue is that "core" is management's lens, and GAAP is what eventually has to fund the business.

10. Falsifiable, next quarter:

1. Q4 core revenue ≥ $264M. Not a forecast — arithmetic from their own raised guide.
2. The GAAP-to-core spread. $29.8M this quarter. If it widens as OpenAI volume ramps, warrant amortization is structural, not transitional.
3. GAAP gross margin off 14%. Core at 41% is the lens. GAAP is the business.
4. AMD in Q4 2026, Bedrock in Q1 2027. Both dated. Both are either revenue or a press release.
5. First non-OpenAI, non-Abu Dhabi contract large enough to move RPO. Six deals over $30M is real progress against a $25.4B number that is essentially one counterparty. Watch whether the third rotation is diversification or just a new address.

Source discipline: Financials from tonight's 8-K filed with the SEC. Management quotes from the earnings call transcript, retrieved directly. OpenAI is simultaneously Cerebras's largest customer, a major shareholder, a lender to it, and effectively its entire backlog — no OpenAI-related datapoint here is independent, including the flattering ones. The speed and throughput claims are the company's own; the HBM trade-off argument is a structural claim I find persuasive but have not independently verified. I was wrong on gross margin and said so before anyone asked.

[Schulz Duggan @Schulz_Research]

@aleabitoreddit The TLA (Three-Letter Acronym) fatigue is real! 😅 Throw ROI and OPEX into the chat and you've got the complete AI infrastructure puzzle.

[Schulz Duggan @Schulz_Research]

$COHR didn't just exit the merchant market. It's still buying in it.

Two sentences past the line everyone's quoting:

"Today we use a mix of internally produced and externally sourced... over the long term, we'll ha[ve a] portion of our Datacom transceivers supported by external sources."

Why it can't stop: "our data center transceiver demand is absorbing every bit of capacity that we have — and then some."

You moved $COHR off the supply side. It belongs on the demand side of the same page.

Supply exiting + demand entering is a far tighter setup than supply exiting alone — and the layer underneath is already printing it.

Does the merchant market actually shrink here, or just reprice?

[Schulz Duggan @Schulz_Research]

The >$100B/GW/year inference revenue model is stunning, but if SpaceX and MSFT are racing to 10GW+, the ultimate chokepoint shifts from power generation to optical interconnects and advanced packaging (CoWoS). 10GW of GB300 clusters implies an astronomical volume of 1.6T/3.2T transceivers. Even if they secure the gas turbines, does your model foresee the optical supply chain scaling fast enough by 2027 to prevent stranded compute?

[Schulz Duggan @Schulz_Research]

@ResearchSfl I will write my in-depth research report after $CBRS's financial report and conference call. Please give me some time to enjoy this gift

[Schulz Duggan @Schulz_Research]

$CBRS beat its own revenue guidance by 8% tonight.

It also missed consensus by 9%.

Both are true. The stock is down 11.5%.

The gap between those two numbers is the whole story, and I have not seen anyone model it.

Core revenue: $209.9M. Guidance was $194M.
GAAP revenue: $180.1M. Consensus was $198.1M.

Non-GAAP revenue came in HIGHER than GAAP revenue.

If that looks backwards to you, it should. Adjusted figures almost always flatter the GAAP number. They do not undercut it.

It gets sharper on a sequential basis:

Core revenue: $191.3M → $209.9M. Up 9.7%.
GAAP revenue: $193.4M → $180.1M. Down 6.9%.

Same company. Same quarter. Opposite directions.

Here is the full split:

GAAP revenue — $180.1M
Core revenue — $209.9M
GAAP gross margin — 14%
Core gross margin — 41%
GAAP operating margin — negative 265%
Core operating margin — negative 16%

GAAP operating loss: roughly $477M in a single quarter.
Core operating loss: $34M.

A 14x gap.

Cerebras states what is inside it: non-cash amortization of customer warrants, stock-based compensation, and data center pass-through revenue and costs.

Those customer warrants are OpenAI's roughly 10-11% stake in the company.

OpenAI committed over $20B across three years for 750MW. It received warrants for about a tenth of Cerebras, and advanced roughly $1B in working capital to build the data centers it will then pay to rent.

That structure is a commitment device. Dixit and Nalebuff's rule: a strategic move becomes credible only when you take a subordinate action that binds you. The equity is what made a $20B purchase order believable.

Tonight that device landed on the income statement. As contra-revenue.

Which means the more Cerebras sells into its anchor relationship, the more GAAP revenue gets clipped on the way in. This is not a cost problem. It is the price of the contract, recognized in the one place nobody puts it in a model.

It also reframes the backlog. RPO is $25.4B and it is real. But some share of the revenue it converts into arrives pre-discounted by warrants already issued. If you are valuing CBRS on RPO times margin, you need a third term.

Now the part where I mark my own work. I published a preview before this print with five falsifiable tests.

I got one wrong.

I said gross margin would print at or below 36-38%, reading "capacity constrained with falling margin" as cost escalation. Core gross margin came in at 41%, above the top of guidance, and full-year margin guidance was raised from 38-41% to 41-43%. That call was wrong. My error was treating a newly public company's conservative guide as an honest description of reality.

What I got right was the shape, not the timing. I argued the full year was entirely back-half loaded. Q3 guided to $214-216M, below my line. But FY26 was simultaneously raised to $880-890M. Rerun it: H1 core is $401.2M. That leaves $479-489M for H2. With Q3 at roughly $215M, Q4 is implied at $264-274M. A 23-27% sequential jump, in one quarter, immediately after a quarter that grew under 10%.

The year did not get de-risked tonight. It got concentrated into Q4.

The rest of the print is genuinely strong and I will not pretend otherwise. Core cloud revenue up 287% year over year. $8.6B in cash. 600MW under contract. Manufacturing capacity scaling more than 10x in 2026. Launch partner for OpenAI's GPT-5.6 Sol at 750 tokens per second. Disaggregated inference with AMD shipping Q4, AWS Bedrock in Q1 2027. New enterprise logos including Block, Figma, AlphaSense, GSK, and CrowdStrike running inline LLM security across enterprise traffic.

The architecture was never the question. Cerebras is the only company at scale that routed around HBM entirely, with weights resident in 44GB of on-chip SRAM at 21 PB/s. It remains the most interesting non-GPU bet in the industry.

The question is what it costs to buy the demand that proves it.

Five things I will score myself on next quarter:

1. Q4 core revenue at or above $264M. Not my forecast. Arithmetic from their own raised guide.

2. The GAAP-to-core spread. $29.8M this quarter. If it widens as OpenAI volume ramps, warrant amortization is structural, not transitional.

3. GAAP gross margin, not core. 14% is the number that has to move. Core is management's lens. GAAP is what funds a business.

4. AMD in Q4 2026 and AWS Bedrock in Q1 2027. Both dated. Both checkable. Both are either revenue or a press release.

5. Abu Dhabi concentration in the 10-Q. Does it fall below 86% because US enterprise arrived, or because OpenAI did? Those are different companies.

Every number above is from the 8-K filed with the SEC tonight, not from a press summary. OpenAI is simultaneously Cerebras's largest customer, one of its largest shareholders, and a lender to it. No OpenAI-sourced datapoint here is independent, including the ones that flatter the company. Speed benchmarks are vendor-published and discounted accordingly.

CBRS $231.93, after hours, down 11.5%.

I own none of it.

[Schulz Duggan @Schulz_Research]

@amitisinvesting The dot-com comparison ignores the actual cash flow generation we are seeing today versus 2000. $MU is a completely different beast now. The risk/reward on this short seems asymmetric in the wrong direction

[Schulz Duggan @Schulz_Research]

The 15x inference speed advantage is a strong narrative, but a 14% hardware gross margin suggests they are essentially buying market share. How durable is this technical moat once competitors scale next-gen custom ASICs? It feels like severe margin compression is inevitable for AI hardware players not named NVDA.

[Schulz Duggan @Schulz_Research]

$COHR beat on everything tonight.

Revenue $2.05B, the top end of guidance.
Non-GAAP EPS $1.74 vs $1.65 consensus.
Q1 guide $2.2–2.4B with EPS $1.85–2.05.

The stock still went from +4.4% to negative after hours.

Here is why, in one number.

A year ago Coherent's GAAP gross margin was 35.7%. Lumentum's was 33.3%. Coherent was ahead.

Tonight:

Coherent 38.5%. Up 280bps.
Lumentum 47.4%. Up 1,410bps.

Five times the expansion. Same sector, same fiscal year ending June, same AI optics demand.

Then there is the line that stopped me:

Lumentum guided next quarter's OPERATING margin to 39.5–40.5%.
Coherent guided next quarter's GROSS margin to 39.5–41.5%.

What Lumentum keeps after paying for all R&D, sales and admin is what Coherent keeps before paying for any of it.

Part of the answer is in the segments. 21% of Coherent's revenue is an industrial business that has now shrunk five quarters in a row, from $511M to $431M. That alone drags group growth from 58.6% down to 33.8%.

But it only explains half. Coherent's datacenter and communications revenue grew 58.6% year over year. Lumentum grew 109.3%.

The other half is structural. That is the more interesting question, and it is not the one the headlines are asking.

Full teardown of both coming, including what each said about CPO, and why Coherent is doubling internal InP output by year-end with CHIPS money behind it.

All figures from tonight's 8-K and investor deck, not press summaries. COHR $339.51 after hours.

[Schulz Duggan @Schulz_Research]

@ResearchSfl I'm about to release a deep research report on it, so you can look forward to it. We can exchange views

[Schulz Duggan @Schulz_Research]

Cerebras's best moat was a U.S. export ban. Washington removed it on July 10.

$CBRS reports after the close tonight. It goes in at $261.93, up 11.5% on the day — after peaking at $350 post-IPO and bottoming near $244. The setup matters: this is not a quiet print.

Here is what almost nobody is looking at.

1. The company guided its own revenue to flat.

Q1 2026 reported revenue: $193.4M.
Q2 2026 core revenue guidance: $194M.

That is roughly zero sequential growth, in the hottest quarter AI infrastructure has ever had — the same three months in which CoreWeave grew 24% sequentially and Nebius grew 46%.

Caliber caveat, stated plainly: "core revenue" and reported total revenue are not strictly the same measure. But even sell-side consensus at $198.1M implies only ~2.4% sequential growth. The direction is not in dispute.

2. The margin tells you it isn't a demand problem.

Q1 GAAP gross margin: 45%.
Q2 core gross margin guidance: 36–38%.
Q2 operating margin guidance: −30% to −32%.

Management attributes the revenue ceiling to capacity constraints. Read those two facts together, because they don't fit:

If you are genuinely capacity-constrained in a supply-starved market, you get pricing power, and gross margin goes UP. Cerebras's is going down 7–9 points.

Capacity-constrained with collapsing margin is not scarcity. It's cost escalation. Building dinner-plate-sized chips at yield is expensive, and the expense is showing up before the volume does.

3. Do the arithmetic on the full-year guide.

FY26 guidance: $855–865M.
H1 actual + guided: $193.4M + $194M = $387.4M.

That leaves $468–478M for H2, or roughly $234–239M per quarter — a required +21% to +23% sequential step-up, twice, starting from a quarter that grew 0%.

The entire year is back-half loaded. Tonight's guide for Q3 is therefore worth more than tonight's Q2 print. If Q3 doesn't guide to ~$234M+, the full-year number is already broken.

4. The concentration didn't go away. It rotated.

• G42 was 87% of revenue in H1 2024.
• G42 fell to 24% of 2025.
• MBZUAI — Mohamed bin Zayed University of AI — became 62% of 2025.
• Two Abu Dhabi entities together: 86% of 2025 revenue.

One analyst put it exactly right: "Concentration rotated, it didn't go away." The denominator changed. The geography didn't.

5. And now the part that actually matters.

Ask the uncomfortable question: why was Abu Dhabi buying wafer-scale chips in the first place?

Partly the technology. But partly because it could not get enough NVIDIA. Every Blackwell shipment to the UAE required per-shipment BIS review — a process that ran roughly six months from approval to arrival for the first batch in May 2026.

Then:

• Nov 19, 2025 — Commerce authorizes advanced chips to G42 (UAE) and HUMAIN (Saudi): up to 35,000 GB300s.
• July 10, 2026 — BIS reclassifies the UAE into Country Group A:5, the highest trust tier. License requirements removed. Framed around the UAE's Major Defense Partner status and its support during the Iran conflict.
• Stargate UAE proceeds: $30B, 5 GW campus, 19.2 km², with OpenAI, Oracle, NVIDIA, Cisco, SoftBank.

So: the single largest structural reason Cerebras's largest customers had to look past NVIDIA was an American export restriction — and that restriction was lifted five weeks ago, on the exact customers who supply 86% of its revenue.

This is the chokepoint that isn't on anyone's supply chain map. Cerebras's moat was partly written in Washington, and Washington just edited it.

I want to be fair about direction: the same policy change expands Abu Dhabi's total compute budget, and a bigger pie can lift Cerebras too. Both readings are live. But only one of them is priced.

6. The OpenAI deal is a financing structure wearing a customer's clothes.
• >$20B over three years, 750 MW delivered in stages through 2028 (announced at >$10B in January, doubled by the first public earnings report).
• OpenAI receives warrants for roughly 10–11% of Cerebras.
• OpenAI advanced Cerebras roughly $1B in working capital to build the data centers OpenAI will then pay to rent.
• That contract is about 23× the midpoint of FY26 core revenue guidance.

Dixit and Nalebuff, in Thinking Strategically, define the mechanism precisely: "To make a strategic move credible, you must simultaneously take an additional or subordinate action. We call that action a commitment."

The equity and the working capital are the commitment device. They make the $20B credible.

But run the logic backwards, which is where it gets interesting: a purchase commitment that requires equity warrants and a $1B cash advance to be believable is telling you something about how believable it was without them.

And note the incentive that creates. OpenAI is now simultaneously Cerebras's largest customer and one of its largest shareholders. A buyer holding 10% of its supplier has a structural reason to negotiate price down — the equity recaptures what the P&L gives up. Watch gross margin on OpenAI volume specifically. That is where this shows up.

7. Reframing the chokepoint — this is the real thesis.

Here is what Cerebras genuinely got right, and it is underrated:

WSE-3 is a full wafer — 46,225 mm², 57× the die area of an H100 — carrying 44 GB of on-chip SRAM at 21 PB/s. Weights live on the silicon. There is no HBM.

That is not a marketing difference. In a market where HBM is the binding constraint and SK Hynix, Samsung and Micron sit on the throat, Cerebras is the only architecture at scale that routed around it entirely. The speed follows from the architecture: 2,522 tok/s vs Blackwell's 1,038 on Llama 4 Maverick; 2,700+ vs ~900 on GPT-OSS 120B; 981 tok/s on Kimi K2.6 with GPU clouds 6.7× behind. (Vendor-published benchmarks — discount accordingly, though third-party runs have directionally corroborated.)

But look at what it traded for.

It escaped the HBM chokepoint and walked into two narrower ones:

• TSMC's wafer-scale packaging capacity. Only one company on earth can manufacture this. "Capacity constraints" in the guidance is that sentence, in accounting language.
• A single sovereign buyer. 86% from two entities in one emirate.

It swapped one chokepoint for two tighter ones. And the second-order conclusion follows immediately: if wafer-scale wins as an architecture, the durable beneficiary is TSMC, not necessarily Cerebras. Cerebras proved the route. It does not own the road.

8. What to actually watch tonight (falsifiable, score me):

1. Q3 revenue guide. Needs ~$234M+ to keep FY26 intact. Below $220M, the full-year guide is broken regardless of the Q2 headline.
2. Gross margin. Does it print inside 36–38%, or below? Below 36% means the cost problem is worse than management modeled one quarter ago.
3. Customer concentration disclosure. Does the Abu Dhabi share of revenue fall below 86% on organic U.S. enterprise demand — or only because OpenAI's ramp arrives? Those are completely different companies.
4. OpenAI revenue recognition timing. 750 MW through 2028 means the 2026 contribution is small. If management pulls it forward in the narrative without pulling it forward in the numbers, that's the tell.
5. Any commentary on NVIDIA availability in the Gulf post-July 10. Management will be asked. The answer is the whole thesis.
Source discipline: Revenue, margin, guidance and concentration figures are from the company's S-1/424B4 and its guidance — filed documents, not adjectives. The speed benchmarks are vendor-published and discounted as such. OpenAI is Cerebras's customer, shareholder and lender simultaneously — treat any OpenAI-sourced endorsement as non-independent by construction. BIS and Commerce actions are official record. Price $261.93, intraday Aug 12.

Cerebras built the most interesting non-GPU architecture in the industry. Tonight is not a test of the architecture. It's a test of whether the architecture has a business underneath it that isn't one emirate and one customer-shareholder.

I own none of it.

[Schulz Duggan @Schulz_Research]

The 'committee tax' is now quantifiable. If DeepSeek is hitting 87.9 on Agentic Coding with ~20% of the active parameters, the implied compute efficiency drastically shifts the unit economics for agentic workflows. Organizational agility is starting to outperform brute-force capex