Cerebras Systems

Phase A — Understand the business

Lens 1 · Company Overview (re-grounded — confirmed)

Cerebras designs and sells wafer-scale AI compute: the WSE-3 processor (a single chip the size of a dinner plate, etched from one wafer) packaged into CS-3 systems, and Cerebras Inference cloud capacity sold as a service. Two business models in one box — lumpy hardware sales (systems to customers/sovereigns) and recurring cloud & other services (tokens/capacity). FY25 split: Hardware $358.4M (70%), Cloud & other services $151.6M (30%) of $510.0M total. Q1-2026 (GAAP): Hardware $110.6M, Cloud & other services $82.8M — cloud growing far faster (see Lens 4/5). The company has "delivered three generations of wafer-scale processors at the 16, 7, and 5 nanometer nodes" and has been "yielding, packaging, delivering, and running software on the industry's first commercial wafer-scale AI systems since 2020". Customer archetypes per the prospectus: hyperscalers (AWS), foundation-model labs (OpenAI), AI-native/digital-native businesses, enterprises, and Sovereign AI initiatives (G42, MBZUAI). The contract structure is the story: take-or-pay-style committed capacity (OpenAI is "contractually committed to purchase" 750MW) wrapped with customer financing and customer warrants — a structure closer to a project-financed datacenter developer than a chip vendor.

Lens 2 · Supply Chain (carried — confirmed by filings, see deep-dive-2026-06-25.md)

Unchanged structurally, now filings-confirmed. Upstream: Cerebras is fabless, single-foundry on TSMC — "We depend on one third-party foundry, TSMC, to manufacture our proprietary processor" and crucially "We have no formalized long-term supply or allocation commitments from TSMC, and TSMC also fabricates wafers for other companies, including certain of our competitors". The single-wafer (no-dicing) flow concentrates yield risk on one supplier with no contractual allocation — the canonical chokepoint. Midstream: Cerebras increasingly operates its own datacenters (PP&E $572.4M at Mar-2026, up from $437.4M Dec-2025; $344.3M of non-cancelable datacenter leases not yet commenced as of Dec-2025) — it is vertically integrating into capacity to serve the inference cloud, which is why the rent-back/start-up margin drag exists. Downstream: end-customers are concentrated — MBZUAI, G42 (Sovereign AI / UAE), OpenAI (foundation-model lab), AWS (hyperscaler, also a partner via the Trainium disaggregation arrangement and a warrant-holder). Names along the chain: TSMC → Cerebras (+ leased/owned datacenters) → {MBZUAI, G42, OpenAI, AWS}. The most dangerous structural feature: AWS is simultaneously partner, customer, and warrant-holder — and runs the competing Trainium silicon (Lens 13).

Lens 3 · Competitive Advantages (moats) (re-judged — architectural moat confirmed, economic moat unproven)

The moat is architectural and latency-driven, not yet economic. Wafer-scale keeps an entire model's weights on a single piece of silicon, collapsing the chip-to-chip interconnect bottleneck that GPU clusters fight — yielding benchmarked low-latency inference that bulls argue is differentiated for "highly responsive model interactions". The genuine moat tells: (1) process/IP — three generations of a hard-to-copy single-wafer flow since 2020; (2) a narrow but real validation — one of only two hardware vendors serving OpenAI's models; (3) switching/integration — co-development with OpenAI gives "a direct view into frontier model development". But the filings expose the moat's economic weakness: Cerebras has no pricing power over TSMC (no allocation commitment) and is paying customers in warrants + accepting customer loans to win/fund contracts — the opposite of supplier/customer bargaining power. The FY26 gross-margin guide (38-41% core, Q2 trough 36-38%) is the first hard number on the economic question, and it sits well below NVIDIA's structural ~70%+ — wafer-scale wins on latency, not (yet) on margin. Moat verdict: real in physics, unproven in economics.

Lens 4 · Segments (re-grounded — geography materially corrected)

By product:

Hardware gross margin FY25 ≈ 43% ($358.4M rev / $204.7M COGS); cloud gross margin FY25 ≈ 30% ($151.6M / $106.2M) — cloud is currently the lower-margin line (datacenter pass-through + start-up costs), inverting the usual "services = high margin" intuition and explaining the guide-down. (Core hardware margin was a flattered 42% in Q1 on performance-incentive pricing that management says normalizes to "low 30s".)

By geography — the prior dossier's "86% UAE" was the wrong frame; the audited number is 63% EMEA, not 86%, and the 86% is customer concentration:

. The swing is the G42/UAE Sovereign-AI ramp: EMEA went from 3% to 63% of revenue in one year. By customer (the real concentration): FY25 MBZUAI 62% + G42 24% = 86% of revenue, both related parties. US revenue actually fell in absolute terms FY24→FY25 ($282.7M → $187.6M) — the "US organic" base shrank while the UAE base exploded. That is the single most important segment fact and it directly contradicts a "diversifying to durable US demand" narrative.

---

Phase B — Measure performance

Lens 5 · Earnings Result (re-grounded — GAAP now sourced; corrects the web "core-only" prior)

The anchor is now the GAAP 10-Q, with the non-GAAP "core" bridge sourced rather than press-reported.

Q1-2026 GAAP: Total revenue $193.4M (+94% YoY); Hardware $110.6M, Cloud & services $82.8M. Gross profit $86.2M (44.6% GAAP GM). Opex $101.2M (R&D $75.5M, S&M $14.7M, G&A $11.0M). Loss from operations $(15.0)M; Other income net $2.5M (after $18.9M non-cash interest on the OpenAI loan); net loss $(14.0)M; net loss/share $(0.22) on 62.8M weighted-avg shares.

Q1-2026 core (non-GAAP): strips $4.1M pass-through revenue, adds back $4.0M pass-through cost, $2.1M customer-warrant amortization, $1.0M SBC → core gross profit $89.1M (46.6%); core operating loss $(3.5)M (−2% margin); core net loss $(2.5)M. (These match the prior dossier's web figures — confirming the press numbers were the core set; the new value is the GAAP statement underneath them.)

vs consensus / vs its own history: +94% YoY revenue is a clear acceleration; the GAAP loss narrowed ($14.0M vs $23.9M Q1-25). But the print is not profitable on any basis that excludes the one-time forward-contract-extinguishment gain, and management guided to deeper core operating losses ahead (Q2 −30/−32%, FY26 −28/−32%). Balance-sheet flags: liquid — cash $1.72B + restricted $1.03B + investments $0.52B = $3.26B (note $1.03B is restricted, tied to the loan/customer arrangements; the headline "$3.3B" is not all free); stockholders' DEFICIT $(194.7)M at Mar-2026 (pre-IPO; the $5.4B raise lands in Q2); deferred revenue $244.3M + customer deposits $368.4M (heavy customer pre-funding); inventory $89.0M (+40% QoQ) outrunning sequential revenue — watch. Operating cash flow +$12.3M but flattered by $68.2M of non-cash add-backs (incl. $18.9M non-cash interest); capex $132.0M in the quarter (~$528M annualized). Market reaction: −10.5% AH then ~−15% to ~$227 on the day, and it has since fallen to ~$182 — the revenue beat was entirely overwhelmed by the margin guide.

Lens 6 · Earnings Calls (sentiment) (re-grounded — first & only call now on the shelf)

The first earnings call (Jun-23, for Q1) is ingested. Management focus: (1) the OpenAI 750MW collaboration as demand proof and a "direct view into frontier model development"; (2) the AWS partnership ("only two hardware vendors that currently serve OpenAI models and we're one of them"); (3) a deliberate financial-framework reframe — CFO walks through "core" metrics, pass-through accounting, and the rent-back capacity strategy. Tone: confident on demand, defensive on margin. The CEO's post-call characterization that the margin guide was "misunderstood" is the tell — management believes the Street mis-read a transitional capacity-shortage drag as a structural-economics problem. The recurring phrases: "ramping capacity rapidly," "rent our own systems back," "core gross margin," "low-latency inference." What's absent: any clean US/organic run-rate or a path to GAAP operating profit. Sentiment trajectory can't be drawn (one call) — but the gap between management's "misunderstood" and a −15% tape is the single most important communications fact: the narrative did not land.

Lens 7 · Comps (re-marked — exact cap, peer multiples sourced)

At ~$181.59 (2026-06-28) on 215.1M shares, market cap ~$39.1B (fully-diluted ~$41B incl. options/warrants). On trailing FY25 revenue $510.0M that is ~76–78x sales; on the FY26 core-guide midpoint ~$860M it is ~45–46x sales. The peer table:

. The comparison that matters: the single closest analog to Cerebras's business + financing model is CoreWeave — a capacity-financed, backlog-driven, OpenAI/Microsoft-anchored AI-infra name — and CoreWeave trades at ~7x forward sales while Cerebras trades at ~46x. Even the richest chip comp (AVGO/NVDA at ~22-25x EV/sales) is less than half Cerebras's multiple, and those names are wildly profitable. Sell-side is undeterred: avg 12-mo PT ~$299 (range $273–340), Strong Buy, 10 buys / 0 sells — UBS $320 (Buy), Wedbush $280 (Outperform), Morgan Stanley $273. The ~$299 sell-side target vs a ~$182 tape (and a ~7x CRWV comp) is the debate.

Lens 8 · Stock-Price Catalysts (updated — but note: ~6 weeks of history, not 5 years)

CBRS has only existed publicly since 2026-05-14, so "5-year catalysts" collapses into the IPO arc:

• 2026-05-14 IPO at $185, the largest semiconductor IPO in history; first-day high ~$386 (ATH).
• Drift down through May–June off the IPO froth.
• 2026-06-23 first earnings: revenue beat (+94%) but Q2 GM guide 36-38% and FY26 −28/−32% op margin → −10.5% AH, then ~−15% (Jun-24) to ~$227; CEO calls it "misunderstood."
• 2026-06-26 all-time low $160.81; 2026-06-28 ~$181.59, below the $185 IPO price. The pattern, even in six weeks, is unambiguous and matches the prior dossier: this name trades on margin and concentration, not the revenue headline. A +94% growth print with a >$20B-headline OpenAI contract still sold off ~15% because the quality/cost of the growth (and now the source of it — UAE) disappointed. Forward catalysts: Q2 print (~Aug-2026) — does GM trough at 36-38% as guided, does OpenAI capacity start recognizing; the staged lock-up (≈28M shares already released post-Q1; the main ~171M-share lock-up expires ~mid-Nov-2026 / 2 days after Q3 earnings ); and the next 10-Q's related-party + revenue-recognition detail on the OpenAI loan/warrants.

---

Phase C — Judge people & books

Lens 9 · Management (carried + filings-confirmed — see deep-dive-2026-06-25.md)

CEO Andrew Feldman + long-tenured founding team (serial chip founders; prior exit SeaMicro→AMD). Track record: built and shipped three wafer-scale generations since 2019-2020 — a genuine from-scratch silicon achievement few teams on earth could execute. Skin in the game: founder-led with a multi-class structure — Class B (185.1M shares, founder/insider) carries super-voting rights vs Class A's one-vote; Feldman signed a continued-employment offer letter dated 2026-03-22. Capital allocation: aggressive growth — financing the buildout with Series G ($1B, Sep-2025), Series H (~$1.0B in Q1-2026), the ~$1.0B OpenAI loan, and a ~$5.4B IPO; capex ~$132M/quarter and $344M of committed datacenter leases. This is a found-then-fund operator, not a capital-returner — appropriate for the stage, but it means dilution + leverage + customer-financing dependence are structural. Red flags to monitor: (a) the customer-warrant practice (paying OpenAI and AWS in equity-linked instruments to land contracts) — defensible as customer alignment, but it depresses reported revenue and dilutes; (b) related-party revenue — MBZUAI + G42 are related parties to each other and dominate the book; (c) post-IPO insider selling into the staged lock-up. Archetype: visionary founder-engineer; the question his tenure now faces is allocation discipline under public scrutiny, not technical capability.

Lens 10 · Forensic Red Flags (re-grounded — the highest-value lens this refresh)

Now sourced from the actual statements rather than press inference.

• Revenue quality / circularity (the central flag). The cash-flow statement discloses "Billed amounts offsetting Working Capital Loan balance $21.6M" — billings to a customer net directly against that customer's loan. Combined with $365.8M of customer warrants issued to OpenAI (amortized against revenue) and the company's explicit guidance that warrant amortization "will reduce reported revenue in future periods", a meaningful slice of "revenue" is entangled with customer-provided capital and equity. This is the textbook circular-AI-financing structure, and it is now a set of line items, not a rumor. Forensic priority for the next 10-Q: how much OpenAI "revenue" is gross vs net of the loan/warrant; whether the loan is on- or off-market (6% stated, waivable — a related-party tell).
• GAAP earnings driven by non-operating marks. FY25 net income $237.8M = $(145.9)M operating loss + $390.7M of non-operating Other income — dominated by a one-time $363.3M gain on extinguishment of the Series F forward-contract liability; FY24 was the inverse ($(378.2)M charge on the same instrument). Any "P/E" on this name is meaningless; the operating business loses money. Non-cash interest of $18.9M already distorts other-income in Q1.
• Pass-through gross-up. Datacenter costs booked gross in both revenue and COGS inflate the top line while compressing the margin — a presentation choice that flatters revenue growth and must be normalized.
• Working-capital signals. Inventory +40% QoQ to $89.0M and prepaids +$46M QoQ outran sequential revenue; receivables +$12.2M. Customer deposits $368.4M + deferred revenue $244.3M are large relative to revenue — favorable (pre-funding) but concentration-linked.
• Capital structure at quarter-end. Stockholders' deficit $(194.7)M, accumulated deficit $(919.3)M, redeemable convertible preferred still outstanding pre-IPO. The IPO repairs the deficit (Q2) but the accumulated-deficit history is real.

Regulatory findings (required sub-section). No regulatory/regulatory-findings.md was generated for this run (the directory does not exist on the shelf; fetch-regulatory-findings.ts was not run this pass — flag for next refresh). On the three sources available:

• SEC LR/AAER: none ingested this run; Cerebras is ~6 weeks public with no known enforcement. (Verify via EDGAR EFTS on the next pass.)
• Non-SEC enforcement (web): none surfaced; the live overhang is CFIUS — the prior dossier noted CFIUS cleared the G42 stake (2025; G42 converted to non-voting then divested), which is why MBZUAI/G42 appear as customers, not controlling holders.
• 10-Q Item 1 (Legal Proceedings): "We are not presently a party to any litigation the outcome of which, we believe, if determined adversely to us, would individually or taken together have a material adverse effect on our business..." — clean as of Q1.
• Net: No material litigation per the company's own disclosure; the real "forensic" exposure is disclosure/related-party risk around the OpenAI loan + warrants and revenue recognition, not active enforcement. Verify SEC EFTS (LR/AAER) on the next refresh.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (re-forecast off the guide; prior forecast resolves)

Prior Brier forecast resolves NO (provisionally). "CBRS FY2026 revenue ≥ $1.0B, p≈0.45" — the company's FY26 core-revenue guide is $855–865M, ~14% below the $1.0B bar; only a large H2 OpenAI-capacity beat closes it. The p=0.45 was too high — the concentration/margin reality was knowable from the S-1. (Resolve formally on the next pass.)

Bottom-up off the guide:

• Base — hit the guide: FY26 core revenue ~$860M (+69%) at core gross margin 38-41% and core operating margin −28/−32% → core operating loss roughly $(240)–$(275)M. GAAP loss wider; non-GAAP net loss deepens vs FY25's $(75.7)M. EPS negative all three years on this path; no positive EPS to forecast — the relevant metric is the loss trajectory and gross-margin recovery, not an EPS line. ~46x guide sales.
• Bull — OpenAI 750MW begins recognizing H2-2026, cloud run-rate exits FY26 >$400M, GM troughs at 36-38% in Q2 then climbs back toward mid-40s as owned datacenters replace rent-back → FY27 core revenue $1.4–1.7B, GM mid-40s, approaching core operating breakeven by late FY27. A software-like re-rate on the recurring mix is the only thing that defends ~46x.
• Bear — GM stays sub-40 (the rent-back/pass-through drag proves stickier and wafer-scale inference is genuinely lower-margin vs Blackwell TCO), OpenAI capacity slips into 2027, US/organic stays shrunk → growth decelerates to ~40%, the market re-rates toward the CRWV ~7x / chip-cyclical 10-18x band and the stock halves-to-thirds from ~$182.
• Forecast to log on commit: "CBRS FY2026 core revenue ≥ $900M," p≈0.45 (resolves ~2027-03) — guide-aware re-anchor of the resolved prior. (Not auto-logged — log on commit per skill.) A second binary worth tracking: "CBRS FY2026 core gross margin ≥ 40% (lands in/above the 38-41% guide top)," p≈0.50.

Lens 12 · Bull vs Bear (re-judged against the filings)

Bull. The only credible non-NVIDIA wafer-scale inference platform, validated as one of two vendors serving OpenAI; cloud +178% YoY (GAAP) shows the high-value recurring line is taking; a 750MW OpenAI committed contract + 1.25GW option + AWS disaggregation partnership is a real multi-year demand pipeline; $3.3B liquidity (post-IPO ~$8B+ pro forma with the $5.4B raise) funds the buildout; management frames the margin hit as a transitional 10-15pt capacity-shortage drag that reverses as owned datacenters come online; sell-side $299 avg. Bear. ~46x FY26-guide sales — roughly 6-7x the CoreWeave multiple for a similar capacity-financed, OpenAI-anchored model — on revenue that is still 86% UAE-concentrated (MBZUAI + G42, related parties) with US revenue shrinking YoY, financed by a ~$1.0B customer loan + customer warrants that net against billings and amortize against revenue; the company has never earned an operating profit and guides to −28/−32% core operating margin; single-source TSMC with no allocation commitment; NVIDIA Blackwell/CUDA TCO remains the default and AWS Trainium is partner-and-rival. Pre-mortem (18mo): GM never clears the low-40s, the OpenAI ramp slips, a 10-Q reveals "revenue" materially net of the customer loan/warrants drawing a related-party comment, the UAE concentration spooks the Street on a single-headline-risk basis → the multiple collapses toward CRWV's and the stock thirds. Contrarian read the market is underpricing (both ways): the bears are underpricing that the margin scare is partly accounting (warrant amortization + pass-through) and partly transitional (rent-back), which management can plausibly reverse — the prior dossier's "structurally buying revenue" verdict is not cleanly supported by the filings. The bulls are underpricing that the realized revenue is not OpenAI at all yet — it's the UAE, and that ~46x vs a 7x CRWV comp leaves no room for the ramp to disappoint by even a quarter. The honest synthesis: real technology, real demand pipeline, deeply concentrated current revenue, accounting-clouded margins, and a price that already credits the bull case twice over.

Lens 13 · Devil's Advocate (short-seller) (reinforced, but with a correction)

The short case has stronger Exhibit A than the prior dossier and a weaker Exhibit B. Exhibit A (stronger): the circular financing is now line items — ~$1.0B OpenAI loan at a waivable 6%, $365.8M OpenAI warrants + a 2.7M-share AWS warrant, and $21.6M of billings netting against the loan. The related party who is also lender-and-warrant-holder is wired into the revenue. And the concentration is worse than the bull narrative admits: it's not "OpenAI diversification," it's 86% UAE (MBZUAI + G42) with US revenue declining — a single geopolitical/sovereign-budget event (or a CFIUS re-look) could gut the realized book. Exhibit B (weaker than the prior dossier claimed): the "structurally low-margin, buying revenue with capex" indictment is partly rebutted by the disclosures — a chunk of the margin compression is warrant-amortization-against-revenue + datacenter pass-through gross-up + rent-back start-up cost, i.e. accounting and transition, not necessarily unit economics. A disciplined short underwrites the valuation (~46x vs ~7x CRWV) and the concentration, not a clean "the economics don't work" story that the filings don't fully support. Most dangerous competitor bulls underrate: AWS Trainium — Cerebras's partner and the prefill/disaggregation counterparty and a warrant-holder, perfectly positioned to learn the inference workload and in-source it; and NVIDIA simply pricing Blackwell TCO low enough that wafer-scale's latency edge doesn't pay for its margin deficit. What must hold for ~$182: GM recovers to mid-40s, OpenAI revenue lands on schedule, and the market keeps paying a 40x+ multiple — three things, all of which the Q1 tape already doubted. Single permanent-impairment scenario: the OpenAI MRA terminates or stalls (OpenAI can seize the loan funds and demand repayment) while the UAE book plateaus — capacity built, customers gone, ~$1.0B repayable. Low probability, catastrophic payoff — the definition of a tail short.

Lens 14 · Management Questions (carried + re-sharpened against the filings — see deep-dive-2026-06-25.md)

Ordered by information value:

1. Of the ~$1.0B OpenAI working-capital loan (6%, waivable) and the $365.8M OpenAI warrants, how much of reported/expected OpenAI revenue is recognized gross vs net of that financing, and are the loan terms on-market (ASC 850)?
2. FY25 revenue was 86% MBZUAI + G42 (UAE, related parties) and US revenue fell YoY — when does a non-UAE, non-loan-funded customer become >10% of revenue, and what is the current US/organic run-rate?
3. Is the 38-41% FY26 gross-margin band a transitional floor (rent-back + start-up) or the structural level for inference capacity against Blackwell TCO? What's the steady-state gross margin at scale on owned datacenters?
4. What are the OpenAI MRA termination triggers, and in which scenarios can OpenAI seize the loan funds / demand immediate repayment — and what is the company's contingency if the 750MW ramp slips?
5. With no long-term TSMC allocation commitment, what guarantees wafer supply for the 750MW+1.25GW ramp, and what is the single-wafer yield trajectory?
6. Why pay AWS (a Trainium competitor) in warrants for a partnership — what stops AWS in-sourcing the inference/decode workload once the disaggregation stack is proven?
7. What is the total committed capex + lease obligation to deliver 750MW (beyond the $344M not-yet-commenced leases), and the funding bridge after the IPO + loan?
8. How should investors think about EPS/GAAP profit when the forward-contract-liability remeasurement/extinguishment swings net income by ±$390M — what is the path to GAAP operating profit, and when?
9. What share of cloud revenue is datacenter pass-through booked gross, and what is the net (ex-pass-through) cloud growth and margin?
10. What is the dilution schedule from OpenAI + AWS warrants + options as payment thresholds vest, at what strikes?
11. How concentrated is the OpenAI relationship's optionality — if OpenAI does not exercise the 1.25GW Additional Capacity, what is the revenue cliff in 2028-2030?
12. What is the competitive win-rate vs NVIDIA in genuinely contested (non-sovereign, non-related-party) inference deals?
13. How do you defend a ~46x sales multiple when the closest capacity-financed comp (CoreWeave) trades at ~7x?
14. What governance protects minority (Class A) holders given super-voting Class B control and a related-party-dominated customer base?
15. What is the staged lock-up release schedule and expected insider-supply through Q3-2026 / mid-November?

---