CBRS › analyze
CBRS · Analyze
Price $183.20 · Market cap $43.5B · EV ~$37.1B · Verdict: WATCH · Conviction 4.5 · Risk: SPECULATIVE IPO 2026-05-14 at $185. Three and a half months of public history.
0. Your framing question first: how do I count this?
You said it "feels more like a small-cap but it's definitely not small-cap." Both halves are right. The resolution is that market cap sizes the position; it does not choose the framework.
| Dimension | Reads as |
|---|---|
| Market cap $43.5B | Large-cap — bigger than 60% of the S&P 500 |
| Revenue $680.7M TTM | Small-cap — the revenue of a $3–6B company |
| 708 employees | Micro-cap |
| 3 customers ≈ 76% of Q2 revenue | Micro-cap risk |
| Public history 3.5 months | Pre-seasoning — no trend to read |
| GAAP profitability | None, no dated path |
| Internal controls | Material weaknesses, unremediated |
So I ran the standard /analyze pipeline and applied the §6 speculative overlay from analysis_notes.md — risks first, default prior that most of these fail. The honest label: a speculative company wearing a large-cap price tag. Size it as speculative (1–2% max), not as a $43B semiconductor position.
1. Your thesis, tested
"What CBRS does is Nvidia but cheaper + more scalable + faster with minimal performance hit."
You flagged your own Hot Chips bias, which is the right instinct. One of those four terms survives.
| Claim | Verdict |
|---|---|
| Faster | ✅ TRUE. Independently corroborated. |
| Cheaper | ❌ Priced cheaper — funded out of a 14% gross margin, not out of architectural efficiency. That is a subsidy, not an advantage. |
| More scalable | ❌ Inverted. A CS-4 rack holds 132GB of SRAM. An Nvidia NVL72 holds ~13.4TB of HBM3e — roughly 100× the capacity per rack. Cerebras scales bandwidth, not capacity. |
| Minimal performance hit | ⚠️ Misframed. The trade is not a performance hit; it is memory capacity, workload generality, MoE efficiency, and data-centre integration. |
What is genuinely real — take this seriously
Wafer-scale integration is a real physical advantage, not marketing. WSE-3: 46,225 mm² of silicon, 4T transistors, 900,000 cores, 44GB on-wafer SRAM, 21 PB/s memory bandwidth. Weights sit in SRAM and cores talk over an on-wafer fabric, so there is no HBM round-trip and no inter-chip hop in the critical path. Autoregressive decode is memory-bandwidth-bound and serial — the one workload where GPUs are structurally weak. Artificial Analysis independently places Cerebras at the top of the throughput leaderboard on open-weight models (2,600 tok/s on Llama 4 Scout, 1,753 tok/s on gpt-oss-120B). The speed claim survives adversarial scrutiny.
The yield objection people raise is also wrong and worth discarding: cores are ~0.05 mm² with 1–1.5% redundancy and reconfigurable routing, making the WSE roughly 100× more defect-tolerant per unit area than a GPU. Yield is a solved problem. The cost problem is that one wafer buys you 44GB.
And there is a genuine supply-chain advantage management stated plainly in the Q2 8-K:
"We do not use HBM memory, CoWoS packaging or 3nm fabrication technology, all of which are currently supply limited."
Cerebras is the one AI accelerator vendor not bidding against Nvidia, AMD, Google and Amazon for scarce HBM. That is real — but note pattern-supply-outage-is-availability-not-quality: the advantage decays exactly as the shortage does.
Where it breaks — four things
(a) "Cheaper" is a statement about the product, not the stock.
| CBRS | NVDA | AMD | |
|---|---|---|---|
| EV / Sales | 54.5× TTM GAAP · 41.9× FY26 core | 19.9× | 18.8× |
| Gross margin | 14.2% GAAP · 41% "core" | 74% | 56% |
| Operating margin | −265% GAAP · −16% "core" | 66% | 17% |
Cerebras trades at 2–2.7× Nvidia's revenue multiple while earning a fraction of Nvidia's gross margin and losing money at the operating line. A cheaper product does not make a cheaper stock — it can make a worse one, because the cheapness is being given to the customer rather than retained. See pattern-asset-light-can-retain-or-distribute-the-economics.
(b) The memory ceiling did not improve this generation, and Cerebras knows it. The August 2026 Hot Chips launch was WSE-3 Turbo, not WSE-4: same TSMC 5nm node, same 4T transistors, same 900,000 cores, same 44GB SRAM. Compute doubled to 250 PFLOPS by raising clocks. ServeTheHome notes explicitly that Cerebras generations traditionally track TSMC process generations — "but not in this case." Memory-per-FLOP got worse by half.
A 1T-parameter model at FP8 needs ~1TB for weights alone, before KV cache — eight-plus CS-4 racks of pure weight storage for one model instance. Off-wafer I/O is 300 GB/s per wafer against 21–43 PB/s on-wafer: anything that does not fit in SRAM crosses a link five orders of magnitude slower. The fix is CS-6 with 3D-stacked DRAM — a 2028-class product. You do not build that to solve a problem you do not have. And CS-5 (2027) targets 10,000 tok/s/user on small models but only 5,000 on frontier models: Cerebras' own numbers show performance halving as model size rises.
(c) The business is drifting from "Nvidia" toward "Nvidia's customer."
| Q2 2026 | GAAP | Ex-warrant | GAAP GM | GM yoy |
|---|---|---|---|---|
| Hardware | $54.1M (−23%) | $82.1M (+17%) | 2% | 34% |
| Cloud & services | $126.0M (+281%) | $142.2M (+331%) | 20% | 26% |
| Total | $180.1M (+74%) | $224.3M | 14.2% | 31.1% |
⚠️ Correction worth stating plainly: the widely-reported "hardware revenue fell 23%" is an artifact. The 10-Q says the $16.2M decline "was primarily attributable to $28.0 million in amortization of customer warrant assets." Ex-warrant, hardware grew ~17% and core hardware gross margin improved 5 percentage points. Hardware is not collapsing.
The real point is the mix: 73/27 hardware-led in FY2024 → 30/70 cloud-led in Q2'26, and management guides cloud "will comprise a significantly higher percentage of total revenue." Cerebras is becoming a capital-intensive AI cloud operator that happens to use its own silicon — 600 MW live/contracted through 2027, H1 capex $549M, plus $318M of lease-financed capacity that never touches the cash-flow statement (pitfall-lease-financed-capex-hides-from-the-fcf-line). Nvidia sells chips at 74% and lets someone else build the data centre. That is a CoreWeave-shaped economic profile, not an Nvidia-shaped one — regardless of how good the wafer is.
(d) The unstated premise — "an Nvidia replacement" — is contradicted by Cerebras' own flagship partnership. The July 2026 AMD–Cerebras disaggregated platform assigns AMD Helios the prefill stage and Cerebras the decode stage only. That is the market's own engineering judgment that a Cerebras system cannot economically serve a complete inference workload alone. Cerebras is a component in someone else's rack. A component supplier in a disaggregated architecture is a price-taker — and 14% gross margin is what a price-taker's income statement looks like.
2. Fundamentals
The growth is real
| FY | Revenue | YoY | Gross margin | Operating income |
|---|---|---|---|---|
| 2022 | $24.6M | — | 11.7% | −$178.8M |
| 2023 | $78.7M | +220% | 33.5% | −$133.9M |
| 2024 | $290.3M | +269% | 42.3% | −$101.6M |
| 2025 | $510.0M | +76% | 39.0% | −$145.3M |
| TTM | $680.7M | +74% | 34.8% | — |
Three-year revenue CAGR 174.6%. This is a genuine commercial ramp, not a story stock.
🚩 The FY2025 earnings line is not real
FY2025 shows net income of +$237.8M against an operating loss of −$145.3M. This is pattern-net-margin-above-operating-margin-is-a-tripwire firing exactly as written. Source, confirmed in the cash-flow statement: +$363.3M "extinguishment of forward contract liability" — non-cash, below the operating line. The same line was −$401.3M in FY2024. Normalized FY2025 income: −$115.1M.
Treat the FY2025 "profit," the trailing P/E, ROE and any net-income CAGR on this name as void.
GAAP vs "core" — the central dispute
| Q2 2026 | GAAP | "Core" |
|---|---|---|
| Revenue | $180.1M (+74%) | $209.9M (+103%) |
| Gross margin | 14.2% | 41% |
| Operating margin | −265% | −16% |
| Net loss | −$450.5M | — |
| Diluted EPS | −$2.98 | — |
Two components, unequally legitimate:
- Customer-warrant amortization — $44.3M in Q2 ($28.0M hardware, $16.2M cloud), recorded as contra-revenue. Cerebras issues equity to its own customers as a commercial inducement; GAAP nets that cost against the revenue it bought. "Core revenue" is therefore revenue before the cost of the equity paid to win it. Treat it as a promotional metric, not a conservative one.
- Data-centre pass-through revenue recognised gross. The 10-Q is candid: "These pass-through revenues are not part of our core technology or service offerings" and pass-through "will increase significantly." Stripping zero-margin pass-through from both sides is defensible presentation — but it also means reported cloud revenue growth overstates economic growth.
Where margin is actually going (GAAP, from the 10-Q MD&A):
| Gross margin | Q2'26 | Q2'25 | H1'26 | H1'25 |
|---|---|---|---|---|
| Hardware | 2% | 34% | 28% | 32% |
| Cloud & services | 20% | 26% | 31% | 46% |
| Total | 14% | 31% | 30% | 36% |
Ex-warrant, hardware GM improved 5pp and cloud GM was roughly flat in Q2 — but cloud gross margin fell from 46% to ~36–40% across H1 even after adjusting for warrants. The segment carrying all the growth is the one whose margin is compressing, driven by pass-through growth and depreciation on capacity that is not yet fully utilized. Adjusting for warrants only, total Q2 gross margin was 31.1% — exactly flat year over year. The advertised "+940bps core gross-margin improvement" requires also removing pass-through from the denominator. On a like-for-like basis there was no gross-margin expansion this quarter.
Cash flow and the funding question
| H1 2026 | |
|---|---|
| Operating cash flow | −$47.5M (from −$123.8M) |
| Capex | −$548.9M |
| Free cash flow | −$596.4M |
| Lease-financed capacity (off the FCF line) | +$318.4M ROU assets |
Capex ran at 147% of revenue. Management plans to "more than triple revenue in 2027." At anywhere near 1× revenue on a ~$2.7B target, that is $2–3B of capex in one year against $8.6B of liquidity and negative operating cash flow.
The balance sheet is genuinely strong, and this is the bull's best point: $6.742B cash + $0.685B restricted + $1.179B investments = $8.606B liquidity, ~13% debt/assets, an $850M undrawn revolver. At the current ~$1.2B/yr burn that is ~7 years. At the burn the 2027 plan implies, closer to 3. Runway here is a function of ambition, not distress.
🚩 Share count — a vendor trap worth naming
Yahoo/fin.py report SharesOut 112.25M (Class A only) alongside MktCap $43.5B (full count). They do not reconcile, and every per-share metric derived from the small number is wrong — P/S 63.9, BVPS 40.27, EPS −4.66, ROE. This is pitfall-yahoo-share-count-dual-class-fpi on a tri-class issuer.
Filed count, 10-Q cover page, 2026-08-05:
| Class | Shares |
|---|---|
| Class A | 112,247,109 |
| Class B | 111,601,424 |
| Class N (non-voting, customer warrants) | 13,715,508 |
| Total | 237,564,041 |
Class N grew from 3,682,000 at 6/30 to 13,715,508 by 8/05 — OpenAI exercised 10,033,508 warrant shares in between. Every figure in this report uses the filed count and primary-source financials.
3. Moat
The quantitative base is not a moat yet
ROIC is not meaningfully computable — operating income is negative and invested capital was negative until the IPO recapitalized the balance sheet. Gross margin is the honest proxy: flat year over year at 31% ex-warrants, against Nvidia at 74%. A moat shows up as pricing power. Three years into a commercial ramp, this one has not shown up in the margin.
🚩 The moat is currently bought, not earned
In six months Cerebras issued $1.022B of equity to its customers — $822.9M to OpenAI, $149.4M to G42, $50.0M to AWS — against H1 revenue of $373.5M. They gave away 2.7× their half-year revenue in stock to secure commitments, and $1.128B of that remains on the balance sheet to be amortized against revenue through October 2031.
That is a customer-acquisition cost paid in permanent dilution. It can be rational at this stage of a ramp. It is not a moat. A moat means you can raise price; Cerebras is paying customers to take the product.
The CUDA argument is weaker than the bear case assumes — the real constraint is different
Cerebras' compiler ingests PyTorch directly and inference is served through OpenAI-compatible APIs, distributed via AWS Marketplace, Microsoft Marketplace, IBM watsonx, Vercel, OpenRouter and Hugging Face. For a developer consuming an inference API, switching cost is roughly a base-URL change. You are right that CUDA lock-in largely does not apply here — and bear theses leaning on porting friction attack the wrong target.
The decisive constraint is weight access. Every Cerebras public benchmark is on an open-weight model — Llama, Qwen, gpt-oss, Kimi K2.6. Cerebras must possess and compile the weights to serve a model. It therefore cannot serve GPT-5.x, Claude or Gemini unless the lab itself elects to deploy on Cerebras. The market is structurally partitioned: open-weight models plus whatever OpenAI chooses to run. Every share point gained by closed frontier models shrinks Cerebras' universe passively, with no competitor taking any action at all. This is the most under-appreciated structural limit in the story, and it has nothing to do with CUDA.
The prospectus concedes the adjacent fragility: "our software offering is currently optimized around certain AI models... Should AI model architectures significantly change, it may take us time to build out the software support." That is an admission the moat is a point-in-time architectural bet. Relatedly, MoE — where the entire frontier has gone — is architecturally awkward for wafer-scale, and Cerebras' own engineering blog concedes the batch-size conflict between attention and expert MLPs.
⚠️ MLPerf: the absence is itself a finding
Cerebras appears in neither MLPerf Inference v5.1 (Sept 2025, 27 submitters) nor v6.0 (April 2026, 24 submitters), and has never submitted in its history. Both lists include AMD, Google, Intel, Nvidia, Oracle, CoreWeave, Nebius and Lambda.
Artificial Analysis measures a live API endpoint — it cannot see what quantization was used or how the endpoint was provisioned. MLPerf's closed division fixes model, weights, numerics and accuracy target, and audits the result. Cerebras has consistently chosen the metric whose conditions it controls. This does not prove anything improper — the speed advantage is architecturally plausible and almost certainly real. It does mean the magnitude is unaudited, and a company with a genuine 10× it could prove cheaply has an odd reason not to.
Adversarial stress test — how do I kill this?
🚩 As Nvidia: I already did the decisive thing, and it is shipping now. Nvidia acquired Groq's inference assets for ~$20B (Dec 2025), taking founder Jonathan Ross and president Sunny Madra. The Nvidia Groq 3 LPX entered full production in August 2026 as a rack-scale low-latency accelerator for the Vera Rubin platform, with Nvidia claiming 35× higher throughput per megawatt on trillion-parameter models vs Blackwell NVL72. In parallel, Rubin CPX (late 2026) handles prefill.
Read what that does. Nvidia's public architecture is now GPUs for training and prefill, SRAM-first LPUs for decode — precisely Cerebras' thesis, validated, funded at $20B, and delivered inside the ecosystem customers already buy. Nvidia spent one-third of Cerebras' market cap to remove "GPUs can't do low-latency decode" from the board. Cerebras' differentiator now has a well-capitalized, ecosystem-native twin. This single development does more damage to your hypothesis than every accounting item in this report.
If I need more: Cerebras is at 14–20% gross margin. Nvidia at 74% can absorb a price war indefinitely; Cerebras cannot survive one quarter of it. And I am a vastly larger TSMC customer.
As Google: I don't need to act. Ironwood/TPU serves Gemini internally, and Gemini is a closed model Cerebras can never serve. Passive and free.
As Amazon: I am already inside. The March 2026 term sheet is binding "with respect to pricing, exclusivity, minimum capacity, and certain other protections in favor of AWS" — read the drafting. Definitive agreements were unnegotiated at the prospectus date; as of 6/30/26 lease conditions were unsatisfied and zero revenue recognized. I learn wafer-scale deployment economics at Cerebras' expense and hold the option to walk.
Cerebras' actual defence, fairly stated: 1. The physics is genuinely different and not licensable. Nvidia bought an SRAM-first LPU; it did not buy wafer-scale integration. The manufacturing know-how — ASE redistribution layers, in-house packaging, thermal and power delivery across 46,225 mm² — is process-embedded, not patent-dependent. This is real. 2. The OpenAI MRA is a genuine binding anchor, not vapor (see §5). 3. AMD needs an answer to Nvidia+Groq and chose Cerebras — the designated anti-Nvidia decode partner for the other GPU vendor. 4. $8.6B of liquidity buys years of being wrong.
Evergreen assessment: NO — 3/10 on durability
The durable-independent outcome requires all seven of: latency becoming a permanent first-class economic input (the one macro trend genuinely in Cerebras' favour, and not a small one) · CS-6 3D-stacked DRAM shipping on time and closing the capacity gap without surrendering the latency edge · flawless OpenAI MRA execution through 2030 · open-weight models holding share · gross margin reaching 45%+ at scale · Nvidia's Groq 3 LPX underdelivering · TSMC allocation holding with no contract.
At least three of those are genuinely uncertain and two are outside Cerebras' control. The likelier terminal states, ranked: (1) acquired — AMD is the obvious buyer and the Groq deal set the template and the price; (2) a profitable specialty component supplier, a fine business but not a $43.5B one; (3) squeezed out; (4) durable independent platform.
4. Valuation
Applicable models. No dividend → DYT and DDM N/A. Negative, non-meaningful earnings → Graham's IV N/A (√(22.5 × EPS × BVPS) on a −$4.66 vendor EPS built from a wrong share count is pure noise) and Bogle N/A without an earnings base. Per analysis_notes.md §3 this is a pre-profit name: EV/revenue versus peers and a reverse-DCF carry the entire weight.
| Basis | EV/Revenue |
|---|---|
| TTM GAAP revenue $680.7M | 54.5× |
| FY26 core guide $880–890M | 41.9× |
| FY27 "more than triple" ~$2.66B (a plan, not guidance) | 13.9× |
The $25.4B RPO is lower quality than it looks
This is the number every bull cites. Four discounts:
- Only ~15% recognizes within 24 months (through Dec 2027); 43% in months 25–48; the remainder after. This is a 2029–2032 story.
- It includes gross-basis pass-through — data-centre rent, power, leasehold improvements — at approximately zero margin. It is not $25.4B of Cerebras economics.
- Growth has stalled. RPO went $24.6B (12/31/25) → $25.4B (6/30/26) while $373.5M was recognized — roughly $1.2B of net new bookings in six months against a $25.4B base. If the MRA were compounding, this would look different.
- OpenAI holds termination rights. If Cerebras misses capacity timelines or service levels, OpenAI "has the right to terminate a portion or all of the agreement" — and can then seize control of the $1.0B loan account and demand immediate repayment. The single largest asset is contingent on flawless multi-year data-centre execution by a 708-person company that has never operated at this scale.
And even taken at face value, it does not rescue the price. Delivered evenly over five years it is ~$5B/year. At $5B revenue and a 20% operating margin, 30× EBIT gives a $30B EV — below today's $37.1B. The backlog is already more than priced in.
Reverse DCF — what must be true at $183?
For a 12% annual return over five years, equity must reach ~$76.7B by 2031. On a 20–25% mature operating margin at 25–40× EBIT, that needs 2031 revenue of $7.7B–$15.3B — a 54% to 77% CAGR sustained for five consecutive years — from a business at −16% core operating margin with flat like-for-like gross margin, facing a newly shipping Nvidia competitor.
Scenario fair value (2028 core revenue, discounted 15%/yr)
| Scenario | 2028 core rev | EV/S | 2028 price | PV today |
|---|---|---|---|---|
| Bear — OpenAI option unexercised, Nvidia takes the decode niche | $1.6B | 4× | $34 | $26 |
| Base — 750 MW deploys roughly on plan, margin stays thin | $2.9B | 7× | $89 | $67 |
| Bull — plan lands, gross margin scales through 45% | $4.2B | 11× | $191 | $144 |
| Blue sky — full 2 GW, ecosystem breaks open | $6.0B | 14× | $337 | $255 |
Fair value range $70–$145. Central ~$105. Spot $183.20 sits above my bull case midpoint and 75% above centre.
After a 53% drawdown, the market is still paying for the bull case. Note principle-down-a-lot-is-not-cheap: the decline is unwinding a 109% first-day pop from a $185 IPO, not creating value.
5. 🚩 Red flags — front and centre
- 🚩🚩 Material weaknesses in internal control, unremediated. Disclosure controls were not effective as of 2026-06-30. The named areas are revenue recognition, inventory management and costing, data-center assets accounting, and equity administration — precisely the four things this entire analysis turns on. A company with a contested GAAP-vs-core revenue presentation and $1.1B of customer-warrant accounting telling you its revenue-recognition and equity-administration controls are not effective is the most serious item in this file.
- 🚩🚩 Nvidia is now shipping a direct competitor. Groq 3 LPX in full production since August 2026, inside the Vera Rubin platform. The differentiator is no longer proprietary in economic terms.
- 🚩🚩 The lockup is not finished, and pending supply exceeds the float. Staggered releases: Sep 2 (~14.6M), Sep 16 (~14.6M), Sep 30 (~19.4M), Oct 14 (~19.4M), Oct 28 (~19.4M) — ~68M shares between Sep 30 and Oct 28 — then all remaining shares at the earlier of two trading days post-Q3 print or ~2026-11-09. Float is 94.1M. Short interest is 14% of float, up 52% month over month.
- 🚩🚩 Founders sold immediately and heavily into the first window. CTO Sean Lie $151.1M and CEO Andrew Feldman $47.3M, both 2026-08-21, with the COO, CAO and two directors alongside — ~$245M across Aug 14–21 at $196–250. Per §6.3 this is the loudest governance signal available and it points down.
- 🚩 Zero supply commitment against gigawatt obligations. Verbatim: "We have no formalized long-term supply or allocation commitments from TSMC... TSMC could reduce or eliminate deliveries to us on short notice, or raise their prices to us." Balance-sheet purchase commitments total $10.3M and are for consulting, not wafers. The WSE process is bespoke and jointly developed — there is no second source.
- 🚩 A ~$2.8B data-centre lease book against $885M of revenue. On-balance-sheet lease liabilities $568.8M, plus ~$1.5B signed in Q2 and ~$753M in Q3 not yet commenced — roughly $2.25B non-cancelable and off the balance sheet. Data-centre contracts are "longer term with limited ability to terminate and often include substantial liquidated damages" while customer agreements are "consumption-based." Fixed, long-dated, liquidated-damages-backed obligations against variable, cancellable, concentrated demand — the classic neocloud maturity mismatch.
- 🚩 Concentration rotated; it did not disperse. FY2023 G42 83% → FY2024 G42 85% → FY2025 MBZUAI 62% + G42 24% = 86% — and the filings state G42 and MBZUAI are related parties with respect to each other. The apparent FY2025 diversification was between two related parties. Q2'26: Customer A 34%, B 32%, C 10% — 76% top three, with one customer at 57% of receivables.
- 🚩 The OpenAI relationship is circular. OpenAI is ~32% of revenue, lent Cerebras ~$1.0B to build the capacity it will buy (repaid by netting against its own invoices — $86.3M offset already), and holds warrants for up to 33,445,026 shares at $0.00001 that reduce Cerebras' reported revenue as they vest. Revenue, financing and equity all run through one counterparty.
- 🚩 Q4 is a hockey stick. Q1 core $191.3M → Q2 $209.9M → Q3 guided $214–216M (+2.4% sequential) → FY guide requires Q4 of $264–274M (+23–27%). Q3 core operating margin is guided worse than Q2 (−25%/−23% vs −16%). A guide that raises the year while guiding next quarter to +2.4% puts the whole raise in the quarter no one can see.
- 🚩 A dated regulatory tripwire in ~8 months. BIS upgraded the UAE to Country Group A:5 in July 2026, with G42/Core42 the only pre-approved commercial entities — but eligibility expires if they do not become U.S. companies on or before 2027-04-10. Separately, the widely-repeated "Cerebras cleared CFIUS" is mechanically wrong: CFIUS permitted withdrawal of the joint voluntary notice after G42 was removed from the cap table; the $335M equity purchase never closed, and the G42 $500M–$5B purchase option was terminated. Any bull case citing that option is citing a dead instrument.
- ⚠️ Dilution pipeline beyond the float. ~23.4M unexercised OpenAI warrant shares, ~2.7M AWS warrant shares, ~21.5M vested employee options — roughly 20% more shares without a single new financing.
- ⚠️ Sell-side coverage is underwriter-anchored. 11 analysts, zero Sells, mean target $291.64, lowest target ($209) still 14% above spot. Coverage initiated at quiet-period expiry; nearly every firm was in the syndicate.
principle-repeated-fair-value-above-spot-is-method-failureapplies to the consensus, not to me. - ⚠️ Copyright class action (N.D. Cal., Oct 2025, amended) over LLMs and datasets. Company cannot estimate a range of loss.
🟢 Golden flags — the bull case stated fairly
- Revenue genuinely compounding — 174.6% three-year CAGR, +103% core in Q2, cloud +281%.
- $8.6B liquidity, ~13% debt/assets, $850M undrawn revolver. No financing distress for years.
- Real, non-licensable technical differentiation — wafer-scale is uniquely Cerebras', corroborated speed leadership, and a credible roadmap to CS-6.
- Supply-chain orthogonality — no HBM, CoWoS or 3nm dependency in a cycle where all three are constrained.
- The OpenAI MRA is binding, not an MOU — 750 MW with delivery milestones, a $1.0B loan, and warrant terms whose own expiry language confirms binding capacity commitments exist.
- AMD chose Cerebras as its decode partner against Nvidia+Groq.
- Customer base broadening: OpenAI, AWS, AMD, CrowdStrike, Cognition, Block, Figma, GSK, AlphaSense.
- Hardware is not collapsing — ex-warrant it grew ~17% with core gross margin up 5pp.
- OCF nearly breakeven (−$47.5M in H1, from −$123.8M).
- Smart money accumulating: Altimeter, Coatue, Atreides (10.5M shares), ARK buying the August weakness at ~$184.
- Honest TAM self-limiting — Cerebras describes "our addressable market within AI inference" at $66B (2025) → $292B (2029), rather than claiming the whole $251B → $672B headline. Use the smaller number; 74% of the headline is the training market they retreated from.
6. Verdict — WATCH · Conviction 4.5 · Risk: SPECULATIVE
A legitimate technology company with a contested niche, no pricing power yet, and a valuation that prices a seven-gate outcome as the base case.
Framework §0: is this a good business? — Partly. Real technology, real customers, a genuinely binding OpenAI anchor, no moat yet in the margin. Has the market priced it in? — Yes, and more. Great-ish company + expensive = wait. Not a short — the growth and the backlog are real enough that shorting into them is its own kind of foolish.
Why not lower. The technology is real and not licensable, the balance sheet is strong, revenue is compounding fast, and the hardware line is healthier than the headline suggests. This is not a fraud and not a nothing.
Why not higher. No margin of safety at any price above ~$110. The moat is purchased with equity rather than earned with pricing power. Nvidia is now shipping the same architectural thesis with better distribution. Concentration rotated instead of dispersing. And the internal-control finding lands squarely on the accounting lines the thesis depends on.
The timing point matters more than usual. Between now and ~November 9, roughly 68M shares release in five tranches and then everything remaining unlocks — into a 94M float, with 14% short interest and founders demonstrably willing to sell. There is no reason to pay up ahead of that. The first clean quarter with nothing left to unlock is the Q3 print in early November.
Action
A portfolio-specific passage was removed from the public build.
What would change my mind — dated and testable
| Upgrade to ACCUMULATE if | Downgrade to AVOID if |
|---|---|
| Gross margin reaches 45%+ at scale — the single cleanest falsifiable test of the bull thesis | Cloud gross margin keeps compressing as cloud scales, confirming price-taker economics |
| Q3 print (early Nov) shows GAAP GM > 30% and the Q4 hockey stick intact | Q4 guide is cut, confirming the raise was back-loaded fiction |
| Material weaknesses formally remediated | A restatement, or a third quarter of ineffective controls |
| OpenAI exercises the 1.25 GW Additional Capacity option | Any top-three customer reduces, defers, or triggers MRA termination |
| Top-three concentration falls below 60% of revenue | Nvidia Groq 3 LPX wins a named low-latency account from Cerebras |
| A Cerebras MLPerf closed-division submission | G42/Core42 miss the 2027-04-10 US-domicile deadline |
| Price reaches $85–110 with the thesis intact | An equity raise below $150 to fund the 2027 buildout |
| Post-Nov-9, insiders stop selling | Founders keep selling after the final unlock clears |
Data provenance
Every material figure comes from primary sources — the Q2 2026 10-Q (filed 2026-08-12), the Q2 8-K EX-99.1, and the 424B4 prospectus. Yahoo/fin.py were used for price, peer multiples and insider Form 4 data only.
⚠️ fin.py/Yahoo is unreliable on this ticker. Share count (Class A only), market-cap reconciliation, P/S, BVPS, EPS, ROE, and the FY2025 balance sheet (pre-IPO, negative equity) are all wrong or stale. The FY2025 income statement is technically correct but contains the +$363.3M non-operating gain that voids the earnings line.
Deliberately excluded as unverified: the Q2 GAAP revenue consensus figure; secondary-source AI-inference TAM statistics; the exact power draw of WSE-3 Turbo (inferred at ~54kW/wafer by ServeTheHome, not disclosed by Cerebras — the non-disclosure is itself a mild signal).