Cerebras Systems
This is the archived full research. Read the story
A maker of one giant AI chip, cut from a whole silicon wafer, that answers questions from AI models faster than anyone and rents that speed, mostly to OpenAI.
The scores
Growth3/4 ↑
Revenue rose 76% in 2025 and 74% in Q2 2026, but two related buyers paid 86% of 2025 and OpenAI now carries the backlog.
Business quality3/4 ↓
Measured fastest per user and chosen by OpenAI, but NVIDIA bought Groq's design and reportedly serves OpenAI's newest fast tier on GPUs.
Financial position3/4 →
About $7.7B of net cash and an undrawn revolver, but the 750 MW build needs capital and leases the filings do not total.
Execution record3/4 →
Beat its core guide in all three quarters as a public company and raised the year; three quarters is too short a record for more.
Per share capture2/4 →
OpenAI's warrant, Amazon's and the founders' units could add about 14.7%, and warrant discounts cut revenue to 2031.
Potential3/4 →
A $26.1B take-or-pay backlog is committed and dated, a third due within two years; the inference market is large, its fast part uncounted.
Risk profile2/4 ↓
One customer holds most of the backlog, two related Abu Dhabi buyers most past revenue, and a short seller disputes the chips' life.
Valuation1/4 ↑
About 24.1 times sales, 308% of a three-stage value near $32.83 at an 8% return: priced for perfection, beyond the 20% margin.
Evidence quality3/4 →
The contract, backlog and margins are filed, but the threats rest on reported claims neither company confirmed.
If it works 1.0× · if it fails you keep about 92%.
The Verdict
We rate Cerebras Unattractive, with medium conviction. Its chip is the fastest at answering AI questions, and the third quarter beat the company's own guide, raised the year and added Microsoft, but OpenAI still fills most of a $26.1B order backlog, NVIDIA is closing the speed gap and the losses run deep, so the business is mixed. On our base path the price is about 308% of our value; only if the backlog lands on its filed schedule does our value pass the price, hence medium conviction. As a bet on the big cloud providers it is defensive: the OpenAI business and the cash hold most of the value.
The long caseIf buyers keep paying more for faster answers, the clouds that serve them will want the fastest chip, and Cerebras could rent through every big cloud rather than one lab.
How we weighed it
The price against our value
Against it
Price at 308% of our value
The price sits far above our value, below only the case where the backlog lands on schedule.
The business
Cuts both ways
Mixed
A mixed business: the fastest chip, but one buyer.
Growth against what the price assumes
In its favour
The backlog grew and a third of it is now due within two years.
What could go wrong
Against it
One buyer carries it, NVIDIA is closing the gap and the chips may wear out early.
Against competitors
Cuts both ways
No filed rival compares directly; it has traded only briefly.
How its group is behaving
Cuts both ways
AI chip names move together on spending news.
The latest news and report
Cuts both ways
The report beat the guide and signed Microsoft, but the call said OpenAI's first tranches are behind schedule, with delay credits already taken off the backlog.
The evidence under this - 7 claims
- Revenue from OpenAI was $56.8M in Q2 2026, small beside the commitment, so delivery is still ahead (S02).
- SemiAnalysis said OpenAI's newest fast tier runs on NVIDIA GPUs (C22).
- On the Q3 call Cerebras said it is behind schedule on the first tranches, and either side may cancel a tranche delayed 90 days or more (C49).
- 2026-11-13editedre-read against the Q3 call: the finance chief restated the 750 MW take-or-pay terms and the 1.25 GW option; the delivery delay is recorded in C49
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
- 2026-11-12editedre-read in the Q3 10-Q concentration note
Core revenue adds back $36.4M of warrant amortisation and removes $16.4M of pass-through revenue; nine-month stock pay was $541.7M.
- Cerebras capacity is still ramping and OpenAI may have used GPUs to launch quickly (S08).
- Cerebras served GPT-5.6 Sol at 750 tokens a second in August (S03).
The base case caps growth well below the backlog's own schedule: on that schedule (the bull case) the value is about $143.65, above the price. At a 12% required return the base is about $20.87.
- 2026-11-10editedre-valued on the owner's rules of 10 November: the return on new capital glides to its terminal level over years six to ten
- 2026-11-12editedre-struck after the Q3 print: trailing sales, the loss, cash, debt and shares from the Q3 10-Q at the 11 November close; the bull case follows the backlog's schedule
- Timing can move at OpenAI's request (C06), and Cerebras has 700 MW of the capacity contracted so far (C27).
- Revenue is not profit: the core gross margin was 41% in Q3 before data-centre wear is tested (C48, C23).
- 2026-11-12editedre-read on the Q3 10-Q's backlog schedule
The company
When someone asks an AI model a question, a computer has to read the model's whole memory for every word it writes back. Graphics chips keep that memory off the chip, so they wait. Cerebras makes one chip from a whole silicon wafer and keeps the memory on it, so an independent tester measures it answering many times faster.
It sells these machines to buyers who run their own data centres, and rents their speed as a cloud service. OpenAI is now the largest renter.
The evidence under this - 4 claims
- The GPU comparison mixes Cerebras's own measurements with the benchmark (S01).
- SemiAnalysis said NVIDIA GPUs at low batch serve OpenAI's newest fast tier (C22).
- Large models must be split across several wafers, which adds complexity (S08).
- SemiAnalysis long judged the design's weaknesses too hard to cover up until fast tokens found buyers (S07).
- Revenue from OpenAI was $56.8M in Q2 2026, small beside the commitment, so delivery is still ahead (S02).
- SemiAnalysis said OpenAI's newest fast tier runs on NVIDIA GPUs (C22).
- On the Q3 call Cerebras said it is behind schedule on the first tranches, and either side may cancel a tranche delayed 90 days or more (C49).
- 2026-11-13editedre-read against the Q3 call: the finance chief restated the 750 MW take-or-pay terms and the 1.25 GW option; the delivery delay is recorded in C49
How it makes money
Two ways. A buyer pays once for machines, which was most of the money until 2026. Or a customer rents capacity by the year: OpenAI pays for tranches of three or four years, so revenue arrives every quarter as data centres come online.
Renting is now most of sales and almost all of the backlog of orders. Cerebras also gave customers cheap shares as part of the deals, and that discount is taken off reported revenue until 2031.
The evidence under this - 5 claims
- Revenue from OpenAI was $56.8M in Q2 2026, small beside the commitment, so delivery is still ahead (S02).
- SemiAnalysis said OpenAI's newest fast tier runs on NVIDIA GPUs (C22).
- On the Q3 call Cerebras said it is behind schedule on the first tranches, and either side may cancel a tranche delayed 90 days or more (C49).
- 2026-11-13editedre-read against the Q3 call: the finance chief restated the 750 MW take-or-pay terms and the 1.25 GW option; the delivery delay is recorded in C49
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
- 2026-11-12editedre-read in the Q3 10-Q concentration note
- 2026-11-12editedre-read in the Q3 10-Q balance sheet
Where the money comes from, stream by stream
| Stream | Share of sales (Q2 2026) | How it earns it |
|---|---|---|
| Cloud and other services | 70.0% | Dedicated take-or-pay capacity, pay-as-you-go tokens, support and custom model work, plus data-centre costs passed through to OpenAI. |
| Hardware | 30.0% | Cerebras systems sold for buyers' own data centres, mostly to G42 and MBZUAI, recognised on delivery. |
Who runs it
Andrew Feldman, a co-founder, has run Cerebras since 2016, after selling his last chip company to AMD. He and Sean Lie, the other co-founder and technology chief, hold super-voting shares, and that class controls almost all the votes.
Their main award pays only if the company is worth $75B, $150B or $250B. Since the listing Lie and Feldman sold about $233M of shares outside pre-set trading plans, and no insider has bought on the market.
The evidence under this - 3 claims
The lock-up let insiders sell to cover tax on shares vesting at the IPO (S02); a missing flag is not proof of a view.
Why it wins, and who could beat it
Why customers come back
Speed is the reason customers pay, and speed now has a price. AI labs sell a fast tier of the same model for more: OpenAI charges about 6 times its standard price for its quickest answers.
Its design avoids the scarce memory graphics chips need, and OpenAI signing for 750 megawatts is outside proof a buyer values the speed. What the moat lacks is time: speed is a lead, not a lock-in, and a customer can move its fast tier when a rival matches it.
The evidence under this - 5 claims
- Large models must be split across several wafers, which adds complexity (S08).
- SemiAnalysis long judged the design's weaknesses too hard to cover up until fast tokens found buyers (S07).
- The GPU comparison mixes Cerebras's own measurements with the benchmark (S01).
- SemiAnalysis said NVIDIA GPUs at low batch serve OpenAI's newest fast tier (C22).
OpenAI's Ultrafast tier reportedly runs at about 300 tokens a second on GPUs (S08), so a speed premium does not require Cerebras.
The price is for OpenAI's own customers; what OpenAI pays Cerebras per megawatt is not disclosed (S02).
- Revenue from OpenAI was $56.8M in Q2 2026, small beside the commitment, so delivery is still ahead (S02).
- SemiAnalysis said OpenAI's newest fast tier runs on NVIDIA GPUs (C22).
- On the Q3 call Cerebras said it is behind schedule on the first tranches, and either side may cancel a tranche delayed 90 days or more (C49).
- 2026-11-13editedre-read against the Q3 call: the finance chief restated the 750 MW take-or-pay terms and the 1.25 GW option; the delivery delay is recorded in C49
The threat
NVIDIA, which sells most AI chips, paid about $2.9B for a licence to Groq's fast-answer design, the approach closest to Cerebras's. A research firm then reported that OpenAI's newest fast tier runs on NVIDIA chips, though neither company confirmed it.
The big cloud companies build their own AI chips too, and some of them, like Oracle, are also Cerebras's partners.
The evidence under this - 4 claims
- Cerebras capacity is still ramping and OpenAI may have used GPUs to launch quickly (S08).
- Cerebras served GPT-5.6 Sol at 750 tokens a second in August (S03).
The rivals: NVDA, CRWV, AMD, GOOG, AMZN, ORCL.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| NVDA | NVIDIA, including the Groq technology it licensedPriced against it: Not checked | Cloud inference capacity, led by OpenAI | On sale | Not checked |
| CRWV | CoreWeave and other GPU cloudsPriced against it: Not checked | Cloud inference capacity, led by OpenAI | On sale | Not checked |
| AMD · GOOG · AMZN | AMD and hyperscalers' own chipsPriced against it: Not checked | AI systems sold for buyers' own data centres | On sale | Not checked |
| AMZN · ORCL | AWS and Oracle serving fast tiers on their own or NVIDIA hardwarePriced against it: Not checked | Hyperscaler and partner clouds: AWS, Oracle and paired inference with AMD | On sale | Not checked |
How big it could get
How big this could get
Cerebras, citing an industry research group, puts the market for answering AI questions at $66B in 2025, rising to $292B by 2029. A second forecaster sees all data-centre spending above $3 trillion a year by 2030.
Trailing sales are about 1% of that 2025 figure. Nobody measures the part that matters most here: how much buyers will pay extra for speed.
The evidence under this - 3 claims
The figure is a vendor forecast chosen by the company, and it counts all inference, not the fast part Cerebras sells.
Dell'Oro says growth depends on investment holding up, power and supply (S10).
- 2026-11-12editedtrailing revenue rolled to Q3
Who has signed
OpenAI has signed take-or-pay terms for 750 megawatts through 2028, and lent Cerebras about $1B to build it. The two Abu Dhabi buyers order machine by machine. Amazon signed a lease and a warrant that vests as it pays, Oracle agreed to buy machines and Microsoft cloud capacity once milestones are met; none of their sizes is disclosed.
The evidence under this - 6 claims
- Revenue from OpenAI was $56.8M in Q2 2026, small beside the commitment, so delivery is still ahead (S02).
- SemiAnalysis said OpenAI's newest fast tier runs on NVIDIA GPUs (C22).
- On the Q3 call Cerebras said it is behind schedule on the first tranches, and either side may cancel a tranche delayed 90 days or more (C49).
- 2026-11-13editedre-read against the Q3 call: the finance chief restated the 750 MW take-or-pay terms and the 1.25 GW option; the delivery delay is recorded in C49
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
- 2026-11-12editedre-read in the Q3 10-Q
- 2026-11-12editedre-read in the Q3 10-Q concentration note
Neither agreement's size is disclosed, the release says more will be shared next year, and the AWS lease had not started at 30 June (C15).
- 2026-11-12editedthe Q3 10-Q names Oracle and Microsoft agreements signed in the second half of 2026
The results
What growth costs
Before the listing, research spending grew slower than sales, so each new dollar of sales cost less to win. On the company's own measure the loss kept narrowing in 2026. The reported loss widened, but mostly because of share pay granted at the listing.
The evidence under this - 3 claims
- GAAP gross margin was 14% in the same quarter (C03); the warrant amortisation is a real discount given to customers and recurs until 2031 (C08).
- Core cloud gross margin includes depreciation struck on a three-to-five-year life that a short seller disputes (C23).
The last report (12 Nov)
Quarter ended June 2026, filed figures.
In the third quarter sales rose 55% to $210.8M. On the company's own measure, which adds back share discounts given to customers, they were $230.9M, above the $214M to $216M guided, at a 41% margin. The year's sales guide rose to $940M to $960M, the backlog grew to $26.1B and Microsoft and Oracle signed deals of undisclosed size.
On the call the company said OpenAI's first batches of capacity are behind schedule because buildings are late. The backlog already subtracts the credits for that, and it expects to be back on schedule in the second half of 2027.
What it did to the thesisIt mostly supported the case: sales and margins beat the bar, the backlog grew and Microsoft signed. But the call said the OpenAI capacity the backlog rests on is running late, and how much the new deals are worth is still unknown.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $-0.67 against $-0.15 expected.
The evidence under this - 7 claims
Core revenue adds back $36.4M of warrant amortisation and removes $16.4M of pass-through revenue; nine-month stock pay was $541.7M.
- Core revenue adds back warrant discounts and removes pass-through revenue, so GAAP revenue runs lower: $210.8M against $230.9M in Q3 (C48).
- SemiAnalysis reported OpenAI's newest fast tier running on NVIDIA GPUs, the customer the 2027 plan rests on (C22).
- 2026-11-12editedthe Q3 print beat the old guide; the guide was raised and a Q4 guide given
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
Neither agreement's size is disclosed, the release says more will be shared next year, and the AWS lease had not started at 30 June (C15).
- 2026-11-12editedthe Q3 10-Q names Oracle and Microsoft agreements signed in the second half of 2026
- 2026-11-12editedre-read in the Q3 10-Q balance sheet
The size of the delay credits is not disclosed beyond 'relatively small', and the chief executive said OpenAI's own data centres are also behind schedule, so the recovery date rests on builders the company does not control.
None of these figures is in the 10-Q or the release tables read in the first round, and the Oracle order's size is still undisclosed.
What management did with the money
The listing raised $6.2B, and first-half capital spending on data centres was $548.9M. Cerebras pays partly in its own shares: OpenAI's warrant could reach about 33M shares and Amazon's about 2.7M, both vesting as they buy.
Insiders in 2026Q3: 77 purchases and 266 sales, from their Form 4 filings.
The evidence under this - 4 claims
- 2026-11-12editedre-read in the Q3 10-Q balance sheet
The 10-Q gives no total cost for building 750 MW, so the remaining spend is unknown (S02).
The price, and where we disagree
In short. Priced for perfection. The price is about 308% of what we think the business is worth, at an 8% yearly return.
The shares listed at $185 in May, opened at $350, and fell to $96.36 by early November after each of the first two results, the report on OpenAI's fast tier and a short seller's attack on the chips' life.
Third-quarter results on 12 November beat the company's guide and raised the year. The finance chief said the lock-up on insiders' shares had already ended on 9 November, freeing up to about 171M shares, and the fourth-quarter report, due by the end of March, is the next test.
The evidence under this - 7 claims
The thesis had read the lock-up as ending on the second trading day after the Q3 results; the call puts the end earlier, on 9 November, and the 10-Q was not reopened here to reconcile the two.
- 2026-11-13editedthe Q3 call says the lock-up releases were complete by 9 November, before the results, not two trading days after them
- 2026-11-12editedthe Q3 results were reported on 12 November; the next dated report is Q4
Core revenue adds back $36.4M of warrant amortisation and removes $16.4M of pass-through revenue; nine-month stock pay was $541.7M.
- Cerebras capacity is still ramping and OpenAI may have used GPUs to launch quickly (S08).
- Cerebras served GPT-5.6 Sol at 750 tokens a second in August (S03).
- Culper's evidence is former employees and customers it does not name, reported second hand (S12).
- Cerebras had not answered publicly when this was written.
How we value it
- Method
- Cash the business will produce, with any shortfall raised in new shares
- Required return
- 8%
- Long-run growth
- 4%
- Moat
- An average moat
| Bear | Base | Bull | |
|---|---|---|---|
| Sales growth, years 1 to 3 | 30%, 30% | 60%, 60% | 200%, 150% |
| Normal operating margin | 8% | 10% | 11% |
| Return on new capital after year ten | 8% | 12% | 15% |
| Value a share | $23.14 | $32.83 | $144 |
- Price
- $101
- Priced on
- 11 Nov
- Price as a share of the base value
- 308%
- Yearly return the price offers on the base case
- 5.2%
- Risk check: valued at 12%, the rate for a business not yet making a profit
- $20.87
Where each figure comes from
| Figure | Source | Where exactly |
|---|---|---|
| Sales, the last twelve months | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Operating margin today | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Tax rate | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Sales per dollar of new capital | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Shares | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Cash and investments | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Debt | Filed report | Read by hand from the filed statements; the period it was read from was not recorded. |
| Sales growth, years 1 to 3 | Analysts' forecasts | FMP's consensus for the years to 31 Dec 2026 and 31 Dec 2027, as it stood when the value was struck on 13 Nov 2026: the average of 7 to 8 analysts' revenue estimates, 60%, 60%, each year held to at most 60%. FMP publishes how many analysts, not who they are. |
| Normal operating margin | Analysts' forecasts | FMP's consensus operating margin for the year to 31 Dec 2027, as it stood when the value was struck on 13 Nov 2026, the average of 8 analysts' estimates, less stock pay: 10% in the base case. FMP publishes how many analysts, not who they are. |
| Return on new capital after year ten | Our method's rule | Set by our reading of the moat: the 8% rate with none, 12% with an average one, 15% with a strong one; the bear case earns the rate. |
| The bear and bull cases | Our judgement | Our reading of the evidence; each case below names it. |
| Price | Market price | FMP's closing price on 11 Nov 2026. |
What each case rests on
Bear caseGrowth runs at half the base path, 30% and 30%, the margin settles a fifth lower at 8.0%, and new capital earns only the 8% rate: the downturn test.
Base caseRevenue grows 60% and 60% over three years, the analysts' path (C46) held to at most 60% a year as on every thesis, then fades toward 4%, closing 30% of the gap each year; the margin moves from -99.3% to 10.0%, the analysts' year-three figure after stock pay. The backlog schedule runs well above this path, which is why the bull case follows it.
Bull caseThe backlog lands on the 10-Q's schedule: about $8.6B in the two years to September 2028 (C06, C34), roughly the analysts' uncapped path (C46), so revenue triples and then grows 150%; the margin reaches 11.0% and new capital earns 15%.
What moves the base value
| Lower | Higher | |
|---|---|---|
| Growth in years 1 to 3, a quarter lower or higher | $29.17 | $38.56 |
| The normal margin, 15% lower or higher | $29.07 | $36.59 |
| Capital per dollar of growth, a quarter more or less | $30.25 | $34.37 |
| Return on new capital after year ten, at the rate or one step up | $26.72 | $35.27 |
| The rate, a point higher or lower | $27.75 | $41.67 |
| Long-run growth a point lower | $30.19 | - |
Value a share as growth and margin change
| Normal margin ↓ · growth → | 50% | 75% | 100% | 125% | 150% |
|---|---|---|---|---|---|
| 8% | $24.94 | $25.89 | $27.81 | $31.11 | $38.57 |
| 9% | $25.98 | $27.53 | $30.32 | $34.84 | $43.35 |
| 10% | $27.02 | $29.17 | $32.83 | $38.56 | $48.23 |
| 11% | $28.06 | $30.82 | $35.34 | $42.29 | $53.22 |
| 12% | $29.10 | $32.46 | $37.84 | $46.02 | $58.30 |
Growth is shown as a share of the base case's. Green: the price is at or under 80% of that value. Red: at or over 120%.
How we read the multiple
There is no profit yet, and future sales sit mostly in one contract. On schedule, that backlog of orders alone brings sales near $5.7B a year by 2030.
The price is fair if OpenAI takes its capacity on time at a normal margin, and dear if either slips.
- Less its net cash, the company is worth 23.3 times last year's sales.
- Take a 9% yearly return, a common rule of thumb, and a sale at 4 times sales in ten years.
- Sales must then grow about 30.0% a year.
Against its own history. Its shares have traded only since May, so the older normal shown beside the price was struck before listing and says nothing about what buyers pay now.
The evidence under this - 5 claims
- The contracted backlog alone would reach several billion dollars a year by 2029 (C34), so early growth far exceeds 29.5%.
- The test ignores the capital still to be spent on data centres (C12) and the OpenAI warrant's unvested shares (C45).
- 2026-11-12editedre-struck at the 11 November close on the Q3 balance sheet and trailing revenue
- Timing can move at OpenAI's request (C06), and Cerebras has 700 MW of the capacity contracted so far (C27).
- Revenue is not profit: the core gross margin was 41% in Q3 before data-centre wear is tested (C48, C23).
- 2026-11-12editedre-read on the Q3 10-Q's backlog schedule
- 2026-11-12editedtrailing revenue rolled to Q3
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
The base case caps growth well below the backlog's own schedule: on that schedule (the bull case) the value is about $143.65, above the price. At a 12% required return the base is about $20.87.
- 2026-11-10editedre-valued on the owner's rules of 10 November: the return on new capital glides to its terminal level over years six to ten
- 2026-11-12editedre-struck after the Q3 print: trailing sales, the loss, cash, debt and shares from the Q3 10-Q at the 11 November close; the bull case follows the backlog's schedule
Priced against its rivals
Its rivals have not yet been checked against their own filings, so there is no group to price it against; the price is judged against its own history instead.
The bet: what it pays if it works
The bet is on Cerebras renting its speed through the big cloud companies, starting with Amazon and Oracle in 2027.
If both run at a solid pace, the business would be worth about what the whole company costs today, so winning adds little. If neither pays, holders keep most of the value, because the OpenAI cloud and the cash already account for it.
| Hyperscaler and partner clouds: AWS, Oracle and paired inference with AMD revenue in two to three years | Dollars of yearly hyperscaler revenue needed | At the stated multiple |
|---|---|---|
| $100.00M | about 100.0M at 100% kept | 0.9× |
| $500.00M (the big win) | about 500.0M at 100% kept | 1.0× |
| $1.00B | about 1.0B at 100% kept | 1.2× |
To pay 2× (a Fair bet) it would need about $3.41B of Hyperscaler and partner clouds: AWS, Oracle and paired inference with AMD revenue (about 3.4B dollars of yearly hyperscaler revenue); to pay 5× (a Cheap bet), about $12.57B of Hyperscaler and partner clouds: AWS, Oracle and paired inference with AMD revenue (about 12.6B dollars of yearly hyperscaler revenue), at the same multiple.
Each row is Hyperscaler and partner clouds: AWS, Oracle and paired inference with AMD alone, plus the rest of the company (Cloud inference capacity, led by OpenAI) counted at about $13.50B.
The steps it has to take
- Done Amazon warrant and AWS lease signed
- Done Paired inference with AMD in production
- In progress Cerebras live on AWS Bedrock due 2027-03-31
- Not started First revenue from other hyperscalers due 2027-06-30
- Not started Cerebras inference on Oracle Cloud due 2027-12-31
The detail behind the label
The evidence under this - 5 claims
Neither agreement's size is disclosed, the release says more will be shared next year, and the AWS lease had not started at 30 June (C15).
- 2026-11-12editedthe Q3 10-Q names Oracle and Microsoft agreements signed in the second half of 2026
- The contracted backlog alone would reach several billion dollars a year by 2029 (C34), so early growth far exceeds 29.5%.
- The test ignores the capital still to be spent on data centres (C12) and the OpenAI warrant's unvested shares (C45).
- 2026-11-12editedre-struck at the 11 November close on the Q3 balance sheet and trailing revenue
- Core revenue adds back warrant discounts and removes pass-through revenue, so GAAP revenue runs lower: $210.8M against $230.9M in Q3 (C48).
- SemiAnalysis reported OpenAI's newest fast tier running on NVIDIA GPUs, the customer the 2027 plan rests on (C22).
- 2026-11-12editedthe Q3 print beat the old guide; the guide was raised and a Q4 guide given
Where we disagree with the market
Analysts, all positive, count the backlog of orders; short sellers, holding a third of the shares that trade, count one customer, wearing chips and NVIDIA.
Both are partly right: the contract is real, one buyer carries it, and today's price is about what it would earn.
The bulls may be missing that a fast tier can move: if NVIDIA's chips match the speed, OpenAI has less reason to take every tranche. The bears may be missing that the contract is take-or-pay and OpenAI lent Cerebras the money to build it.
The evidence against our middle view is the margin. If the cloud earns its 41% after the chips are paid for, the price is cheap; if the chips wear out early, it is dear.
What analysts say, and the evidence
The evidence under this - 7 claims
- The contracted backlog alone would reach several billion dollars a year by 2029 (C34), so early growth far exceeds 29.5%.
- The test ignores the capital still to be spent on data centres (C12) and the OpenAI warrant's unvested shares (C45).
- 2026-11-12editedre-struck at the 11 November close on the Q3 balance sheet and trailing revenue
- Timing can move at OpenAI's request (C06), and Cerebras has 700 MW of the capacity contracted so far (C27).
- Revenue is not profit: the core gross margin was 41% in Q3 before data-centre wear is tested (C48, C23).
- 2026-11-12editedre-read on the Q3 10-Q's backlog schedule
- Cerebras capacity is still ramping and OpenAI may have used GPUs to launch quickly (S08).
- Cerebras served GPT-5.6 Sol at 750 tokens a second in August (S03).
- Culper's evidence is former employees and customers it does not name, reported second hand (S12).
- Cerebras had not answered publicly when this was written.
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
What could go wrong, and how we would know
The risks
- One buyer carries it. Most of the backlog of orders is OpenAI's, its first batches of capacity are running late, and before that two related Abu Dhabi buyers paid 86% of sales.
- NVIDIA closes the speed gap. NVIDIA licensed Groq's fast design and reportedly serves OpenAI's newest fast tier, which would leave Cerebras's speed less special.
- The chips may wear out early. A short seller claims the wafers last under two and a half years in service; the company writes them off over three to five.
- Margins nobody has seen. The reported gross margin was 14%; the 41% figure removes share discounts to customers that recur until 2031.
- More shares are coming. Unvested warrants and founders' units could add about 15% to the share count, and the lock-up on insiders' shares ended on 9 November.
The evidence under this - 10 claims
- Delivery timing may change at the customer's request, the filing says.
- The figure excludes most future pass-through costs, which are billed on top.
- 2026-11-12editedre-read in the Q3 10-Q: the backlog grew and more of it is scheduled inside two years
- 2026-11-12editedre-read in the Q3 10-Q concentration note
- Cerebras capacity is still ramping and OpenAI may have used GPUs to launch quickly (S08).
- Cerebras served GPT-5.6 Sol at 750 tokens a second in August (S03).
- Culper's evidence is former employees and customers it does not name, reported second hand (S12).
- Cerebras had not answered publicly when this was written.
- GAAP gross margin was 14% in the same quarter (C03); the warrant amortisation is a real discount given to customers and recurs until 2031 (C08).
- Core cloud gross margin includes depreciation struck on a three-to-five-year life that a short seller disputes (C23).
- 2026-11-12editedre-read in the Q3 10-Q balance sheet
Some OpenAI tranches vest only if OpenAI buys the optional 1.25 GW (S02), and the executive units need a market value far above today's.
- 2026-11-12editedre-read on the Q3 10-Q's contingently issuable shares
The thesis had read the lock-up as ending on the second trading day after the Q3 results; the call puts the end earlier, on 9 November, and the 10-Q was not reopened here to reconcile the two.
- 2026-11-13editedthe Q3 call says the lock-up releases were complete by 9 November, before the results, not two trading days after them
The size of the delay credits is not disclosed beyond 'relatively small', and the chief executive said OpenAI's own data centres are also behind schedule, so the recovery date rests on builders the company does not control.
What we still do not know
- Whether OpenAI's newest fast tier really runs on NVIDIA chips, and whether that changes how much Cerebras capacity it takes.
- How long the wafer systems last in service, which sets what the cloud really earns.
- Whether any big cloud company signs paid capacity, rather than a lease and a warrant.
- What the six new deals above $30M each are worth, since no customer has disclosed one.
The evidence under this - 4 claims
- Cerebras capacity is still ramping and OpenAI may have used GPUs to launch quickly (S08).
- Cerebras served GPT-5.6 Sol at 750 tokens a second in August (S03).
- Culper's evidence is former employees and customers it does not name, reported second hand (S12).
- Cerebras had not answered publicly when this was written.
Neither agreement's size is disclosed, the release says more will be shared next year, and the AWS lease had not started at 30 June (C15).
- 2026-11-12editedthe Q3 10-Q names Oracle and Microsoft agreements signed in the second half of 2026
None of these customers filed or announced a contract value, and none appears among the 10% customers (C10).
What would change our view
- Unattractive→NeutralThe price falls back within 20% of fair value
It needs:
- ○The share price falls to about $39.40, back within 20% of what we think the business is worth.about ≈$18.04 today, 81% below the last close
- Unattractive→AttractiveThe price falls 20% under fair value
It needs:
- ○The share price falls to about $26.26, 80% of what we think the business is worth.about ≈$0.13 today, 100% below the last close
- Unattractive→AttractiveProof the backlog is not one customer
It needs all of:
- ○A customer other than OpenAI signs $2B or more of take-or-pay capacity.watched as it happens
- ○The company's own gross margin stays at 38% or more for two quarters.checked at the next results, due by 31 Mar
We would also be wrong if
- OpenAI cuts, delays or renegotiates its 750 megawatts.
- Cerebras shortens the life it writes its data-centre systems off over to under three years.
- No big cloud company or second large customer joins the backlog of orders by the end of 2027.
What is coming
- 16 November 2026 Lock-up ends: How many of the roughly 171M locked shares holders sell.
Every dated event ahead
- 16 November 2026 Lock-up ends: How many of the roughly 171M locked shares holders sell.
- 31 March 2027 AWS Bedrock goes live with Cerebras: The first hyperscaler revenue, and whether Amazon's warrant starts vesting.
- 31 March 2027 Q4 2026 results: FY2026 against the raised $940M to $960M core guide, the 47% to 49% core margin, and the 2027 plan restated.
Since we wrote this
- 13 NovEarningsCall says OpenAI capacity is late; Unattractive heldThe finance chief said the first OpenAI tranches are behind schedule on late buildings, with delay credits already deducted from the backlog. Our value does not rest on that schedule, so the view holds.
- 12 NovEarningsQ3 beat and raised; Unattractive held, conviction cut to mediumRevenue and margins beat the bar and Microsoft signed, but our capped value sits far below the price; only the backlog landing on schedule would reach it.
- 11 NovThesis reviewUnattractive, high conviction: the view holds even in our best caseConviction re-graded on the new anchors: how robust the view is, not whether anything is still open.
- 10 NovNewsBilt picks Cerebras for its AI assistant, with no size or price disclosed. View heldBilt selects Cerebras to power its Neighborhood OS and Concierge assistant.
- 9 NovThesis reviewUnattractive: priced at 221% of a value near $40.51Re-valued in three stages at an 8% return, the method now drawn on every thesis: the price sits beyond the 20% margin of safety.
- 3 NovThesis writtenNeutral, low convictionWritten with the shares at $96.36, below the $185 IPO price, nine days before the Q3 2026 results and the end of the lock-up.
Sources 19
Sources - 19
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-05-14business, competition, market sizing, FY2024 and FY2025 statements, customer concentration, useful lives, share classes, ownership, executive grantsbacksC01C02C07C10C13C14C18C19C23C24C26C33C38C39C42
- S02primarycompany filing2026-08-12results, OpenAI MRA, working capital loan, warrants, remaining performance obligations, customer concentration, cash, leases, lock-upbacksC01C03C04C07C08C12C15C28C32C40
- S03primarycompany release2026-08-12core measures, Q3 and FY2026 guidance, capacity and customer highlights
We have made rapid progress in key areas required to deliver exceptional growth against our remaining performance obligations of $25.4 billion, and plan to more than triple revenue in 2027.
backsC04C19C27C37 - S04primarytranscript2026-08-12the 2027 plan, data-centre capacity, hyperscaler timing, disaggregated inference, analyst questions on concentration
And with all of this progress, I think it's important to keep in mind that our $25 billion in RPO does not reflect any backlog of business from AWS or any other hyperscaler at this time.
backsC16C36C37 - S05secondaryvendor data2026-11-03Research pack, FMP and FINRA market datacontent/CBRS/research-pack.jsonprice history, analyst split, short interest, earnings date, insider feed
Cerebras Systems Inc. is a leading innovator in artificial intelligence infrastructure.
backsC28C29C30C31C32 - S06primaryindustry data2026-11-03measured output speed and price on Cerebras
For output speed, the fastest models are gpt-oss-120b (low) (1,746 t/s) and gpt-oss-120b (high) (1,684 t/s).
backsC18 - S07secondaryanalyst2026-05-13why buyers pay for speed; Nvidia's Groq deal described as SRAM-based
with frontier labs releasing fast, priority, standard and batch tiers of the same model weights, the world has revealed their preference for fast tokens with their wallets.
backsC20C21 - S08secondarynews2026-09-30the SemiAnalysis claim, Ultrafast pricing, OpenAI's mixed fleet
OpenAI has not said publicly which hardware serves the 6.1 Sol Ultrafast tier, and Cerebras has not commented on the claim at the time of writing.
backsC22C44 - S09primarycompany filing2026-08-26the payment for the non-exclusive Groq licencebacksC21
- S10primaryindustry data2026-08-18data center capex outlook and the growth of AI-specialised clouds
The newly added AI-specialized cloud segment, comprising AI model builders and neocloud service providers, is projected to grow at nearly a 60 percent CAGR, outpacing the growth of other customer segments.
backsC25 - S11primarycompany release2026-10-27the Oracle offering and its date
Cerebras Inference on OCI is planned to be available in calendar year 2027, giving OCI customers another option for AI applications where speed is critical.
backsC17 - S12secondarynews2026-10-29Culper Research's short claims, as reported
In draft IPO filings, Cerebras had said that its chips could last between three and five years, but Culper said the wafer-scale chips last just 1.4 to 2.3 years.
backsC23 - S13primarycompany filing2026-10-28open-market sales and the Rule 10b5-1 flagbacksC35
- S14primarycompany filing2026-11-09trailing revenue, operating profit, stock pay, tax, debt, cash and the diluted share countread, but no claim rests on it
- S15secondaryvendor data2026-11-09the analysts' average revenue and operating profit for the next three fiscal years, and the closebacksC46
- S16primarycompany filing2026-11-12results, remaining performance obligations and their schedule, customer concentration, the OpenAI loan, warrants, cash, leases, the Oracle and Microsoft agreements, shares outstanding at 5 NovemberbacksC06C09C10C11C13C16C33C34C42C43C45C47C48
- S17primarycompany filing2026-11-12core revenue and margins, the Q4 and FY2026 core guide, 700 MW of capacity, the hyperscaler cloud dealbacksC05C11C16C27C41C48
- S18secondaryvendor data2026-11-12the close and the analysts' average revenue and operating profit for fiscal 2026 to 2028backsC33C47
- S19primarytranscript2026-11-12the OpenAI delivery delay, delay credits and cancellation terms, the lock-up end, segment core margins, the Q4 mix, Oracle outside the backlog, disaggregated inference with AMD in production, the CS-5 schedule; questions from Barclays, Morgan Stanley, Citi, Needham and TD Cowen
After signing the agreement with OpenAI, it took us longer than we would have liked to find contracts for and bring online the substantial increase in data center capacity required to meet the aggressive requirements in the agreement. And as such, we find ourselves running behind schedule on initial tranches.
backsC07C32C36C49C50