Cerebras

IThe business

Cerebras builds the largest chip ever sold in volume. A normal processor is a square cut out of a silicon wafer, and a wafer yields dozens. Cerebras keeps the whole wafer as one processor, 46,225 square millimetres with 900,000 cores and 44 gigabytes of fast memory on the chip.² The company says this removes the data movement between chips that slows down GPU clusters, and that answers come up to 15 times faster on open models.² That is the company’s claim, not a result I could test.

It sells that chip in three ways. It sells whole systems to customers who run them in their own data centers; that was 70% of 2025 revenue, or $358.4M. It rents capacity in its own cloud, either by the token or as dedicated capacity under take-or-pay contracts; that was $151.6M. And it sells services to build and tune models.² The mix is reversing. In the second quarter of 2026 cloud was 70% of revenue, up from 32% a year earlier R (reported).³

The reason is OpenAI. In December 2025 Cerebras signed a master relationship agreement under which OpenAI is contractually committed to buy 750 MW of inference capacity, delivered in tranches from 2026 to 2028, each lasting three or four years and extendable to five.³² OpenAI also holds an option on a further 1.25 GW by the end of 2030, for 2.0 GW in all.³ OpenAI lent Cerebras about $1.0B in January 2026 at 6%, to be repaid mostly by service credits, and received a warrant for up to 33.4M shares at a thousandth of a cent.³ Revenue allocated to unfinished contracts was $25.4B at 30 June, up from $25.0B at 31 March; the company says a significant amount of it is OpenAI’s R (reported).³¹¹

That number needs care. Divide it by 750 MW and each megawatt is about $34M of contract over three to five years, roughly $7M to $11M a year C (calculated). It includes some data-center costs passed through to OpenAI at no profit, and it leaves out the pass-through for capacity beyond what leases already exist.³ The company expects 22% of it by June 2028, 43% in the next two years, and 35% after that.³

Cerebras is a fabless company: TSMC makes the wafers, and Flex, Sanmina and Rocket EMS assemble the systems.²¹⁰ It leases data-center space rather than building it. At 30 June it had $568.8M of operating lease liabilities on its balance sheet and had signed leases for about $2.25B more that had not yet started R (reported).³ The company said in August that more than 600 MW was live or under contract for delivery by the end of 2027.¹⁰

All figures are in US dollars under US GAAP. The fiscal year is the calendar year. The company has not yet filed an annual report as a public company; its first 10-K, for 2026, is due in early 2027.¹² It reports a second set of numbers it calls core, which add back stock pay, the non-cash cost of warrants given to customers, and pass-through items. I use core only where I say so.

How much history exists. Four audited years: 2022 and 2023 from a 2024 registration statement that the company later withdrew, and 2024 and 2025 from the prospectus.¹³² Add two unaudited 2026 quarters. The 2022 and 2023 statements describe a business that sold chips to a few buyers and lost money; they say little about a business that rents 750 MW. So a ten-year model here rests on the signed contracts and my assumptions about renewals, not on a long record. I say so wherever it matters.

Customers are few. In the second quarter one customer was 34% of revenue, a second was 32% and a third was 10%; in 2025 MBZUAI, a university in Abu Dhabi, was 62% of revenue and its affiliate G42 was 24%, and in 2024 G42 was 85% R (reported).³² The filing does not name them, but OpenAI’s revenue under the agreement was $56.8M of $180.1M in the quarter, which is 32%, so Customer B is OpenAI C (calculated).³

IIThe moat

I ask what protects the profits, and then what would let someone else take them. Here the answer has three layers, and only one is solid.

The manufacturing process. Nobody else has sold a wafer-sized processor in volume. The company worked with TSMC to build a way to connect the dies across a wafer and route around flaws, and it says it protects the method with patents.² That took ten years and about $1.9B of preferred stock raised by the end of 2025 R (reported).² A new entrant would need the same time and money. This is real, but it protects the chip, not the business of renting it out.

The OpenAI contract. The commitment is firm for the committed 750 MW, with OpenAI paying as tranches are delivered, and the $1.0B loan is secured on a lockbox account.³ It is also the weak point, because it is one buyer. If the agreement ends for any reason other than OpenAI’s own uncured breach, the loan can become due at once.³ Beyond the committed 750 MW, the extra 1.25 GW is an option that OpenAI may or may not take.

Supply. The company says its design avoids high-bandwidth memory, advanced CoWoS packaging and 3 nanometre wafers, which are the parts in short supply for GPUs, and that it has secured the TSMC wafers it needs.¹⁰ That is an advantage in a shortage. It is also a dependence: TSMC makes every wafer, there are no long-term capacity commitments with suppliers, and most orders are on a purchase-order basis.²

Where it is thin. Software lock-in is small. The compiler takes PyTorch models directly and customers can run models not trained on Cerebras hardware, which makes it easy to adopt and, for the same reason, easier to leave.² Hyperscalers and labs are building their own chips and some are customers too.² The cloud is a business where scale and cost of money decide who wins, and Cerebras competes with companies that borrow far more cheaply.

So what protects the profits is a process lead that is real and a contract that is large, with almost nothing that makes either permanent. The test is whether the price holds when the first customer’s tranche comes up for renewal. That first test comes with OpenAI’s earliest tranche, delivered in 2026 and ending between 2029 and 2031 E (estimate).

IIIThe competitive advantage

The advantage is speed for one user. A language model produces one token at a time, and on a GPU each token waits on memory. The chip holds the model’s working memory on the silicon, so each token comes sooner. Cerebras puts it at up to 15 times faster than leading GPU solutions and says OpenAI’s GPT-5.6 Sol runs on its capacity at 750 tokens per second.²¹⁰ I did not test those numbers, and speed is only worth money when a user will pay for it. Coding assistants, voice agents and security filters are the early cases, and the company names Cognition, Lovable, CrowdStrike, Block, Figma and AlphaSense as customers.¹⁰

The second part is cost. Cerebras sells itself as cheaper to run per answer because it needs less networking and less power for moving data.² I cannot check that either. What I can check is what the company earns on the capacity it rents. Its core gross margin was 41% in the second quarter and it guides 41% to 43% for 2026, with the third quarter at 38% to 40%.¹⁰ That is after depreciation of its own systems. The core definition leaves out stock pay and the cost of warrants, so I do not read it as a final margin; it is the company’s best case.

The company is also widening the offer. In June 2026 it signed a global hardware leasing agreement with Amazon Web Services for a joint solution in AWS data centers, and it has built a disaggregated inference design with AMD that it says lifts throughput up to five times, to reach production in the fourth quarter and Amazon Bedrock in the first quarter of 2027.³¹⁰ Neither had produced revenue at 30 June: the AWS leases had not commenced.³

Who can take it away. Nvidia’s GPUs get faster and its software is the default. Hyperscalers, including AWS, Microsoft, Google and Oracle, are named competitors, and so are CoreWeave and other neo-clouds.² Several labs build their own accelerators. If a model maker can get most of the speed from its own chip or from a rival inference design, the price Cerebras can charge for speed falls, and the contract with OpenAI would be tested at renewal.

The advantage is therefore a lead measured in months to a few years, backed by a contract that locks it in for the first term. My model assumes the lead holds for the first term and that renewals come at lower prices.

IVManagement

Andrew Feldman, 56, co-founded Cerebras in 2016 and is chief executive and a director. Sean Lie, 46, the other co-founder in the executive team, is chief technology officer. Dhiraj Mallick is chief operating officer, and Robert Komin has been chief financial officer since March 2024; he was CFO of Sunrun from 2015 to 2020.² Feldman and Lie came from AMD after it bought their earlier company, SeaMicro. The record that matters here is that they built a chip nobody else had shipped, kept the company alive through years of losses, and then in 2025 and 2026 raised about $9.3B from preferred stock, an OpenAI loan and an IPO C (calculated).²³

The prospectus also discloses that in 2008 Feldman settled an SEC civil case from his time at Riverstone Networks, without admitting or denying it, and that in December 2007 he pled guilty to one count of circumventing accounting controls and was sentenced to three years of probation.² I mention it because the company now has material weaknesses in its own financial controls, and the same prospectus says so.³ I draw no more from it than that.

How they are paid. Feldman’s salary was $450,000 in 2025 and is $600,000 in 2026. In 2025 his total pay was $11.75M, almost all of it a grant of 400,000 restricted stock units valued at $10.8M.² In February 2026 the board approved three more packages for him and for Lie, advised by a compensation consultant: a make-whole grant of 500,000 units for Feldman, an annual grant of 243,902, and 5.7M performance units. Lie received 312,500, 182,926 and 3.3M.² The performance units vest in three equal parts if the market value of the company, averaged over 90 trading days, passes $75B, $150B and $250B, from six months after the IPO.³ The company values them at $531.2M for accounting, because it estimates the chance of the hurdles.³

What that pays them to do is plain: raise the market value of the company, by a lot, not necessarily by earning money. The first hurdle is $75B, which is about 2.9 times the market value of today’s outstanding shares C (calculated). The size matters because the founders already own large blocks. At 31 March 2026 Feldman beneficially owned 5.4% of the shares and Lie 2.9%; the Class B shares carry 20 votes each, and the holders of Class B shares controlled about 99.2% of the votes after the IPO.² That means the people who write the targets also control the vote on them. Outside investors who own more than 5%, Fidelity (11.0%), Benchmark (9.5%), Foundation Capital (8.3%), Eclipse (7.3%) and Alpha Wave (6.5%), hold Class B shares as well; three of the five also sit on the board, through Benchmark’s Eric Vishria, Foundation’s Steve Vassallo and Eclipse’s Lior Susan, so three of six directors represent early investors.²

The sales are the part I weigh most. A founder-run company in a lock-up can have good reasons to sell: taxes on stock units, diversification, the first chance after ten years. The stock units vested in a way that created large tax bills, and the company withheld shares to pay them, which cost $416.7M in cash in the first half.³ So not all selling is a signal: $26M of the $415M was sold to cover taxes. Still, most was sold under trading plans, and the two co-founders and the chief operating officer sold $368.6M between them: the chief technology officer $200.7M, including 710,247 shares for $147.6M on 20 August, the chief operating officer $103.6M and the chief executive $64.3M C (calculated).¹⁹ The chief operating officer sold 396,000 shares for $80.8M in one day, more than 100 times his 2025 pay of $690,000. His average price was $204; the stock closed at $108.37 four weeks later, 47% lower C (calculated).⁴²¹ The one sale since the price fell below $120 is Mr. Lie’s 60,000 shares on 27 October. I draw no conclusion about why.

What they did with capital. There is only one cycle to judge, and it is short. The company raised preferred stock at $36.23 a share in September 2025 and $89.02 in January 2026, then sold stock to the public at $185.00 in May 2026; it bought back employees’ shares at $36.23 in a December 2025 tender.² The price at the tender was a fifth of the price at the IPO six months later. The $6.2B of IPO money, and the $1.0B from OpenAI, are now spent only a little: capital spending was $548.9M in the first half.³ The real capital allocation test is whether $8.6B of cash is turned into rented capacity that earns more than it costs. That test has barely begun.

VFinancials

Ten years of history is not possible. The company has four audited years, 2022 to 2025, and each looks like a different business. Revenue went from $24.6M to $78.7M to $290.3M to $510.0M, while the operating loss stayed between $101M and $179M.¹³² The first two years are from a registration statement that was withdrawn, and the lines for expenses are grouped differently from the later years, so I do not compare R&D across them closely.

What changed in 2026 is not visible in the earlier years. In the first half, cloud revenue was $208.8M, 232% above the year before, and hardware revenue was $164.7M, up 18% R (reported).³ Gross margin fell from 36% to 30% on a reported basis, because $46.3M of the warrants Cerebras gave to customers was charged against revenue and $16.3M of stock pay went into cost of revenue.³ The research and development line jumped from $113.5M to $395.6M, of which $228.0M the company attributes to stock pay and payroll tax caused by the IPO.³ These are one-time catch-up charges for units that vested on the IPO date, not a run rate. Even so, the loss from operations of $492.3M in the half is a real cost to shareholders, because the units were real shares given away.

Core against GAAP. The company’s core operating loss was $33.6M in the second quarter against $477.2M on a GAAP basis. The difference is $377.0M of stock pay, $44.3M of customer-warrant amortization and $22.8M of payroll tax on the IPO.³ Core is a fair guide to how the cloud operates; it is not what shareholders keep. Stock pay of $377.0M in a quarter is partly catch-up, but even the earlier annual rate of $50M will rise: there are 16.5M unvested units and 9.0M founder units to expense.³

The warrants are a price cut. Cerebras gave OpenAI the right to buy up to 33.4M shares for a thousandth of a cent, and gave G42 and AWS smaller warrants. The accounting treats the value of vested warrants as an asset, $1.13B at 30 June, written off against revenue as the related revenue is earned through October 2031.³ The charge was $46.3M in the half. The grant-date value was $82.02 a share for the first OpenAI warrant, so a full vest is worth about $2.7B at that price and about $3.7B at today’s price, shares OpenAI could sell.³ I treat the warrant as shares, not as a cost: it dilutes the shareholders instead of lowering revenue, and my share count includes the part I expect to vest.

Cash and obligations. At 30 June the company held $6.74B in cash, $1.18B in Treasury securities and $0.68B of restricted cash, $8.61B in all.³ Against that stand the OpenAI loan of $918.2M, repaid so far only by $86.3M of service credits and no cash; $418.3M of deferred revenue and $242.7M of customer deposits, which are services still owed; and the leases.³ Operating lease payments recognized come to $690.3M undiscounted, and leases signed but not started add about $2.25B C (calculated).³ An $850M revolving facility with Morgan Stanley is undrawn, costs SOFR plus 2.25%, matures in April 2031 and requires $150M of minimum liquidity.³

Cash is not going to be enough for the whole build if OpenAI orders more than the committed 750 MW, and the company says it expects to need significant additional capital for cloud.² In my base case the cash falls to about $4.4B by the end of 2027 before it recovers, so no new equity is needed for the committed capacity E (estimate). The first half’s equipment spending of $548.9M, with $604.0M still in construction in progress, gives a sense of the pace: if I take what the company has put into data-center and computer equipment, $403.6M, against the second quarter’s annualized core cloud revenue of $511M, each dollar of yearly revenue so far has needed about 80 cents of equipment C (calculated).³¹⁰

Shares. At 5 August there were 237.6M shares in three classes: 112.2M Class A, 111.6M Class B and 13.7M Class N, which has no vote.³ At 30 June the company excluded 76.0M potential shares from its loss per share: 26.1M warrants, 23.8M options, 16.5M unvested units, 9.0M founder performance units and 0.6M early-exercised shares.³ Adding what I expect to arrive, my base share count for valuation is 275M, 16% above the 237.6M outstanding. IPO-year dilution is large: the share count was 57.9M Class A at the end of 2025, plus 113.3M preferred, against 228.2M of common at 30 June.³

Controls. The company says its disclosure controls were not effective at 30 June because of material weaknesses it had already reported: weak general controls over the systems that feed its financial reports, too little separation of duties, and too few accounting staff with the right experience. These affected revenue recognition, inventory, data-center assets and equity records.³ None led to a misstatement the company has found. They matter because the accounting here is complicated, with warrants, a loan repaid in services and pass-through revenue, and the company has not yet published an audited annual report as a public company; the 2025 audit was done before the IPO.

VIValuation

Before the numbers, three ways this investment goes badly wrong.

One. OpenAI does not renew when its first tranches end in 2029 to 2031, or renews at much lower prices. Cerebras would be left with data-center leases it must keep paying, about $2.9B of lease payments already signed, and with systems three to five years old.³ The loan could also fall due if the agreement ends for a reason other than OpenAI’s own breach.³ Two. The build costs more than the cash on hand. The 10-Q says the cloud needs significant capital and that equity, debt and convertibles may all be used.³ Raising equity at a lower price than the $185 of the IPO would hand more of the business to new buyers. Three. Someone else gets the speed for less, so Cerebras can charge for capacity only at the price of a GPU. In every one of these the company still has the cash it raised, which is why the bear value is above zero.

What an owner can take out. Not what the company reports. In my base case for 2030, EBITDA is $3,475M, but stock pay takes $550M, tax $265M and equipment $1,339M. The equipment is not growth. It is the cost of replacing the systems that wear out every four years, and the cost of keeping up with a customer who will not accept old chips. What is left, $1,321M, is 38% of EBITDA C (calculated). Any figure that adds back stock pay and counts only the early growth capital will look three times better.

What I cannot judge. Three inputs. First, what happens to OpenAI’s pricing when the first tranche ends; nothing in the filings says. Second, how many megawatts beyond 750 OpenAI will take; the option is for 1.25 GW more, but I include only part of it, in the bull. Third, how much the equipment will cost per dollar of revenue at scale; I anchor to the 80 cents so far, which is a guess about the future from an early year. The range is wide because these three are large and unknown.

How the model works. I start from the 30 June balance sheet and the second half of 2026. For 2027 to 2036 I build core revenue from the company’s outlook and the signed contract, subtract the cash cost of running the cloud, add the hardware business, and take off research, sales and administration, stock pay, tax and the equipment needed. Stock pay is a cost, because the units are shares given away. Equipment is built one year ahead of the revenue it serves. Cash above a reserve is paid out from 2029. Value is the present value at 11.5% of what is paid out plus the cash and business left at the end of 2036, divided by shares, with the OpenAI warrant and the options counted as they vest or come into the money E (estimate).

The base is $39 a share. About $27 of it is cash and the rest, about $13, is the business. The bear is $11: OpenAI takes less and renews at lower prices, the company burns $3.4B in its first two and a half years, and it runs off the fleet from 2031 as revenue falls 15% a year. The bull is $109: all 750 MW is delivered, part of the extra 1.25 GW is bought, core gross margin reaches about 57%, above anything the company has reported, and the business grows 5% a year after 2031 E (estimate). The margin of safety follows my other pages: $14, which is where the bear value is three quarters of the price paid.

Reverse DCF. The price asks for about my bull case. Holding the base margins, revenue must be 1.9 times my base path, or about $14B in 2030, to reach $109.18. Margins alone cannot do it: cutting the cash cost of running the cloud from 31% to 20% of revenue lifts the base only to $57. Or look at it as a return: at $109.18 my base cash flows yield about 5.5% a year, barely above the 4.94% ten-year Treasury C (calculated).¹⁶

Quality and price are separate. The chip is a real invention and the OpenAI contract is large. The speed advantage and the contract count in its favour; the capital needed, the single customer and the short life of the equipment count against it. So quality is mixed and not yet proven. The price is a different question: it is at the top of my range. A good business bought at the bull case’s value has no cushion if the base is right. The next use of the same dollar is Broadcom, whose page here has a base value of $261 against a price of $326.08, 25% above the base and below the bull case of $431. Cerebras starts further out: 178% above base.

Where this could be in 5 and 10 years

For my base to hold at five years, OpenAI has taken delivery of most of the 750 MW by 2029, core revenue is near $7B, and the company has paid out its excess cash rather than spent it: about $20 a share by 2031. At ten years the base needs the early tranches to be renewed, at lower prices, so revenue holds near $7B and owner cash near $1B a year E (estimate). Both years assume no new equity.

The cases differ on three numbers: how much revenue there is after 2030, the equipment needed per dollar of it, and the cost of running the cloud. The bear loses revenue as the first tranches end and holds little after 2036. The bull has revenue above $14B and equipment at 70 cents per dollar. The discount rate matters less than usual because cash is so large a share of value.

The single most likely way my base is wrong is the renewal price. If OpenAI renews at half the price, the base falls toward the bear. If it renews at the same price and takes more capacity, the base moves toward the bull.

A reader’s guide: these are estimates of value, not forecasts of the share price, and the market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three facts cut in Cerebras’s favour and get less attention than the share price. First, cash is large against the price. It is $8.6B, and after the OpenAI loan and the prepaid services about $7.4B, or $27 a share, which is a quarter of $109.18 C (calculated).³ Second, the signed revenue is 50 times last year’s revenue: $25.4B against $510.0M C (calculated).³² Third, the company’s own core margins improved quickly. Core operating margin was minus 16% in the second quarter, 26 points better than a year ago, and core gross margin was 41%.¹⁰

The bull case

The strongest case in plain words: AI is moving from training to serving, serving is limited by the speed at which each answer arrives, and Cerebras is the only company selling a chip built for that. OpenAI has committed to 750 MW and holds an option on 1.25 GW more, so the first contract may be a fraction of what follows. Cerebras can build the systems at cost, avoids parts in short supply, and has $8.6B of cash to do it. At the bull case’s numbers the business earns about $2.7B of owner cash in 2030 and is worth what the stock costs.

What the company says. In the 12 August 2026 results release, Feldman said core revenue had more than doubled to $210M and the cloud business nearly quadrupled, and that the company was securing more data-center capacity and expanding manufacturing. Komin said the company had made rapid progress toward delivering against the $25.4B of remaining performance obligations and planned to more than triple revenue in 2027.¹⁰ In the June release the company described the OpenAI agreement as worth more than $20 billion.¹⁵ Those are management statements, the best case, and they are all in the filings.

Others I looked for. I searched for statements about Cerebras by Gavin Baker, Brad Gerstner, Chamath Palihapitiya, David Sacks, Bill Gurley, Bill Ackman and Cathie Wood. I found none I could verify against a transcript or filing, so I attribute nothing to them. The 13F of Situational Awareness LP, filed 14 August 2026, shows 10,000 Cerebras shares worth $2.21M at 30 June out of $20.24B reported, or about 0.01%. I draw nothing from that.⁹

Against reality

If OpenAI takes the full 750 MW on time, the second half of 2026 and 2027 revenue lands near the company’s outlook, and renewals hold the price, the business is worth between my base and my bull. The price is possible only at the top of that range. Today the numbers are on the right side for delivery: revenue is $373.5M for the half, core margins are improving and $25.4B is signed. They are far from the owner cash the price needs: the company spent $596M more than it took in from operations in the half C (calculated), and a loss from operations of $492M on a GAAP basis.³

Multiply it out. OpenAI’s 750 MW is delivered on schedule: 80%. The first tranches renew at close to today’s price: 50%. Other customers reach a quarter of revenue by 2028: 50%. The equipment cost per dollar of revenue stays near 80 cents: 60%. All four together are 12%, and the bull case needs most of them and more. The 80, 50, 50 and 60 are my guesses E (estimate). They show why a price that equals the bull case has little room for even one miss.

What a 10x would take. Ten times $109.18 is $1,091.80 a share, or about $300B on 275M shares. The founders’ last performance tranche pays at $250B, so the company has put that outcome in its own pay plan.³ At 20 times owner cash, $300B needs $15B a year. My base turns 18% of revenue into owner cash and my bull 25%, so that is $60B to $83B of revenue C (calculated). The company reported $0.5B in 2025. I cannot name a large company that went from $0.5B of revenue to $60B in under a decade E (estimate). A tenfold move inside five years is not plausible, and inside ten it needs an outcome far past my bull case.

Where I come down

The contract and the cash moved me. Without a firm 750 MW and $8.6B I would not give a bull case. The insider selling, the single-customer risk and the fall from $311 to $109.18 in five months did not move my value, which comes from the filings; they only tell me the market is unsure too. What would raise my values: renewal terms for the first tranches, a second customer with more than $1B of annual revenue, or equipment cost per dollar of revenue below 70 cents. What would lower them: delays in OpenAI deliveries, a new equity raise, or a margin below 38%.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• Customer concentration. OpenAI was 32% of revenue in the second quarter, and 34% and 10% were two other customers; in 2025 MBZUAI and G42 were 86% together. The MRA is the larger share of $25.4B.³²
• Control. Class B shares have 20 votes each and carry about 99.2% of the votes. Class B converts to Class A on a sale or transfer, and six months after the last founder leaves.³
• Lock-up and supply of shares. The lock-up ends at the earlier of the second trading day after the third-quarter results or 180 days after the 13 May prospectus date, around 9 November; up to about 171.1M shares could be released early.² At 237.6M outstanding that is large.
• Material weaknesses in controls. Disclosure controls were not effective at 30 June.³
• Supply. TSMC makes every wafer; most orders are on a purchase-order basis; export licences are needed for some sales.²
• Warrant and loan terms. The OpenAI warrant vests as capacity is delivered. The loan becomes due on a termination that is not OpenAI’s breach.³
• Litigation. A putative copyright class action filed in October 2025 in the Northern District of California; the company cannot estimate a loss.³
• Leases. Operating leases of $690.3M, plus about $2.25B signed and not started.³
• Dilution. 76.0M potential shares were excluded from the loss per share at 30 June, 32% of the 237.6M outstanding.³ New employee grants follow each year.

Geopolitical risk, and what it means for the value

Three places matter. Taiwan. TSMC makes every wafer, there is no long-term capacity commitment, and most orders are by purchase order. The prospectus names a China–Taiwan conflict among the risks.² The UAE. In 2025 UAE customers gave a significant majority of revenue: MBZUAI 62% and G42 24%.² In the second quarter Customer A was 34% of revenue, and it and Customer D are related parties; I read them as MBZUAI and G42, and EMEA was 59% of first-half revenue C (calculated).³ The United States decides who may buy the chips and what imports cost.

G42 and CFIUS. G42 and Cerebras filed with CFIUS in July 2024 for G42 to buy preferred stock. They dropped G42 as a party, CFIUS let the notice be withdrawn on 27 March 2025, and the G42 purchase option ended.² The ownership question is closed. The sales question is not: Cerebras ships to G42 and MBZUAI under BIS licences with security conditions, and the prospectus says a licence can be revoked even after the systems are built.² A BIS rule effective 10 July 2026 moved the UAE to Country Group A:5 and approved G42 and Core42 for licence-free advanced computing, but that approval lapses on 6 April 2027 unless BIS extends it. I did not find MBZUAI in the rule’s table.¹⁷

Tariffs. A January 2026 proclamation puts 25% on a narrow list of advanced chips but not on chips imported for use in US data centers, where Cerebras’s fleet is.¹⁸ A second phase of broader duties is left open. I assume a duty would fall on the imported chip, about 60% of a system’s cost, so 25% of that is roughly 15% of equipment cost. The 60% is my guess E (estimate).

What the numbers say. Net cash of about $27 a share sits under every row. The OpenAI row matters most: it takes the base to $21, and the loan of $918M can fall due if the agreement ends other than for OpenAI’s breach.³ I use the page’s own 50% for renewal at close to today’s price. The other odds are guesses. Licences are revocable and one approval lapses in April 2027, but Washington moved the other way in July. One foundry makes a year’s delay possible, and I know of no year-long gap at TSMC. The tariff text exempts US data centers so far. I give no odds for a war.

None of this is in the price. The $109.18 sits at my bull case, which assumes licences, wafers and renewals all hold. A wafer stop of more than a year would leave $2.9B of leases to pay with little revenue, and these cases cannot value it.

Cerebras and the Culper Research short report

Who and what. Culper Research is a short seller. Its thread opens “We are short” Cerebras and says the shares are headed lower, so it profits if they fall.²¹ I could not open its 28 October 2026 report: the site returns an access error to automated requests, and the Internet Archive holds only error pages.²⁰²⁶ This section rests on Culper’s X thread and two third-party summaries, so it is secondhand, and I could not confirm the title.²¹²²²⁸

What Culper says. Culper says wafers break down after about two years in use; its own heat model gives 1.4 to 2.3.²¹ The summaries headline 1.5 to 2.5 years.²²²⁸ Culper says draft filings through December 2025 gave equipment lives of 1.5 to 3 years, then 3 to 5 in an April 2026 S-1, with no explanation.²¹²² The summaries add that at 2.25 years, first-half gross margin is 24.0%, not 29.9%.²²²⁸ Activ8 puts the OpenAI agreement between break-even and minus $13.8B at 2.5 years, and between minus $6.3B and plus $4.1B over its full term.²² Culper also alleges constant “dummy operations”, quiet swaps of failed cloud systems booked as capex, and a connector material that is a 40-year-old outside product.²¹ Some of its evidence is unnamed customers and former employees, whom I cannot check.²¹

The company’s response. I found no 8-K since 12 August.²⁷ On 5 November the chief executive, Andrew Feldman, replied in a blog post. He says under half of one percent of CS-3 wafers deployed since 2024 are decommissioned, TSMC qualifies the wafer for 10 years at 105°C, and junction temperature never exceeds 85°C. He says Culper’s heat-model inputs are wrong and that scaling a GPU’s economic life is a logical error. On accounting, he says the first S-1’s life reflected lab equipment, 3 to 5 years followed the private cloud build, other cloud providers book 5 to 6 years, and KPMG audits the estimate. He says dummy operations are one tool to smooth power, the connector never shipped, and the margin and OpenAI conclusions rest on the two-year figure. He ends by saying he is long the stock.²³

What is established. The change in life is in the filings: 1.5 to 3 years for computer equipment in the withdrawn 2024 S-1, 3 to 5 years for a wider equipment line in the May 2026 prospectus.¹³² I found no note in the prospectus or the two 10-Qs that explains or sizes the change; the company’s reason, the private cloud, is only in the blog post.²¹¹³²³ Culper’s margin reproduces. Depreciation in cost of revenue was $28.4M in the half. If all of it is on assets booked at a 4-year midpoint, a 2.25-year life adds about $22M and takes 29.9% to 24.0% C (calculated).³ That is arithmetic, not evidence on the wafer. Not established: how long a wafer runs. Culper’s two years is a model, and Feldman’s 0.5% has no fleet age behind it. The fleet is young: equipment at cost rose 45% in the half.³ No court or regulator has ruled. A law firm, Levi & Korsinsky, announced an investigation on 5 November. That is a notice, not a lawsuit, and I found no filed case.²⁵

Where it touches my model. The life the company books changes reported profit, not cash. My value is built on cash, so the booked life alone moves nothing. The margin triggers under “What would change my view” are measured after depreciation, so a longer life makes them easier to hit. The physical life moves value, through replacement. My base uses 4 years, the middle of the range. I reran the model changing the life, with revenue, prices and equipment per revenue dollar held, and replacement from 2028, a year earlier than my base. That earlier start alone takes the base from $39 to about $37. At 3 years the base is about $28. At 2.5 years it is about $18, near the $21 where OpenAI does not renew. At 2.25 years it is about $12, and the $109 bull case about $74 E (estimate). At 1.4 years the base and bear go below zero; I read no number off that. OpenAI’s tranches run three to four years, so a two-year wafer is replaced inside one; whether the price covers that, the filings do not say.³² Swapping failed cloud systems is the same input, replacement spending. Power for dummy operations sits in cloud running costs, which the filings do not split from rent; I cannot size it. The connector claim touches no input.

So what. A buyer needs to know how many years a wafer earns its keep, and nobody outside the company can measure that today. Culper is right if wafers wear out in about two years and the OpenAI price does not cover replacing them; my base is then near $12. The company is right if wafers outlast the 3 to 5 years booked and newer chips retire them first; then my 4 years stands. Culper profits if the price falls. Management and insiders are paid in the stock, and Feldman says he holds it. Nothing I could read gives fleet-age data for either side. Third-quarter results come after the close on 12 November,²⁴ and the 10-Q that follows will show depreciation in cost of revenue and any change in equipment life.

What would change my view. Up: OpenAI exercising part of the 1.25 GW option; a non-OpenAI customer above 25% of revenue; core gross margin above 45% at $5B of annual revenue. Down: a missed tranche; a second equity raise; renewal pricing in an 8-K or 10-Q that is lower than current; the loan being called.