Cerebras Fell 16% After Hours. Its Core Revenue Grew 103%.
Wall Street saw a revenue miss. Cerebras saw a beat. Both are true—and the difference reveals its cloud inference business, OpenAI exposure and valuation risk.
The short version: the cloud business is compounding at rare speed, but the current price leans on a fourth-quarter acceleration nobody has seen yet, and customer concentration, warrant dilution and a lock-up release due within three months all raise the cost of any slip.
What Cerebras does and where the money comes from
Cerebras builds wafer-scale AI processors, single chips the size of a dinner plate, and makes money two ways: selling the hardware, and running AI models for customers on its own machines as a cloud service.
Cloud and services brought in $126.0M, 70% of the quarter’s revenue, up 281% from a year ago, at a 20% gross margin (what’s left after the direct cost of delivering the service). Hardware brought in $54.1M, the other 30%, down 23%, at a 2% gross margin. Both are GAAP figures, the standard accounting rules public companies report under.
Three customers supplied 76% of the quarter’s revenue; OpenAI alone was $56.8M, about 32%.
The quarter, in two sets of numbers
The official (GAAP) numbers: revenue $180.1M, up 74%. Gross margin 14.2%. It lost $477M on operations and $451M overall: the company spent far more than it took in.
“Core,” the company’s own adjusted measure: revenue $209.9M, up 103%, gross margin 40.6%, operating loss $33.6M. The gap comes from two edits.
Edit one: customer warrants. A warrant is a right to buy stock; Cerebras granted them to OpenAI, G42, and Amazon’s AWS alongside purchase deals. Accounting treats them like a discount, reducing official revenue as those customers buy: $44.3M this quarter. Core adds it back.
Edit two removes $14.5M of pass-through billings, data center charges Cerebras collects and passes along.
Core also leaves out $377M of stock compensation (employee pay given in shares rather than cash), $274M of it a one-time catch-up triggered by the IPO, plus $22.8M of related payroll tax.
The company’s case is real: the warrant charge involves no cash, and pass-through is mostly not Cerebras technology. The cost of accepting it: those warrants bought revenue with equity, and $1.13B of them still sit on the books, to be charged against future revenue piece by piece through October 2031.
The company’s own cross-check points the other way. Adjusted EBITDA, another common measure of underlying profit that strips out non-cash and one-off items, showed the loss getting bigger: $53.1M this quarter versus $38.3M a year ago.
Why you saw both a big miss and a big beat on the same quarter
The $194M consensus matches, to the dollar, the company’s own core revenue guide from June 23. How any vendor built its estimate is not knowable from outside, so treat the match as an arithmetic coincidence rather than an explanation.
The “missed EPS by $2.80” headline measured nothing. It set the GAAP loss, which carries the stock compensation and payroll tax that core leaves out, against an estimate built on an adjusted basis. Two different yardsticks, printed as though they were one. The roughly 12 cent beat is directionally the more relevant comparison, with two caveats: Cerebras never published a core loss per share, that figure is ours, and the consensus vendor’s own adjustment method is not visible from outside. A reader who saw only the $2.80 number was misinformed either way. Whether either headline moved the stock is not established here.
The forecast is the real news
Q2 beat the company’s own prior guidance on all three measures, and beat consensus on the core measure while missing it on GAAP revenue (above): 40.6% gross margin versus 36 to 38% promised, and operating margin, the business’s profit or loss per dollar of revenue, of minus 16% versus a promised minus 30 to minus 32%. Minus 16% means it spent about $1.16 for every core-revenue dollar.
The full-year forecast rose to $880 to 890M core revenue with better margins; most of the $25M raise was already banked in Q2’s beat, with only about $9M added to the second half.
The shape is the story. Q3 is forecast to grow only about 2.4% compared with the quarter before, with margins going backwards, and management calls Q3 the margin low point (from the call). To hit the full year, Q4 then needs $263 to $275M of core revenue, a jump of 22 to 28% over Q3 (our arithmetic). Against last year, the growth rate roughly halves in the second half: Q2’s headline was +103%, while on the company’s own guide Q3 is about +58% and the Q4 that guide requires is about +57% (our arithmetic). Nothing is shrinking in absolute terms; the comparison is getting harder because 2025 itself accelerated from $99.5M and $103.3M in its first half to $135.7M and $171.4M in its second. Those 2025 quarters appear in the IPO prospectus, not in a quarterly filing.
OpenAI wears three hats
Customer: $56.8M of the quarter’s revenue, plus “a significant amount” of the $25.4B of remaining performance obligations, contracted work not yet delivered, a figure that includes some data center pass-through charges and excludes others. The deal: 750MW of committed computing capacity (AI contracts are sized by the data center power they use), and OpenAI holds an option on another 1.25GW.
Lender: OpenAI lent Cerebras $1.0B. $918M is still outstanding, $736M of it due within a year, and it can be repaid in computing services instead of cash: $86.3M credited that way so far, none in cash. If the deal ends for most reasons, the whole balance can come due at once and OpenAI can take control of the pledged account holding the loan funds.
Shareholder: warrants for up to 33.4M shares; 10.0M already exercised in July 2026.
The bull case, at full strength
The cloud business nearly quadrupled: core cloud revenue $127.7M, up 287%. Demand is so far ahead of supply that Cerebras is renting back systems it had already sold to customers, which cost about 5 points of gross margin (from the call).
$25.4B of remaining performance obligations (with the pass-through caveat above), roughly $5.6B of it expected to turn into revenue within 24 months, though timing may vary. Management expects core revenue to more than triple in 2027 (from the call).
New sales channels are opening: a pairing of Cerebras chips with AMD server racks in its cloud is expected in production in Q4 2026, and availability on Bedrock, Amazon’s AWS service for AI models, in Q1 2027; more than 600MW of data center capacity is live now or under contract for delivery by the end of 2027.
$8.6B of cash and investments in the bank, $685M of it restricted to the OpenAI build-out; management says it covers at least the next twelve months.
The bear case, ranked by what decides the next year
1. The Q4 cliff, and a growth rate already halving. The raised forecast depends on a Q4 acceleration nobody has seen, right after a nearly flat Q3, and even if it lands, growth against last year still falls from +103% to roughly +57%. Second-half operating margin works out to minus 23 to minus 28%, spending roughly $1.23 to $1.28 per revenue dollar, versus minus 9% in the first half (our arithmetic).
2. A wave of stock becomes sellable. The lock-up (the post-IPO rule that keeps insiders and early investors from selling) ends at the earlier of two trading days after the Q3 report or about November 9, 2026, freeing up to roughly 171.1M shares, up to 15M of them held by directors and officers: 72% of all shares outstanding (our arithmetic). Eligible to sell, not sold.
3. Hardware is shrinking: core hardware revenue fell 26% compared with the quarter before, and the CFO had already said declines continue for several quarters.
4. Cash is going out fast. Free cash flow, the cash a business generates after paying for its equipment, was roughly negative $596M in the first half (our arithmetic). Cerebras has also signed about $2.9B of data center lease commitments, about $690M under leases that have commenced and about $2.25B under leases that have not, against that $8.6B in the bank, and the filing says significant additional capital will be required.
5. The company’s own bookkeeping warning. Cerebras identified material weaknesses in its financial reporting controls: IT access controls, separation of duties, and too few qualified accounting staff, affecting revenue recognition, inventory, data center assets and equity administration. Management judged its disclosure controls not effective as of June 30, 2026. The filing also states the deficiencies did not result in a material misstatement.
What settles it
The Q3 report: a Q4 forecast at or above the needed $263 to $275M supports the ramp; below it weakens the 2027 tripling story; a margin low point as promised supports the cost case.
The lock-up window, by about November 9, 2026: heavy insider selling from that director and officer pool weakens the alignment story; restraint supports it; eligibility alone settles nothing.
Dated milestones: the AMD pairing in production in Q4 2026 and Bedrock availability in Q1 2027 support a customer base beyond OpenAI; slippage weakens it.
Bottom line
Both stories are true: a quadrupling cloud business with $25.4B of remaining performance obligations, built partly by handing customers equity that the company’s own measure adds back.
The market moved both ways on August 12, 2026: up 11.6% in the regular session, before results were public, then down 16.6% after hours to $218.53. The first headlines carried the GAAP miss; no cause is established for either move.
The price assumes a lot: $218.53 values the company at 58.7 times this year’s expected core revenue, about $59 paid today for each $1 of revenue expected this year. Subtract the $7.9B of unrestricted cash and investments from market capitalization, then add the $918M loan, and enterprise value is still about 51 times this year’s expected core revenue (our arithmetic). Either way, a price that only works if much goes right.
The next two quarters decide it: the Q4 acceleration is unverified, and up to 72% of the shares become sellable before it is reported.
The OpenAI relationship is the structural risk: a third of revenue, a significant amount of those obligations, a loan repayable in computing services, and a growing shareholder. One wobble hits every channel.
At this price it is a prove-it stock: the growth is real, but the next leg of the cloud ramp is not yet proven, and a buyer today is already paying for a fourth quarter that has not happened.
Primary sources: Cerebras Systems Inc. Form 10-Q for the quarter ended June 30, 2026 and the earnings press release (exhibit 99.1 to the Form 8-K), both filed with the SEC on August 12, 2026; the IPO prospectus filed May 14, 2026 for the 2025 quarterly comparison; the Q2 2026 earnings call held August 12, 2026; and the Q1 2026 earnings call held June 23, 2026. Prices from Interactive Brokers, retrieved August 12, 2026. Chart data emitted by scripts/cbrs-q2-2026-charts.mjs. Not investment advice.
See More
To understand the infrastructure bottlenecks Cerebras is designed to address:
Memory Chips: An Investor Primer
How HBM and DRAM feed AI accelerators—and why memory bandwidth matters.
Wiring the AI Factory – Networking Primer
Why connecting thousands of accelerators is difficult, and how networking becomes the bottleneck.
Questions, disagreements or feedback? I read every message.
Disclaimer:
This article is for informational and educational purposes only and does not constitute financial, investment, tax, or legal advice. Any opinions, scenarios, price targets, or market observations reflect my personal views and may change without notice. Investing and trading involve substantial risk, including the possible loss of principal. You are solely responsible for your own investment decisions, position sizing, risk management, and trades. Conduct your own research and consult a qualified professional where appropriate.