- What happened: Cerebras's GAAP revenue for July–September 2026 rose 55% year over year, with cloud and other services driving the increase.
- Why it matters: If a non-excused capacity shortfall for OpenAI persists 30 days after the deadline, credits are deducted from later months' fees. The term of the affected capacity block also shortens by the number of days delivery was delayed.
- What to watch next: Whether the contracted facilities come online, how much compute capacity is delivered to customers, and the operating profitability once services begin.
U.S. AI chip company Cerebras Systems announced on November 12 (U.S. time) that its revenue for the July–September 2026 quarter was $210.809 million under U.S. generally accepted accounting principles (GAAP), up 55% from a year earlier. Cloud and other services accounted for about two-thirds of revenue, and the company raised its full-year outlook for "core revenue," its own adjusted metric. Meanwhile, compute capacity delivered to OpenAI as of the end of September fell short of the schedule, before adjusting for delays that may be excused under the contract. In the cloud business that underpinned revenue growth, when compute capacity can be delivered also affects the fees Cerebras receives and the period over which it provides the service.
Cloud growth offset the decline in hardware
Revenue from cloud and other services rose 265% year over year, lifting its share of total revenue from about 29% to about 67%. Hardware, which accounted for roughly 70% of revenue a year earlier, fell 29%.
| Revenue (thousands of USD) | Jul–Sep 2025 | Jul–Sep 2026 | Change YoY |
|---|---|---|---|
| Cloud and other services | 38,920 | 141,928 | +103,008 |
| Hardware | 96,794 | 68,881 | −27,913 |
| Total | 135,714 | 210,809 | +75,095 |
The $103.008 million increase in cloud and other services exceeded the $75.095 million increase in total revenue. Subtracting the $27.913 million decline in hardware from that gives the total increase. The year-over-year change was calculated by subtracting the prior-year amount from consolidated GAAP revenue for July–September 2026. Cerebras quarterly report (Form 10-Q, MD&A revenue table)
The growth in cloud and other services was supported by the OpenAI contract, new inference service contracts, and revenue from installing, operating, maintaining and supporting clusters introduced since the year-earlier period. Revenue deductions tied to warrants granted to customers partly offset the services growth. In hardware, similar deductions of about $17.1 million and about $10.8 million fewer equipment deliveries under existing contracts were factors in the decline. Hardware core revenue, with the warrant-related deductions added back, also fell about 11%, from $96.794 million to $85.959 million.
Raised full-year outlook versus the distance to operating profit
The core revenue the company uses for its full-year outlook is GAAP revenue minus data center costs passed through to customers, plus the non-cash amortization of assets related to warrants granted to customers. For July–September, taking GAAP revenue of $210.809 million, excluding $16.36 million in pass-through costs and adding back $36.424 million in amortization deducted from revenue gives $230.873 million. The prior-year quarter had no adjustments for these two items, and core revenue growth was 70%.
The core revenue outlook issued in August for July–September was $214 million to $216 million, and the actual result exceeded the top of that range by $14.873 million. The full-year core revenue outlook was raised from $880 million–$890 million to $940 million–$960 million. Comparing midpoints, that is an upward revision of $65 million, or about 7.3%. For October–December, the company expects $310 million to $330 million; the $320 million midpoint is about 39% above the July–September result.
At the operating level, the loss widened. The GAAP operating loss for July–September grew from $26.469 million a year earlier to $224.149 million. In calculating core operating income, the company excludes the net amount of data center pass-through revenue less related costs, and adds back amortization of warrant-related assets and stock-based compensation of $155.105 million. After these adjustments, the core operating loss also increased, from $15.123 million to $33.108 million. Net income was a GAAP loss of $161.218 million, but a core profit of $29.823 million. Net non-operating income was about $63.9 million, driven mainly by higher interest and dividends. The company expects a core operating margin of negative 1% to positive 1% for October–December, and negative 10% to negative 8% for the full year.
What happens if OpenAI deliveries are delayed
Revenue Cerebras recorded from the OpenAI contract in July–September was $72.4 million after deducting $3.2 million in amortization tied to warrants. That was 34% of total GAAP revenue. Combined with 38% from Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in the United Arab Emirates, the two reach about 72%. A year earlier, MBZUAI accounted for 70% and G42, described in the 10-Q as a related party of MBZUAI, for 14%, and the top two customers combined for about 84%. The structure of relying on a few customers for most revenue continues. If the fees received from OpenAI fall because of capacity-shortfall credits, total revenue would be affected as well.
Under the contract the two companies signed in December 2025, OpenAI committed to purchase 750MW of compute capacity for AI inference. In January 2026, OpenAI indicated a plan to deploy it in stages through 2028. There is also an option to purchase an additional 1.25GW. As a rule, Cerebras must meet minimum capacity at the required number of data centers by specified dates and deliver compute capacity in sequence.
As of September 30, delivered capacity fell short of the contractual schedule, before reflecting delays that may be excused due to specific dependencies and other factors. Based on its current data center deployment forecast, Cerebras expects the delays could continue into parts of 2027. If a non-excused capacity shortfall is not resolved 30 days after the specified deadline, service credits accrue for the capacity block with the shortfall and are deducted from subsequent monthly fees. The company disclosed that, before excused-delay adjustments, credits began accruing on October 1.
If the shortfall in a capacity block is not resolved 90 days after the deadline, either Cerebras or OpenAI can terminate that block. Under the company's current forecast, blocks that become terminable in 2026 amount to less than 5% of the capacity OpenAI committed to purchase. The term of the affected capacity block also shortens by the number of days delivery was delayed. In this 10-Q, Cerebras did not disclose any capacity actually terminated or the final amounts to be deducted from fees.
Cerebras's remaining performance obligations at the end of September were $26.1 billion. This is future revenue allocated to contracted services and other items not yet delivered, and the OpenAI contract accounts for a substantial portion. It also includes estimates for some data center cost pass-throughs, and the company expects to recognize about 33% in the 24 months through the end of September 2028, 44% in the following 24 months, and the remainder after that. If delivery dates change at a customer's request or otherwise, the timing of revenue recognition shifts, and if capacity blocks are terminated under the contract, the amount recognized itself falls. Cerebras also cites the risk that, if it cannot offer terminated blocks to other customers, it may be unable to cover already committed spending with other revenue.
The 700MW facility contracts and revenue terms by customer
The 700MW Cerebras cited in its November earnings announcement is the total power capacity of contracted data center facilities scheduled for delivery by the end of 2027. The 750MW is the inference compute capacity OpenAI committed to purchase, planned to be deployed in stages through 2028. Cerebras also signed new contracts to secure data center capacity in Norway and Finland. In this announcement, the company did not disclose how facility capacity is allocated by customer or how much is already operational.
Spending to add facilities has already swelled. Cash paid for purchases of property and equipment in January–September 2026 was $1,046.059 million, about 3.47 times the $301.139 million of a year earlier. The main driver was systems and related equipment to expand cloud data center capacity. According to the lease note in the 10-Q, non-cancelable leases for additional facilities contracted as of the end of September and scheduled to begin in 2026 or 2027 total about $8.5 billion in undiscounted future minimum payments over multiple years. As the leases had not commenced as of the end of September, they will be recorded on the balance sheet at each lease's commencement.
In its November earnings announcement, Cerebras said it had delivered "disaggregated inference," which splits processing across different types of chips, at production scale. It did not disclose the combination of chips actually used. Earlier, in July, the company and AMD announced a plan to deploy a configuration in which prefill, which reads the input, runs on AMD Helios GPUs, and decode, which generates output token by token, runs on Cerebras's Wafer-Scale Engine (WSE).
In this plan, the WSE, which puts compute circuits and memory on the same wafer-scale chip, handles decode, which requires large data read bandwidth. In a model calculation by the two companies using Kimi 2.6 1T, at comparable responsiveness, tokens generated per second per unit of power reached up to 5 times that of a WSE-only configuration. It is a metric for how much can be processed at the same responsiveness in facilities with limited available power.
What Cerebras provides differs by customer contract. The Oracle contract is a hardware sale, with Cerebras inference on Oracle Cloud Infrastructure planned for 2027. The June collaboration with AWS is aimed at jointly configured deployments in Amazon's data centers, initially leasing Cerebras systems over multiple years. The Microsoft contract is for cloud inference services, with conditions that must be met before it begins.
Under the general accounting policies Cerebras described in the 10-Q, equipment revenue is recognized when the customer obtains control of the equipment. That is usually at shipment or delivery depending on shipping terms, or when contractual acceptance conditions are met. Revenue from dedicated cloud compute services, including hosted inference, is recognized over the service period.
In the next earnings report, when equipment comes online at contracted facilities and how much compute capacity can be delivered to customers will be clues to progress against the revenue outlook. Once cloud delivery begins, whether operating profit emerges after subtracting the costs of operation will show the effect of expanding the business.