One headline number is a future obligation; the other is revenue already earned
On November 12, 2026, Cerebras Systems reported a record $210.8 million of GAAP revenue for the quarter ended September 30, up 55% from a year earlier. Its same-day SEC Form 10-Q disclosed $26.1 billion in remaining performance obligations, or RPO, as of that quarter's end. The contrast is dramatic, but it does not mean Cerebras has already made $26.1 billion in sales.
RPO represents transaction value allocated to contractual promises still unsatisfied or partially unsatisfied. It includes revenue expected to be invoiced and recognized in future periods. A significant share relates to a multi-year OpenAI relationship. The $210.8 million figure measures one quarter of recognized revenue; RPO extends across years and includes estimates that can change. Dividing one by the other would not produce a meaningful growth rate or return.
The first operating question, then, is not how many times larger RPO is than a single quarter. It is how contracted demand becomes delivered AI compute, how much of that delivery is recognized as revenue, and what costs and contractual adjustments arise before Cerebras can keep the economic benefit.
OpenAI committed to 750 MW. Another 1.25 GW is optional—not purchased capacity
Cerebras signed a Master Relationship Agreement with OpenAI in December 2025. Under its November 2026 10-Q, OpenAI is contractually committed to purchase 750 megawatts of AI inference compute capacity and related services, delivered in tranches. Individual services run for three or four years, with extensions available up to five years in total.
The agreement also gives OpenAI an option to purchase another 1.25 gigawatts, for a possible combined 2.0 GW with deployments in tranches by the end of 2030. An option is not the same as a binding purchase of that additional capacity. Using the 2 GW ceiling as an already-sold amount would overstate what the contract establishes.
OpenAI funded an approximately $1 billion secured working-capital loan in January 2026 to support build-out, and the arrangement includes Cerebras equity warrants linked to specified milestones. Financing and equity incentives strengthen the relationship but cannot be relabeled as earned inference revenue. Cerebras recognized $72.4 million of revenue under the OpenAI arrangement during Q3, net of $3.2 million in customer-warrant amortization.
The SEC filing reveals delayed delivery—and credits accruing on a conditional basis
The most significant qualification is not in the headline earnings numbers. Cerebras states that as of September 30, 2026, it had delivered less capacity to OpenAI than contemplated by the agreement, before considering delays that could qualify for contractual excuses. Management expects some delivery delays may continue during part of 2027.
Under the agreement, a capacity tranche that remains below its milestone for 30 days without an excused or cured delay causes service credits to accrue, offsetting later monthly fees payable by OpenAI. Cerebras says that, without adjustment for excused delays, such credits began accruing on October 1, 2026. This is not proof that OpenAI has collected a final cash penalty, or that every shortfall is unexcused. The exact credit amount and any permitted exemptions matter.
If a particular shortfall is not cured for 90 days after its milestone, either party may elect to terminate that tranche. Cerebras currently expects capacity potentially terminable during 2026 to be less than 5% of total committed capacity, not the entire 750 MW. A right to terminate is not an actual termination. Nevertheless, termination could reduce contractual value and RPO, and service credits may lower cash receipts or fees on future deliveries.
700 gross MW secured is not a like-for-like subtraction from 750 MW committed
Cerebras announced that it had secured 700 gross megawatts of data-center capacity under contract for delivery by the end of 2027, including additional agreements in Norway and Finland. It is tempting to subtract this figure from OpenAI's 750 MW commitment and declare a fixed 50 MW deficit. That calculation is not supported by the disclosures.
The 700 MW figure is gross data-center capacity contracted for future delivery across Cerebras's needs and locations, not a disclosure of the exact net installed inference capacity handed to OpenAI at a single date. The 750 MW figure measures OpenAI's contractual inference capacity across several scheduled tranches. Timing, power overhead, deployment status and allocation differ; neither disclosure provides the reconciliation needed for a precise shortfall.
Instead, separate at least four milestones: space and utility capacity under contract, completed and commissioned infrastructure, working AI systems available for the specified workloads, and capacity accepted and billed under each customer's contract. A facility lease can be signed long before a customer receives production-ready inference capacity.
The second side of a giant contract: $8.5 billion of uncommenced leases
The September 30 10-Q discloses non-cancelable agreements for additional data-center leases scheduled to commence in 2026 and 2027. Their aggregate undiscounted future minimum payments are approximately $8.5 billion. They had not commenced at that date and therefore were not yet recognized as corresponding right-of-use assets and lease liabilities on the September balance sheet.
The $8.5 billion is not an immediate cash bill, a discounted present-value liability, or an extra $8.5 billion of earned revenue. It is a multi-year minimum payment schedule under committed leases that has to be considered alongside the RPO. When the leases commence, recognized accounting liabilities and assets will change. Cash requirements and timing still depend on individual agreements.
Cerebras also reported $1.0 billion of cash purchases of property and equipment during the first nine months of 2026, versus $301.1 million in the same 2025 period. This illustrates that turning AI demand into deliverable service requires material infrastructure spending. If customer deliveries slip while facilities and equipment commitments continue, the economics can change even if future revenue obligations remain large.
RPO is not all high-margin inference income
The $26.1 billion RPO includes estimated variable consideration and pass-through costs. Under the OpenAI agreement, pass-through items may include data-center rental, leasehold improvements, security, power and utilities tied to committed-capacity delivery. Some have been included once the underlying right-of-use and lease liabilities were recognized; other uncertain pass-through charges tied to remaining capacity were excluded until their amounts can be determined.
Where pass-through costs are recorded as revenue and offsetting costs, reported revenue can increase without producing the same proportional improvement in gross profit. Cerebras expects approximately 33% of the current RPO to be recognized in the first 24 months through September 30, 2028, another 44% in months 25–48, and the remainder later, subject to customer timing and other adjustments. Those percentages are a recognition forecast, not cash already received.
Investors and business customers should therefore distinguish contract consideration, gross pass-through accounting, ongoing cost of running cloud sites, delivery credit offsets and the revenue that remains after the contract's actual operating economics. No single backlog number establishes the final margin.
GAAP and Core results show why the accounting basis matters
In Q3 2026 Cerebras reported $210.8 million in GAAP revenue, a 20% GAAP gross margin and a negative 106% GAAP operating margin. Its separately defined non-GAAP Core revenue was $230.9 million, with a 41% Core gross margin and a negative 14% Core operating margin. Both sets are company-reported, but they are not interchangeable.
Core metrics exclude specified items including non-cash amortization of customer warrants, stock-based compensation, data-center pass-through revenues and costs, and certain other adjustments. Removing those items produces a useful alternate operating view, but it does not mean Cerebras has achieved GAAP operating profitability, or that the cost of future data-center leases disappears.
The contract structure has another timing consequence: the OpenAI customer warrant is amortized as a reduction to GAAP revenue as related services are recognized. This complicates comparisons between gross contract demand, management's Core outlook and the actual GAAP revenue seen each quarter. A large RPO does not eliminate the need to examine the quality and timing of recognized revenue.
Manufacturing capacity and routes to customers are expanding—but those are separate milestones
Cerebras says it is scaling manufacturing with Flex and Sanmina and expects capacity at the end of 2026 to exceed ten times the level at the beginning of the year. In July it described an expansion at Flex's Milpitas facilities intended to support about seven times CS-3 system production capacity during 2026, including new assembly, integration, testing and skilled manufacturing work. The 7x CS-3 expectation and companywide 10x forecast refer to different scope and baselines.
Commercial routes are also widening. On October 27 Cerebras and Oracle announced that Cerebras inference would become available on Oracle Cloud Infrastructure during 2027. The November earnings release added a deal with a major hyperscaler and wider neocloud distribution. These announcements point toward a broader demand base, but announced routes and future deployment dates are not the same as recognized Q3 revenue.
This matters for concentration risk. A significant portion of current RPO is tied to OpenAI, while factory capacity, cloud-host access and customer distribution are being expanded in parallel. The ability to produce more machines does not itself guarantee that the required data-center power, installations, customer acceptance or economic returns arrive on the same schedule.
The decision test: measure delivered capacity, not just contracted demand
An AI infrastructure operator should track the sequence behind each major contract: binding demand, factory and supply-chain readiness, secured site and power, energized and tested capacity, customer-delivered capacity, recognized revenue, and cash margin after pass-through costs and service credits. These measures answer different questions and should not be combined into one headline growth number.
For Cerebras, the public milestones are unusually clear. OpenAI has 750 MW committed, the additional 1.25 GW is optional, 700 gross MW has been secured for delivery by the end of 2027, and the SEC filing identifies delayed tranche delivery and conditional credits from October 1. The outstanding $8.5 billion of uncommenced leases shows another obligation attached to that delivery path. The precise net customer capacity delivered and the final credit effect are not disclosed in the same simple terms.
The broader shift in AI infrastructure is from securing demand to industrializing fulfillment. A binding contract can be a major commercial achievement and still leave difficult obligations in financing, power, data-center commissioning, manufacturing, uptime and revenue recognition. Cerebras's next evidence should be evaluated against those milestones rather than the size of RPO alone.
BANSEOG VIEW
BANSEOG VIEW | An AI contract is a pipeline of operational obligations, not a revenue screenshot
$26.1 billion in RPO indicates substantial future contracted work; $210.8 million is Q3 recognized GAAP revenue. They differ in time span, accounting treatment and operating obligations, not merely company growth.
The 750 MW OpenAI commitment is binding, while 1.25 GW of additional capacity remains optional. The 700 gross MW contracted for future data-center delivery is a different measure. September's delivery shortfall and October's conditionally accruing service credits show why scheduled tranches matter more than comparing MW headlines.
Evaluate the conversion sequence: contracted demand → commissioned infrastructure → actual customer delivery → revenue after credits and pass-through accounting → durable cash returns. The $8.5 billion of not-yet-commenced lease obligations is part of the execution equation, not an offset to ignore.
SOURCES
Primary sources and references
- SEC EDGAR — Cerebras Form 10-Q, Quarter Ended Sep 30 2026 (filed Nov 12)
Primary authority for 750MW committed and 1.25GW option, OpenAI revenue/loan/warrants, RPO composition/recognition, delivery shortfall, credit accrual/termination rights, and $8.5B uncommenced leases.
- Cerebras Investor Relations — Third Quarter 2026 Results (Nov 12)
Company source for GAAP/Core revenue and margins, 700 gross MW contracted for end-2027 delivery, manufacturing scale-up, and named channel/customer announcements.
- Cerebras — OpenAI High-Speed Inference Partnership (Jan 14, 2026)
Original public partnership announcement; phased 750MW deployment beginning 2026, distinct from Nov 10-Q's as-reported delivery status.
- Cerebras and Flex — U.S. Manufacturing Expansion (July 9, 2026)
First-party planned roughly 7x production expansion in CS-3 systems through 2026, new lines/testing/manufacturing roles.
- Cerebras and Oracle — OCI Inference Partnership (Oct 27, 2026)
First-party announcement of future Oracle Cloud Infrastructure inference availability in 2027; not Q3 recognized revenue.
- Cerebras Investor Relations — Q2 2026 Results (Aug 12)
Earlier official comparison of 600MW secured capacity and prior manufacturing-scale guidance; scope and timing checked.
- Reuters — Cerebras Q3 Outlook and Customer Deals (Nov 12, 2026)
Independent earnings coverage of scale-up and risk context; not used in place of the SEC source.
Reviewed Nov 16, 2026 KST. SEC 10-Q filed Nov 12 for quarter ended Sept 30: OpenAI MRA executed Dec 2025, 750MW committed, additional 1.25GW option by 2030, $1B secured working-capital loan funded Jan 2026, milestone-linked customer warrants, Q3 $72.4M revenue under OpenAI deal net of $3.2M warrant amortization; RPO $26.1B includes estimates of pass-through variable consideration, expects recognition about 33% within initial 24 months through Sep 30 2028, 44% in months 25–48 and rest later. As of Sep 30 delivered less than contracted milestone (before possible excused delays); SEC states service credits commenced accruing Oct 1 without adjustment for excused delays. No claim made that credits were finally paid or tranche terminated; rights after 30/90 days conditional; estimated potentially terminable in 2026 under 5% of 750MW. Future uncommenced leases ~$8.5B undiscounted as of Sep 30, not yet recognized on that balance sheet. Nov 12 Q3 IR: GAAP $210.8M, Core $230.9M, 700 gross MW capacity secured under contract for delivery by end 2027, manufacturing growth targets. July Flex/October Oracle official releases corroborate manufacturing and route to market. Gross MW vs committed inference capacity are different populations and dates; no inferred 50MW shortfall or final effective margins. Current sales, RPO, financing and lease obligations must not be conflated.