Cerebras adds Microsoft as cloud buyer while capacity delays test expansion

The new deal follows an existing distribution route: the filing already lists Microsoft Marketplace among the places customers can access Cerebras’ inference services. Microsoft is also a competitor to Cerebras Cloud and has its own inference chip. In a January 26 announcement, Microsoft introduced Maia 200 and said it was deployed in its US Central data-center region. The agreement therefore establishes another supply relationship, without establishing that Cerebras will replace Microsoft’s other hardware.

Cerebras’ filing identifies Nvidia as the dominant hardware competitor and also names AMD, Intel, cloud operators and specialized AI clouds. It describes Cerebras’ offering as a premium solution whose competitors may cost less by some measures. Neither the Microsoft contract’s pricing nor a comparable cost-per-token measure is disclosed in the announcements; the deal alone cannot establish an economic advantage.

The filing distinguishes the recent agreements: Oracle is buying hardware, while Microsoft is buying cloud-inference services. Both receive preferential commercial terms, and Microsoft’s service depends on milestones that the filing does not specify. The earnings release says Cerebras will become available through Oracle Cloud Infrastructure in 2027.

Amazon Web Services signed a separate agreement in June for a joint computing solution in Amazon data centers. That arrangement includes an initial multiyear systems lease, pricing commitments, exclusivity and minimum manufacturing-capacity guarantees. Cerebras warns that these provisions could constrain its ability to meet other customers’ demand—or leave excess capacity if Amazon does not make additional purchases. More routes to market do not necessarily mean less dependence on powerful counterparties.

Microsoft’s Maia 200 inference accelerator, introduced in January 2026. Image: Microsoft. Source: Microsoft.

Services drive growth; accounting changes the picture

Cerebras’ earnings release reported the following results for the quarter ended September 30:

The core revenue calculation adds back $36.4 million of amortization associated with customer warrants—rights to acquire shares—and removes $16.4 million of data-center pass-through revenue. Core profit measures also exclude stock-based compensation and other specified items. These adjustments help explain why the two presentations diverge; they do not eliminate the commercial significance of customer equity incentives or infrastructure costs.

Adjusted earnings of 12 cents a share beat an expected 14-cent loss, and core revenue exceeded the $214.7 million LSEG consensus cited in the earnings report. Core revenue growth nevertheless slowed from the second quarter’s 103% year-over-year rate to 70%. Under GAAP, net loss widened to $161.2 million from $22.2 million a year earlier, and gross profit fell 24% to $42.1 million. Stock-based compensation was $155.1 million, versus $11.3 million a year earlier.

Reported hardware revenue fell 29% to $68.9 million from $96.8 million a year earlier. The filing attributes $17.1 million of the decline to warrant amortization and $10.8 million to lower hardware deliveries. On the core basis, hardware revenue was $86.0 million, down about 11%, calculated against the same prior-year revenue. Preferential pricing for a large strategic customer, higher materials costs and inventory charges also weighed on core hardware margins.

Cloud and other services supplied 67% of reported quarterly revenue, up from 29% a year earlier. That category includes support, management, AI modeling and pass-through costs as well as cloud computing. The filing says dedicated-capacity contracts generally require customers to pay regardless of utilization; on-demand inference is sold through token consumption or prepaid bundles. Those are different revenue mechanisms, and the disclosed category does not isolate Microsoft’s future contribution.

For the fourth quarter, Cerebras forecasts core revenue of $310 million to $330 million, a core gross margin of 47% to 49% and a core operating margin between negative 1% and positive 1%. The revenue range exceeds the $271 million analyst consensus cited in the interview report. Its full-year core revenue outlook is $940 million to $960 million, and management expects revenue to triple in 2027. The interview report put the implied 2027 revenue at about $2.85 billion, below analysts’ $2.92 billion expectation.

Cloud and other services became the larger share of GAAP quarterly revenue. The category also includes support, management, AI modeling and pass-through costs. Chart: Magica; data: Cerebras. Source: Cerebras Form 10-Q.

Capacity commitments outlast the announcement

Customer concentration has shifted rather than disappeared. The filing says Mohamed bin Zayed University of Artificial Intelligence, a UAE university focused on AI, accounted for 38% of third-quarter revenue, OpenAI for 34%, and UAE technology group Group 42 Holding Ltd and its affiliates, known as G42, for 7%. Together, those disclosed shares amount to 79%. The university and G42 supplied 84% a year earlier, adding their reported 70% and 14% shares, and are related parties with respect to each other under the accounting rules cited in the filing.

These UAE ties predate the recent cloud deals. Cerebras and G42 announced their Condor Galaxy AI-training supercomputer partnership in July 2023, offering its first system as a cloud service. That announcement identified the university as G42’s research partner. The newer OpenAI relationship changes the customer mix while retaining dependence on a few large accounts.

OpenAI is also financing part of the expansion. A December 2025 agreement commits it to purchase 750 megawatts of inference computing capacity over an initial three- or four-year service term, extendable to five years. OpenAI funded a roughly $1 billion working-capital loan in January 2026 for the required infrastructure. The filing says $175.5 million of billings had been applied against the loan by September 30, leaving $829.1 million outstanding; such noncash repayment reduces amounts Cerebras would otherwise collect in cash. OpenAI also received a warrant with vesting tied to funding, capacity and other milestones.

A significant portion of Cerebras’ $26.1 billion in remaining performance obligations came from the OpenAI agreement. The total includes estimated pass-through costs and is subject to revision. Cerebras expects to recognize about 33% over the 24 months ending September 30, 2028, with most of the rest later. It is neither current revenue nor solely a measure of sales of Cerebras technology.

Delivery difficulties already have contractual consequences. Cerebras says it delivered less capacity to OpenAI than scheduled at September 30, before adjustments for excused delays, and expects delays could continue into part of 2027. Delayed delivery shortens the affected tranche’s contract term day for day. Unexcused shortfalls left uncured for 30 days after specified milestones generate service credits; the filing says credits began accruing October 1 before adjustments for excused delays.

Shortfalls left uncured for 90 days after the relevant milestone let either party terminate the affected tranche. Cerebras forecasts less than 5% of committed capacity will be terminable in 2026. Termination would reduce contract value and remaining performance obligations, while some infrastructure costs could remain. Certain termination or trigger events can also accelerate repayment of the OpenAI loan.

The company says it has secured 700 gross megawatts of data-center capacity for delivery by the end of 2027 and expects to finish 2026 with more than ten times its manufacturing capacity at the start of the year. Those future capacity figures do not measure computing already serving customers. At September 30, noncancelable data-center leases that had not begun carried approximately $8.5 billion in undiscounted future minimum payments over their lease terms. Further leases signed in the fourth quarter added $217.8 million in such payments. The filing warns that data-center commitments generally outlast customer contracts, leaving costs if customers reduce or terminate purchases.

Cerebras reported $7.8 billion in cash, cash equivalents, restricted cash and short-term investments, following an IPO that raised about $6.2 billion net. It also had an undrawn $850 million revolving facility. Management said available resources should cover anticipated needs for at least 12 months from the filing, while Feldman said in the interview that Cerebras was evaluating financing options. In the first nine months of 2026, the company used $388.9 million in operating cash and paid about $1.05 billion for property and equipment, versus $134.2 million and $301.1 million, respectively, a year earlier.

The filing also says financial-reporting control weaknesses remained unresolved, including in revenue recognition, inventory and data-center accounting. Cerebras says the deficiencies had not caused a material misstatement and describes remediation work underway.

Selected customers’ shares of revenue for the quarter ended September 30, 2026; this is not a complete revenue breakdown. Chart: Magica; data: Cerebras. Source: Cerebras Form 10-Q.

What must happen before Microsoft revenue arrives

For Microsoft, the next evidence is the service’s commencement, the milestones required to reach it and the agreement’s financial scale. Cerebras’ earnings release promises more information next year; the available disclosures do not provide a contract value or identify which models and workloads Microsoft will run. The filing establishes conditions, but does not disclose a Microsoft delay.

GPU availability also remains relevant even for this Nvidia challenger. Cerebras’ filing describes a disaggregated inference design that uses GPUs for prefill—processing the input—and Cerebras systems for decode, the generation of output tokens. It warns that acquiring those GPUs and financing them could constrain deployment. No disclosed detail establishes that Microsoft will use that particular design.

Hardware deployment is another test. Cerebras said its CS-4 systems would soon enter full production and targeted early summer 2027 for CS-5, its next generation. Those schedules, available data-center power and the GPU supply needed for disaggregated inference constrain delivery. The next results will test the forecast of roughly break-even fourth-quarter core operating margin. Early 2027 will then test whether Cerebras can start Microsoft service while reducing the OpenAI shortfalls it expects may persist into that year.