Skip to content

Business

Cerebras signs Microsoft cloud deal, expects revenue in first half of 2027

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. Cerebras expects to share details of its Microsoft cloud agreement in the first half of 2027.
  2. Oracle is buying Cerebras equipment and aims to offer inference on it in summer 2027.
  3. Cerebras reported $210.8 million in third-quarter GAAP revenue, with cloud and other services up 265% year over year.
PracticalDevelopers
A purple silicon wafer beside a folded sheet of paper and a pen
AI-generated conceptual image: A wafer beside paperwork represents an AI chipmaker’s cloud agreement. · AI Ping

Cerebras Systems said on its November 12 earnings call that it signed a cloud deal with Microsoft. CEO Andrew Feldman expects to give more details in the first half of 2027, when he also expects the deal to start generating revenue. Cerebras left the value and capacity of the agreement undisclosed, along with the date when Microsoft customers might be able to use a service that runs on its systems.

Microsoft joins Amazon Web Services and Oracle as the third major cloud provider Cerebras says it has signed up. The arrangements take different forms. Oracle is buying Cerebras equipment and has announced when it aims to offer inference on that hardware. Developers looking for an equivalent option through Microsoft will have to wait for product details.

Cerebras reported $210.8 million in revenue for the quarter ended September 30, up 55% from a year earlier under U.S. accounting standards. Cloud and other services contributed $141.9 million on that basis, an increase of 265%. Services, including inference delivered on infrastructure Cerebras has secured, now account for a bigger share of its sales alongside sales of the machines themselves.

Oracle buys systems. AWS pairs two kinds of chips.

Oracle and Cerebras announced on October 27 that they planned to bring Cerebras inference to Oracle Cloud Infrastructure. On the call, Feldman said Oracle was buying the equipment and would install it in its U.S. data centers. He expects the inference service to launch in summer 2027. Delivering equipment to Oracle and opening the service to OCI customers are separate milestones.

AWS described a different setup in March. Its Trainium chips handle the first stage of inference, processing a request and its supporting material. Cerebras systems take over for the second stage, generating the tokens that form the response. The two stages place different demands on the hardware, particularly when a user submits a long document or asks for a long answer. AWS said it planned to offer the combined service through Amazon Bedrock. Cerebras also says in its quarterly filing that it needs a supply of separate GPUs when it deploys its own systems with that division of labor.

A cable joining a board of small chips to a larger compute module
AI-generated conceptual image: Two connected types of hardware illustrate how input processing and response generation can run on separate chips. · AI Ping

Capacity arrives in stages

Cerebras says it has contracted for 700 megawatts of gross data-center power capacity due by the end of 2027. Gross capacity includes the power needed to operate the facilities themselves. Computing capacity reaches customers as the equipment is installed and brought into service. Cerebras disclosed additional facility contracts in Finland and Norway and gave no allocation by region for Microsoft.

That timing already matters for an existing customer, OpenAI. Under its agreement, an unexcused shortfall that lasts beyond 30 days after a delivery milestone starts accruing credits against future bills, according to the quarterly filing. If the shortfall remains after 90 days, either side can cancel the affected portion. Cerebras estimated that less than 5% of the total capacity committed to OpenAI could become eligible for cancellation in 2026 under that condition. The filing said some delays could extend into 2027, based on the company's data-center deployment forecasts.

A transformer and an empty equipment pad outside an imagined data center
AI-generated conceptual image: A transformer beside an empty equipment pad represents the staged delivery of data-center power. · AI Ping

The company reported $26.1 billion in remaining performance obligations as of September 30. That figure represents contracted revenue still to be recognized as Cerebras delivers goods and services. The company said much of the balance comes from OpenAI. The filing says the total can change as estimates of pass-through costs and other variable amounts change. Agreements reached after September 30 fall outside the reporting date.

Cerebras lifted its 2026 forecast for core revenue to $940 million to $960 million. The company’s core measure adjusts for items such as pass-through data-center revenue and the amortization of warrants it issued to customers. For the final quarter of the year, Cerebras projects core revenue of $310 million to $330 million.

Sources