The Chip Startup’s Third Large Customer

Andrew Feldman has spent most of the past decade explaining why a dinner-plate-sized slab of silicon is a better way to run artificial intelligence than a rack of ordinary processors. On Nov 12 he was able to stop explaining for an afternoon.

Cerebras Systems, the company he runs, reported third-quarter revenue of $210.8 million, up 55% from a year earlier on a standard accounting basis. On the non-GAAP measure the company favors, core revenue reached $230.9 million, a 70% increase that cleared the $214.7 million analysts had expected. Cloud and services revenue, the line that shows whether customers are renting its machines rather than buying them, rose 265% to $141.9 million on a GAAP basis and 272% on the core measure. Core gross margin was 41%.

The shares rose 18% in after-hours trading. They had spent much of the year trading on the argument that a single-product company selling an unusual architecture to a handful of buyers should not be valued like a supplier to everyone.

The quarter’s real disclosure was buried in the commentary rather than the statements. Mr. Feldman confirmed that Cerebras signed a cloud inference agreement with Microsoft in October, and that the arrangement will begin contributing revenue in the first quarter of 2027. Microsoft is the third large cloud operator to commit to Cerebras in six months, following Oracle and Amazon Web Services.

That sequence is the answer to the question that has followed the company since its listing. The bear case was never that its chips were slow. It was that the market for accelerators had consolidated around one vendor, that the large clouds were building their own silicon to escape it, and that a startup selling an unfamiliar chip with an unfamiliar programming model would win pilots and lose contracts.

Three signed cloud customers in two quarters does not refute that case so much as complicate it. A person familiar with the Microsoft arrangement said the agreement covers inference rather than training, and that the cloud company intends to use Cerebras for workloads where response time matters more than raw throughput. Inference is the cheaper half of the artificial-intelligence business per unit of computing, and the half that grows with usage rather than with model development budgets.

Cerebras also said it has contracted 700 megawatts of data-center capacity. That figure is a commitment as much as an achievement. Megawatts are the binding constraint on every company in this industry, and a company of Cerebras’s size securing that much power implies obligations to landlords, utilities and equipment suppliers that will be paid whether or not the customers arrive.

Mr. Feldman told analysts he expects 2027 revenue to be roughly three times this year’s level. The company raised its full-year core revenue guidance to $940 million to $960 million from a range of $880 million to $890 million, and guided the fourth quarter to $310 million to $330 million. Those numbers imply a business that will roughly double between the last three months of this year and the same period next year, before the Microsoft contract contributes anything.

The margin line is where the skepticism concentrates. Core gross margin of 41% is low for a semiconductor company and low for a cloud service.

Nvidia’s gross margins run in the mid-70s. A person close to Cerebras said the figure reflects the cost of building out owned data-center capacity ahead of demand, and that margins should expand as utilization rises. Analysts who have modeled the business said the same thing in less generous terms: the company is buying growth with capital, and the payoff depends on utilization rates it does not control.

The company’s route to the public markets was itself a study in the politics of computing. Its offering was delayed for more than a year while a national-security review examined the stake held by its largest Middle Eastern backer, a process that forced it to restructure its ownership and disclose customers it would have preferred to keep quiet. It emerged with a valuation that assumed hyperscale adoption and a customer base concentrated enough that any single loss would be material.

What has changed this year is the shape of demand. Training a frontier model is a concentrated, episodic purchase made by a handful of organizations with unlimited budgets. Serving that model to millions of users is a continuous, distributed purchase made by companies that care about cost per query and latency. The second market is larger and less loyal, and it rewards architectures that can hold an entire model in on-chip memory rather than shuttling data to and from external storage.

Whether that advantage survives the next generation of competing chips is unresolved. Nvidia has been improving inference economics with each release, and the cloud companies building their own silicon are aiming at exactly this workload. A person familiar with one of the three cloud contracts said the evaluation came down to a benchmark the customer would not describe publicly, and that the decision was close.

Mr. Feldman has said he expects the company to be profitable on a core basis at some point in 2027.

That forecast now rests on three customers, 700 megawatts and a margin structure that has not yet been tested at scale. The after-hours move suggests investors are willing to wait for the test. The fourth quarter will show whether the waiting is being financed by customers or by the company itself.

  • Related Posts

    • November 14, 2026
    • 51 views
    Anthropic Packages Claude for the Factory Floor

    In a machine shop that supplies aerospace brackets, the quoting process has looked the same for two decades. An engineer opens a spreadsheet, pulls costs for material and machine time,…

    • November 14, 2026
    • 50 views
    Chipmakers and Cloud Providers Form an Alliance Around SGLang

    At the SGLang summit on Nov. 12, the slide that mattered was the one with eight logos on it. LMSYS, the research group behind the Chatbot Arena model leaderboard, used…