Cerebras Revenue Jumps 55% as Demand for Fast AI Inference Grows

Cerebras reported third-quarter revenue growth of 55% year over year, with cloud and services sales more than tripling. Its fourth-quarter revenue outlook was $320 million, which host Ed Ludlow described as well above consensus estimates. Bloomberg Technology’s on-screen chart showed Cerebras shares rising during the morning after the results, though Ludlow noted they had pulled back from the session high.

CEO Andrew Feldman said Cerebras was expanding both hardware sales and cloud capacity to meet demand. Asked whether the stronger outlook reflected improving supply or opportunities the company could not serve in the prior quarter, Feldman pointed to momentum at Cerebras and in the market more broadly. He said demand for inference was growing alongside a better understanding of the value of fast inference, which is the area where Cerebras specializes. Selling more hardware and bringing more cloud capacity online were both part of removing the constraints, he said.

Feldman’s pricing argument is that customers will pay more when speed matters: “We can charge more for fast tokens than you can for slow tokens.” He tied that willingness to pay to the value of getting answers sooner, not to a general case for consuming more tokens.

Ludlow raised the problem of “token maxing,” or companies spending heavily on tokens without clear discipline. Feldman called it wasteful and distinguished indiscriminate use from assigning tokens to difficult problems where AI is useful. In those domains, he said, delivering tokens more quickly and getting answers in less time can be valuable.

Feldman compared enterprise budgeting for AI with the early period of cloud adoption, when companies were uncertain how to plan for a new category of spending. He said CFOs and business leaders are working out how to budget for AI, and argued that spending should be thoughtful and planned rather than “willy-nilly.” The case for fast-inference pricing power, as he presented it, depends on speed helping with particular workloads—not on unlimited token consumption.

Cerebras signed AWS and Oracle as customers and announced Microsoft the day before the interview. A Bloomberg Technology graphic identified all three as Cerebras’ new customers. Feldman said the relationships give Cerebras reach into the enterprise: the companies can help it deliver compute services to customers large and small around the world.

Taken together, the three providers reach “just about every enterprise worldwide,” Feldman said, offering immediate global distribution at a scale Cerebras could not have built by itself. He described that reach as a way to meet customers where they want to use the company’s services, whether in the cloud or on their premises.

Asked whether Jensen Huang’s resignation could create an opening for Cerebras, Feldman first praised Huang’s record. He called Huang’s path—from starting Nvidia to building it into the world’s most valuable company—one of the great entrepreneurial runs in history, and said other entrepreneurs aspired to that kind of achievement. He also said Nvidia had a “deep bench” and would be fine.

Ludlow asked whether the news might bring Cerebras more calls or help it win market share, referring to its relationship with AMD and the possibility of combining their technologies for fast inference. Feldman said the companies’ products had not changed. He maintained that Cerebras products are superior in some segments, including fast inference, but said he expected customers to decide on product merits.

His broader expectation is that the opportunity will grow beyond what any one, two, or three companies can capture. He anticipated good numbers across infrastructure providers and said Cerebras was focused on building what he called the fastest inference solutions and meeting customers in the cloud or on their premises. The potential competitive opening, in his account, mattered less than product performance and a growing market.

Feldman said Cerebras’s modeling had been fairly accurate and that it would be bringing up about 75% of its data-center capacity outside the United States in 2027. He cited large builds underway in Canada, with additional builds in France, Norway, and Finland.

Asked whether policy changes after the midterms could affect data-center expansion and Cerebras’s business, Feldman did not identify a specific policy change or spell out its effects. He said the company’s modeling still appeared accurate and pointed to a trend that, in his view, people working on the ground had seen well before the midterms.