TL;DR — Key Takeaways
- Cerebras reported record third-quarter revenue of approximately $211 million, up 55%, and CEO Andrew Feldman forecast that revenue would triple in 2027.
- A new Microsoft cloud inference agreement gives Cerebras three major hyperscaler deals in approximately six months, alongside Oracle and AWS.
- Data center shortages have delayed some OpenAI deployments, even as Cerebras expands manufacturing and secures additional compute capacity for 2027.
Cerebras Systems reported record third-quarter revenue Thursday and announced a new cloud agreement with Microsoft. During the company’s earnings call, CEO Andrew Feldman said he expects revenue to triple in 2027.
Cerebras shares climbed nearly 18% in after-hours trading following the announcement, reversing a 3.3% decline earlier in the day. The company reported approximately $211 million in quarterly revenue, up 55% from a year earlier. Revenue from its cloud and related services business nearly quadrupled on the company’s core reporting measure, while overall core revenue exceeded its forecast.
With its customer base expanding, Cerebras faces a significant challenge in supplying enough compute capacity to meet its commitments. The company sells AI accelerator systems directly to customers and provides inference services through its own cloud, two businesses whose growth depends on how quickly it can manufacture hardware and bring additional data centers online. Feldman said the company is adding manufacturing and data center capacity to support its anticipated growth, along with signing agreements with some of the world’s largest cloud providers.
Microsoft Deal Adds to Cerebras’ Hyperscaler Wins
The new Microsoft agreement calls for Cerebras to provide cloud inference services, according to the company’s quarterly SEC filing. Feldman disclosed the agreement during Thursday’s earnings call but gave few details about its scope or financial terms. He said the company will share more information next year.
Microsoft joins Oracle, which announced plans in October to deploy Cerebras hardware in its data centers and offer inference services through Oracle Cloud Infrastructure beginning in 2027. Together with an earlier agreement involving Amazon Web Services, Cerebras has secured three major hyperscaler deals in approximately six months.
Cerebras also reported progress on its next generation of AI hardware. Its fifth-generation Wafer Scale Engine, or WSE-5, returned from fabrication ahead of schedule and is already running AI models in system testing. The chip will power the forthcoming CS-5 accelerator, which Feldman said will more than double the speed of its predecessor when it enters production in early summer 2027.
CFO Bob Komin said Cerebras expects relatively flat cloud revenue in the fourth quarter because additional data center capacity will not become available soon enough to contribute significantly. Feldman also acknowledged that capacity shortages have delayed some initial deployments under a Cerebras agreement with OpenAI.
The company is working on the shortage. Cerebras said it has secured approximately 700 gross megawatts of data center capacity for delivery by the end of 2027, including additional facilities in Europe and North America. The company is also expanding manufacturing with plans to finish 2026 with more than 10 times the production capacity it had at the beginning of the year.
Feldman Defends Cerebras’ Wafer-Scale Technology
Feldman also used Thursday’s call to challenge criticism of Cerebras’ hardware, following an Oct. 28 report from short seller Culper Research that alleged its wafer-scale chips wear out after approximately two years. He separately rejected claims that Cerebras suffers from poor manufacturing yields or cannot support very large AI models.
Feldman explained that Cerebras builds spare processing tiles into its wafers, allowing the hardware to bypass defective regions. The company also adjusts voltage independently across different sections of each wafer to compensate for variations in electrical performance. He attributed lower margins on some hardware shipments to manufacturing expansion and older customer agreements, not poor yields.
In a Nov. 5 rebuttal to the Culper report, Feldman disputed the allegations about wafer lifespan, explaining that Cerebras uses specialized cooling, power management and flexible connections to limit thermal stress. He said fewer than 0.5% of deployed CS-3 wafers had been decommissioned since the systems were introduced in 2024.
The rebuttal also challenged the report’s interpretation of equipment depreciation. Feldman said Cerebras’ accounting estimates of equipment life are not predictions of when its chips will wear out and argued that newer processors offering better performance for the cost are more likely to determine when existing hardware is replaced.
The dispute over Cerebras’ hardware is a reminder that the long-term reliability and economics of emerging AI accelerator architectures are still largely unproven. With major cloud agreements in place and ambitious revenue projections, Cerebras now faces the challenge of scaling production and data center capacity while proving it can compete with the industry’s heavyweights on more than inference speed alone.
