Cerebras shares jump 17% on upbeat annual forecast, easing AI jitters

Nov 12 (Reuters) - Cerebras Systems raised its 2026 revenue forecast and signed Microsoft as a customer, easing investor concerns about AI spending and sending the chip and cloud company's shares more than 17% higher in extended trading on Thursday.

The forecast signaled enduring demand for AI computing despite concerns driven by US voter opposition to data centers and uncertainty from Jensen Huang's exit from Nvidia.

The stock had fallen 47.2% from its May IPO price as investors grew wary of the massive infrastructure spending required to support AI growth.

"As we look to the months ahead, we see two certainties: rapid change and compute shortages," Cerebras CEO Andrew Feldman said on a call with analysts. He said Cerebras' job is to make sure it grows no matter how AI develops, whether customers want hardware or cloud capacity.

Cerebras is racing to support a deal to provide computing power for OpenAI, which underpins Cerebras' remaining performance obligations of $26.1 billion as of September 30.

Feldman also said the company has signed a cloud deal with Microsoft, with details to be shared next year.

Cerebras now expects 2026 adjusted revenue of $940 million to $960 million, up from its previous expectations of $880 million to $890 million.

Its adjusted cloud and services revenue, which reflects the OpenAI deal, more than tripled to $144.9 million for the third quarter.

The chip firm posted an adjusted profit of $29.82 million in the third quarter, compared with an adjusted loss of $10.86 million a year earlier. Analysts, on average, had expected an adjusted loss of $38.4 million, according to data compiled by LSEG.

Third-quarter adjusted sales rose 70% to $230.9 million, beating analysts' average estimate of $214.70 million.

In July, Cerebras announced a partnership with AMD to combine AMD's Helios systems with its wafer-scale engine for inference. Feldman said its cloud now uses AMD graphics processors for prompt processing.

Cerebras' dinner-plate-sized WSE packs trillions of transistors on a single semiconductor and keeps memory on-chip, which it says is more efficient than linking many smaller GPUs that rely on costly high-bandwidth memory.

Last month, Oracle expanded its cloud AI chip deals to add offerings from Cerebras. Feldman said the hardware will run in Oracle data centers, with the inference service expected to launch in summer 2027.

He also said OpenAI's business with Cerebras grew from the second quarter, but that Cerebras lacked the capacity to support a broad rollout of OpenAI's Ultrafast, a faster service tier the AI firm previewed in August for its GPT-5.6 Sol model.

He said data center availability remains a constraint across the industry and that US policy changes could create greater challenges for development.

(Reporting by Anhata Rooprai in Bengaluru; Editing by Sahal Muhammed)