The race to build the fastest AI infrastructure is reshaping the semiconductor industry, with inference speed emerging as the defining competitive dimension of the AI era.

As AI model wars intensify across OpenAI, Anthropic and Google, the underlying compute layer is under pressure to keep pace — and the companies that moved early on inference performance are now reaping the rewards. Cerebras Systems Inc. recognized that opportunity years before the market validated it, according to Andrew Feldman (pictured), founder and chief executive officer of Cerebras Systems.

“The number of people who said you can’t build a chip this big, it’ll never work, here are the 20 reasons why it’s going to fail,” Feldman said. “They told us that for years, and now it works. We’re the fastest inference in the industry, not by a little bit, but 10, 15, 20, 30 times faster than GPUs. And everybody wants fast results.”

Feldman spoke with theCUBE’s John Furrier and Dave Vellante at the RAISE Summit, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed Cerebras’s IPO milestone, the importance of inference speed, European data center buildout and the company’s plans to scale manufacturing by 8x to 10x. (* Disclosure below.)

AI inference speed and the agentic demand curve

The case for fast inference goes well beyond user experience. As agentic AI workflows proliferate, models are making many sequential calls, reasoning over longer context windows and driving dramatically higher compute consumption — dynamics that create lasting demand advantages for Cerebras. According to Feldman, this is fundamentally changing who is buying inference and why.

“Speed can become better answers,” he said. “The longer your reasoning, the more iterations, the better answer you get. And so you can convert speed into better answers. We have a customer, AlphaSense, and they use our speed to search over more documents and that gives their customers better answers.”

Cerebras’s wafer-scale architecture sidesteps memory constraints that slow conventional GPU systems by keeping model weights in on-chip SRAM, giving it a structural advantage as agent workloads balloon memory demand. Coding agents have been the initial beachhead, with customers including Cognition AI Inc. and OpenAI Group PBC using Cerebras in OpenAI’s coding flows, Feldman noted. The company’s customer base now spans financial agents through Block Inc., enterprise deployments at GlaxoSmithKline PLC and European high-performance computing centers. To support that global expansion, Cerebras plans to scale its manufacturing capacity by eight to 10 times this year.

“We’re traveling the world watching data centers be built, standing up our data centers,” Feldman said. “Our hardware engineers are designing next-generation chips and systems. I think it’s going to be a great several years for us.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the RAISE Summit:

(* Disclosure: TheCUBE is a paid media partner for the RAISE Summit event. Neither Solidigm, the headline sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE