Interview

Cerebras CEO Andrew Feldman: We're 10–20x faster than GPUs and 75% of our 2027 data centers will be outside the US

Nov 17, 2026 with Andrew Feldman

Key Points

  • Cerebras claims its wafer-scale chip runs 10 to 20 times faster than GPUs on inference, positioning speed as the undervalued capability in AI infrastructure.
  • By 2027, approximately 75% of Cerebras' data centers will operate outside the US, reflecting anticipated permitting slowdowns and new economics in Brazil, Europe, and Asia.
  • Cerebras deliberately built its supply chain around 5-nanometer capacity and avoided HBM memory to compete where Nvidia's scale advantage is weakest, not strongest.
Cerebras CEO Andrew Feldman: We're 10–20x faster than GPUs and 75% of our 2027 data centers will be outside the US

Cerebras: Speed as Strategy

Andrew Feldman's pitch for Cerebras is straightforward: fast inference is the most undervalued capability in AI infrastructure, and nobody delivers it faster. The company's wafer-scale engine — the largest chip in computer industry history, manufactured at TSMC on a 5-nanometer node — runs 10 to 20 times faster than GPUs on inference workloads. Feldman builds the full stack around it: the chip, the liquid cooling system, the server, and the API delivered to the customer. Bring power, and Cerebras handles the rest.

Supply chain as competitive moat

Competing with Nvidia means avoiding the fights Nvidia wins. Feldman deliberately built Cerebras' supply chain to minimize overlap: 5-nanometer capacity at TSMC rather than 3-nanometer, and no HBM memory, where Nvidia is one of the largest buyers on earth. The logic is explicit — force Nvidia to compete where Cerebras is fast, not where Nvidia's scale is decisive.

On hardware releases, Cerebras is running roughly annual cycles with performance doubling each generation for the next several years.

“We're not a little bit faster than GPUs. We're 10, 15, 20 times faster. In '27, about 75% of our data centers will be out of The US. The data center industry did ourselves no favors — we could have gone in as a community and been good to local communities. An opportunity was missed across the board.”

Geography: 75% outside the US in 2027

The sharpest disclosure from Feldman is geographic. By 2027, approximately 75% of Cerebras' data centers will be located outside the United States. That number reflects a deliberate bet made ahead of what Feldman sees as a US permitting slowdown following the midterms. Data centers coming online in 2027 were already permitted before the elections — concrete is being poured, bulldozers are moving. The permitting environment today will show up in 2028 builds, not 2027.

Active or planned markets include the US, Europe, Asia, and Brazil, where recent tariff reductions on equipment imports made the economics workable. Canada is also attracting builds, with local politicians explicitly positioning themselves to capture data center investment the US may lose.

Feldman is candid that he'd prefer more US capacity — he wants data centers close to the customers they serve — and says Cerebras engaged directly with local officials in communities near Austin before committing. For inference specifically, 50 to 200 megawatt facilities work well, and geographic distribution actually improves latency. The gigawatt-scale campuses being built for training aren't the right form factor for what Cerebras sells.

The PR failure the industry won't admit

The data center industry, in Feldman's view, badly mishandled community relations during the first build wave. Rebranding facilities as "AI factories" made things worse — nobody wants a factory next door. The opportunity to offer tax revenue, construction jobs, and community investment was largely missed. Data centers have become a physical target for AI anxiety precisely because software is invisible and buildings are not. Feldman argues the industry failed to tell a credible story about AI's benefits — world-class tutors, medical access, educational tools — and created fear instead.

Cybersecurity and financial markets

On cybersecurity, the core argument is that damage happens in seconds, so detection and response need to happen faster than any human workflow allows. Cerebras positions its inference speed as the enabling layer: scrubbing traffic, recognizing attack patterns, and triggering responses without imposing login friction on end users. Speed at 10 to 20 times GPU performance changes what's operationally possible.

In financial markets, Feldman sees a parallel problem — information diffuses instantly but processing lags, producing the kind of headline-driven whipsawing visible in markets today. Fast inference, he argues, can synthesize noisy signals into thoughtful recommendations rather than knee-jerk reactions.

Model architecture: generalist vs. specialist

Feldman is agnostic on whether the AI ecosystem converges to one large model or an ensemble of specialists — and says the question may not resolve cleanly. The generalist-versus-specialist tension plays out in every mature software market, from security vendors to development tools. His read is that both will coexist, with routing layers directing traffic to the right model for each task. Cerebras is positioned to serve either architecture, since the speed advantage is valuable regardless of whether the model is large or small.

Every deal, every interview. 5 minutes.

TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.