Cerebras CEO building wafer-scale AI processor chips
Andrew Feldman
Profile
Andrew Feldman is the co-founder and CEO of Cerebras Systems, the company betting that the way to beat Nvidia is not to build a better GPU — but to stop building GPUs altogether. Cerebras makes the Wafer-Scale Engine: a single chip carved from an entire silicon wafer, roughly the size of a dinner plate. The current WSE-3, launched in March 2024, packs 4 trillion transistors and 900,000 cores. It is the largest chip ever built, and it is not even close.
Feldman is a serial hardware entrepreneur, not an AI researcher. Before Cerebras he co-founded SeaMicro, a low-power server startup AMD bought in 2012 for $334 million — where he then ran the server business as a corporate VP before walking out the door to do it again. He founded Cerebras in 2015 with four colleagues (Gary Lauterbach, Michael James, Sean Lie, Jean-Philippe Fricker), and they spent four years in stealth before shipping the WSE-1 in 2019. Stanford undergrad, Stanford MBA. He is unusually direct in interviews, which is good for listeners and occasionally bad for his PR team.
The bet is simple: memory bandwidth, not raw flops, is what bottlenecks modern LLMs. A wafer-scale chip keeps weights on-die with ~21 PB/s of memory bandwidth — roughly 7,000× an H100’s off-chip HBM. In August 2024 Cerebras launched an inference service claiming 10–20× the tokens-per-second of H100 clusters on popular open models. Whether the bet pays off commercially is the next chapter. In January 2026 they signed a $10B+ contract with OpenAI to deliver 750 MW of compute through 2028, and they filed for a Nasdaq IPO at a $22–25B valuation. That is the moment — for a decade Cerebras was “interesting hardware”; now it’s an actual alternative to buying GPUs from Jensen Huang.
For developers learning AI, Feldman matters because he is running the most serious non-GPU architecture bet in the industry. Jonathan Ross at Groq is doing something similar on a different axis (custom LPUs for inference); Lisa Su at AMD is the GPU challenger. Cerebras is the architectural challenger. If you care about where inference infrastructure goes after the current HBM-constrained equilibrium breaks, Feldman is the person to follow.
Videos
Controversies
Cerebras’ 2024 IPO was postponed by a CFIUS review of the investment by UAE-based G42 — then both Cerebras’ largest customer and a strategic investor. Critics in Washington worried G42 could become a conduit for advanced US AI hardware reaching China. Clearance arrived in March 2025; the refiled S-1 no longer lists G42 as an investor, though G42 remains a major customer. Feldman has defended the relationship publicly as fully compliant with US export controls.
YouTube
Spotify Podcasts
