Cerebras emerged from its IPO at a $60 billion valuation—roughly three times what private markets had anticipated for a company that spent a decade as the semiconductor industry's most visible contrarian bet. Shares closed at $280 on Friday, and the chipmaker had just signed a $10-20 billion deal with OpenAI, the largest AI infrastructure commitment since the compute gold rush began.
The question now is whether this validates CEO Andrew Feldman's long-running thesis that the future of AI computing is custom silicon—or simply reflects a market desperate for alternatives to NVIDIA's dominance.
Cerebras built its entire company around the Wafer-Scale Engine (WSE), a single massive chip that abandons GPU conventions entirely. Where NVIDIA strings together thousands of smaller dies, Cerebras stacked 850,000 AI cores, 1.5 terabytes of on-chip SRAM, and 20 petabytes per second of memory bandwidth onto one wafer-scale package. The architecture eliminates the communication bottlenecks that cripple GPU clusters during both training and inference.
The numbers behind the OpenAI deal reveal why: the WSE-3 can train and serve trillion-parameter models while keeping all weights on a single chip. CFO Bob Komin pushed back explicitly on the "small models only" narrative, telling CNBC that Cerebras currently serves OpenAI's internal models 5.4 and 5.5, with "no limit" to model size. That claim matters. If accurate, it positions Cerebras as the only architecture that can handle frontier-scale models without the latency penalties of multi-GPU interconnects.
This IPO also signals something broader about where the market sees inference heading. NVIDIA paid $20 billion for Groq six months ago, and now values Cerebras at $60 billion—figures that suggest custom silicon is no longer a niche bet but an infrastructure category. The Cerebras thesis has always been that general-purpose GPUs are a transitional architecture, and that specialized designs optimized for AI workloads will eventually command the market. The IPO doesn't prove that thesis correct. But it does prove the market is now pricing it as a serious scenario.
For investors who held through Cerebras's difficult years—the pulled S-1, the questions about commercial viability—the payoff is real. The inference infrastructure cycle that Apoorv Vyas linked to Stanford's compute scarcity discussions appears to be materializing, with Cerebras positioned as both beneficiary and validation. Whether the $60 billion valuation represents a ceiling or a floor depends entirely on whether the WSE architecture can sustain its performance advantages as models keep growing—and whether OpenAI's $10-20 billion commitment translates into recurring revenue rather than a one-time deal.