The Cerebras story is one of the more extraordinary in recent technology history, and not because the company succeeded — because it almost didn't, twice.

The first near-death was the CFIUS investigation. Cerebras had filed confidentially for an IPO in 2024, but its financial relationship with G42, an Abu Dhabi AI group with reported connections to Chinese technology companies, triggered a national security review. The IPO was withdrawn. The company spent most of 2025 untangling that relationship.

The second near-death was the market itself. By early 2026, the DeepSeek narrative had raised serious questions about whether AI infrastructure was as capital-intensive as everyone had assumed. If you could train competitive models for $6 million, why did you need Cerebras's wafer-scale chips?

The answer, as it turned out, was inference. Training efficiency and inference efficiency are different problems. Cerebras chips are extraordinarily fast at running already-trained models — the use case that scales with enterprise deployment. As enterprise AI deployment accelerated through 2026, Cerebras's inference story found its audience.

The 108% first-day gain was not a meme-stock moment. It was institutional buyers who had watched the company for two years concluding that the CFIUS risk was resolved and the inference thesis was validated. At $8.8 billion, it remains small relative to Nvidia. But it is public, it is real, and it is the benchmark for what an AI chip IPO looks like when it actually works.