What's this about?
Cerebras announced the CS-4, the next generation of its wafer-scale computing system, days after its partnership with OpenAI put GPT-5.6 Sol at 750 output tokens per second under the Ultrafast tier.
What exactly is wafer-scale?
Chips are normally cut from a large silicon wafer into hundreds of small dies, packaged individually, then wired together. Cerebras doesn't cut. It uses the entire wafer as one chip roughly the size of a dinner plate. The advantage is that data moving between parts of a computation stays on-silicon instead of crossing between packages, and that inter-chip movement is the bottleneck for generating text one token at a time. The trade-offs are manufacturing yield, power delivery and heat across an object that size — problems the industry avoided by cutting wafers up in the first place.
Why is this worth noticing?
- The timing follows commercial validation rather than preceding it: the OpenAI Ultrafast partnership demonstrated the architecture doing something GPUs can't match on latency.
- It's a genuine architectural alternative in a market where the conversation is otherwise entirely about who can buy the most Nvidia — and this week's memory-price story is a reminder of what that concentration costs.