Cerebras make a chip the size of a whole silicon wafer. Where everyone else cuts the wafer into hundreds of chips, they leave it whole.
The idea is old and simple: the slowest thing in a computation is the conversation between chips. You remove the conversation by removing the borders.
Fact
The facts: on 18 August 2026 Cerebras announced CS-4, the fourth generation of its system. It is built around three Wafer Scale Engine 3 Turbo processors, in a completely redesigned rack and a platform called Nexus. Per the company, the system delivers up to 30 times faster inference than GPU systems, with better economics and a path to scaling to gigawatt capacity. The company points to three groups: developers who want a fast response from reasoning and agentic applications; data centre operators who measure throughput per gigawatt; and providers who need parts that can be manufactured and installed quickly.
How I read the number 30
It is theirs and it is about inference, not training. The two are different jobs and are measured differently. There is no independent check yet.
I write it the way they say it: up to 30 times. At the top end, in a suitable case, by their methodology.
Response speed is no longer a convenience. It is the difference between an agent that does work and an agent you sit and wait for.
What interests me is where the market is pushing. For a long time the measure was how big a model you could train. Now more and more it is how fast you answer, because agentic applications make dozens of round trips for a single task.
Their other sentence is not about chips: component by component no longer works, compute, power, cooling and I/O have to move together. NVIDIA says the same thing this month, in different words.
For us, the small ones, this changes nothing tomorrow. But it explains why the price of a fast answer is falling rather than rising.
Official primary sources
Cerebras: Introducing Cerebras CS-4 (18.08.2026)The visual is generated code-art, no third-party images.