Cerebras' CS-4 Hits 750 PFLOPS With Three Wafers and a Better Cooling System

Cerebras unveiled the CS-4 out of its Supernova conference, with a release timed 20:00 ET on 18 August — 00:00 UTC on 19 August. The system is rated at 750 PFLOPS and 129.6 petabytes per second of memory bandwidth, and Cerebras claims speeds up to thirty times faster than GPU-based solutions in certain configurations.

What is in the box

Three WSE-3 Turbo processors, each carrying four trillion transistors and 900,000 AI-optimised cores, fabricated by TSMC on a 5nm process. Cerebras describes the system as double the speed of the CS-3. First deliveries are stated for the third quarter of 2026, and no pricing was disclosed.

What the common telling gets wrong

The first error is treating 750 PFLOPS as a chip specification. It is three wafers combined. Anyone comparing it directly against a single-wafer predecessor, or reporting a sixfold jump over the CS-3, is multiplying a per-wafer improvement by a package count.

The second is assuming new silicon. The process node is unchanged. Electronics Weekly is explicit that Cerebras "increases the performance of its existing silicon by running more power through it," achieved by "coming up with a more effective cooling system." This is a thermal and power-delivery result on a die that already existed, not a fabrication advance.

The third is the comparison metric. "Up to 30x faster than GPUs" is tokens per second per user — how fast a single interactive stream returns, demonstrated at more than 4,400 tokens per second per user on GPT-OSS-120B. It is not throughput per rack or per dollar, which is where GPU fleets are optimised, and the vendor itself notes results vary with model architecture, context length and configuration.

The interesting implication

If a two-year-old wafer can double its output on cooling and power delivery alone, the binding constraint on inference silicon has shifted away from lithography and toward thermals and power conversion — the same place the data-centre buildout is already constrained. Nothing has shipped yet, which makes this a launch rather than revenue.