Cerebras Unveils CS-4, a Three-Wafer Rack That Targets 4,400 Tokens Per Second Per User

The Californian manufacturer assembles three processors at the wafer scale in a single rack and claims up to 30 times the speed of GPU solutions on large models.

Cerebras Systems presented on 18 August 2026 the CS-4, the first system of its new Nexus rack platform. The machine brings together three copies of the WSE-3 Turbo, an accelerated version of the largest AI processor ever manufactured: four thousand billion transistors, 900,000 cores and 44 GB of SRAM engraved directly on 46 225 mm2 of silicon. At the rack scale, the set announces 750 PFLOPS calculation, 129.6 petabytes per second of memory bandwidth and 7.2 terabits per second of input-output.

The commercial argument is less about gross flow than about latency per user. On GPT-OSS-120B, Cerebras claims more than 4,400 tokens per second per user, i.e. up to 30 times the speed of GPU-based solutions on identical queries – a figure that the company adds to a reserve on dependence on model architecture, context length and service configuration. Its technical director Sean Lie defends the interest in agent systems: at constant runtime, the margin earned is converted into more reasoning, verification or tool calls.

The rest of the winnings come from mechanics. The calculation module becomes a plug-in "backpack" that brings together energy conversion, direct liquid cooling and electronic control around the wafer, with half less components than the previous generation. Bringing power conversion closer to about 0.5 mm from processors, compared to about 50 mm on a conventional GPU card, doubles the energy actually delivered to silicon. Finally, the Direct Wafer Links connect the wafers without a switch with a latency down to two microseconds, aiming at clusters capable of serving models of more than 50,000 billion parameters. The first deliveries are announced for the current quarter.