Cerebras Launches CS-4 AI Accelerator, Claiming 30x Faster Inference Than GPUs on an Overclocked WSE-3

Overview

Cerebras Systems has unveiled the CS-4, a new rack-scale AI inference system that the company describes as “the fastest AI accelerator in the industry,” claiming it delivers “up to 30 times faster inference than GPU systems.” The system was unveiled at the company’s Supernova event and is shipping this quarter, marking Cerebras’s first new hardware release since its Nasdaq debut in May.

What We Know

According to Cerebras, CS-4 is “built from three new Wafer Scale Engine 3 Turbo processors” and pairs them with a redesigned rack system called Nexus. The company says the design “reduces wafer-to-wafer interconnect latency to as low as 2 microseconds” and “can deliver more than 1,000 tokens per second on models exceeding 10 trillion parameters.” Cerebras also claims CS-4 delivers “up to 10 times more throughput per watt than CS-3,” its prior-generation system.

According to The Next Web, the three WSE-3 Turbo wafers in each CS-4 combine for “750 petaflops of sparse FP16 compute” and “129.6 petabytes per second of memory bandwidth,” with support for models above 50 trillion parameters. The outlet reports that Cerebras’s 30-times performance claim “measures tokens per second per user on a single model, gpt-oss-120b, against unnamed GPU systems.” Cerebras CEO Andrew Feldman framed the launch around latency: “In AI, speed is productivity,” he said, while chief technology officer Sean Lie tied the speed claim to agentic workloads, saying it “gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use.”

On the hardware itself, The Next Platform and The Next Web both report that the WSE-3 Turbo carries “the exact same 900,000 cores” and “44 GB” of on-wafer SRAM as the original WSE-3, built on the same TSMC 5-nanometer process Cerebras has used since it first delivered WSE-3 engines in March 2024. Both outlets describe the new chip as a clock-speed increase rather than a new design: The Next Platform estimates the clock “has doubled to 2.8 GHz from the 1.4 GHz used in the plain vanilla WSE-3 engine,” and The Next Web reports that The Register concluded it is “not new silicon so much as the existing die pushed from roughly 1.4GHz to 2.8GHz.” Cerebras itself confirms in its announcement that CS-4 is not launching a new compute engine — the improvement instead comes from “moving power conversion 100 times closer to the processors compared to conventional GPU boards,” which “enables the delivery of twice as much power to the WSE-3 Turbo, enabling higher operating frequencies and faster token generation.”

The Nexus rack also targets manufacturing speed: Cerebras says the redesigned “Wafer-Scale Backpack” assembly has “50% fewer components and uses 60% more automated manufacturing,” which The Next Platform confirms independently, adding that the rack “can be deployed up to 3X faster than prior systems.” For deployment strategy, Cerebras says CS-4 is built for “disaggregated inference,” pairing its own decode performance with “complementary prefill platforms, including AMD Helios and AWS Trainium,” rather than requiring an all-Cerebras stack. On power draw, The Next Web reports The Register’s estimate that a CS-4 rack consumes “120 to 140 kilowatts,” roughly half what comparable AMD and Nvidia rack systems draw.

On availability, Cerebras says “the first CS-4 shipments begin this quarter,” and The Next Platform reports that select customers already have early access, with “general availability… later in the third quarter of this year.” Cerebras has also outlined a roadmap of roughly 2x annual throughput improvements through 2029 built on the same Nexus rack design, according to The Next Platform.

What We Don’t Know

Neither Cerebras nor the outlets covering the launch disclosed pricing for CS-4 systems or racks. The exact clock-speed figures (1.4 GHz to 2.8 GHz) are estimates from outside analysts rather than numbers Cerebras itself published, and the 30x inference-speed claim is based on a specific benchmark — tokens per second per user on the gpt-oss-120b model against unnamed GPU competitors — rather than a universal figure across all workloads.

CS-4 arrives a few months after Cerebras priced its Nasdaq IPO at $18.50 a share, closing with a $5.55 billion raise, and follows earlier production use of Cerebras hardware, including OpenAI’s deployment of GPT-5.3 Codex Spark on Cerebras systems at 1,000 tokens per second.