Cerebras CS-4 Claims Up to 30x Faster GPU-Class Inference

Cerebras introduced CS-4 on 18 August 2026, the fourth generation of its wafer-scale system. Angela Yeung and Eric Gardner wrote the launch note. The rack uses three new Wafer Scale Engine 3 Turbo processors and a redesigned Nexus platform.

The headline number is vendor-facing. Cerebras says inference is up to 30 times faster than GPU systems, citing Artificial Analysis and internal benchmarking from August 2026. Treat that as a scoreboard, not a promise for your model.

What Cerebras says CS-4 is for

CS-4 is aimed at interactive tokens and at total tokens per gigawatt. The company wants developers to feel 30x-class agent loops, operators to get more throughput per watt, and neoclouds to install a modular rack instead of a custom maze.

It claims more than 1,000 tokens per second on models larger than 10 trillion parameters, using wafer-to-wafer links as low as 2 microseconds. That last pair is extrapolation from internal tests. File it as a design target.

Versus CS-3, Cerebras claims up to 10x more throughput per watt and up to 2x faster performance, from internal benches and projections. The footnote on the page is the honest line: observed speed versus GPUs varies by workload, configuration, date, and model.

If you already moved agents to Gemini 3.7 Flash for cheap loops, CS-4 is the other side of the same itch: latency, not list price.

Disaggregated inference is the architecture, not a slogan

CS-4 is built to decode, not to do every stage. Prefill can sit on a GPU or ASIC. Cerebras named AMD Helios and AWS Trainium as complementary prefill platforms. State then moves to CS-4 for low-latency decode.

That split is how you should write a skill around hardware: name the prefill home and the decode home, then stop if either side errors twice. Do not bake “30x” into a customer SLA.

Nexus is the rack story. Cerebras says 50 percent fewer components, a rear-mounted Wafer-Scale Backpack that folds power conversion, liquid cooling, and I/O around the wafer, and 60 percent more automated manufacturing. Power conversion is said to sit 100 times closer to the processors than on conventional GPU boards. Deployment, in the vendor line, moves from days to hours.

I/O is dual-mode: RoCE v2 RDMA over Ethernet for mixed rooms, and Direct Wafer Links for switch-free Cerebras-to-Cerebras paths.

What actually ships, and when

First CS-4 shipments begin this quarter. Cerebras did not name a month. Do not invent one.

Save hardware notes in the library the same way you version a model ID. Compare serving options on tools before you rewrite a stack around a wafer you cannot buy yet.

This is an industry rack, not a creator laptop. The reason it belongs next to prompt work is simple: agent products are becoming latency products. A code-review skill that waits on a 10T-class model cares about decode, not about the launch video.

What to watch

Watch independent decode numbers on named models, not only the 30x chart. Watch whether Helios and Trainium prefill actually show up in customer deployments. Watch the first public cluster that claims 1,000 tokens per second above 10T.

Until a third party reruns the table, quote Cerebras as Cerebras.

FAQ

When was CS-4 announced? 18 August 2026, in a Cerebras blog by Angela Yeung and Eric Gardner.

What is inside? Three Wafer Scale Engine 3 Turbo processors on the new Nexus rack-scale platform.

How fast is it? Cerebras says up to 30x faster inference than GPU systems, from Artificial Analysis plus internal August 2026 benches.

When do systems ship? First shipments begin this quarter. No calendar day was given.

Can it run with other chips? Yes, as described: prefill on GPU or ASIC, including AMD Helios and AWS Trainium, decode on CS-4.

Sources

• Introducing Cerebras CS-4: The Fastest AI Gets Faster — Cerebras, 18 August 2026