Cerebras CS-4 debuts, claiming up to 30x faster AI inference than GPUs

Cerebras has unveiled the CS-4, a rack-scale AI system built on three WSE-3 Turbo wafer-scale chips that the company claims delivers up to 30x faster inference than production GPU systems and up to 10x more throughput per watt than its prior CS-3. The pitch centers on interactivity at extreme scale: Cerebras says the design sustains more than 1,000 tokens per second on models exceeding 10 trillion parameters, thanks in part to switchless wafer-to-wafer links with latency as low as two microseconds. As with any vendor launch, the eye-catching multipliers are the company’s own figures and warrant independent benchmarking before they’re taken at face value.

The more concrete story is packaging and deployment. The CS-4 is the first product on Cerebras’s new Nexus Platform Architecture, which splits the system into modular Compute, Power, and I/O elements. A self-contained “Wafer-Scale Backpack” folds the wafer, power conversion, direct liquid cooling, high-speed I/O, and control electronics into one 3D assembly with roughly 50% fewer components. Placing power delivery about 0.5mm from the processor—versus roughly 50mm on conventional GPU boards—cuts board-level power loss and lets Cerebras push more power into the chip for higher clocks.

Strategically, the design targets hyperscale operators frustrated by long buildout times. Because the PowerRack (power, cooling, network) can be installed and facility-qualified before any compute arrives, operators can rack the stable infrastructure first and later slide in compute backpacks, cutting deployment from days to hours and easing future upgrades. It’s a direct bid to position wafer-scale hardware as a practical alternative to GPU clusters for frontier-model inference, not just a lab curiosity.