Cerebras Unveils CS-4: Three Wafer-Sized Chips, Up to 30x Faster Inference Than GPUs

Cerebras has announced the CS-4, which the company calls "the fastest AI accelerator in the industry." While most AI chips connect many small silicon dies, Cerebras has stuck to an unusual approach: using an entire wafer as a single processor. The CS-4 bundles three of its newest wafers into one rack-scale system and is the first product built on a new platform architecture called Nexus.

What was announced

The CS-4 is built from three new WSE-3 Turbo (WSE-3T) processors. As a full system it delivers 750 PFLOPS of AI compute, 7.2 terabits per second of I/O, and 129.6 petabytes per second of memory bandwidth. Total compute-fabric bandwidth jumps to 160.5 PB/s, and wafer-to-wafer latency drops as low as two microseconds — enabling large clusters and support for models with more than 50 trillion parameters, the company says. The core message is that speed equals productivity. "Historically, fast inference meant using smaller and less capable models," said CEO Andrew Feldman. "Cerebras CS-4 delivers industry-leading speeds on the largest frontier models."

What the numbers say

Cerebras' headline benchmark is on GPT-OSS-120B. Given identical prompts, the CS-4 produces more than 4,400 tokens per second per user, which the company says is up to 30x faster than GPU solutions. "Being 30 times faster doesn't just make a response feel fast," said CTO Sean Lie. "It gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use in the same wall-clock time." Cerebras notes in a footnote that actual throughput varies by model architecture, context length, precision, and serving configuration. It is worth keeping in mind that the benchmark is a vendor measurement based on a single open-weight model.

Chip and system design

The CS-4's engine, the WSE-3 Turbo, remains "the largest AI processor ever built." It packs four trillion transistors and 900,000 AI-optimized cores across 46,225 square millimeters of silicon, with 44GB of SRAM integrated directly on the wafer. Versus the prior generation, per-wafer AI compute doubles to 250 PFLOPS and memory bandwidth doubles to 43.2 PB/s. The system design also changed. A rear-mounted "Backpack" module decouples compute from the power supplies, cutting deployment time from days to hours and reducing component count by 50%, the company says. Cerebras also moved power conversion roughly 100x closer to the processor — from about 50mm on conventional GPU boards to around 0.5mm — nearly eliminating board-level power loss.

Why it matters now

As AI spending shifts from training toward inference, the opening for alternative architectures in the inference market Nvidia has effectively dominated is widening. Cerebras' bet is to use a whole wafer instead of connecting many small chips, arguing that this reduces chip-to-chip communication bottlenecks and wins on speed. But figures like 30x and 10x are vendor benchmarks, not independent results; real-world performance across workloads and model sizes will be tested by outside parties after Q3 shipments. Cerebras counts OpenAI, G42, and AWS among its major customers, and how much the CS-4 cracks the inference competitive landscape is the key thing to watch.