← Back to Latest

Cerebras CS-4: Wafer-Scale AI Computing Arrives

On August 18, 2026, Cerebras Systems unveiled CS-4 — the fourth generation of its wafer-scale AI system and, by the company's claims, the fastest AI accelerator in the industry. CS-4 is built from three new Wafer Scale Engine 3 Turbo (WSE-3T) processors, delivering up to 30× faster inference than GPU-based systems while doubling throughput per watt compared to its predecessor.

The numbers are striking: 750 PFLOPS of AI compute, 129.6 PB/s of aggregate memory bandwidth, and 160.5 PB/s of on-chip fabric bandwidth — all in a single rack-scale system. On GPT-OSS-120B, Cerebras reports exceeding 4,400 tokens per second per user in their own testing, with speeds on Gemma 4 31B reaching 3,400 output tokens per second in third-party benchmarking by Artificial Analysis.

Being 30 times faster doesn't just make a response feel fast. 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.

— Sean Lie, CTO and co-founder, Cerebras

What is wafer-scale computing?

Most AI accelerators — NVIDIA's Blackwell GPUs, AMD's Instinct line — are cut from a silicon wafer into dozens of smaller chip dies, then wired back together on a board or across a network. That is standard chip manufacturing: yields improve when dies are small because a single defect only ruins one small die instead of an entire wafer.

Cerebras does the opposite. Its Wafer-Scale Engine (WSE) keeps an entire 300mm silicon wafer as one chip — no dicing, no die-to-die packaging. The payoff is bandwidth: instead of data hopping across a board or a network switch between separate chips, it stays on one piece of silicon with on-die interconnect, which is orders of magnitude faster than off-chip communication.

The tradeoff is everything else — power delivery, cooling, and yield all become harder problems at that scale. Cerebras has addressed these through redundant compute cores, redundant routing, and a fail-in-place architecture that shuts flaws down and routes around them. CS-4's redesigned Wafer-Scale Backpack — a vertically mounted module integrating the processor with power conversion, direct liquid cooling, high-speed I/O, and control electronics — is largely about making these challenges manageable at a new scale.

What is actually new in CS-4

CS-4 is the first system built on Cerebras's Nexus Platform Architecture — a modular design split into three parts: Compute, Power, and I/O.

The WSE-3T processor itself retains the same 4 trillion transistors, 900,000 AI-optimized cores, and 44 GB of on-wafer SRAM as the original WSE-3, but doubles AI compute to 250 PFLOPS per wafer and doubles memory bandwidth to 43.2 PB/s. Each wafer is roughly 2× the speed of the previous generation.

Why speed matters for agentic AI

Agentic AI creates two distinct computing challenges: efficiently processing enormous amounts of context and generating tokens with extremely low latency. When an agent reasons, uses tools, and interacts with other systems, it generates responses one token at a time — causing even tiny delays to multiply across complex chains of work. Faster token generation gives agents more time to inspect files, write and test code, call tools, verify results, and iterate — all while maintaining a responsive user experience.

Cerebras specifically cited this as the design target: CS-4 is built to extend inference performance for the era of agentic AI, where workloads are interactive and latency-sensitive rather than pure bulk throughput.

Ecosystem and availability

CS-4 is designed to interoperate with other ecosystem architectures — Cerebras cited AMD Helios and AWS Trainium as partners for disaggregated inference configurations, where a purpose-built prefill engine processes an incoming prompt and then hands it to Cerebras for ultra-low latency decode.

First CS-4 shipments are scheduled to begin in Q3 2026. The company has not published public pricing. Independent benchmarks beyond Artificial Analysis's Gemma 4 measurements were not yet available at the time of writing.

The broader context: Cerebras filed for an IPO and signed a multi-year deal with OpenAI for a 750 MW AI inference rollout. OpenAI's GPT-5.6 Sol already runs on Cerebras silicon in "Ultrafast mode," delivering 750 tokens per second ahead of the CS-4 generation.