Cerebras Introduces CS-4 — Up to 30x Faster Inference, 1,000+ Tokens/sec for 10T-Parameter Models

Key Takeaways

→ Up to 30x faster inference than GPU systems across the model spectrum

→ 1,000+ tokens/sec on models exceeding 10 trillion parameters — enabled by 2-microsecond interconnect latency

→ Up to 10x more throughput per watt and 2x faster than CS-3

→ Nexus modular rack: 50% fewer components, deployment cut from days to hours

→ Native disaggregated inference: pair GPU/ASIC prefill (AMD Helios, AWS Trainium) with Cerebras decode

Cerebras has introduced the fourth generation of its AI system: CS-4. Built from three new Wafer Scale Engine 3 Turbo processors with a completely redesigned rack, CS-4 delivers up to 30 times faster inference than GPU systems and a simple path to deploying hyperscale capacity.

Key Performance Claims

Up to 30x faster than GPU systems — a new record for production inference
1,000+ tokens per second for models exceeding 10 trillion parameters, enabled by wafer-to-wafer interconnect latency as low as 2 microseconds
Up to 10x more throughput per watt than CS-3, with up to 2x faster performance
Native support for disaggregated inference — pairing purpose-built prefill engines with CS-4’s ultra-low-latency decoding

The system targets a broad audience: developers wanting responsive reasoning at 30x speed, data center operators wanting higher throughput per gigawatt, and neoclouds/hyperscalers needing modular systems that can be manufactured, installed, and upgraded at gigawatt scale.

The Nexus Platform Architecture

CS-4 is the first system built on Cerebras’ new Nexus rack-scale platform, which rethinks the rack around compute, power, and I/O:

Wafer-Scale Backpack — a pluggable compute design that folds power conversion, liquid cooling, I/O, and control electronics into a self-contained package with 50% fewer components, reducing deployment time from days to hours
High-density power delivery — moving power conversion 100x closer to the processors, delivering twice the power to the WSE-3 Turbo and enabling higher operating frequencies
New I/O subsystem — doubles bandwidth with standards-based RoCE v2 RDMA plus Direct Wafer Links for switch-free connections between systems

Disaggregated Inference

A notable architectural choice: CS-4 natively supports splitting inference into prefill (processing the incoming prompt) and decode (generating the response). Operators can use GPU or ASIC infrastructure — including AMD Helios and AWS Trainium — for efficient prefill, while CS-4 handles ultrafast decoding. This heterogeneous approach lets providers combine infrastructure economics with Cerebras decode speed.

Availability

The first CS-4 shipments begin this quarter. The launch extends Cerebras’ momentum in the inference speed race, weeks after revealing how it serves GPT-5.6 Sol at up to 750 tokens per second in its OpenAI ultrafast partnership.

Frequently Asked Questions

What is the Cerebras CS-4?

The Cerebras CS-4 is the fourth generation of Cerebras' AI system, built from three new Wafer Scale Engine 3 Turbo processors. Announced August 18, 2026, it delivers up to 30x faster inference than GPU systems and over 1,000 tokens per second on models exceeding 10 trillion parameters.

How fast is the Cerebras CS-4?

CS-4 delivers up to 30x faster inference than GPU systems, over 1,000 tokens per second for 10T+ parameter models, and up to 10x more throughput per watt than the previous CS-3 generation — with wafer-to-wafer interconnect latency as low as 2 microseconds.

When does the Cerebras CS-4 ship?

The first CS-4 shipments begin in Q3 2026 — the quarter it was announced. Cerebras positions it as a foundation for frontier AI, hyperscale capacity, and heterogeneous inference deployments.

What is disaggregated inference on CS-4?

CS-4 natively splits inference into prefill (processing the prompt) and decode (generating the response). Operators can use GPU or ASIC infrastructure — like AMD Helios or AWS Trainium — for efficient prefill while CS-4 handles ultrafast decoding.