← Back to feed

Cerebras Systems launched the CS-4 on August 18, its fourth-generation AI accelerator and the first product in the new Nexus rack-scale platform. The headline number: 4,400 tokens per second per user on GPT-OSS-120B, which Cerebras claims is 30 times faster than comparable GPU systems in production.

The Hardware

The CS-4 is a three-wafer rack unit. Each wafer is a WSE-3 Turbo (WSE-3T) — an overclocked version of the WSE-3 that doubles clock speed to approximately 2.8 GHz, enabled by redesigned power delivery that brings voltage conversion 100x closer to the die. The result is double the compute and double the memory bandwidth per wafer, without changing the underlying 5nm TSMC silicon.

Metric	CS-3 (1 wafer)	CS-4 (3 wafers)
AI compute	125 PFLOPs	750 PFLOPs
Memory bandwidth	21.6 PB/s	129.6 PB/s
On-chip fabric bandwidth	26.7 PB/s	160.5 PB/s
System I/O bandwidth	1.2 Tb/s	7.2 Tb/s
Wafer-to-wafer latency	5 microseconds	2 microseconds

Each WSE-3T contains 900,000 AI cores, 4 trillion transistors, and 44GB of on-wafer SRAM across 46,225 mm² of silicon. The rack draws 125–135kW TDP.

What Actually Changed

The WSE-3T is the same silicon as the WSE-3, clocked harder. SemiAnalysis estimated the original WSE-3 ran at 1.4 GHz; the Turbo version pushes roughly 2.8 GHz. Cerebras was able to do this because the N5 process node is significantly more mature in 2026 than when the WSE-3 was first deployed in March 2024, and the company redesigned power delivery to feed twice as much current to each wafer.

The Nexus rack itself is a bigger architectural change. Cerebras split compute from power and networking: each wafer sits in a modular “backpack” that plugs into a shared power shelf at the front of the rack. Three backpacks per rack. The separation means power modules, networking, and wafers can be upgraded independently — a deliberate design to accommodate future WSE generations without rebuilding the entire rack. Cerebras has committed to roughly 2x performance per year through 2027.

What did not change: the 44GB of SRAM per wafer. That capacity limit is the persistent constraint on how large a model any single Cerebras system can hold in on-wafer memory. For frontier models, Cerebras continues to rely on pipeline-parallel inference across multiple wafers, with all MoE experts for a given model distributed across the rack.

Disaggregated Inference and Partners

The CS-4 adds a modular I/O interface designed for disaggregated inference setups. Cerebras positions itself as the decode chip, pairing with HBM-based accelerators for prefill. Current partners include AMD and AWS Trainium. The field-upgradeable NIC module means Cerebras can swap in partner-specific network interfaces without redesigning the compute backpack — which Nextplatform noted appears calibrated specifically for AWS’s EFA networking standard on Trainium.

Wafer-to-wafer latency at 2 microseconds (down from 5 on CS-3) enables clusters supporting models with more than 50 trillion parameters.

Economics

Cerebras claims 10x more throughput per watt than the CS-3. On the interactivity side, SemiAnalysis estimates the CS-4 will hit roughly 4,000 tok/sec/user on frontier models versus the CS-3’s 2,000 tok/sec/user, while Blackwell GPUs realistically deliver around 100 tok/sec/user at normal concurrency levels — making the effective gap closer to 20-40x on interactive workloads.

CS-4 shipments begin Q3 2026. Early access is available for select customers now.

Context

Cerebras went public on Nasdaq (CBRS) in late 2025 with a $20 billion OpenAI contract disclosed in its IPO filing. The CS-4 is the first hardware release as a public company. The 30x speed claim is measured at token throughput per user — the company’s strongest metric — not compute throughput, where GPU systems compare more favourably on raw FLOP counts.