Cerebras CS-4 · Wafer-Scale System · Nexus Platform
The all-new Cerebras CS-4 is a revolutionary rack-scale solution built from three WSE-3 Turbo wafer-scale engines — delivering up to 30× faster inference than GPU systems, enhanced economics, and a simple path to deploy hyperscale capacity. It is the architecture for frontier AI.1
01 — The machine
CS-4 is the first system built on the Cerebras Nexus rack-scale platform. Each system packs three WSE-3 Turbo processors into pluggable Wafer-Scale Backpacks, with all-new power, cooling, and I/O engineered around the wafer — not bolted onto it.2
Every wafer delivers up to 2× the speed of the previous generation, and the system delivers up to 10× more throughput per watt than CS-3 — while generating tokens up to 30× faster than production GPU systems. The result: a machine designed to deliver both throughput and interactivity.2
02 — The wafer at its heart
At the center of CS-4 is the newly announced WSE-3 Turbo — like the WSE-3, the largest AI processor ever built. One wafer integrates 4 trillion transistors and 900,000 AI-optimized cores across 46,225 mm² of silicon, with 44 GB of SRAM directly on the wafer.1
The Turbo doubles what matters: AI compute to 250 PFLOPS per wafer, memory bandwidth to 43.2 PB/s, on-chip fabric bandwidth to 53.5 PB/s, and off-chip I/O to 2.4 Tb/s. Because memory bandwidth determines decode speed and throughput, this is a step-change in both.1
Why wafer scale? A single wafer holds the model weights and the entire interconnect on-chip, so inference never bottlenecks on GPU-to-GPU links. Compared with NVIDIA's flagship B200, one WSE-3 carries 19× more transistors and delivers 28× more AI compute — and it does the work that would otherwise require racks of GPUs.4
Wafer-scale integration replaces the interconnect fabric of an entire GPU cluster with on-chip wires. Squares below are drawn with areas in proportion to transistor count — 4 trillion on one WSE-3 versus 208 billion on an NVIDIA B200.4
19× more transistors · 28× more AI compute per device. Illustration; areas proportional to transistor count.
03 — Architecture
The next leap in AI infrastructure cannot come from improving one component in isolation. Nexus rethinks the rack around three modular elements — each self-contained, each field-upgradeable, and together cutting deployment time from days to hours.3
Wafer-Scale Backpack
Each Wafer-Scale Backpack folds the wafer, power conversion, direct liquid cooling, high-speed I/O, and control electronics into one compact, self-contained 3D package — with 50% fewer components than the prior generation. Hover or tap a marker to inspect each subsystem.
Compute
Cerebras fundamentally re-imagined the server. Each backpack is a self-contained assembly built directly around the wafer, decoupled from power supplies — simplifying manufacturing and letting compute slide into a pre-qualified rack in hours, not days.3
Power
Power conversion sits just 0.5 mm from the processor — roughly 100× closer than the ~50 mm of conventional GPU boards — nearly eliminating board-level power loss. Twice as much power reaches the WSE-3T, enabling higher frequencies and faster tokens.2
I/O
A new I/O subsystem doubles bandwidth and slashes latency. Standards-based RoCE v2 RDMA over Ethernet plugs into existing infrastructure, while Direct Wafer Links connect wafers within and across racks with no switch — wafer-to-wafer latency as low as two microseconds.1
04 — Performance
In head-to-head serving on GPT-OSS-120B with identical prompts, CS-4 delivers more than 4,400 tokens per second per user — up to 30× faster than GPU solutions. And the advantage holds from small efficient models to the largest models in the world.1
Values approximate, read from Cerebras's published chart. Source: Artificial Analysis and internal benchmarking, August 2026.5 Actual throughput varies by model architecture, context length, precision, and serving configuration.
Direct Wafer Links cut interconnect latency from 5µs to as low as 2µs. Serving a model across wafers no longer costs interactivity — CS-4 sustains more than 1,000 tokens per second on models exceeding 10 trillion parameters, and scales to models beyond 50 trillion parameters across clusters.1
CS-4 shifts the inference Pareto frontier: up to 10× more throughput per watt than CS-3 while generating tokens up to 30× faster than GPU systems. Fast tokens are worth more than slow tokens — operators get more capacity and more value per token within the same power budget.2
Agentic and reasoning workloads live or die on decode speed. At over 1,000 tokens per second per user on 10-trillion-parameter models, a CS-4 cluster keeps frontier-class models fully interactive.2
06 — Workloads
"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 & co-founder, Cerebras1
Agents chain dozens of model calls per task. At thousands of tokens per second, multi-step reasoning, self-verification, and tool use complete in the time a GPU system spends on a single reply — making genuinely interactive agents practical.
CS-4 natively supports disaggregated serving: a purpose-built prefill engine — including AMD Helios or AWS Trainium — processes the prompt, then hands state to CS-4 for ultra-low-latency decode. GPU or ASIC for prefill, wafer scale for speed.3
With 2µs Direct Wafer Links, massive CS-4 clusters serve models beyond 50 trillion parameters at interactive speeds — capacity for the next generation of frontier models without waiting for new silicon.1
Wafer-scale systems have long served both training and inference with single-device simplicity — chosen by leading corporations, research institutes, and governments. Cerebras counts organizations such as OpenAI, AWS, G42, and MBZUAI among its customers and partners.14
"In AI, speed is productivity. Historically, fast inference meant using smaller and less capable models. Cerebras CS-4 delivers industry-leading speeds on the largest frontier models, fundamentally changing the paradigm."
Andrew Feldman — CEO & co-founder, Cerebras1
"CS-4 makes dramatic improvements in system deployability, reliability, and networking, which enables scaling performance to larger models for large scale token factories."
Dylan Patel — Founder & CEO, SemiAnalysis1
07 — By the numbers
| Metric | CS-3 (one wafer) | CS-4 (three 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 Tbit/s | 7.2 Tbit/s |
| I/O latency | 5 µs | 2 µs |
Source: Cerebras CS-4 announcement, August 18, 2026.1
First CS-4 shipments begin this quarter.1
08 — FAQ
CS-4 is Cerebras's fourth-generation AI system: a rack-scale solution built from three WSE-3 Turbo wafer-scale processors, delivering 750 PFLOPS of AI compute, 129.6 PB/s of memory bandwidth, and 7.2 Tb/s of I/O. It is the first system on the Nexus rack-scale platform architecture, announced August 18, 2026.1
CS-4 is up to 2× faster per wafer and delivers up to 10× more throughput per watt than CS-3. The WSE-3 Turbo doubles AI compute (250 vs. 125 PFLOPS per wafer), memory bandwidth (43.2 vs. 21.6 PB/s), on-chip fabric bandwidth (53.5 vs. 26.7 PB/s), and I/O (2.4 vs. 1.2 Tb/s), while wafer-to-wafer latency drops from 5µs to 2µs. The rack itself is new: modular Nexus compute, power, and I/O instead of a monolithic design.1
In a head-to-head comparison on GPT-OSS-120B with identical prompts, CS-4 delivers more than 4,400 tokens per second per user — up to 30× faster than GPU solutions — and sustains over 1,000 tokens per second on models exceeding 10 trillion parameters.1
CS-4 delivers up to 10× more throughput per watt than CS-3 while generating tokens up to 30× faster than production GPU systems — shifting the Pareto frontier so operators no longer choose between interactivity and total capacity.2
Agentic workloads chain many model calls — reasoning, verification, tool use — for every user request. At up to 30× GPU speed, CS-4 gives agents more than an order of magnitude more compute inside the same wall-clock budget, turning batch-feeling agents into interactive ones.1
The backpack is CS-4's compute module: a rear-mounted, self-contained assembly that folds the wafer, power conversion, direct liquid cooling, high-speed I/O, and control electronics into one compact 3D package. It has 50% fewer components than the prior design and attaches vertically to the power array, so compute can be installed, serviced, or upgraded independently.3
Nexus separates the stable infrastructure layer — power, cooling, network — from modular compute. The PowerRack is installed and facility-qualified first; compute backpacks slide in later and connect to power, cooling, and data. Deployment drops from days to hours, and future upgrades no longer require redesigning the whole system.2
CS-4 serves today's frontier open and proprietary models — Cerebras's published results span Llama 3.3 70B, GPT-OSS-120B, GLM 4.7, and Kimi K2.7-class trillion-parameter models — and clusters support models beyond 50 trillion parameters. It natively supports disaggregated inference, pairing GPU or ASIC prefill (including AMD Helios and AWS Trainium) with Cerebras decode.35
First CS-4 shipments begin this quarter. Talk to Cerebras about hyperscale capacity, cloud access, or a proof of concept on your own models.
