Cerebras CS-4 claims 30x GPU inference speed

CS-4 is an AI accelerator of three wafer scale engines

The Cerebras CS-4 is the fourth generation of the Cerebras System, a rack-scale AI accelerator built from three newly released Wafer Scale Engine 3 Turbo processors. Cerebras calls it the fastest AI accelerator in the industry. It is also the first system on the Nexus platform architecture, the rack design that the same announcement describes.

A wafer scale engine is Cerebras's processor built on a whole silicon wafer rather than a cut-down chip. The release describes the WSE-3 Turbo, shortened to WSE-3T, as the largest AI processor ever built. All of these figures are Cerebras's own claims, from its press release and blog post of August 18, 2026.

The WSE-3 Turbo chip doubles bandwidth for inference

The release says the WSE-3 Turbo doubles the WSE-3's AI compute to 250 PFLOPS per wafer and doubles its memory bandwidth to 43.2 petabytes per second. Each wafer holds four trillion transistors and 900,000 cores, with 44GB of SRAM on the wafer itself. Cerebras calls memory bandwidth the determining factor in speed and throughput, so the release presents the bandwidth doubling as a gain in both.

Two more per-wafer figures double as well: on-chip fabric bandwidth to 53.5 petabytes per second and off-chip I/O to 2.4 terabits per second. I/O latency falls from five microseconds to as low as two.

The release also gives a system-level comparison. Read it with one fact in mind: the CS-3 column is one wafer and the CS-4 column is three. The compute and bandwidth rows are totals, so each reflects both the move from one wafer to three and the doubling above, while the per-wafer figures show the change in the chip itself.

What 30x AI inference speed means in tokens per second

The 30x is measured in tokens per second per user. In the release's head-to-head run on GPT-OSS-120B with identical prompts, the CS-4 reaches more than 4,400 tokens per second for each user, which the release puts at up to 30 times the speed of GPU solutions. Against its own CS-3, Cerebras claims up to 2x, and the release says that 2x brings the CS-4's advantage over GPUs to up to 30x.

A token is a small piece of text that a model reads or writes. Tokens per second per user measures how fast one person's answer streams out. The release makes a separate claim about efficiency: up to 10x more throughput per watt than the CS-3.

Sean Lie, the CTO and co-founder of Cerebras, links speed to what an agent can do. He says the extra speed lets an agentic system fit more than an order of magnitude more reasoning, checking or tool use into the same elapsed time. A footnote in the release adds that real throughput changes with the model, the context length, the numeric precision and how the system is served.

Cerebras vs NVIDIA: the release names only GPU solutions

Both documents compare the CS-4 with unnamed hardware. They say "GPU solutions", "GPU systems", "production GPU systems" and "conventional GPU boards", and neither names a vendor, NVIDIA included, or a GPU model. So neither says which company's product the 30x is measured against.

The 30x is also not a comparison with the CS-3. The CS-3 comparison is the separate 2x in the release.

Which figures are measured and which are projected

The blog prints a source line under each chart, and the lines differ. The "Source lines printed under the CS-4 figures" table lists each figure with the line Cerebras attaches to it.

Source lines printed under the CS-4 figures

The blog's text does not list the models in the 30x chart. A footer under the whole post adds that comparisons rest on third-party benchmarking or internal testing, and that observed speed gains over GPU-based systems may vary by workload, configuration, date and models tested.

Where the CS-4 speed comes from: power delivery and I/O

Cerebras credits the whole system rather than one part. The blog says faster links between wafers, dense power delivery and the modular Nexus rack platform work together. Cerebras splits the Nexus platform into three elements: compute, power and I/O.

On power, the release says conversion sits about 0.5 millimeters from the processor, 100x closer than the roughly 50 millimeters of conventional GPU boards. Cerebras says this nearly eliminates board-level power loss and delivers twice as much power to the WSE-3T. That, it adds, allows higher operating frequencies and faster token generation.

On I/O, a programmable Wafer I/O Module supports RoCE v2 RDMA over Ethernet, a standard way to connect to existing infrastructure. It also supports Direct Wafer Links, a switch-free way to join wafers within and across racks, with latency as low as two microseconds. Off-wafer bandwidth is 2.4 terabits per second per wafer and 7.2 per CS-4 rack.

The backpack cuts installation from days to hours

The CS-4 uses a Wafer-Scale Backpack, mounted on the rear and attached upright to the power array. It is a self-contained assembly. Inside it, the electronics for power conversion, I/O and control, plus the direct liquid cooling, wrap around the wafer.

Cerebras says that separating compute from the power supplies simplifies manufacturing and cuts deployment from days to hours. Compared with the previous generation, the backpack has 50% fewer components and uses 60% more automated manufacturing.

The modular design also lets compute, power and I/O each scale on their own, which Cerebras says gets innovations to market faster. The blog adds that an improvement in one part of the system can reach customers without waiting for every other part to be redesigned.

Disaggregated inference with AMD Helios and AWS Trainium

Disaggregated inference gives each of the two phases of inference to a platform built for it. In prefill, the system processes the incoming prompt and prepares the model state. In decode, it generates the response.

The blog describes a purpose-built prefill engine that hands the model state to the CS-4 for ultra-low-latency decoding. It names AMD Helios and AWS Trainium as complementary prefill platforms and says providers can use GPU or ASIC infrastructure for prefill.

The release lists the same two as examples of Cerebras ecosystem partners. See also OpenAI Jalapeno Chip: Inference Benchmarks 2026.

Is there a CS-4 price, and when does it ship?

Neither document gives a CS-4 price or a general availability date. Cerebras says the first CS-4 shipments begin this quarter, and the release points to the CS-4 datasheet for full system specifications. The two documents also leave out any named customer and the name of the quarter.

One more detail is easy to blur. The blog says the CS-4 can reach over 1,000 tokens per second on models above 10 trillion parameters, a figure the blog itself labels an extrapolation. The release says the architecture supports models with over 50 trillion parameters.

The 1,000 tokens per second is a speed figure at one model size, while the 50 trillion is a size the architecture supports, so they should not be read as two versions of one limit. The Cerebras CS-4 headline of 30x is best read with its conditions attached: GPT-OSS-120B, tokens per second per user, and the footnote on throughput.