Cerebras unveiled the CS-4 on Tuesday at its Supernova event, packing three dinner-plate-sized wafer-scale chips into a single rack for the first time. The company claims up to 30x faster AI inference than GPU-based systems, with a benchmark of 4,400 tokens per second on gpt-oss-120b running on one CS-4. That is roughly 12 times the throughput of the fastest GPU inference service today, according to independent benchmarking firm Artificial Analysis.
What is a wafer-scale chip? Most AI chips are individual dies cut from a silicon wafer — typically the size of a postage stamp. Cerebras instead keeps the entire wafer intact, producing a processor the size of a charcuterie board with 900,000 cores and 44GB of on-chip SRAM. The idea is that by keeping all that memory on the chip, you avoid the bottleneck of shuffling data back and forth between the processor and separate memory chips.
🔍 THE BOTTOM LINE
The CS-4 is less about new silicon and more about everything around it. Cerebras doubled the clock speed of its existing chip, redesigned the power delivery and rack architecture, and cut wafer-to-wafer latency to two microseconds. The result is a system that can run models above 50 trillion parameters interactively — but the real test comes in 2027, when a genuinely new chip generation is due.
Same Chip, Double the Clock
Here is what gives the game away. The WSE-3 Turbo carries the same 4 trillion transistors, the same 900,000 cores, the same 44GB of SRAM, and the same TSMC 5nm manufacturing node as the WSE-3 it replaces. Per-wafer compute and memory bandwidth both exactly double. As The Register concluded, that is the signature of a clock bump rather than a redesign.
Cerebras is now running the silicon at an estimated 2.8GHz, up from 1.4GHz. That is a significant engineering feat — pushing a wafer-scale chip to double its clock requires dramatically better power delivery, which is where most of the CS-4 innovation actually sits. The company moved power conversion to within 0.5 millimetres of the processor, roughly 100 times closer than conventional GPU boards, and packed it into a removable “backpack” at the rear of the chassis.
The 30x claim also comes with caveats. It measures tokens per second per user on a single model — gpt-oss-120b — against unnamed GPU systems. The 250 petaflops per wafer is a sparse FP16 number; dense FP16 performance is closer to 25 petaflops, which is still impressive but not the headline figure. Sparsity benefits vary by workload and do not always translate to real LLM inference gains.
The Architecture That Actually Matters
Where the CS-4 does look genuinely differentiated is the rack design. Three wafers per system, with wafer-to-wafer latency of just two microseconds, means Cerebras can use simple pipeline parallelism — each chip processes a portion of the model sequentially — rather than the complex tensor and expert parallelism that GPUs require. The company ditched external switches entirely, connecting chips in a 2D torus mesh topology.
Each rack carries 132GB of SRAM and draws an estimated 120 to 140 kilowatts. That sounds enormous until you compare it to the 240 to 250kW rack systems coming from AMD and Nvidia later this year. Cerebras claims 10x more throughput per watt than its predecessor, and the power efficiency story has a plausible starting point.
The modular design also separates the stable infrastructure — power, cooling, networking — from the compute. A data centre can install and qualify the rack before the compute backpacks arrive, then slide them in. Deployment time drops from days to hours.
Why This Is an Agent Story
CTO Sean Lie framed the speed argument around agents, not chatbots. “Being 30 times faster gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use,” he said at the launch.
That is the part that matters for where AI is heading. A chatbot that answers in one second versus three seconds is a convenience improvement. An agent that can run ten reasoning steps, verify each one, and still return an answer in seconds rather than minutes is a capability shift. The CS-4 is built for that workload — sustained, high-throughput decoding on large models where latency per token determines whether multi-step agent loops are practical.
CEO Andrew Feldman told Reuters the company expects to get “four times as fast between now and the end of 2027, and 20 times more throughput.”
The Concentration Question
The financial backdrop is mixed. Cerebras reported $180.1m in second-quarter revenue, up 74 per cent year on year, with cloud revenue nearly quadrupling. But revenue fell sequentially from $193.4m in the first quarter, and the quarter produced a GAAP net loss of $450.5m. The company guides to $880m to $890m of full-year revenue.
G42 and the Mohamed bin Zayed University of Artificial Intelligence together accounted for roughly 86 per cent of 2025 revenue. The $10bn-plus OpenAI contract signed in January is Cerebras’s answer to that concentration risk, alongside second-quarter additions including Cognition, Lovable, CrowdStrike, Block, and Figma. OpenAI’s Ultrafast mode, which runs GPT-5.6 Sol roughly 14 times faster on Cerebras silicon, went live just five days before the CS-4 launch.
No CS-4 customer agreements or pricing were disclosed. First shipments are due before the end of the quarter. The harder test comes in 2027, when the next generation will need to arrive on new silicon rather than a faster clock.
❓ FAQ
Is the CS-4 actually a new chip? No. The WSE-3 Turbo uses the same die as the WSE-3, manufactured on the same TSMC 5nm node with the same transistor count and SRAM capacity. Cerebras doubled the clock speed through improved power delivery, not new silicon.
What does 4,400 tokens per second mean in practice? It means a single CS-4 system can generate roughly 3,300 words per second on a 120-billion-parameter model. For agentic workloads that require multiple reasoning steps, that speed allows more verification and tool-use loops within a reasonable response time.
How does this compare to Nvidia’s rack systems? Nvidia’s NVL72 and AMD’s Helios racks draw 240 to 250kW each. The CS-4 draws an estimated 120 to 140kW, though the comparison is not straightforward because the systems use different architectures and the 30x claim is per-user throughput on a specific model, not a head-to-head rack comparison.
Does this matter for New Zealand? Not directly yet. Cerebras has no announced New Zealand customers. But the inference speed trajectory matters for any NZ company building agentic AI systems — faster inference means more capable agents at lower latency, which changes what is practical to deploy locally or via cloud providers.
