What is Cerebras CS-4?
Cerebras announced CS-4 on August 18, 2026, according to its official release. It is a rack-scale system with three Wafer Scale Engine 3 Turbo (WSE-3T) processors, built on the modular Nexus platform. Each wafer contains 900,000 cores and 44 GB of integrated SRAM, according to the manufacturer. Editorial correction: the previous version dated the announcement to August 19; the company's release and blog both state August 18.
Up to 30x faster: compared with what?
The release reports over 4,400 tokens per second per user on GPT-OSS-120B, with identical prompts in the comparison, and an advantage of up to 30x over GPU-based solutions. The blog attributes its charts to Artificial Analysis data and internal benchmarking from August 2026. These are supplier-published claims, not independent tests conducted by Max Assistant. The release itself notes that performance varies with model architecture, context length, precision and serving configuration.
Efficiency and the Nexus architecture
Cerebras also claims up to 10x more throughput per watt and up to 2x the speed of CS-3. Those are different comparisons from the up-to-30x advantage over GPUs. The modular design separates compute, power and I/O. According to Cerebras, the Wafer-Scale Backpack compute module uses 50% fewer components and 60% more automated manufacturing than the prior generation; these figures refer to the module, not a verified count of every rack component. The manufacturer describes wafer-to-wafer latency as low as two microseconds and the ability to build clusters supporting models above 50 trillion parameters, not a demonstration of such a model in commercial use.
Shipments: a forecast is not confirmation
On August 18, Cerebras said first shipments would begin that quarter. This is the original announcement's forecast, not confirmation that every shipment or general availability has already happened. This article does not independently establish which delivery milestones were met.
What to measure before buying inference
Our recommendation: compare the same model and workflow, including time to first token, generation speed, end-to-end task latency, concurrency, quality and cost per completed task. An agent also waits for tools, network calls and validation; faster tokens do not automatically multiply business outcomes. Hardware energy efficiency does not guarantee a lower API price either. To distinguish speed from completion, read AutomationBench-AA and the limits of agent metrics. To design a test for your support or operations workflow, see how to apply AI in your business.
Primary sources
Cerebras release, August 18, 2026; launch blog and benchmark conditions; Hot Chips architecture details, August 25, 2026. Sources consulted October 4, 2026.