Cerebras CS-4: three wafers in one rack, inference speeds up (but independent benchmarks are missing)

On 18 August 2026 Cerebras Systems announced the CS-4, a rack-scale inference system and the first member of the new "Cerebras Nexus" platform. The structural news: three Wafer Scale Engine 3 Turbo (WSE-3T) processors in a single rack, the first time the company has put more than one wafer in the same cabinet (source: official press release; confirmed by DigiTimes and ServeTheHome). In the release the chief executive summed up the goal: "Speed is productivity... CS-4 delivers industry-leading speeds on the largest frontier models." (source: Cerebras).

On the silicon side, ServeTheHome reports the following for a single WSE-3T: 900,000 AI cores, 44 GB of on-chip SRAM, roughly 4 trillion transistors and a TSMC 5 nm process; the stated performance is 250 PFLOPS at sparse FP16, twice that of the WSE-3, while staying on the same process node (so the gain does not come from a new process). At rack level, Cerebras claims 750 PFLOPS of AI compute, 132 GB of total SRAM, 129.6 PB/s of memory bandwidth, 7.2 Tbit/s of I/O and 2 microseconds of latency; wafer-to-wafer latency drops from ~5 to ~2 microseconds (ServeTheHome; implicator.ai). The structural constraint of on-chip memory remains: 132 GB per rack. More broadly, the structural constraint of the wafer-scale approach remains, compared with GPUs, which use external memory.

The performance promises are ambitious but, as of today, self-reported: Cerebras talks about "up to 2x" tokens/sec per single user versus the CS-3 and "up to 30x" versus GPU-based solutions, with over 4,400 tokens/s on gpt-oss-120b; plus "up to 10x" throughput per watt and "50%" fewer components in the rack (source: press release). Implicator.ai clarifies that the 4,400+ tokens/s and the 30x figure come from Cerebras internal testing, and that the GPU-side comparison rests on public Artificial Analysis data; no independent lab has published direct benchmarks on the CS-4, and the 30x is measured on a single model, gpt-oss-120b, not on frontier models. In current Artificial Analysis data for gpt-oss-120b (high profile), Cerebras comes out as the fastest provider at ~1,670 tokens/s, ahead of SambaNova (~704) and Groq (~479), but those figures reflect infrastructure already in service, not the CS-4. On why push per-user speed, the chief technology officer stresses: "Being 30 times faster gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use." (source: Cerebras).

On power draw, no figure is given; ServeTheHome estimates ~54 kW per wafer system, while implicator.ai puts a fully populated rack at 120–140 kW. First shipments are indicated for the current quarter (Q3 2026); pricing, volumes and customers have not been disclosed. On raw compute, ServeTheHome calculates roughly six times a single CS-3 system and roughly three times a CS-3 rack: a gap between raw compute and user-perceived speed that the company does not explain. On deployment and networking experience, an outside comment in the release: "CS-4 makes dramatic improvements in system deployability, reliability, and networking." — attributed to SemiAnalysis founder Dylan Patel. The claims about "50% fewer components" and installations in hours rather than days also remain, for now, company statements unverified by third parties.

The CS-4 pushes where it currently matters for agents and tool use — low latency and tokens/sec per user — just as the bottleneck shifts to power, cooling and deployment time. The technical foundations are solid and well documented; on the claimed performance, we're waiting for independent measurements and transparent numbers on power draw and capacity. If those arrive, we'll really understand the step change. — Pixie