“OpenAI selected Cerebras WSE-3 to serve GPT-5.6 Sol at 750 tokens per second — the highest publicly disclosed inference throughput for a frontier model. This is not a benchmark artifact: 750 tok/s on a frontier model is 8–10× faster than NVIDIA H100-based inference clusters at comparable model quality. The OpenAI deployment validates WSE-3 as production-grade at scale and creates a reference architecture that every enterprise evaluating real-time AI applications must now benchmark against.”

“Cerebras WSE-3 contains 4 trillion transistors on a single wafer — 57× larger than the largest conventional GPU die. The architectural advantage is latency: a single WSE-3 holds an entire 70B-parameter model in on-chip SRAM without HBM bandwidth bottlenecks. For inference workloads, on-chip memory eliminates the memory-bandwidth wall that limits GPU clusters. This is a structural throughput advantage that cannot be replicated by stacking more H100s.”

“WSE-3 draws up to 23kW per chip — substantially more than a single GPU but dramatically less than the equivalent GPU cluster delivering the same throughput. A 750 tok/s inference job on Cerebras consumes 23kW; the GPU cluster needed to match that throughput would require 8–16 H100s drawing 80–130kW total. Cerebras power-per-token at peak inference is 3–5× more efficient than GPU clusters, a metric that is now entering hyperscaler TCO conversations as power costs escalate.”

“Cerebras' wafer-scale architecture is structurally HBM-free — the WSE-3 uses on-chip SRAM instead of external HBM stacks. This makes Cerebras revenue completely independent of the HBM supply constraint that gates every NVIDIA, AMD, and Google TPU shipment. As HBM remains sold out through 2026, Cerebras is the only high-performance AI accelerator that can ship without waiting for SK Hynix or Micron allocation.”