Cerebras CS-4 claims 30x GPU inference speed

Cerebras says its new CS-4 rack delivers more than 4,400 tokens per second per user on GPT-OSS-120B — up to 30x the speed of GPU solutions — from three WSE-3 Turbo wafers. The company’s announcement also claims 750 PFLOPs and first shipments this quarter, while warning that actual throughput varies by model, context, precision, and serving configuration. For builders, the number to file away is not a benchmark victory but a new latency envelope for agent loops; OpenAI’s ultrafast model work points to the same product pressure from the software side.