Cerebras and Gimlet Labs route each AI phase to different chips

Cerebras and Gimlet Labs announced an inference cloud that is already serving tokens in private deployments with joint customers. The setup splits each workload by phase instead of pushing everything through one type of chip: Cerebras hardware handles latency-sensitive work, while GPUs take on high-throughput processing.

That routing is the real change, with software matching each phase to the silicon suited for it. The companies are targeting peak speeds of up to 3,000 tokens per second and plan a first data center later in 2026.