AMD and Cerebras pair Helios racks with Cerebras WSE for inference

AMD and Cerebras announced a disaggregated inference setup in which Helios handles prompt processing and the Cerebras wafer-scale engine handles token generation. The first availability is planned through Cerebras Cloud in the second half of 2026. The 5x tokens-per-second-per-watt figure is a modeled estimate for one Kimi model configuration, not a measured result.

• Helios runs prompt processing; Cerebras WSE runs token generation
• Claimed up to 5x higher tokens per second per watt vs a Cerebras-only setup (modeled, Kimi 2.6 1T)
• First availability expected through Cerebras Cloud in H2 2026