What changed
The planned service uses AMD Helios for prompts and long-context processing, then Cerebras hardware for latency-sensitive decoding.
Why it matters
Prefill and decoding optimize for different constraints. Disaggregated inference can route each stage to hardware designed for it.
What to do
Measure context-processing time, first-token latency, decode rate, energy, reliability and cost per completed task separately.
Caveat
Availability is planned for the second half of 2026, and the claimed efficiency gain comes from company modelling rather than deployed third-party testing.