Nebius buys Inferize to cut idle GPU tax

Nebius acquired Inferize on October 1 and folded its inference-optimization technology and team into Token Factory. Nebius says the system attacks cold-start idle time—GPUs sitting unused while models load, scale or update—and Inferize had a working prototype within three months of its January 2026 founding; terms were not disclosed. The claim is a capacity-utilization thesis, not a published savings rate, so teams should ask for cold-start, burst and SLA traces before paying for a similar optimization layer; the archive’s goodput analysis explains why a faster first token is not automatically more useful work.