Fewer GPUs, Same SLO.

Most GPU fleets are sized for peak demand, one deployment per GPU, and sit idle for much of the day. We evaluate Nodus Elastic Inference, a pooled serving layer, against the solutions operators use today: one dedicated GPU per tenant, NVIDIA MIG and MPS partitions, cold starts, unmanaged memory overcommit, HAMi-core memory caps, and idle reserved capacity. On NVIDIA H100 GPUs with replayed multi-tenant traffic, Nodus served 32 tenants on 4 GPUs instead of 32 at ≥ 98.5% SLO with zero errors across 3 seeds, an 8× reduction in GPUs and a 7.8–8.1× reduction in cost per million SLO-met tokens. At 2× traffic it planned its own layout on 5 GPUs at 99.98% SLO, 6.4× cheaper than dedicated. Model activation took 0.16–0.39 s instead of 34.5–49.9 s. This report lists only results that were measured on real GPUs and beat the existing solution.