Nodus: Useful AI Output per Paid GPU-Hour through Multi-Tenant Inference Consolidation

GPU utilization is not the same as useful inference output: a busy accelerator can still produce responses too late to count toward a service commitment. We present Nodus, a GPU virtualization and multi-tenant inference system evaluated by useful AI output per paid GPU-hour. Nodus consolidates tenants onto a shared pool and reassigns GPU capacity on demand. In a single-node H100 SXM evaluation with 32 synthetic tenants and identical replayed traffic, Nodus uses 4 GPUs against a dedicated per-tenant baseline of 32 GPUs: 8× fewer GPUs across 3 independent traffic seeds. The reported improvements in SLO-met requests per GPU-second are 7.6×/ 8.0×/ 8.0×; reported cost-per-SLO-met-token reductions are 7.8×/ 8.1×/ 8.1×. Lowest per-class SLO attainment is 99.7%/ 99.0%/ 98.5%, rather than the baseline's 100%. Component measurements cover shared serving, tenant reassignment, suspend/resume, batch harvesting, and memory-cap behavior. These findings support a bounded consolidation claim, not production validation or universal isolation guarantees.