Deployable Self-Planning: GPU Fleet Layouts from Past Traffic at 2× and 3× Load

A consolidation system is only deployable if it can choose its GPU layout without knowing the future. We evaluate Nodus as a black box that plans its own fleet layout from observed past traffic, on a single 8× H100 SXM node serving 32 tenants, replaying the same trace to Nodus and to one dedicated GPU per tenant. At 3× traffic, Nodus runs on 6 GPUs at 99.78% lowest per-class SLO attainment with 0 errors, versus 32 dedicated GPUs: 5.333× fewer GPUs, 5.329× more SLO-met requests per GPU-second, and 5.384× lower cost per million SLO-met tokens ($2.1088 vs $11.3528). At 2× traffic, planning from past traffic only—with no foresight—chooses 6 GPUs (one more than a perfect-foresight plan's 5) and delivers 5.33× fewer GPUs and 5.34× lower cost than dedicated. Across 1×, 2×, and 3× traffic, GPU demand grows sub-linearly: 4, 5, and 6 GPUs versus 32 dedicated.