For GPU clouds, neoclouds, enterprise GPU platforms and data-center operators
Energy is wasted when computation is repeated, moved unnecessarily, executed inefficiently, admitted too early, poorly timed, or disconnected from the power system supporting it. LuxiEnergy is a four-layer stack for data-center deployment and scheduling, from the GPU kernel to the site power limit.
01 / The stack
Layer 1 · Execute
Deterministic, efficient GPU math. Up to 16 to 50% less GPU energy per token than vLLM batch-invariant mode, and 14 to 19% less on prompts than SGLang deterministic mode, at 1.5 to 2.7x and 1.18 to 1.24x the speed.
Prefill on H100: 1.18x the tested vLLM throughput at batch 16 with about 12% lower GPU-board energy per position; under 1% run variation; about 44,860 positions/s at about 0.01532 J/position at batch 72.
Layer 2 · Shape
Schedules when AI work draws power so rack and hall power stop lunging. Same average work, calmer aggregate kW.
On 8x RTX PRO 6000 GPUs: about 9x lower aggregate power variance with about 5% throughput delta. 16 of 16 model gates pass. Complements UPS, batteries and power caps; not a battery replacement.
Layer 2 · Shape
For AI load next to on-site generation: phase control of the AI load and a deterministic plant-side power schedule the generator can follow ahead of time.
16 of 16 gates pass; multi-GPU timing about 4.7% throughput delta.
Layer 3 · Schedule
GPU cluster packer and placement layer. Packs multi-GPU jobs onto nodes, admits work only after dependencies are ready, scores idle waste in dollars, and sends placement recommendations to Slurm or Kubernetes.
Synthetic scale from 8 to 256 nodes; 21 of 21 gates green (v0.4.0).
Layer 4 · Coordinate · Roadmap
Matches compute demand to the site: power generation, electrical system, cooling and reliability limits.
Roadmap
Guarded local demand management for smaller commercial sites.
02 / Supporting evidence
Memory movement
Identical NORMAL and WNSM receipts, 0.00e0 maximum output difference on Linux x86-64.
Long context
Attention and long-context memory scaling (O(N) vs O(N²)). H200 and H100 research snapshots, June and July 2026.
Resident inference
GPU-resident weights and KV in BF16. Active-batch output matched serial BF16 on p32 and p128 gates.
Try it
luxi-tools energy and ate commands are in LuxiDemo releases (illustrative, not facility energy). python3 scripts/verify_v99_pack.py recomputes published throughput and GPU-board energy ratios.
LuxiDemo releases03 / Deployment
A control surface (scheduler phase or power caps) and a meter better than node-only, such as a PDU.
GPU-board energy measured with NVML. Site PDU, utility and demand-charge results are not yet measured.
Talk to us
Email eric@luxiedge.com or use the form on the contact page.