Fit more GPUs per megawatt

Power — not GPUs — is the binding constraint on AI capacity. Phase-aware power capping reclaims 15–25% of GPU power during memory-bound inference with minimal throughput loss. That reclaimed power is sellable capacity. See how much.

Provisioned power: 5 MW

GPU model

Power reclaimed: 22%

Verified band from published A100/H100 studies.

PUE (facility overhead): 1.00

In 5 MW of H100 SXMs, you can fit

+2,015

more GPUs (28% capacity expansion)

At full power7,142 GPUs

With NemulAI9,157 GPUs

Added annual revenue*$36.9M/yr

Talk to us about Operator Edition →

* At H100 SXM ref. rate $2.99/GPU-hr, 70% utilization. Illustrative.

The evidence

Microsoft (ASPLOS '24): frequency locking reclaims up to 20% of power for ≤7% performance loss; the curve is superlinear (<2% drop at 7% below max clock).
Patel et al. (arXiv 2310.03003): A100 capped 250W→175W on LLaMA inference — 23% energy saved for +6.7% time.
throttLL'eM (arXiv 2408.05235): SLO-aware frequency scaling, 1.71–1.78× tokens/Joule vs NVIDIA Triton.