Data-center GPUs are built for throughput, not efficiency — so every AI fleet burns power doing nothing and overpays for the work it does. GPU Optimizer recovers both. And through Serial Alice, it hands you a certificate that proves how much — verifiable by anyone, trusting no one.
Two wastes hide in every fleet — invisible on today's dashboards, and impossible to report credibly.
An idle H100 burns ≈72.5 W; an idle H100 server ≈2–3 kW. Merely keeping a model's CUDA context alive costs a fixed 26–66 W — serving 0 or 1,000 tokens/sec. With average GPU utilization near 5%, most fleet energy does no useful work.
Cards run at the full board limit when a lower clock delivers the same throughput at far less energy. On H100 decode, the efficiency clock cut energy 35% while throughput went up. Nobody was capturing that — or proving it.
The customer proves the waste on their own fleet before anything is ever touched.
Read-only, no privileges. Learns each card's idle floor per serial and flags waste. Ships JSON / CSV / HTML.
Prometheus metrics as a no-privilege Kubernetes DaemonSet. Installed first, left running.
The actuator. Idle-park + efficiency-clock. Persist-before-apply, release-on-exit, dry-run by default.
And we measured the ceiling honestly — then signed every result into a certificate.
Ghost recovery ladder — one lever does it; extra privileges add nothing (all within a 3 W band).
| Lever on the 119 W ghost | Recovery |
|---|---|
| Context reset — free, no root | 40.6% |
| SM-clock-lock — non-destructive | 41.1% |
| -ac / -pl / -lmc alone | ~0% |
| Every privileged lever stacked | 40.5% |
No combination beats one lever. We could have claimed "more with root." We measured it, and didn't.
| Decode clock (Qwen2.5-7B) | Thrpt | Energy |
|---|---|---|
| 1980 MHz (unlocked) | base | base |
| 1485 MHz | −0.25% | −13% |
| 1185 MHz | +1.45% | −35% |
Serving idle-gap (120 s): −32% energy, latency preserved, 0 errors. Training gap: −35%, active step preserved.
Every number here is backed by a signed certificate — verify at api.serialalice.pt/v2/certificates/<id>/verify →
NVIDIA now ships its own capping (WPPS, ~15%) — but Blackwell B200 only, inside Mission Control + SLURM. Anyone can type nvidia-smi -pl. Capping is table stakes. Our defensibility is coverage, idle reclamation, and proof.
1,000× H100 · 80% utilization · PUE 1.54 · 7% saving · throughput preserved.
| Scenario | Annual saving |
|---|---|
| US, GPU-only | $29,600 |
| US, facility (incl. cooling) | $45,600 |
| EU, facility (incl. cooling) | €105,200 |
Post-quantum signed, Merkle-anchored, offline-verifiable certificates — in partnership with Serial Alice, a working attestation platform with two provisional patents filed (INPI Portugal).
The mechanism is proven end-to-end on a single H100. The next unlock is a fleet-scale, multi-GPU, production pilot — with attested, at-scale savings. If you run GPUs, we can prove what you're wasting.