The energy of AI,
measured — not declared.
240 measurements taken on 1 October 2026 on one NVIDIA H100, each with a signed Serial Alice certificate you can verify yourself. Five open models from 3B to 32B parameters, swept across batch size and prompt length — next to our earlier H100, B200 and Intel TDX-attested H200 runs.
Six findings. Every one checkable.
From batch 1 to 32 with a 1,024-token prompt, energy per token fell 23× on Qwen2.5-3B and 16× on Qwen2.5-14B.
At batch 1: 985 µWh per token for Qwen2.5-3B, 3,596 for Qwen2.5-32B.
At batch 32, Phi-3.5-mini (3.8B) used 159 µWh per token; Qwen2.5-3B used 43. Parameter count alone does not predict energy.
Qwen2.5-14B at batch 8: 692 µWh per token for 4,096 tokens in and 128 out, against 262 for 128 in and 128 out.
On Llama-3-8B (H100, batch 128) AWQ drew 37% less power than BF16 but used 145% more energy per token — it was slower. GPTQ used 25% less.
The GPU's energy counter and the integral of its sampled power agreed within 0.33% (median). That is internal consistency, not accuracy against an external meter.
Serve more at once.
Pay less per token.
Every curve falls steeply: the fixed cost of running the GPU is shared by more requests. The slope differs by model — Phi-3.5-mini flattens early, the Qwen models keep improving.
NVIDIA H100 80GB · 1,024-token prompt, 512-token answer · median of 2–3 repeats
Batch 1 and batch 8 · 1,024 in / 512 out
Batch 8 · tokens in / out · Qwen2.5-3B and Qwen2.5-14B
BF16 · GPTQ · AWQ · batch 1 → 128 · log scale · earlier H100 benchmark run
Measured,
not folklore.
On Llama-3-8B, GPTQ used 25–44% less energy per token than BF16 at every batch size. AWQ drew fewer watts — and still used 2.5–4.5× more energy, because it generated tokens more slowly. Power on its own tells you very little.
H200 inside Intel TDX.
Bound to an attested VM.
Seven certificates from an H200 in a Phala confidential VM: each measurement bound to an Intel TDX quote, an allow-listed TD image, and a platform TCB that was up to date.
| Workload | Power | Energy | Certificate |
|---|---|---|---|
| H200 idle (0% util) | 80.3 W | — | sa-c020e48c… |
| BF16 matmul load | 592.8 W | — | sa-40cb8908… |
| vLLM Phi-3-mini BF16 · b=1 | 309.8 W | 5.30 Wh | sa-ee4cd7b3… |
| vLLM Phi-3-mini BF16 · b=16 | 354.8 W | 6.06 Wh | sa-1a051cf4… |
| vLLM Phi-3-mini BF16 · b=64 | 451.5 W | 7.83 Wh | sa-1751b088… |
| vLLM Phi-3-mini BF16 · b=128 | 516.9 W | 8.81 Wh | sa-d64b2e20… |
| FP8 matmul | 661.7 W | 11.52 Wh | sa-0244226d… |
Every point on the batch chart
has a certificate.
Open any of them: the verifier recomputes the hash, checks the Ed25519 signature and shows the evidence tier. No account needed.
| Model | Batch | µWh / token | Tokens / s | Certificate |
|---|
- GPU energy per generated token. Each certificate states exactly what was measured and on which machine — open any of them.
- What ran. One NVIDIA H100 80GB, 1 October 2026: 246 attempts, 240 completed. The other 6, all Phi-3.5-mini at batch 32 with 4,096-token prompts, ran out of GPU memory and are not in the data. 222 points used the default 700 W limit; 18 Qwen2.5-7B points used 300–500 W limits.
- What kind of evidence. The H100 certificates are signed and anchored, and read the GPU's own energy counter through our agent. They are
self_reported: not attested in a TEE and not checked against an external meter. The seven H200 certificates arehardware_attested: each is bound to a verified Intel TDX quote, but the energy still comes from the GPU's own reporting. - How repeatable. Across the 84 configurations with repeats at the default limit, variation was 0.50% at the median and 3.05% at worst (coefficient of variation). With power limits it reached 9.07%.
- What is not counted. GPU only: no CPU, RAM, cooling or network.
- Stronger evidence, on other runs. Two H100 training runs were measured in a sealed collector on an attested GPU, and a gate checked the evidence before issuing: sa-prod-b2-0003 (538 Wh, one hour) and sa-prod-b2-0004 (6.6 Wh, 75 seconds). Both are GPU only, are training rather than inference, and are not part of the charts above.
- Compare, don't copy. Absolute values depend on the serving software and hardware; the differences between configurations are what carry over.