LLM inference may already beat human inference on energy per token
jd_pressman · x · 2026-07-21
The post argues that LLM inference may already be more energy-efficient than human inference on a per-token basis, at least for short-context generation.
- A rough estimate uses Epoch’s inventory of about 48,000 NVIDIA NVL72 racks, roughly 5.8 GW of rack power, and divides by output capacity to get about:
- 0.3 J/token for short context
- 1 J/token for medium context
- 12 J/token for long context
- After accounting for cooling and power overhead, the author says the figures rise somewhat, but the core comparison still stands for short outputs.
- Using the tweet’s 12 W brain figure and 3–4 tokens/s gives about 3–4 J/token; using the more conventional 20 W whole-brain budget gives 5–7 J/token.
- Their conclusion: modern batched inference is already more energy-efficient per short-context token than the proposed human baseline, while very long-context inference is still somewhat worse.
- They also note an important caveat: Kimi K2.6 has 1T stored parameters, but only 32B active per token, not the roughly 360B active parameters assumed in the quoted comparison.
Related event: Epoch AI: LLM Query Energy Consumption Drops Below Human Brain's(2 posts)→
More from Infra
- HBM demand shifts: low-stack inference plans shelved, Rubin to go 12-high only — BenBajarin · 2026-10-26
- CircuitHub unveils The Grid, a factory-scale robot making PCBs in the US twice as fast as Shenzhen — garrytan · 2026-10-26
- Ubicloud publishes "Fixing Tool-Call Accuracy in vLLM" — petrusenko_max · 2026-10-26
- Meta approves nonvoting Class C shares as 2026 capex guidance hits $135B–$145B — thione · 2026-10-26
- Unsloth Ships MiniMax-M3.1-Flash Quants: 2-bit Retains 91.5% Accuracy, Runs in 150GB RAM — danielhanchen · 2026-10-26
- Engineer with thousands of fabricated devices pours cold water on photonic crystal modulator yields — jwt0625 · 2026-10-26