The question this page answers
What does a million AI tokens cost in electricity and in depreciating silicon, using a single GPU's watts, tokens per second, PUE, and a grid price? The numbers below are the tool defaults, written in HTML so they are on the page before any script runs.
Worked example
A 1.2 kW GPU at 60 tok/s with PUE 1.25 uses about 6.9 Wh per 1,000 tokens. At $0.06/kWh that is $0.42 per million tokens of electricity; a $30,000 GPU over 5 years at 60% utilization adds about $5.28 of depreciation — hardware, not energy, dominates. Total floor about $5.70 per million tokens. First-order estimate, not an engineering or investment sign-off, and not affiliated with any chip vendor.
Formula
Serving one million tokens takes 1,000,000 / (tokens per second × 3600) GPU-hours. Energy in kWh is GPU-hours × kW × PUE. Energy cost is that kWh times the electricity price. Hardware cost per GPU-hour is installed cost divided by (lifetime years × 8,760 × utilization). Hardware cost per million tokens is GPU-hours times that rate. The two components sum to the physical floor; an API price stacks margin on top.
Constants
- GPU draw 1.2 kW — illustrative default, replace before citing.
- Inference speed 60 tok/s — illustrative default, replace before citing.
- PUE 1.25 — site writing: typical 1.1–1.5.
- Electricity $0.06/kWh — site writing: $0.06/kWh in the joules-per-token example.
- GPU installed cost $30,000 — illustrative default, replace before citing.
- Lifetime 5 years, utilization 60% — illustrative default, replace before citing.
Limitations
Batching, networking, storage, staff, idle draw, and vendor margin are left out. This isolates the energy-plus-silicon floor.
Matching essay: joules per million AI tokens. Related models: data-center planner, solar + battery sizing.