Ground AI training & inference costs in real commodity prices
GPU pricing last verified: 2026-03-15. Actual pricing varies by volume and configuration.
| Market | Training Cost | Electricity | $/M Tokens |
|---|---|---|---|
| N. Virginia | $2.7K | $66.73 | $2.51 |
| DFW | $2.6K | $48.43 | $1.84 |
| Phoenix | $2.7K | $55.97 | $2.11 |
| C. Ohio | $2.7K | $51.66 | $1.95 |
| Atlanta | $2.7K | $53.81 | $2.03 |
| Chicago | $2.7K | $59.20 | $2.23 |
| SV | $2.7K | $80.72 | $3.02 |
| Portland / Hillsboro | $2.6K | $40.90 | $1.56 |
| Denver / Aurora | $2.6K | $45.20 | $1.72 |
| Reno / Las Vegas | $2.6K | $43.05 | $1.64 |
GPU count estimated as ceil(modelParams * 2 / gpuMemory) for bf16 training. Training cost = GPU amortization (3-year) + electricity + networking (8%) + facility (5%). Cloud pricing uses a 3x multiplier over owned hardware; colocation uses 1.5x.
Inference cost per million tokens derived from GPU throughput (FP16 TFLOPS / 2 * modelParams) plus power cost at market rates. Water usage from DOE/industry data by market. Carbon at 0.4 tCO2/MWh (US grid average).
Electricity prices from CBRE Q3 2025 DC Market Report, EIA-861, and ISO LMP data. All figures are estimates for comparison purposes.