Where a GPU-hour's price actually goes

Ask why an H100-hour costs three dollars and change, and most answers wave at power bills and water-hungry data centers. So we did the arithmetic with verified public inputs — the EIA’s industrial electricity rate (8.71¢/kWh, May 2026), the Uptime Institute’s industry-average PUE (1.54), and NVIDIA’s own published system power for an 8-GPU DGX H100 (10.2 kW — that’s 1.275 kW per GPU at the wall, counting the CPUs, memory, and fans around it).

An H100-hour consumes about 1.96 kWh and roughly $0.17 of electricity — about 5% of its market price. Water, priced at a deliberately generous $5/m³ and an evaporative-cooling upper bound of 1.8 L/kWh, adds at most a penny. The remaining ~94% of the price is the machine itself: capital recovery on the server and networking, the building, staff, and margin.

The proportion inverts as the silicon gets cheaper. An RTX 4090 rents for around $0.56/hr on marketplaces, and its 450 W TDP works out to roughly 11% of the price in electricity — double the H100’s share. Cheap GPUs are power businesses; expensive GPUs are financing businesses.

That framing explains the market’s behavior better than any input cost: rental prices are dominated by hardware amortization under competition, so they fall as supply arrives and financing terms improve — which is exactly what the Price of Compute Index exists to measure.

The full decomposition — with every input, source, and date — lives inline on every GPU page, computed against that SKU’s live median. We deliberately publish hardware cost as a residual, not an estimate: vendors don’t publish server street prices, and we don’t print numbers we can’t stand behind.