price of compute

The anatomy of a GPU-hour

Before you can rent one GPU for one hour, a fab has to etch it, a memory maker has to stack it, someone has to build it into a machine, capital has to finance it, and a building has to power and cool it. Pick a GPU — the chain and the arithmetic change with the silicon. Every number below is verified against a public source.

The chain · NVIDIA H100 SXM
The foundry. Nearly every datacenter GPU is fabbed here — the tightest bottleneck in the chain.TSMC4N process · packagingThe memory feeding this SKU — for HBM parts, the second great bottleneck.SK hynix · Micron · Samsu…HBM3 · 80 GBDesign and the software moat — and most of the margin in the whole chain.NVIDIAGPU design & softwareHow this silicon becomes a rentable machine.Supermicro · Dell · Foxco…HGX/DGX & OEM serversFleets are financed; capital recovery is the largest slice of the hourly price.Capital marketsdebt & equityColo and hyperscale facilities; PUE lives here.Data centersland · shell · coolingGrid electricity including cooling overhead (PUE) — priced on the cost stack below.Power gridkWh per GPU-hourEvaporative-cooling upper bound — environmentally real, economically small.Waterliters per GPU-hourThe providers listing this SKU right now — the layer this site prices hourly.Clouds & marketplacesAzure · DataCrunch · Hypersta…The number this site tracks.You$3.99/GPU-hr median

Dashed boxes are commodity inputs; solid boxes are companies and layers. Hover any box for context. The provider box is live — it lists who priced this SKU in the latest daily aggregate.

One H100 SXM hour, decomposed

Today’s cross-provider median for NVIDIA H100 SXM is $3.99/hr (2026-08-11). Power basis: published full-system wall power (DGX H100: 10.2 kW ÷ 8). The verifiable operating inputs explain little of the price:

Electricity$0.17/hr4.3%1.96 kWh from the grid
Water (upper bound)$0.011/hr0.3%≤ 2.3 L, priced high on purpose
Capital · facility · network · staff · margin$3.81/hr95.4%everything the operating inputs don't explain

Electricity. 1.275 kW at the wall × PUE 1.54 × 8.71¢/kWh = 1.96 kWh costing $0.171/hr 4.3% of the market price. A best-in-class fleet (PUE 1.09) pays $0.121.

Water, bounded high. At 1.8 L/kWh and a deliberately expensive $5/m³: 2.3 L and $0.011/hr. An environmental and siting constraint — not a line item.

The other 95% is the machine itself. $3.81/hr is capital recovery on the hardware and networking, the building, staff, and margin — hardware amortization under competition. That’s why prices fall as supply arrives, and why the H100 SXM price history is worth watching.

Verified inputs
InputValueSourceChecked
NVIDIA H100 SXM power basis
DGX H100: 10.2 kW ÷ 8 — includes CPUs, memory, NICs, fans
1.275 kWNVIDIA specifications2026-08-10
US industrial electricity, average retail0.0871 $/kWhEIA Electric Power Monthly (May 2026, preliminary)2026-08-10
Industry average PUE1.54 ×Uptime Institute Global Survey 2025 (weighted avg; ~flat six years)2026-08-10
Best-in-class fleet PUE (Google, 2025)1.09 ×Google data center efficiency (2025 fleet average)2026-08-10
Water usage effectiveness, evaporative-cooling upper range
Used as an upper bound; many fleets report far lower.
1.8 L/kWhOperator-reported WUE ranges (~0.2 hyperscaler best → ~1.8 evaporative)2026-08-10
Industrial water price, generous upper bound
Chosen high on purpose: even at $5/m³ the water line stays under two cents.
5 $/m³Upper-bound assumption (US industrial rates are typically well below this)2026-08-10

What this page is: reference context, computed from public, dated sources, with the arithmetic shown. What it is not: our measured data. Hardware and facility costs are deliberately presented as a residual rather than estimated — vendors don’t publish server street prices, and we don’t print numbers we can’t stand behind. Nothing on this page feeds the Index or any published series (see methodology).