GPUs / NVIDIA L40S / Anatomy
The anatomy of a NVIDIA L40S-hour
One NVIDIA L40S rents for $1.67 per hour today. This is where that money goes — the parts that can be sourced, and the residual that can’t.
Anatomy: why it costs this
Everything that has to exist before one L40S-hour can be rented — then the arithmetic on the parts that can be priced honestly.
Electricity$0.05/hr2.8%0.54 kWh from the grid
Water (upper bound)$0.003/hr0.2%≤ 0.6 L, priced high on purpose
Capital · facility · network · staff · margin$1.62/hr97.0%everything the operating inputs don't explain
Power basis: vendor GPU TDP only — full-server draw runs meaningfully higher × PUE 1.54 × 8.71¢/kWh (0.54 kWh from the grid). Water bounded at 1.8 L/kWh and $5/m³ — deliberately high. The residual is hardware amortization, facility, staff, and margin — presented as a residual because vendors don’t publish server street prices, and we don’t print numbers we can’t stand behind.
Verified inputs & sources
- NVIDIA L40S power basis: 0.35 kW — NVIDIA specifications
- US industrial electricity, average retail: 0.0871 $/kWh — EIA Electric Power Monthly (May 2026, preliminary) (checked 2026-08-10)
- Industry average PUE: 1.54 × — Uptime Institute Global Survey 2025 (weighted avg; ~flat six years) (checked 2026-08-10)
- Best-in-class fleet PUE (Google, 2025): 1.09 × — Google data center efficiency (2025 fleet average) (checked 2026-08-10)
- Water usage effectiveness, evaporative-cooling upper range: 1.8 L/kWh — Operator-reported WUE ranges (~0.2 hyperscaler best → ~1.8 evaporative) (checked 2026-08-10)
- Industrial water price, generous upper bound: 5 $/m³ — Upper-bound assumption (US industrial rates are typically well below this) (checked 2026-08-10)
- Reference context with the arithmetic shown — never fed into the Index or any published series (see methodology).