What is a GPU-hour and why is it the unit?

A GPU-hour is one GPU rented for one hour — the natural unit of compute pricing. Providers quote whole instances (an 8×H100 node at ~$31/hr), so we divide by GPU count to make everything comparable: that node is $3.87 per GPU-hour.

• Per-GPU normalization is what makes a 1× marketplace listing comparable to an 8× reserved node.
• Pricing tiers (on-demand, spot/interruptible, community) are different products and are never blended into one number.
• Today's reference point: the H100 SXM medians $3.87/GPU-hr across 11 providers.
• Electricity is a small share of the price — roughly 5% for an H100 — the rest is mostly hardware amortization (see the anatomy).

This answer is computed from live data on every load — per-provider daily medians, per single GPU per hour, pricing tiers never mixed (see methodology). Quote it with attribution: “Data: Price of Compute — priceofcompute.com” (cite).