H100 Rental Benchmarks Show Mixed Signals on AI Compute Costs
A completed-transaction index fell 1.8% over seven days, while a separate advertised-rate measure had a provisional median of $3.53 per GPU-hour for Oct. 12–18.

Rental prices for NVIDIA’s H100 data-center GPU present different readings across two measures, leaving no single market-wide rate: a transaction index fell over seven days but rose over 30, while advertised prices were higher in a weekly snapshot.
A transaction-based index for H100 SXM settled at $2.74 per GPU-hour at 4 p.m. ET (20:00 UTC) on Oct. 16. It was down 1.8% over seven days and up 4.6% over 30 days. The index reflects completed neocloud rental transactions, weighted by volume and adjusted to limit outliers.
A separate index of advertised on-demand rates put the provisional median at $3.53 per H100 GPU-hour for Oct. 12–18. It drew on 36 listed rates from 34 providers as of Oct. 16 at 11:40 p.m. ET (Oct. 17 at 3:40 a.m. UTC), before the weekly period ended. The finalized median for Oct. 5–11 was $3.49, based on 33 rates from 31 providers.
One reading reflects completed rentals and the other published offers. Their inputs and time windows differ, so they cannot be combined into a single market price or used as one common measure of change. Both are hourly rental measures; neither measures what a buyer pays to acquire a physical H100 card on the resale market.
NVIDIA CFO Colette M. Kress said on May 20, “The price of renting an H100 has risen 20% year-to-date, while A100 cloud pricing is up nearly 15%.” Her figures describe rental rates at that point, not card resale values.
NVIDIA’s Data Center revenue was $75.2 billion for the fiscal quarter ended April 26, up 92% from a year earlier. That growth adds company context but does not establish what caused rental prices to change or quantify any resale-price movement.
Comparisons also depend on GPU configuration, location and contract terms. A short hourly rental is not equivalent to a long reservation, and hardware in one region may not match capacity elsewhere, complicating efforts to standardize prices. Earlier TokenPost coverage of GPU compute financing addressed those differences.
NVIDIA announced the H100 in 2022 and said it was in full production by Sept. 20 that year. The current figures show rental rates at different market points without establishing a resale-price trend.