GPU price trends

The pattern repeats every generation: a new flagship launches into scarcity at eye-watering rates, supply catches up, and the price finds a floor set by older cards and marketplaces. The H100 is the textbook case, from $8+ per hour in 2023 to under $3 by 2025. The strategic read: last generation's flagship is usually the value play, and the floor under old-GPU prices is firmer than spec sheets suggest because HBM supply and inference demand prop it up.

• 2023, the famine: H100 launch met frontier-lab demand with no supply; anything with HBM rented at panic prices, and even the A100 re-inflated.
• 2024, supply arrives: CoWoS packaging capacity scaled, neoclouds racked hundreds of thousands of Hoppers, and on-demand H100 rates roughly halved in a year.
• 2025, the floor forms: Blackwell took the frontier, pushing Hopper toward inference duty; marketplaces and redeployed fleets set the new floor.
• 2026, stabilization: inference demand growth absorbs supply; medians flatten, and previous-generation prices firm rather than collapse, held up by HBM scarcity and structurally growing token demand.

How much have H100 prices dropped?

From roughly $8 or more per GPU-hour at 2023 launch scarcity to the $2-4 range by late 2025, one of the fastest infrastructure price corrections on record. Since mid-2025 H100 pricing has stabilized and drifted slightly upward, with market medians around $3.35/hr in late 2026.

Are GPU rental prices going up or down?

It depends on the tier. Previous-generation cards (A100) drifted down for years and have recently firmed as inference demand absorbs supply. Current-generation cards (H100, H200) fell steeply from launch, then flattened in 2026. Consumer cards on marketplaces keep grinding lower.

Where does this price history data come from?

Two layers: historical milestones assembled from public market reports, index publications, and archived provider pricing pages (cited on the page), and from September 2026 onward our own index, which refreshes daily across every provider we track.

Pre-2026 milestones are monthly-resolution approximations of the typical GPU-cloud on-demand rate, assembled from Silicon Data's H100 price history analysis, GetDeploying's price index (weekly observations since July 2024), Thunder Compute's market trend reports, and archived provider pricing pages. They are directional, not tick data. From September 2026 the series extends automatically from our own index: 123 tracked rates across 24 providers, refreshed daily, methodology on the about page. Current per-provider rates live on the GPU pricing index.