GPU Rental Price Index

Every week we record the on-demand price of every GPU listing we track, converted to US dollars and normalized to per-GPU, per-hour: an 8×H100 instance at $40/hr counts as $5.00 per GPU per hour. The data is our own: 41,524 weekly price observations recorded since July 2024, across 69 providers and 96 GPU models, each distilled from raw price scrapes that run as often as every 15 minutes. Spot and reservation pricing is measured separately and does not enter the index.

The index is the median of each model's price relative to its own price in the base week, taken across a fixed panel of models, with every model weighted equally. The better weight would be each model's deployed capacity or rental volume, but providers do not publish those, and any proxy we could compute, such as listing counts, would quietly favor whatever gets listed most rather than what gets rented most. Jan. 6, 2025 = 100. Weeks in which our pipeline captured only part of the market are excluded, and a single model sits out any week in which fewer than four providers priced it, rather than let a thin median move the panel — the "GPU models priced" count on each point shows how many models stand behind it.

The panel is reviewed every January and July. A model joins once it has been priced broadly and consistently for about three months (13 qualifying weeks), and leaves only when it has stopped being priced for a sustained stretch. When membership changes, the new panel is linked to the old one so the level runs continuously. A recent launch therefore waits a few months before it can move the headline; until then it appears in the per-model charts, measured the same way.

Each model's line is the median listed price a buyer faces, not any one provider's price list. It moves both when providers change their prices and when cheaper or pricier offerings enter or leave the market. The band around the index line spans the middle 60% of panel models: when it widens, models are diverging rather than the market moving together.

One caveat on the early history. Coverage has grown substantially since the index began: from roughly 30 providers and 300 tracked prices in early 2025 to around 60 providers and 2,000 tracked prices by February 2026, and most prices now refresh daily rather than every few weeks. The chart shades earlier stretches by the market breadth we tracked at the time; treat index levels there as approximate.

The panel: AMD MI300X, Nvidia A10, Nvidia A100, Nvidia A40, Nvidia A4000, Nvidia A5000, Nvidia RTX A6000, Nvidia B200, Nvidia B300, Nvidia H100, Nvidia H200, Nvidia L4, Nvidia L40, Nvidia L40S, Nvidia P100, Nvidia RTX 3070, Nvidia RTX 3080 Ti, Nvidia RTX 3090, Nvidia RTX 4000 Ada, Nvidia RTX 4080, Nvidia RTX 4090, Nvidia RTX 5090, Nvidia Quadro RTX 6000, Nvidia RTX 6000 Ada, Nvidia RTX Pro 6000, Nvidia T4, Nvidia V100. The panel has changed at January 2026 and July 2026. Departed: Nvidia GH200 (July 2026 review).

Accruing history: Nvidia RTX 3080 (11 of 13 qualifying weeks); Nvidia RTX 5080 (12 of 13 qualifying weeks); Nvidia RTX Pro 4500 (1 of 13 qualifying weeks). Each can join at the January 2027 review if it keeps being priced.

The underlying weekly snapshots are available through the GetDeploying API, so everything on this page can be reproduced independently.