H100 & H200 price tracker

What this tracks

The board lists the cheapest currently-rentable single-GPU H100 and H200 instances on three clouds with fully public pricing, sorted by price per GPU-hour. It deliberately tracks 1× instances because that is what most fine-tuning and inference experiments actually rent; multi-GPU cluster pricing follows different supply and is not comparable per GPU. Spot rows are interruptible capacity — the provider can reclaim the machine — while on-demand rows are dedicated instances that run until you stop them.

Where the prices come from

Vast.ai prices are live marketplace offers from its public search API, filtered to verified, rentable hosts; the spot price shown is the current minimum bid on an interruptible offer, which is the price you would pay to take the machine right now. Verda (formerly DataCrunch) prices come from its public API, including its true spot tier and dynamic on-demand pricing. Lyceum publishes flat list prices and has no spot tier, so its rows are always on-demand. All prices exclude storage and egress. We are not affiliated with any of the three providers; refresh happens through our server every 30 seconds, and nothing about you is sent anywhere.

Spot vs on-demand: which should you rent?

Spot capacity is 40 to 80% cheaper and perfectly good for anything that can survive an interruption: fine-tuning with regular checkpoints, batch inference, evaluations, and development. On-demand is for work that must not die mid-run — production endpoints, long uninterruptible training, demos. A common pattern is to develop and train on spot, then serve on the cheapest reliable on-demand instance. If you are estimating how much GPU memory your job needs first, our LLM VRAM calculator and fine-tuning memory calculator answer that.

Why the same GPU costs 10× more elsewhere

The chips are identical; the business models are not. Marketplace hosts compete each other's prices down and sell interruptible capacity for whatever the market bears, specialist GPU clouds price for reliability and support, and hyperscalers price H100s at $6–12/hr for enterprise integration and compliance. Form factor matters too: SXM parts with NVLink command a premium over PCIe cards, and the H200's extra memory (141 GB vs 80 GB) adds roughly a dollar per hour. If a price here looks too good, remember what spot means — a $0.60 H100 is real, but it is a minimum bid on a machine someone can outbid you for.

Bid guidance: how much should you bid?

On marketplaces like Vast.ai, the spot price is a minimum bid: you name a price, you hold the machine while your bid stays above the market, and you lose it when someone outbids you. So the practical question is not "what is the cheapest price" but "what should I bid so the cheapest machine stays mine". The chart answers it from history: the blue line traces the cheapest observed price — never an average, never other offers — drawn from every roughly 30-second market sample on the 1H, 6H, and 24H views. The 7-day view aggregates to hourly points but shades each hour's full low-to-high range, so no price movement is hidden at any zoom. The dashed guides mark the bid levels that would have beaten the cheapest price in 50%, 75%, and 90% of those raw samples; they are computed from the samples, not the drawn line, so zooming never changes the guidance. Bid at the aggressive (50%) level and you win half the time but get interrupted often; bid at the 90% level and you pay a few cents more per hour for a bid that almost always held. The suggested bid above the chart is the 90% level, recomputed live for whatever range and filters you have selected. History is retained for 30 days.

How much does an H100 cost per hour in 2026?

The live board above is the source of truth, but the ranges are fairly stable. These are the single-GPU prices we observe across the three tracked providers as of July 2026, against typical hyperscaler list prices for the same chip:

In short: renting an H100 costs from about $0.60/hr on interruptible spot capacity, $2–4/hr on-demand at specialist GPU clouds, and $6–12/hr at the big three clouds. An H200 adds roughly $0.40–1/hr over the H100 at the same tier for 76% more memory.

Frequently asked questions