GPU Rental Price Explained: What Sets the Price of an H100 Hour

EXPLAINED · Mechanism · Last reviewed: 30 Oct 2026 · Next review: after the October month-end index readings (2 Nov 2026)

A GPU rental price is what a cloud provider charges to use one graphics processing unit, or GPU, for one hour. It is quoted in US dollars per GPU-hour.

Why an investor cares: the hourly rent on a GPU is the top line of every neocloud and the number every GPU lender checks before it lends. When it rises, spare capacity is worth more; when it falls, a GPU bought on borrowed money takes longer to pay back. Our guide to neoclouds covers the companies that sell these hours; this page covers how the price is set.

What is a GPU rental price?

It is the price of one GPU for one hour. AI companies rarely buy single chips to train or run a model; they rent servers full of them, usually eight GPUs to a server, and the bill is worked out per GPU per hour. A cluster of 1,000 H100s rented for a day is 24,000 GPU-hours.

The price depends on four things: which chip, how long the buyer commits for, how many GPUs are joined together, and how fast the network between them is. An H100 rented for one hour on a single server is a different product from an H100 inside a 4,000-GPU cluster signed for three years. Our explainer on GPU as a service covers what the buyer gets for the hourly fee.

How much does it cost to rent an H100 per hour?

About $2.90 to $3.40 at the specialist clouds in late October 2026, depending on which number you read. Ornn, which records rentals actually agreed between buyers and sellers, settled its H100 index at $2.92 on 29 October. That is $70.08 per GPU per day, or about $2,102 a month.

Silicon Data’s neocloud benchmark read $2.90 the same day. The two sites that read prices listed on provider websites came in higher: GPU Economy’s weekly median was $3.39 across 38 providers, and GPU Rental Prices put the typical on-demand H100 at $3.30 for October. The cheapest listed H100 it found in the month was $1.80.

Each index’s own published reading, on the dates shown. Full source list at the end of the page.

Hyperscalers charge far more on their list prices. GPU Economy put the median listed H100 at the four big clouds at $9.90 an hour in the week of 26 October, against $3.23 at specialist GPU clouds. Our arithmetic: that is about three times the specialist rate.

Why do the price indices disagree?

Because a listed price is an asking price and a trade is what someone paid. Listed rates sit on a provider’s website for anyone to click; they are often for small amounts, with no commitment, and they change only when the provider edits the page. A trade records a real deal, which can be larger, longer or done at a discount to the list.

The gap is clearest on the newer chips. On 29 October Ornn’s executed-trade index had the H200 at $5.96 an hour, Silicon Data’s standardised reading had it at $3.48, and GPU Rental Prices’ October list average was $4.73. Three honest measures, three numbers, because each one counts a different slice of the market.

Our read: use executed trades for what the market is paying today, standardised benchmarks for comparing like with like over time, and list prices for what a small buyer would pay on the day. Never compare a trade index with a list index and call the difference a price move.

What sets the price of a GPU hour?

Supply of live GPUs that are not already sold, and demand from buyers who need them now. Silicon Data says H100 rents move with new clusters and provider capacity coming online, the pace at which work moves to newer chips, and the mix of on-demand and reserved deals.

Most GPU hours never reach the daily market. Large clusters are usually signed for one to several years before they are switched on, often on take-or-pay terms where the buyer pays whether it uses the hours or not. Our explainer on take-or-pay GPU contracts covers how those deals work. What is left over, the spare hours, is what gets rented on demand and what the indices mostly track.

So a small change in spare supply can move the price a lot. A new cluster that comes online without a buyer adds spare hours and pushes the rate down. A large buyer that signs up everything a provider has left removes them and pushes the rate up. Our explainer on reserved vs on-demand GPUs covers who carries the idle hour in each kind of deal.

Why are GPU rental prices rising in October 2026?

Ornn’s H100 index rose 7.7% in October to date after an 8.8% fall in September, ending the month near August’s level. The bigger moves were on the newer chips: its H200 and B200 indices each rose about 21% in a month, to $5.96 and $9.83. GPU Rental Prices’ listed averages show the same direction, with the H200 up 10.9%, the B200 up 14.3% and the B300 up 15.0% on September.

Executed-trade prices from Ornn’s Compute Price Index; listed on-demand averages from GPU Rental Prices’ October reading. Years are when each chip first shipped in volume, approximately.

Old chips rose too. The A100, first shipped in 2020, traded at $1.09 an hour on 29 October, up 18.5% in a month on Ornn’s index. When the oldest GPUs in the market get dearer, the usual reading is that buyers are taking whatever is live. Our Wire on Nebius raising on-demand GPU prices covers one provider that moved its list prices up from 1 October.

A longer view is calmer. GPU Economy’s weekly median listed H100 rate fell from $3.92 in the week of 7 September to $3.39 in late October, as more providers listed the chip at lower prices. Over the same weeks the number of providers in its count rose from 18 to 38, and cheaper sellers joining pull a median down. The list price of an older chip can drift lower even while the rate actually traded is rising.

How much does one GPU earn in a year?

Our arithmetic: a GPU-hour price times the hours sold. At $2.92 an hour and every hour of the year sold, one H100 bills about $25,600. At 70% of hours sold, it bills about $17,900. A 1,000-GPU cluster at full use bills about $25.6 million a year.

The sensitivity is simple. Each 10 cents an hour is worth $876 a year per GPU at full use, or $876,000 a year on 1,000 GPUs. That is why the share of hours sold matters as much as the price. Our explainer on GPU utilization covers why an idle hour earns nothing.

The owner then has to cover the chip, the server, the power and the hall over the GPU’s useful life, and pay its lenders. Our guide to GPU depreciation covers how long that life is assumed to be, and our guide to GPU financing covers the loans that draw down as the GPUs go live.

Who gets paid, and where is the stuck step?

The suppliers are paid first: Nvidia for the chip, the server builders, and the firms that build and power the hall. They are paid when the GPUs are bought, before any hour is rented, whatever the index does next.

Next is the owner of spare hours. A neocloud or reseller with live GPUs that are not yet sold captures a rising index straight away, through on-demand sales or short trades. Our guide to who the neoclouds are covers the companies with the most of these hours.

Hours already sold on multiyear contracts are paid last for a price rise. They keep the agreed rate until the contract ends. If the index is higher at renewal, the next contract starts higher; if it is lower, the owner takes the cut then.

The stuck step is live, unsold capacity. A higher rent pays only on GPUs that are installed, powered, networked and not already spoken for, and new halls take a year or more to come online. Our Neoclouds layer tracks how each company turns capacity into billed hours.

What would change the view?

• A wave of new capacity. If large Blackwell clusters come online without buyers already signed, spare hours rise and the daily price falls, starting with the newest chips.
• Big buyers sign up the spare hours. If AI labs keep taking capacity on long contracts, less reaches the daily market and the index can keep rising on thin trade.
• Cheaper ways to run models. If inference work moves to custom chips or gets more efficient per token, demand for rented GPU hours grows more slowly than the halls being built.

What to check next

The October month-end readings come next: Ornn’s last October settle on Friday 30 October at 4 p.m. New York time, which is 4 a.m. on Saturday in Singapore, and GPU Rental Prices’ final October figures after its 31 October reading. Watch whether the H200 and B200 hold their 21% gains or give them back. Our explainer on AI tokens vs GPU hours covers the other way AI capacity is priced, by the token.

Common misconceptions

• People assume there is one GPU price. Actually the same H100 hour was $2.92 as an executed trade, $3.39 as a median list price and $9.90 at a hyperscaler in the same week.
• People assume a rising index lifts every neocloud’s revenue. Actually hours already sold on multiyear contracts keep their rate; only spare and renewing hours reprice.
• People assume older GPUs only get cheaper. Actually the A100 traded 18.5% higher in a month to 29 October, six years after it first shipped.
• People assume the list price is what big buyers pay. Actually large clusters are signed on negotiated multiyear terms that rarely appear on any website.

FAQ

Related: reserved vs on-demand GPUs; take-or-pay GPU contracts; GPU utilization; GPU cloud providers; neocloud vs hyperscaler; Nvidia cloud partners.

Stack doors

Layer: Neoclouds. Also: CoreWeave, Nebius and Lambda. Markets: US.

Get the Wire — short daily notes when GPU rents and neocloud contracts move.

Not investment advice. Do your own research.

References