How GPU cloud pricing works
Renting the same GPU can cost wildly different amounts depending on who you rent it from and what you are willing to give up. This page normalises those quotes so they can be compared honestly, and this section explains exactly what the numbers mean and how they are produced.
Why everything is quoted per GPU-hour
Providers do not sell the same unit. A marketplace rents you a single card. Most large clouds only rent whole machines — an 8-GPU box billed as one instance. Oracle and Google bill the accelerator separately from the machine it attaches to. Vultr sells slices of video memory rather than cards at all.
The only figure that survives those differences is dollars per GPU-hour: the price of the thing you rent, divided by the number of GPUs inside it. An 8×H100 instance at $32/hour and a single H100 at $4/hour are the same rate. Every number on this page is computed that way, and where a provider sells fractions of a card, only whole-card configurations are included — pricing a 4GB slice as though it were a GPU would not be a comparison, it would be an error.
Operators and marketplaces are different businesses
The tag beside each provider says which kind it is, because the same price does not mean the same thing across them.
- Operator — Owns and runs the hardware, and sets the price itself. If something breaks, the company you paid owns the machine.
- Marketplace — A venue where independent hosts list their own machines. It owns no GPUs, but the price forms here — it is a first-hand quote, not a reseller's markup on someone else's rate. Supply ranges from professional racks to a single consumer card in someone’s flat, which is why the cheapest rates on this page live here.
- Mixed — Sells both its own datacenter capacity and peer-hosted machines, at different prices and different guarantees. The tier shown under each price tells you which one a given rate came from.
What is deliberately absent is the third category: resellers and aggregators that republish another cloud’s rate card. Their prices are second-hand, so every rate here is taken from the venue that actually rents the hardware.
List prices and market prices are not the same data
A large cloud publishes a rate card that changes when the vendor decides to change it. Tracking that over time produces a change-log, not a market. A marketplace price moves continuously with supply and demand, and an auction-cleared spot price (AWS) is a genuine market print. Both appear here, but they are different animals, and a chart that treats a rate card as a market is telling you something untrue.
For marketplaces we publish the median of what is actually listed at snapshot time, with the cheapest live listing beneath it as “from”. The median is the going rate; the minimum is one host’s idle card and moves far too much to be a series. When fewer than three offers are listed the median is still shown, marked with its offer count, but it is left out of the Best price and the price ranges — one or two listings are not a market rate. When fewer than three offers are listed the median is still shown, marked with its offer count, but it is left out of the Best price and the price ranges — one or two listings are not a market rate.
On-demand and spot
On-demand runs until you stop it. Spot (also called preemptible or interruptible) is the same silicon at a discount, with the catch that it can be reclaimed at short notice — on AWS by an auction, elsewhere by the provider simply taking the capacity back. They are genuinely different products, so they get a table each above; only the providers that publish a spot price appear in the spot table, since it is not sold everywhere. The gap between on-demand and spot at one provider is often larger than the gap between providers.
Cheapest is not the same as best. A low headline rate can mean interruptible capacity, a consumer card on a domestic connection, or a region far from your data. The point of these tables is to make those trade-offs visible, not to declare a winner.
How the numbers are produced
- Prices are collected once a day, at a fixed time, directly from each provider’s own public pricing interface.
- Instance rates are divided by the GPU count in that instance to give $/GPU-hour. GPU model names are normalised across providers, so one card means one row no matter what each vendor calls it.
- Each cell shows the cheapest rate that provider offers for that GPU. Where a provider sells the same card in several sizes, tiers or regions, the cheapest is the one shown; the “Best” column is the cheapest across all providers.
- Reservations, committed-use discounts, licence add-ons and non-hourly products are excluded — they are not comparable with an hourly rate.
- History accumulates forward. Almost no provider publishes what a GPU cost last month, only what it costs now, so these series begin when tracking began and lengthen daily. The exception is AWS, whose spot history is available for a rolling window and has been backfilled.