Why the same GPU has different prices

The same GPU costs different amounts at different providers because the rate buys more than the card: the host, the network, the terms and the region.

INFRATOAPP · PUBLISHED 27 OCTOBER 2026 · 5 MIN READ · 4 SOURCES · PRICES LAST CHECKED 19H AGO

One card, a range of rates

The same GPU is priced differently by different providers because the hourly rate pays for more than the card. It also pays for the host it sits in, the network around it, the storage attached, the support behind it and the terms of the rental. Across the cards with more than one seller (cards compared: 29), the median ratio of the dearest on-demand rate to the cheapest is 3.7×. Some of that gap is those differences; some is just price.

Our longer analysis of the spread goes through the data. This post is about what a renter should do with it.

The short version: a spread is a reason to read more carefully, not a reason to pick the low end or to avoid it. What the rate covers decides whether the cheap seller fits your job, and only the provider's own page can tell you that.

What the gap looks like on one card

The widest spread on record is on the RTX-A4000-16GB: cheapest $0.10 (TensorDock), dearest $1.04 (Runcrate), a ratio of 10.4×. That is one card, on-demand, per GPU-hour, so the three usual sources of false difference (another card, another commitment, another unit) are already removed.

For the H100 SXM 80GB: cheapest $1.68, dearest $6.16, sellers with a listed rate 13. Every seller and the page each rate was read from is on the H100 SXM page, so you can open the dearest and the cheapest side by side and see what each says it includes.

Do the same for the card you actually need. The spread differs from one card to the next, and a card with few sellers can show a narrow range simply because there is little to compare it with.

What the dearer rate might be buying

Bundled resources: more CPU cores, more system memory or a large fast local disk in the instance. A rate that includes these can be cheaper in practice than a lower rate that charges for them separately.

Network: some machines sit in clusters with fast links between nodes for multi-node training, others are single hosts. For one card running inference, that link is paid for and unused; for many linked nodes, it may be the only option that works.

Operation and terms: a managed fleet with support, a status page and a provider who replaces failed hardware may cost more to run than a host listed on a marketplace. Stock held for buyers who need it today may too. We do not measure any of these, so the gap cannot be divided between them.

What it might not be buying

Some of the spread is plain pricing. Providers set rates for their own reasons, regional rows can lift the top of a range, and a page that has not been updated in a while can sit high while the market moves. A higher rate is not evidence of anything on its own.

Rates older than 21 days are hidden on this site, so a stale figure does not widen the range here, but a provider's own page can still be out of date. Check the read date beside each rate, and if a rate looks wrong, the corrections page says how it gets fixed.

There is also timing. A provider that has just added a card may price it high while stock is short and lower it later, and one with spare machines may price low to fill them. We do not record why a rate is set where it is, so treat this as a question to ask, not an explanation.

Reading a card page from the bottom up

A card page lists every seller with a rate, the page each rate was read from, and the date. Read it from the dearest end as well as the cheapest. Whatever a dearer seller's page says its rate includes is a list of things to look for, or ask about, at the cheaper end.

Look at the number of sellers too (for the H100 SXM 80GB: 13). On a card with few sellers, one seller leaving or changing price moves the whole range. Use the GPU table to see how many sellers each card has before you settle on one.

Check the dates. A spread built from rates read on the same week is a fair picture of the market at that time; one rate a fortnight older than the rest may have changed since.

Using the spread when you rent

Start at the low end and ask what is missing. Read the cheapest provider's page for what is bundled, how billing works, what happens to a failed host and whether the card is the exact one you need. If all of that fits your job, the low rate is the right rate.

Otherwise, move up the list one seller at a time until the terms fit. The rental cost calculator shows every seller's month for your hours, which turns a small hourly gap into the figure you will actually see.

Then revisit it. A seller that was dearest last quarter may have cut its rate, and the cheapest may have raised it. The card page shows the current read and its date, so a ten-minute check before each new rental keeps the choice honest.

What we could not answer

⚑The site holds rates, not what each rate includes, so the gap cannot be split between bundled resources, network, support and plain pricing.
⚑Availability is not measured, so a low rate on a card that is rarely in stock looks the same as one that is.

What this rests on

infratoapp, “Why the same GPU has different prices”, read 27 October 2026. median gap between dearest and cheapest seller of a card: 3.7×. https://www.infratoapp.com/blog/why-the-same-gpu-has-different-prices

Questions people ask about this

Why do GPU prices vary so much by provider?

The hourly rate pays for more than the card. It also covers the host hardware, bundled CPU, memory and disk, the network between machines, support and the provider's terms. Providers also price for their own reasons. The spread on one card mixes all of these and cannot be split between them from rates alone.

Is the cheapest GPU provider a bad choice?

Not necessarily. A low rate may simply leave out things your job does not need. Read the provider's page for what is included, how billing works and what happens when a host fails. When those fit your workload, the low rate is a fair price for it.

What is a GPU price spread?

It is the highest listed on-demand rate for one exact card divided by the lowest, across every provider that lists it. It shows how far apart sellers price the same hardware. It only exists where more than one seller lists a rate for that card.

These render as FAQPage structured data · answers are filled from the same live figures