Loading
InfratoApp / GPU prices
GPU cloud prices
51 GPU models priced across 44 providers, read from each provider's own pricing page within the last 21 days. The lowest on-demand list rate on record is $0.09 per GPU-hour for RTX-3090-24GB at Salad.
Per GPU-hour, read from each provider’s own public pricing page, with the date it was read. Prices over 21 days old are hidden rather than shown stale — so a provider priced this week may have none next week.
The hunt · H200-SXM-141GB, per GPU-hour
Most listed · cheapest against the median
Bar: the cheapest rate as a share of the median; the number after the price is how many providers list it.
51
GPUs priced
44
Providers
302
Readings
2026-10-27
Most recent
Every model we hold a current price for
TERMS
MEMORY
TIER
GPUMEMORYARCHITECTURE7-DAY
7-day compares the cheapest reading we hold on the oldest and newest day in the window · a card read on fewer than two days at least three days apart says first reading rather than a number · a change can be a correction to our own reading as much as a move in the market
Frequently asked
What's the cheapest GPU for LLM inference today?
Per GB of VRAM, the RTX-4090-24GB at $0.12 an hour — $0.0050 per GB. The lowest hourly rate on record is $0.09 for the RTX-3090-24GB, but the cheapest card is not the cheapest answer: a model that does not fit in its memory has to be split across two, and two cheap cards can cost more than one larger one. All 51 priced cards are listed with their VRAM so the comparison can be made on the row that matters to you.
Why don't you show spot prices?
Because a spot rate changes faster than we can honestly re-read it. A spot price shown with a read date from this morning implies it is still available, and it very often is not — so the figure would be accurate at capture and misleading on the page. Every rate here is a posted on-demand price from the provider's own page, read across 44 providers.
Why does the same GPU cost different amounts?
Because the card is a commodity and everything wrapped around it is not. Region, the network and storage included in the rate, support, contract terms, and how recently the seller last looked at a competitor all move the number. The spread between the cheapest and dearest seller of an identical part is the most useful thing on this page, and it is shown per card rather than averaged away.
Are these the prices people actually pay?
No, and the page says so at the top. These are posted rates — what a provider publishes on its own pricing page. Committed and negotiated deals are struck privately, nobody outside the exchanges knows what GPU hours clear at, and a cleared-trade figure is not something we could honestly publish.
What does a region flag mean next to a rate?
That the provider stated where that rate applies. Many publish one price everywhere and say nothing about place, so no flag means they did not say rather than that the rate is unavailable. Where a provider prices the same card in several regions, the rate shown is the cheapest they charge anywhere and the page names which region that is.
Can a provider pay to appear higher in this list?
No. The list has one stated sort at a time and nothing in it is for sale — a sponsor cannot enter the table, reorder a row or change a figure. How the site is paid for is set out under money and independence at /our-data#sec5.
Why this is its own page
WHY THIS IS ITS OWN PAGE · EVERY NUMBER HERE ALSO LIVES ON THE PROVIDER AND LAYER PAGES, SCOPED TO THAT CONTEXT. THIS VIEW EXISTS BECAUSE THE COMPARISON ACROSS PROVIDERS IS THE THING PEOPLE ACTUALLY SEARCH FOR.