How much does it cost to rent a GPU in the cloud?

The short answer

For a single hourly budget, use three bands from the Nov 14 board. Cards start below a dollar an hour: RTX 4090 at $0.34, RTX A6000 at $0.35, RTX 5090 at $0.69, and L40S or L40 at $0.79. The A100 band starts just above a dollar: A100 80GB at $1.09 and A100 40GB at $1.70. The current data-centre flagships cost several dollars an hour: H200 at $3.59, B200 at $5.98, B300 at $7.40 and GB300 at $14.27. Those are floors, not averages and not quotes for every buying shape. A floor can belong to a community tier, carry a minimum, or describe one variant while a similarly named variant costs more. The rest of this guide is about not mistaking the floor for the bill.

TL;DR — the cheapest verified offer for each card below (checked Nov 14, 2026)

All guide-table prices are USD per GPU per hour and on-demand unless a row says otherwise — from GPUJournal's verified board snapshot, checked Nov 15, 2026.

No L4 or A10 price appears above because the Nov 14 snapshot contains no offer under either name. L40 and RTX A6000 are separate cards and are priced under their own names, not used as substitutes. No spot price appears in the table: every row above is an on-demand offer. The snapshot also carries offers labelled GB200 NVL72, H100 with no variant split, A100 80GB NVLink and H100 NVLink (PCIe); they are not separate rows in this table, and none of them changes a cheapest figure shown above.

Why the same GPU can cost several times more

The spread is not a rounding error. In the Nov 14 snapshot, the range on a single card runs to several times its floor, before the buying shapes are even considered:

Some of that spread is the provider and service tier. Some is the machine wrapped around the GPU. Some is a per-GPU figure divided out of a whole node or VM. The name of the card alone does not make two offers equivalent.

What moves the price — the minimum setup: compare totals, not per-GPU stickers

The most expensive misunderstanding in GPU rental is also the simplest: a per-GPU rate is not the amount leaving your account when the offer is sold as a node, a VM or a multi-GPU minimum. GPUJournal prices viable multi-GPU offers by their minimum setup total wherever the snapshot gives that total. Date note: the TensorDock and CoreWeave rows below were re-read on Nov 14, 2026. The Oracle OCI, Google Cloud and Microsoft Azure rows are snapshot rows carried from the board's Oct 27–28, 2026 verification; they are used here to show buying shape, not as fresh Nov 14 cheapest offers.

TensorDock is the cleanest example. Its H100 SXM sticker is the lowest in the snapshot, and for a buyer who genuinely needs eight GPUs it may be the right comparison. For a buyer who needs one GPU, the meaningful TensorDock number is not $2.25; it is the $18.00 hourly floor. The next H100 SXM rate in the snapshot, RunPod Community's $2.69 per GPU per hour, is the cleaner per-GPU comparison when the eight-GPU commitment is not the shape you intend to buy.

What an hour becomes over time

An hourly rate feels small until the GPU stays on. The table below is simple arithmetic on selected Nov 14 per-GPU rates: 24 hours, one 168-hour week, and a 730-hour always-on month. It excludes storage, data transfer, host charges and every other extra in the next section.

Minimums scale the same way. TensorDock's $18.00 H100 SXM hourly floor becomes $432 over 24 hours, $3,024 over 168 hours and $13,140 over 730 hours — before any extra. That does not make the offer bad. It makes the commitment visible.

The real bill beyond the sticker

What moves the price — the buying basis, the region, the provider type and the date

The buying basis: on-demand, spot or term. Every price in the TL;DR table is on-demand. That matters because on-demand is the flexible reference rate: it is the price shape our snapshot can compare across providers without hiding a duration or interruption condition inside the number. Spot and interruptible capacity are a different bargain. The Nov 14 snapshot contains no offer labelled spot, so this guide prints no spot price and does not borrow one from another page. If you price spot elsewhere, compare it only after adding the cost of checkpointing, restarting and losing unfinished work. A lower interruptible rate is not cheaper if the job has to buy the same hours twice. Term and reserved prices are also not on-demand prices. The snapshot used here contains no verified term or reserved figure, so none is mixed into the table. A term quote belongs in a comparison only with its duration, minimum spend and cancellation or expiry terms printed beside it. Leaving those terms out makes a commitment look like a discount with no cost. Region is part of the price. A regional meter is not a worldwide quote. The snapshot identifies Microsoft Azure's H200 figure as an East US 2 meter: $10.60 per GPU per hour (a carried hyperscaler row verified Oct 27–28, 2026), divided from an $84.80 eight-GPU VM. It does not present that figure as Azure's price in every region, and neither does this guide. Before comparing two offers, put the region in the row. A cheaper card in the wrong place can also create a second cost when data has to move to it or results have to move back. The snapshot does not give us a verified cross-region price table, so this guide does not invent one. Provider type changes the shape of the offer. The snapshot's provider labels describe different buying shapes. Hyperscalers — Oracle OCI, Google Cloud and Microsoft Azure in this snapshot — often price the GPU as part of a node or VM; Google Cloud's A100 80GB figure goes further and prices an attachable GPU while billing the host VM separately. Specialist GPU clouds — including CoreWeave, Nebius, Hyperstack, Verda, Thunder Compute and TensorDock — include both per-GPU offers and whole-node divisions, so the provider category alone still does not settle the comparison. RunPod Community supplies the cheapest verified offer for many cards in the TL;DR table, while RunPod Secure prices the same card family as a separate tier; on Nov 14, for example, H200 is $3.59 at RunPod Community and $5.29 at RunPod Secure. The lesson is not that one provider type is always cheaper. It is that the words around the number — Community, Secure, node, VM, attachable GPU, minimum — are part of the price. The date is part of the price. GPU prices are moving while you read. In the Nov 14 check, Verda rose for the 16th consecutive time, with all nine tracked cards up about 1% from Nov 13. Its GB300 has risen from $11.82 on Oct 19 to $14.27 on Nov 14 — 20.7% — in our dated record. That is why this guide prints the check date in every TL;DR row instead of presenting November's floor as a permanent market fact. It is also why stale comparison tables mislead. Vast.ai published a guide with this question as its title on Nov 13; our daily research note records that its competitor rows are stamped “checked July 20, 2026.” This guide does not reuse those rows. It answers from GPUJournal's Nov 14, 2026 snapshot.