Affiliate disclosure: Some links on this page may be affiliate or sponsored links. We may earn a commission at no extra cost to you. This does not change our editorial criteria.

GPU racks beside subtle power meters with soft amber indicator lights in a muted slate hall. — Custom editorial hero · Powered Lands Imager

GPU cloud pricing index

This page is a beginner guide to GPU cloud pricing. It is the living index we will keep as quotes change. A GPU is a graphics processing unit: a chip that is very good at the math in AI, simulation, and rendering. GPU cloud means you rent those chips by the hour, then shut them off.

Who it is for: the person comparing quotes who was shown a "from" rate and does not know what is included. You do not need to be an engineer to follow this page. You do need to know which job you are buying hours for.

There is no honest single dollar-per-hour number here. We will not invent one. List prices move with GPU type, region, and whether you pay only when the machine is on. When a public band is live on this page, it will carry an as-of date and a reminder to re-check the provider the same day. Until that dated band exists, the row stays empty on purpose.

If the same GPUs would stay busy for years, the next problem is usually electricity and a site, not a cheaper hour. That is powered land.

GPU cloud pricing varies. Verify live rates.

A cloud menu is not a sticker on a car. Two lines that both say "H100" can be a whole GPU, a slice, an 8-GPU box, or a plug-in card, and they may not start on the same day. Quoted "from" rates often skip storage, data copies, idle time, and people. Treat every public figure as stale until you re-check the provider. That is the rule this index will follow when bands exist.

How a live band will look on this page

When we publish a band, it will name five things:

The GPU (H100, H200, B200, or A100) and the memory size if we know it.
The shape: one GPU vs several in one server, SXM (dense tray) vs PCIe (plug-in card) if we know it.
How you pay: on-demand (pay while it is on), reserved (you commit to a block of time), or spot (interruptible, can vanish).
A region class if we know it (for example a U.S. region vs "unspecified").
An as-of date.

It will not be a promise. It will tell you to verify the provider the same day. Refresh intent, when bands are live: weekly, or sooner if public menus move. Until a dated band exists, we publish method, not dollars.

Placeholder bands (no dollar figures)

These rows are the index. They are empty of money on purpose.

GPU	What a live band will name	As-of	Public band
NVIDIA A100	40 GB vs 80 GB, whole GPU vs slice, on-demand vs reserved vs spot	DATE TBD	Not published. We will not invent a dollar-per-GPU-hour.
NVIDIA H100	Memory size, SXM vs PCIe, whole GPU vs slice, pay style	DATE TBD	Not published. We will not invent a dollar-per-GPU-hour.
NVIDIA H200	141 GB class, SXM vs PCIe, pay style	DATE TBD	Not published. We will not invent a dollar-per-GPU-hour.
NVIDIA B200	GPU vs 8-GPU node vs rack-scale quote, pay style, start date	DATE TBD	Not published. We will not invent a dollar-per-GPU-hour.

What the chips are, in plain English, lives on the SKU pages: A100, H100, H200, B200. Those pages also refuse invented hourly rates.

How to read a quote

Ask what unit you are buying. Per GPU-hour means one chip for one hour. Per node-hour means the whole server, which may hold eight GPUs.

Ask whether you get a whole GPU. MIG (Multi-Instance GPU) means the provider can slice one GPU into smaller isolated pieces. A slice is not the card you saw on a datasheet.

Ask whether you can shut it off. On-demand is a meter that can hit zero. A reservation keeps billing even if your job died. Spot is cheaper because the machine can disappear. Spot is fine if you save checkpoints (snapshots) often. It is a bad fit for a deadline.

Ask when the hours actually exist. A quote for chips next quarter is not capacity now. Ask what else is metered: storage, egress (the fee for copying data out), extra CPUs, support. The GPU line is often not the whole bill.

What actually drives the bill

GPU generation and memory come first. An H200 hour that finishes on fewer chips can beat a cheaper H100 hour that needs twice as many machines. An H200 hour that sits idle is still wasted. NVIDIA's product pages list memory sizes. Cloud menus that only print "H100" are incomplete.

Form factor is next. NVLink is NVIDIA's fast link between GPUs in the same server. Training code that expects NVLink will crawl on loosely connected cards.

Region matters because the files have to live somewhere. Keep working files next to the chips. Idle time is the silent line. Utilization means how many hours the machines actually run. If you reserve a week and train for two days, you paid for five days of nothing. CUDA is NVIDIA's software layer. A machine without working drivers is not a discount.

How to compare two quotes without a made-up rate

Hold four things equal, then look at everything else.

Same GPU name, same memory size, same whole-GPU vs slice, same count in the server. Same region, or a region you can actually use. Same interruptibility. Same rule for turning the meter off.

Then compare start date, storage, egress, and who you call at 2 a.m. Do not compare a spot slice in one country to a reserved 8-GPU box in another and call that a price winner.

If the honest description is "we would rent the same shape all year," you are no longer shopping for a bargain hour. You are making a cash-flow choice. See rent GPUs vs build a data center.

GPU cloud partners (placeholder)

Partner rows below are placeholders until terms are signed. No live rates. No outbound URLs. When links go live, they will use rel="sponsored".

Best for	GPU types	Notes
Partner TBD - on-demand hours you can shut off	NVIDIA data-center GPUs	Placeholder. Confirm the unit is per GPU-hour or per node-hour. `rel="sponsored"` when live.
Partner TBD - reserved matching training nodes	Multi-GPU NVIDIA (often H100-class)	Placeholder. Ask NVLink, memory, and what you still pay if the run dies. `rel="sponsored"` when live.
Partner TBD - spot experiments you can replay	Interruptible NVIDIA GPUs	Placeholder. Checkpoint first. Not for a customer endpoint. `rel="sponsored"` when live.

This table is not a ranking. It will reflect partner relationships once partners exist. Partner quotes, when they exist, will be labeled. They still will not replace a live check with the provider.

When GPU rental stops working

Rental stops working when utilization is high and predictable, the hardware mix is stable, and the next constraint is megawatts and a site, not another login. A megawatt is a large unit of electrical power. At that point a prettier hourly rate does not fix the problem.

The honest alternative is powered land: a development site with a credible path to large-load electricity. Browse the U.S. data center directory, then read when GPU clusters need powered land. Hybrid is normal. Rent the bursts. Site the work that never turns off.

Find powered land when you need a site. Rent GPUs when you need capacity now.

FAQ

How much does an H100 cost per hour?

We will not invent that number. When a dated band for H100 is live on this index, it will sit in the placeholder table above with an as-of date. Until then, re-check the provider. See what the chip is on the H100 page.

Why do two quotes for the same GPU disagree?

They may not be the same product. Memory size, whole GPU vs slice, SXM vs PCIe, region, spot vs on-demand, and whether the price is per chip or per server all change the number. Hold those equal before you compare.

What is a "from" rate missing?

Often storage, egress, idle reservation time, and the start date. Sometimes it is a slice, or a different memory size than the card you need.

Should I pick the cheapest GPU on the menu?

No. Pick the chip that fits the job. A cheaper A100 that cannot load the model will bill you for failed starts. A newer GPU that sits idle is also a waste.

When is owning cheaper than renting?

When the same shape would be busy most hours for a planning horizon in years, not for one run. Count real hours. This page still will not publish a made-up break-even dollar figure. Use rent vs build for the decision, then re-check live quotes.

Where do I verify a number I saw on this site?

With the provider, the same day. Dated bands here are a reading aid. They are not a contract.

Sources

Need land or power instead? See the map or contact.