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 cloud providers

This page is a beginner guide to GPU cloud providers. A provider is the company that rents you the machines. 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: someone shopping for a GPU cloud who needs types and buyer questions, not a ranking of unsigned brands. You do not need to be an engineer to follow this page.

We do not list named partners until terms are signed. There are no outbound partner URLs on this page. There are no invented hourly rates. When partner links go live, they will use rel="sponsored". Until then, learn the kinds of providers, then ask the same questions of whoever sent you a quote.

You rent a provider when you need capacity now. If the same machines would stay busy for years, the next problem is usually electricity and a site. That is powered land. A reserved login is still not a utility interconnection.

Four kinds of GPU cloud (plain English)

Most catalogs mix four products. They are not interchangeable.

Large public clouds (often called hyperscalers) sell many services. A GPU is one SKU among storage, databases, and networking. That breadth is useful if your data already lives there. It is not always the fastest path to a matching 8-GPU training box.

GPU-focused clouds (sometimes called neoclouds) sell data-center GPUs as the main product. The pitch is training nodes, fast links between chips, and people who talk about CUDA all day. CUDA is NVIDIA's software layer that lets programs use NVIDIA chips. These clouds can still run out of a generation. Ask when the reservation actually starts.

Marketplaces let you rent a GPU that may sit in someone else's machine. That can be a cheap way to try a chip for an afternoon. Quality, uptime, and data isolation vary more than on a dedicated node. Fine for a throwaway experiment. A poor home for a restricted dataset or a customer-facing assistant.

Inference platforms sell answers, not GPU hours. You pay for tokens (pieces of text in and out) or for an endpoint that already hosts a model. That is a different product from "give me an H100 I can log into." Do not compare a token price to a GPU-hour unless you know how many hours the endpoint actually burns.

NVIDIA, the chip maker, also documents its own cloud-style offerings around DGX systems. That still does not make every "NVIDIA GPU" listing the same machine. Ask what you are logging into.

What to check before you pick a provider

Start with the job, not the brand. Training (teaching a model from lots of data) wants matching GPUs and fast links. Inference (using a trained model to answer questions) wants memory, a region near the people who wait, and a machine that does not vanish. Fine-tunes and science campaigns have stop dates. See rent GPUs for LLM training if teaching is the job.

Confirm NVIDIA drivers and CUDA. Confirm memory size on the GPU you will actually get. Cloud menus that only say H100 are incomplete. Ask whole GPU vs MIG slice. MIG (Multi-Instance GPU) means one chip can be cut into smaller isolated pieces.

Ask how GPUs in one server talk to each other. NVLink is NVIDIA's fast link inside a box. SXM is the dense tray. PCIe is the plug-in card. An "8x" machine with slow links behaves like a much weaker box.

Ask tenancy. Shared means neighbors on the same host. Dedicated or single-tenant means the box is yours for the window. Restricted data and deadline training usually want the second.

Ask about spot (interruptible) capacity. Spot is cheaper because it can disappear. Save checkpoints (snapshots) if you use it. Do not put a live customer on it.

Ask region, data residency, and egress. Egress is the fee for copying data out. Ask whether you can shut the meter off. Ask who you call when the node dies at 2 a.m.

Ask the start date. A famous GPU next quarter is not capacity now. An older chip you can start tonight may be the better rental. The A100 is still widely rented for jobs that fit. H200 and B200 only help if the provider actually has that memory class or generation.

GPU cloud partners (placeholder)

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

This table is not a ranking. It will reflect partner relationships once partners exist. A paid row will be labeled. It will not become a fake star rating.

This is not a "best GPU cloud" scoreboard

"Best" depends on the job. A marketplace GPU can be right for a two-hour notebook. It is wrong for a private corpus. A large public cloud can be right because the files already live there. It is wrong if you cannot get matching GPUs for the window. A GPU-focused cloud can be right for an 8-GPU train. It is wrong if the waitlist is the real product.

When live partner pages exist, each one will still have to answer the checklist above. Until then, use this index to interrogate whoever is selling.

How quotes work, and why we refuse a single dollar-per-hour number, lives on the GPU cloud pricing index. Bands on that page will be dated when they exist. We will not invent rates here to fill empty partner cells.

GPU cloud pricing varies. Verify live rates.

There is no honest single dollar-per-hour number on this page. List prices move with GPU type, region, form factor, and whether you pay only when the machine is on. Quoted "from" rates often skip storage, data transfer, and idle reservations. See the GPU cloud pricing index for how we talk about rates. We will not invent hourly prices here. Treat every public figure as stale until you re-check the provider.

Compare hours you will actually run. A provider with a pretty menu and a three-month queue is not cheaper. It is later.

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 provider login. A megawatt is a large unit of electrical power. Switching clouds does not create electricity.

A reserved instance is still rental. Colocation (renting space in someone else's building) is closer to owning the hardware, and it still depends on a hall with power and cooling headroom. When that headroom is gone, you are shopping for land and load.

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. For the rent-versus-build fork, see rent GPUs vs build a data center. 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.

Related

• Rent NVIDIA H100 GPUs
• Rent GPUs for LLM training
• When GPU clusters need powered land

FAQ

What is a neocloud?

A GPU-focused cloud: a provider whose main product is data-center GPUs, not a full supermarket of IT services. The word is informal. Ask about chips, links, start date, and whether you can turn the meter off. Do not treat the label as a quality score.

Is a GPU marketplace the same as a dedicated cloud?

No. A marketplace often rents you a GPU in someone else's machine. A dedicated node is a box (or a slice of a box) reserved for you. Isolation, uptime, and support are different products. Match the product to the data and the deadline.

Is the large public cloud always the safe choice?

It is familiar, and it is where a lot of data already lives. It is not automatically the only way to get CUDA GPUs, and it is not automatically faster to a matching training box. Run the checklist. This page does not rank unnamed clouds.

Do you rank GPU cloud providers?

Not today. The table is Partner TBD until terms are signed. When rows go live they will reflect partner relationships, labeled, with rel="sponsored" on outbound links. That is not a five-star review.

How much do providers charge per GPU-hour?

We will not invent that number. Use the pricing index to learn how to read a quote, then re-check the provider. Two catalogs that both say H100 may not be selling the same machine.

When do I need powered land instead of another provider?

When the cluster is the product, the mix of GPUs is stable, and the limiter is electricity and a site. Start with when GPU clusters need powered land and what powered land is.

Sources

1. NVIDIA H100 GPU
2. NVIDIA H200 GPU
3. NVIDIA A100 Tensor Core GPU
4. NVIDIA DGX B200
5. NVIDIA NVLink
6. NVIDIA CUDA Toolkit
7. What is powered land
8. U.S. data center directory
9. Powered Lands data center map

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