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.

Closeup of a generic dual-slot data-center GPU with blower-style cooler. — Custom editorial hero · Powered Lands Imager

Rent NVIDIA A100 GPUs

This page is a beginner guide to the NVIDIA A100, a data-center graphics chip that GPU clouds still rent by the hour. A GPU is a graphics processing unit: a chip that is very good at the same math thousands of times at once. GPU cloud means you rent those chips, then shut them off when the job ends.

Who it is for: teams who have been offered A100s instead of H100s, or who already have A100 code and want extra capacity. You do not need to be an ML engineer to follow this page.

The A100 is Ampere, the NVIDIA generation before Hopper (the family that includes the H100 and H200). It is older, and it is the chip many labs and clouds still have on the floor. You rent A100s when the job fits, the quote is available, and you can turn the meter off. You do not rent A100s to win a "newest GPU" argument. If the same A100s would stay busy for years, the next problem is usually electricity and a site. That is powered land.

What an A100 actually is

NVIDIA's A100 page describes a Tensor Core GPU for AI and HPC (high-performance computing: big scientific jobs). Tensor Cores are special units that do matrix math quickly. TF32 (Tensor Float 32) is a number format those cores use so many training jobs run faster than full 32-bit math without rewriting the model. CUDA is NVIDIA's software layer that lets programs use NVIDIA chips. PyTorch, many Hugging Face models, and a long list of scientific CUDA codes already know A100.

HBM is high bandwidth memory: fast memory on the GPU package. A100 uses HBM2e, an older generation than the HBM3 on H100 and the HBM3e on H200. NVIDIA ships A100 in 40 GB and 80 GB sizes. The 80 GB SXM version (the dense tray) is listed with about 2 TB/s of memory bandwidth. Cloud menus that only say "A100" are incomplete. 40 GB and 80 GB are different products. A model that loads on 80 GB may never start on 40 GB.

NVLink is NVIDIA's fast link between GPUs in the same server. A100 NVLink is the previous generation versus Hopper. SXM trays in 4-GPU or 8-GPU boxes still train together. PCIe cards are the plug-in version; NVIDIA documents an NVLink Bridge for up to two PCIe A100s. MIG (Multi-Instance GPU) can split one A100 into as many as seven isolated slices. A slice is useful for a shared notebook. It is not a whole 80 GB GPU.

You rent A100s for fine-tunes and models that already fit

Fine-tuning means teaching an existing model a smaller job, often with LoRA (a method that trains a small extra set of weights). Many of those jobs fit on one 80 GB A100. That is a classic rental: copy the examples, run for days, save a checkpoint (a snapshot), shut the GPU off. See GPU rental for model fine-tuning.

Inference means using a trained model to answer questions. If the model fits in 40 GB or 80 GB, A100 still serves it. Serving engines need CUDA drivers. If people are waiting, keep the GPU near those people. The checklist is GPU rental for LLM inference.

Full training of a large language model from scratch is a different size of job. It usually wants more memory per GPU and faster GPU-to-GPU links than A100 was built for. Teams still train on A100 clusters. Newer work often moves to H100-class machines. Read rent GPUs for LLM training before you rent a pile of A100s because they were the leftover inventory.

You rent A100s for CUDA science that already runs

A100 spent years as the default scientific GPU in the cloud. GROMACS (a molecular-dynamics program), climate and airflow solvers, and other CUDA HPC codes often have an A100 path already. NVIDIA's A100 datasheet calls out HPC and analytics as first-class jobs, not only chatbots.

You rent A100s for a campaign with a stop date: a parameter sweep, a backlog of simulations, a methods paper. Confirm the solver's GPU build and the CUDA toolkit version. Several matching 80 GB GPUs in one server is a different product from one 40 GB card in a notebook. Physics, climate, and CFD live on GPU cloud for physics, climate, and CFD. Gene and molecule campaigns live on GPU cloud for life sciences.

Rendering can use A100 through CUDA. It is rarely the cheapest render farm. Rent A100s for frames only if that is already the card the pipeline uses.

What to check before you rent

Ask 40 GB or 80 GB first. Then ask SXM or PCIe, NVLink inside the box, and whole GPU vs MIG. Confirm CUDA drivers. A "cheap A100" that is a 40 GB slice will not load the 80 GB job you tested on a workstation.

Spot (interruptible) A100s are common and can vanish. They are fine if you checkpoint. They are a bad fit for a deadline or a customer endpoint. Egress is the fee for copying data out. Re-downloading a training set onto a bargain GPU is how the bill leaves the GPU line.

If the provider is steering you to A100 because H100s are sold out, that can still be the right rental for a fine-tune or a CUDA campaign. It is the wrong rental if the model physically does not fit.

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 - 80 GB A100 fine-tunes	NVIDIA A100 80 GB	Placeholder. Confirm whole GPU vs MIG and CUDA version. `rel="sponsored"` when live.
Partner TBD - A100 CUDA science campaigns	NVIDIA A100 with a CUDA build of the solver	Placeholder. Matching GPUs and toolkit version. `rel="sponsored"` when live.
Partner TBD - leftover capacity and experiments	On-demand NVIDIA A100 40 GB or 80 GB	Placeholder. Memory size on the invoice, not only the name A100. `rel="sponsored"` when live.

This table is not a ranking. It will reflect partner relationships once partners exist.

GPU cloud pricing varies. Verify live rates.

There is no honest single dollar-per-hour number on this page. A100 list prices move with memory size, region, and whether the machine is on-demand, reserved, or interruptible. Quoted "from" rates often skip storage, data transfer, and idle time. 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.

A lower A100 hour is not automatically cheaper than H100. Count the hours the job actually takes, and whether it can start this week. An idle A100 you left on overnight is still a full night of rent.

When GPU rental stops working

Rental stops working when utilization is high and predictable, you are still running Ampere on purpose, and the next constraint is megawatts and a site, not another login. A megawatt is a large unit of electrical power. Some teams will keep A100s as a stable generation they already understand. Others will only keep them as overflow next to Hopper or Blackwell. Either way, a floor that never turns off is a site 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 A100 bursts. Site the cluster that is the product.

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

FAQ

Is A100 still worth renting if H100 exists?

Yes, when the job fits and A100s are the machines you can actually get. Fine-tunes, inference of models that fit in 40 GB or 80 GB, and CUDA science campaigns are still A100 work. Newest-generation training of huge language models usually is not.

What is the difference between A100 40 GB and A100 80 GB?

Memory. The model, the batch, and the working grid have to fit in the GPU's own memory. 80 GB is not a luxury trim. It is often the difference between "the job starts" and "the job dies." Ask which one is on the invoice.

Will my A100 code run on H100?

Often yes, because both use CUDA. Confirm drivers and toolkit. Going the other way is harder: an H100 job sized for 80 GB Hopper features may not fit or may not use A100 Tensor Core paths the same way.

Should I use MIG slices of an A100?

For a shared notebook or several small jobs, yes if the provider isolates memory honestly. For a production train or a model that needs the full 80 GB, rent a whole GPU.

When is a dedicated A100 cluster cheaper than renting?

When the same 40 GB or 80 GB shape would be busy most hours for a planning horizon in years. Count real hours. We do not publish dollar-per-hour rates. Use the pricing index, then re-check the provider.

What if the limiter is power and a site, not another A100?

Then rental is the wrong product. Use the powered-land GPU guide and what powered land is.

Sources

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