NVIDIA A100 80GB cloud prices

The cheapest A100 SXM 80 GB hour on 16 October 2026 was an interruptible bid of about $0.53 on a Vast.ai host in Iowa, and the cheapest on-demand listing was the same host at $0.937 per GPU-hour 2. Those are marketplace prices: storage, at $0.333 per GB-month on that host, and bandwidth are billed separately, and the machine and its reliability depend on the host. The cheapest price for a whole managed machine was GPUBank's $0.95, with 16 vCPU, 125 GB of RAM, a 1 TB NVMe disk, an IPv4 address and traffic included, followed by CloudRift at $1.05 and Thunder Compute at $1.09 1 5 6. GPUBank's interruptible tier costs $0.57 and its top-up bonus brings the effective on-demand hour down to $0.59. Many cheaper-looking rows on aggregators are 40 GB or PCIe cards, so check the variant first.

An A100 SXM 80 GB cost between about $0.94 and $5.07 per GPU-hour on demand in October 2026, depending on the kind of offer. Marketplace hosts start around $0.94, managed single-GPU clouds charge $0.95 to $2.00 (GPUBank $0.95, CloudRift $1.05, Thunder Compute $1.09, Beam $1.36, Massed Compute $1.38, Jarvislabs $1.49, Hyperstack $1.60, Runpod $1.79, Verda $1.996), and eight-GPU instances cost more per card, with Lambda at $2.79 2 1 5 11 12 14. The large clouds are highest: AWS $3.43, Oracle $4.00, Azure $4.10 and Google Cloud $5.07 for its single-GPU shape 19 25 23 24. GetDeploying's median across all A100 variants and 39 providers was $1.88 on 15 October 26.

The SXM4 module mounts on an HGX baseboard, runs at up to 400 W, has 2,039 GB/s of memory bandwidth in its 80 GB version and talks to the other cards on the board through NVLink at 600 GB/s, which matters most for multi-GPU training 27. The 80 GB PCIe card fits an ordinary server slot, has a lower power limit, slightly less bandwidth and usually no full NVLink mesh, and for a single-GPU job that fits in memory it performs close to the SXM. The 40 GB versions have half the memory and less bandwidth and are a third product. Many listings say only "A100 80GB": Thunder Compute, Jarvislabs, Koyeb and Modal do not state the form factor on their pricing pages, and Lambda's single-GPU A100 is a 40 GB card 6 9 17 18 14. Compare SXM with SXM.

Yes, for jobs that need 80 GB of fast memory but gain little from FP8 or raw compute. QLoRA and LoRA fine-tuning, BF16 batch inference on older stacks, long-sequence serving, notebooks left open for hours and MIG partitioning are the strongest cases, and at $0.95 for a whole machine the A100 is the cheapest 80 GB card we found per hour 1. It is weaker for compute-heavy work: in MLPerf Training v2.1 eight H100s trained BERT 2.65 times faster than eight A100s, and in E2E Networks' Llama 3.1 8B test the H100 was 2.62 times faster on short sequences, which makes it cheaper per job at GPUBank's $1.85 39 42. On ResNet-50 and long sequences the ratio was close to 2 and the two cards tied per job. Run your own job on both for half an hour before committing.

On GPUBank the H100 SXM is cheaper per job whenever it runs more than 1.95 times faster than the A100 SXM, because that is the ratio of their hourly prices, $1.85 against $0.95 1. On measured results that is true for BERT training, $1.57 per run on eight H100s against $2.14 on eight A100s, and for short-sequence LLM serving, $0.032 against $0.044 per million tokens 39 42. It is not true for ResNet-50 training, $3.62 against $3.63, or for long-sequence serving, $0.069 against $0.070 per million tokens, where the cards tie. The H100 gains most where it can use FP8 and at high concurrency; the A100 holds its own on memory-bound work and wins whenever the GPU sits idle part of the time, because idle hours are billed at the hourly rate on both cards.

The cheapest interruptible A100 SXM 80 GB we found on 16 October 2026 was a Vast.ai minimum bid of about $0.53 per GPU-hour on an Iowa host, where a higher bid can take the machine at any time 2. GPUBank's interruptible tier costs $0.57 with five minutes of notice through the metadata endpoint, a webhook and the console, a checkpoint hook, and a stop rather than a deletion, and a missed notice credits the last 30 minutes 1. TensorDock advertises spot from $0.67 on a mix of SXM and PCIe hosts 16. On the large clouds, Azure's eight-GPU ND96amsr A100 v4 spot price is $8.44 an hour in East US, about $1.06 per GPU, AWS's p4de spot is about $2.59 per GPU on Vantage, and Google's single-GPU a2-ultragpu-1g spot is $3.04 23 19 24.

AWS sells the A100 80 GB only as the p4de.24xlarge instance, with eight GPUs, 96 vCPU and 1,152 GiB of RAM, at $27.447 an hour on demand in Vantage's default view, which is $3.43 per GPU 19 20. AWS's own page lists the instance's specifications but not its current price, and the figure reflects the price reduction AWS announced for its GPU instances in June 2025; in 2022 the same page showed $40.96 an hour 21 22. The older p4d.24xlarge costs $21.958 an hour but carries A100 40 GB cards and should not be compared with 80 GB offers. Vantage shows p4de spot at $20.736 an hour and a three-year reservation at $11.718, about $1.46 per GPU. You cannot rent a single A100 on AWS, and data transfer out and block storage are billed on top.

You can rent a single A100 SXM 80 GB from many specialised clouds, but several well-known providers sell only eight at a time. GPUBank (machines of one to eight cards), CloudRift, Massed Compute, Beam, Hyperstack, Verda, Runpod and the Vast.ai marketplace all rent one card 1 5 8 7 10 12 11 2. Lambda sells its A100 SXM 80 GB only as an eight-GPU instance at $2.79 per GPU, $22.32 an hour, and its single-GPU A100 is the 40 GB version 14. AWS, Azure's ND series and Oracle also sell eight-GPU instances only, while Google Cloud's a2-ultragpu-1g rents a single A100 80 GB at $5.07 an hour 24. When a per-GPU price looks low, check the minimum unit and multiply by it before comparing.

One A100 SXM running a full 730-hour month costs $693.50 on GPUBank on demand, $554.80 on a GPUBank monthly reservation and $416.10 on its interruptible tier, before any persistent disk at $0.03 per GB-month 1. With the 60 percent bonus on 1,000 dollar top-ups, the on-demand month costs about $433 of cash. CloudRift costs $766.50 on demand and $649.70 on its three-month rate, Thunder Compute $795.70 and Runpod Secure Cloud $1,306.70 5 6 11. The cheapest Vast.ai host would cost $684.01 of GPU time plus storage at its own rate, about $33.30 for 100 GB 2. On AWS you must rent eight GPUs, and a p4de instance running all month costs about $20,036 19.

Most keep charging for storage while an instance is stopped, and the rates differ widely. On GPUBank a stopped machine costs only its persistent disk at $0.03 per GB-month, $6.00 a month for 200 GB 1. Runpod charges $0.20 per GB-month for an idle volume disk, twice its running rate, so the same 200 GB left idle for a month costs $40.00, or $14.00 on its network storage 11. Vast.ai bills storage for as long as the instance exists, running or stopped, at a rate each host sets, and a stopped Vast instance does not keep its claim on the GPU, so restarting can mean waiting or moving hosts 4. On the large clouds, attached block storage and reserved public addresses keep billing while the VM is deallocated, so delete what you do not need before leaving a job overnight.

It depends on the provider, and this is one of the largest real differences between a marketplace and a managed cloud. On GPUBank a launch that fails its health check is not billed, and if a host fails mid-job the machine is restarted on a healthy host with the same disk and the interrupted session is credited in full 1. On a peer-to-peer marketplace a host that disappears ends your session, whatever was not checkpointed to storage you control is lost, and whether any time is refunded depends on the platform's rules, which is why Vast.ai publishes a reliability score for each listing 2. Spot capacity on any cloud carries the same risk by design in exchange for its lower price. Whatever you rent, write checkpoints to a persistent disk or external storage often enough that a failure costs minutes rather than hours.

Rent, unless you already run a data-centre rack and can keep eight cards busy most of every day for two years. AmCompute valued a used HGX A100 node with eight 80 GB cards at $95,200 in October 2026, and used module prices fell about 35 percent over the year to August 32 31. Assuming that decline continues, the server is worth about $40,000 after two years, so depreciation costs $2,300 a month, and power at 5 kW and 18.19 cents per kWh adds $0.91 per machine-hour 44. That breaks even with GPUBank's $0.95 on-demand rate at about 344 hours per GPU per month, with its $0.76 reserved rate at 445 hours and with its $0.59 effective rate at 599 hours, before rack space, cooling, networking, spares and staff time, which all push the break-even higher.

A used A100 SXM4 80 GB module sold for about $4,700 to $5,400 over the twelve months to August 2026 according to AmCompute's appraisal of realised sales, down about 35 percent on the year before 31. Asking prices are higher and scattered: Caladan Semi recorded US asks of $8,300 to $12,200 with a median of $9,925 on 9 September 2026, from only four listings, and Jarvislabs' guide gives $5,000 to $9,000 used and $10,000 to $15,000 new 33 34. The 80 GB PCIe card asks more, a median of $18,500 in Caladan's data, because it fits ordinary servers, whereas an SXM module needs an HGX baseboard. A whole used eight-GPU HGX A100 node was valued at $95,200 by AmCompute and $80,000 to $120,000 for a used DGX A100 by Jarvislabs 32 34.

Yes, on GPUBank crypto is the only way to pay for now: you top up a prepaid US dollar balance with Bitcoin, Ethereum, USDT, USDC, Monero or more than a dozen other coins, from $25 to $5,000 at a time, and cards are not accepted 1. The balance never expires, and each top-up carries a bonus of 20 percent from $25, 30 percent from $50, 35 percent from $100, 40 percent from $250, 50 percent from $500 and 60 percent from $1,000, which brings the A100 SXM's $0.95 on-demand hour to about $0.79, $0.70 and $0.59 at the $25, $100 and $1,000 levels. Every account and every team member can have a monthly spend cap. The bonus is worth having only on credit you will actually spend, because the money is committed up front. A team that must pay by card should look at CloudRift, Thunder Compute or Runpod, the nearest managed offers on price 5 6 11. GPUBank is run by GPU Compute OÜ in Tallinn, Estonia, registry code 17984370.