Where to rent an A100 SXM at the cheapest price

Cheap per hour is not the same as cheap per job, and the measured ratios decide which card to rent. On GPUBank the H100 SXM costs 1.95 times the A100's rate, so the A100 stays the sensible choice wherever the H100 is less than about twice as fast: long-sequence and memory-bound inference, ResNet-style training, BF16 stacks that cannot use FP8, QLoRA runs that spend real time loading data and writing checkpoints, notebooks that sit idle, and any session billed while you think. The H100 is cheaper per job on short, compute-heavy work and on anything that can run FP8, as the BERT result shows, and its advantage grows with concurrency. The practical test costs little: run your own job for half an hour on each card at per-second rates, divide the price by the throughput you measured, and keep the cheaper card.

A managed cloud sells a fixed machine, the GPU with its vCPUs, memory and local disk, at one hourly price that may or may not include the public IP, the persistent storage and the traffic. For the A100 SXM that group runs from GPUBank's $0.95 through CloudRift at $1.05, Thunder Compute at $1.09, Beam at $1.36, Massed Compute at $1.38, Jarvislabs at $1.49 and Hyperstack at $1.60 to Runpod Secure Cloud at $1.79 and Verda at $1.996^{12316171218519}. A peer-to-peer marketplace lets independent hosts post their own GPU price and bills storage and bandwidth as separate lines set by each host, which is why Vast.ai's $0.937 and a managed $0.95 are not two prices for the same product even though they sit a cent apart⁴.

The other three kinds of offer trade something away for the price. Spot or interruptible capacity can be taken back: on Vast.ai you bid and can be outbid, GPUBank gives five minutes of notice and stops rather than deletes the machine, and on Azure the eight-GPU ND96amsr A100 v4 spot instance is evicted when the capacity is wanted elsewhere⁴¹²⁰. Reserved capacity is paid for whether you use it or not, and for this card it ranges from GPUBank's $0.76 for a week or a month and CloudRift's $0.89 for three months to Hyperstack's $1.36¹²¹⁸. Serverless offers bill only the seconds in which a request or job runs, at an hourly equivalent of $1.60 to $2.72, which pays off only when the GPU would otherwise sit idle most of the time¹³¹⁴⁵.

Two traps recur on this card. The first is the minimum unit: Lambda, Crusoe and every large cloud except Google sell the 80 GB SXM only as an eight-GPU instance, so a per-GPU figure of $2.79 is in practice a bill of $22.32 an hour⁶²¹. The second is the static price page: Vast.ai's A100 SXM4 page showed "No current offers" when we fetched it, while its live offers API listed eight rentable 80 GB machines at the same moment, and Lium showed a reference price of $0.70 with no pod available, so a marketplace figure is only real at the moment you can actually launch on it²²⁴²³.

The captures below were taken on 16 Oct 2026 from each provider's own A100 or pricing page, so every figure can be read next to the conditions printed beside it.

GPUBank rents the A100 SXM at $0.95 per GPU-hour on demand, $0.000264 a second or $693.50 for a 730-hour month, with 16 vCPU, 125 GB of RAM and a 1 TB local NVMe disk per GPU, in machines of one to eight cards joined by NVLink at 600 GB/s¹. The price includes the vCPU, the RAM, the local NVMe, one public IPv4 address and all inbound and outbound traffic, and the only other line a bill can carry is a persistent disk at $0.03 per GB-month. Billing is per second with no minimum and starts only once the machine has passed its health check, a failed launch is not billed, a stopped machine costs only its persistent disk, and a host that fails during a job restarts the machine on a healthy host with the same disk and credits the interrupted session in full. The interruptible tier costs $0.57 with five minutes of notice, and a week or a month reserved costs $0.76, prepaid and cancellable before it starts¹.

That makes GPUBank the cheapest A100 SXM we found that comes as a whole machine, a cent above the cheapest marketplace listing, and its effective hour falls to $0.79 with a 25 dollar top-up and $0.59 with a 1,000 dollar top-up thanks to a balance bonus on credit that never expires¹. It loses in three places. A Vast.ai interruptible bid of about $0.53 undercuts its $0.57 spot tier for a buyer who accepts being outbid without notice⁴. For throughput-bound jobs its own H100 SXM at $1.85 is often cheaper per finished job, as the measurements further down show. And GPUBank takes only prepaid top-ups in crypto, from 25 to 5,000 dollars, with cards not accepted for now, so a team that must pay by card should look at CloudRift, Thunder Compute or Runpod instead¹.

Beyond the six offers shown as cards, several managed clouds sell a single A100 80 GB at prices between $1.36 and $2.00. Beam lists an A100 80 GB SXM4 machine from $1.36 an hour with 16 vCPU, 128 GB of RAM and 1 TB of NVMe, one of the most generous machines in this price range¹⁶. Jarvislabs charges $1.49 with 16 vCPU and 112 GB of RAM, billed per minute, without saying SXM or PCIe on its pricing page, and GetDeploying showed Geodd's A100 SXM with 16 vCPU and 120 GB and Sesterce's unlabelled A100 80 GB, both at $1.49 and in stock on 15 October⁹; and Hyperstack lists its A100 SXM at $1.60 on demand and $1.36 reserved, with up to 24 pCPU and 120 GB of RAM, next to a PCIe A100 at $1.35¹²¹⁸. Verda, formerly DataCrunch, now shows $1.996 per GPU-hour for its A100 SXM4 80 GB in one, two, four and eight-GPU sizes, up from the $1.79 we recorded at the end of September, with 22 CPUs and 120 GB of RAM listed against the card¹⁹. Above that range, and only by the eight-GPU instance, Crusoe lists an A100 80 GB SXM at $2.30 per GPU-hour next to a PCIe version at $2.00, and Lambda sells its A100 SXM 80 GB as an eight-GPU instance at $2.79 per GPU with 240 vCPU, 1,800 GiB of RAM and 19.5 TiB of SSD²¹⁶.

Marketplace floors also say little about a reliable host: Thunder Compute's review of 9 October priced three Vast.ai A100 80 GB hosts at a median of $1.63 per GPU-hour, between $1.21 and $1.76, with 100 GB of storage counted and bandwidth on top, against the $0.94 floor we found a week later; the same review puts Paperspace's A100 80 GB at $3.18 and CoreWeave's eight-GPU node at $2.70 per GPU²⁴. On the marketplace and serverless side, TensorDock advertises A100 80 GB hosts from $1.42 to $1.63 an hour depending on which line of its page you read, and spot from $0.67, on a mix of SXM hosts in Prague and Oregon and PCIe hosts in Devon, so the variant depends on the listing²⁵. Lium had no A100 SXM pod available when we checked and showed only a reference price of $0.70, and Runpod's Community Cloud figure of $1.39 appears in its page's structured data but not in the visible table²³⁵. For per-second work, Koyeb lists an A100 80 GB at $1.60 an hour and an A100 SXM at $2.15, both with 15 vCPU and 180 GB of RAM, Modal charges $0.000694 a second, or $2.50 an hour, with CPU and memory billed on top, and Runpod Serverless charges $2.72 an hour for an A100 80 GB worker¹³¹⁴⁵. On Runpod's own prices a serverless worker beats a Secure Cloud pod left running only while it is busy less than about 66 percent of the time, $1.79 divided by $2.72.

Storage is the line that bills while nothing runs. A 200 GB persistent disk costs $6.00 a month on GPUBank at $0.03 per GB-month and $14.00 on Runpod's network storage at $0.07, while a Runpod volume disk costs $0.10 per GB-month while the pod runs and $0.20 while it sits idle¹⁵. Thunder Compute includes 100 GB of disk and charges $0.219 per GB-month beyond it, so its 200 GB would add about $21.90³. On Vast.ai each host sets its own storage rate and Vast bills storage for as long as the instance exists, running or stopped; the cheapest 80 GB host on the day charged about $0.33 per GB-month, so the same 200 GB kept all month would cost about $66.60, eleven times the GPUBank disk, and its bandwidth was priced at $0.004 per GB⁴²⁹. A stopped marketplace instance also keeps its data but not its claim on the GPU, so if another renter takes the card you either wait or move your data to another host.

Billing granularity matters most for short jobs. A job of 61 minutes on a per-second meter at $0.95 costs $0.97, while a provider that bills each started hour charges two hours, $1.90 at the same rate; Thunder Compute and Jarvislabs bill per minute, CloudRift, Runpod and GPUBank per second³¹²²⁵¹. On an eight-GPU-only cloud the unit itself is the cost: a quick test on Lambda's A100 80 GB instance bills $22.32 an hour for all eight cards even if the code uses one⁶. Prepaid models add their own terms: GPUBank's balance never expires and its top-up bonus lowers the effective hour, but the money is committed up front, so the bonus is worth having only for credit you will actually spend¹.

The last cost is lost time, and the question is who pays for it. On GPUBank a launch that fails its health check is not billed and a host failure during a job is credited for the whole interrupted session¹. On a marketplace a host that disappears ends your session and the work since the last checkpoint is yours to lose, which is why Vast.ai publishes a reliability score for every listing and why the best-scored hosts are rarely the cheapest⁴. A reclaimed spot instance costs whatever was done since the last checkpoint: with GPUBank's five-minute notice and checkpoint hook that can be close to nothing, while an outbid marketplace instance with no checkpointing can cost the whole run.

The A100 SXM is the sensible rental when a job needs 80 GB of fast memory but not FP8, when it is memory-bound or spends much of its time idle rather than saturating the tensor cores, or when the software stack predates Hopper: QLoRA and LoRA fine-tuning of 7B to 70B models, BF16 batch inference, notebooks left open for hours and MIG slices for small services. It is the wrong tool for FP8 serving at high concurrency, for pretraining at scale and for any long throughput-bound job where the H100's speed repays its higher rate. The measured gap depends heavily on the workload. In MLPerf Training v2.1, eight A100 SXM 80 GB took 16.88 minutes to train BERT-large against 6.38 minutes for eight H100 SXM, a ratio of 2.65, while ResNet-50 took 28.69 against 14.68 minutes, a ratio of only 1.95³⁰³¹. The H100's BERT entry used FP8 through the Transformer Engine, which is part of why the gap is wider there³².

Inference shows the same pattern. E2E Networks' TensorRT-LLM offline benchmark of Llama 3.1 8B on one GPU gives 6,051.62 tokens per second on the A100 against 15,861.63 on the H100 with 128 input and 128 output tokens, a ratio of 2.62, but only 3,747.46 against 7,425.07 with 1,000 input and 1,000 output tokens, a ratio of 1.98³³. That page does not state the precision, the software version or the A100's form factor, and it shows a third, unlabelled H100 column at 27,006.16 tokens per second that may be an FP8 run, so we use the like-for-like column and treat the figures as indicative. The results therefore agree on direction and disagree on size: the H100 is about twice as fast on bandwidth-limited or long-sequence work and two and a half to nearly three times as fast on short, compute-heavy work, and more again when it runs FP8 and the A100 cannot.

On GPUBank the H100 SXM costs $1.85 an hour against $0.95 for the A100 SXM, 1.95 times as much, so the H100 is cheaper per job whenever it runs more than 1.95 times faster. On the measurements above that threshold is crossed on BERT, where a training run costs $2.14 on eight A100s and $1.57 on eight H100s, and on short-sequence LLM serving, where a million tokens cost $0.044 on the A100 and $0.032 on the H100³⁰³³¹. It is not crossed on ResNet-50, at $3.63 against $3.62, or on long-sequence serving, at $0.070 against $0.069 per million tokens, where the two cards tie. The top-up bonus and the interruptible tier apply the same factor to every card, so neither changes this ranking, and the result holds as long as the job keeps the GPU busy; idle time is billed at the hourly rate on both cards, which favours the cheaper one.

We found no measurement from a named source that ran the H100 PCIe, the L40S or the RTX PRO 6000 against the A100 SXM on the same test, so we do not compute a cost per job for them. From the specifications alone, the L40S at $0.55 has 48 GB of GDDR6 at 864 GB/s, less than half the A100's bandwidth, so it is unlikely to win on memory-bound decoding, but it has FP8 and costs 58 percent of the A100's rate for models that fit in 48 GB. The RTX PRO 6000 at $1.15 has 96 GB at 1.8 TB/s, more memory for 1.21 times the price, and the H100 PCIe at $1.35 has the same 80 GB of HBM2e at 2.0 TB/s with Hopper's FP8 but no NVLink between cards¹.