Where to rent an H100 PCIe at the cheapest price

NVIDIA sells three H100 boards and they are not interchangeable. The SXM module sits on an eight-GPU HGX board with 80 GB of HBM3 at 3.35 TB/s, up to 700 W and 900 GB/s of NVLink through NVSwitch; the NVL is a dual-slot PCIe card with 94 GB of HBM3 at 3.9 TB/s and 350 to 400 W; the PCIe card has 80 GB of HBM2e at 2.0 TB/s, 350 W and an NVLink bridge to one partner card at 600 GB/s³¹⁹. In dense FP8 that is about 1,979 TFLOPS for the SXM, 1,671 for the NVL and 1,513 for the PCIe card, half the sparse figures NVIDIA prints. The PCIe card loses 40 percent of the SXM's bandwidth and half its power budget, which is why measured throughput falls by about a third, and it cannot scale past two cards on NVLink.

That is why it matters when a listing says only 'H100'. On the same Hyperstack page the on-demand row reads 'NVIDIA H100' and the spot row 'NVIDIA H100 PCIe', and a third row called 'NVIDIA H100 NVLink' is a machine of bridged PCIe cards, not the NVL⁴⁷. Koyeb, HyperAI and fal.ai print 'H100' with no variant at all, and Massed Compute's 'H100 (80GB)' is the PCIe card only by elimination^36373819. A buyer who compares a $2.50 unlabelled row with a $3.20 SXM row, or pays an SXM price for a PCIe card, loses about a third of the throughput. Ask the provider, or run nvidia-smi in the first minute: it reports 'NVIDIA H100 PCIe', 'NVIDIA H100 NVL' or, for the SXM, 'NVIDIA H100 80GB HBM3'.

Five kinds of offer sell this card, and they are not the same product even when the price is close. Managed on-demand clouds such as GPUBank, Beam, Civo, Hyperstack, Runpod Secure Cloud and Lambda rent a fixed machine with stated vCPU, RAM and disk at one hourly price you can start and stop at will. Marketplaces such as Vast.ai, Runpod Community Cloud and GPU.ai pass through a host's own price, and the CPU, memory, disk and network around the card change from one listing to the next, with disk and bandwidth usually billed on top of the headline¹¹. Spot or interruptible capacity is the same hardware at a discount that can be taken back.

Reserved capacity is where the terms vary most. GPUBank reserves a single card for a week or a month at $1.08, Hyperstack shows reservations 'starting from' $1.75 without stating the term, and Civo's $1.99 and Oblivus's $2.56 both require 36 months¹⁴³⁷. Longer terms buy less than you might expect: ComputePrices listed Impossible Cloud at $2.09 per GPU-hour on 29 September, but for a whole eight-GPU server held for 12 months⁸. Serverless is the fifth kind: Beam bills the H100 PCIe at $0.000972 per second of GPU time, about $3.50 an hour, with CPU and memory charged separately, which suits spiky traffic and costs too much for a model that serves all day². The comparison table keeps these groups apart, because a marketplace headline at $2.00 and a managed all-in price at $2.00 buy different things.

The captures below were taken on 16 Oct 2026 from each provider's own H100 PCIe or pricing page, so each figure can be read next to the conditions printed beside it.

GPUBank is cheapest where the comparison is like for like: a managed single card with the machine, the disk, the IP address and the traffic included. At $1.35 on demand it undercuts the next managed offer we could read at source, Beam from $1.83, by 26 percent, and Lambda's $3.29 by more than half; its reservation at $1.08 for a week or a month is the cheapest reserved hour we found and the shortest commitment, where Hyperstack starts at $1.75 on an unstated term and Civo asks 36 months for $1.99¹²²⁶⁴³. Per-second billing from the health check, a full credit when a host fails mid-job and a persistent disk at $0.03 per GB-month make the monthly bill close to the rate times the hours¹.

It loses in four places. Vast.ai interruptible bids have been well below its $0.81 interruptible rate, so a fine-tune that checkpoints often and tolerates host churn can cost less there. Its own H100 SXM at $1.85 finishes saturated training and serving jobs about 1.5 times faster, which makes it cheaper per job whenever the speed-up exceeds the price ratio of 1.37²⁷²⁸. It offers this card only in Frankfurt and Ashburn, so teams that need a machine in Canada, London, Paris or Warsaw should look at Hyperstack, Civo or Scaleway. And it takes only crypto top-ups from $25 for now, with no card payments, which some finance departments cannot use¹.

Beyond the six providers above, the rest of the market is a long tail of smaller clouds whose prices we could read only through trackers. ComputePrices showed QuantaCloud at $2.59 in the US Midwest, Omega Gradient at $2.63 in Canada, OVHcloud at $2.99 for its h100-380 flavour, which pairs the card with 30 vCPU, 380 GB of RAM and 3.84 TB of passthrough NVMe and reports itself to nvidia-smi as 'NVIDIA H100 PCIe' in Spare Cores' test¹², and AtmosCompute at $3.01, with Thunder Compute at $3.20 in an older reading dated 5 September, and GetDeploying showed Scaleway's single-card H100-1-80G at $3.21 in stock in Paris and Warsaw, Oblivus at $3.50 out of stock and Enverge at $4.00⁸¹³⁷. Scaleway's own page confirms that both its flavours carry the PCIe card, with 24 vCPU and 240 GB of RAM on the single-GPU one, but did not show a price when we read it¹⁴. io.net's own price guide, dated April 2026, gives $1.49 for an H100 PCIe on its decentralised network, a snapshot we could not refresh¹⁵.

The marketplace readings move by the hour, so their history says more than any single number. GridCompute, which tracks Vast.ai hourly, counted four on-demand H100 PCIe listings with eight GPUs between them on 8 October, with a lowest ask of $1.25 and a median of $2.78, and the median had ranged from $1.63 to $3.52 since tracking began on 2 October¹⁶. On the interruptible side GetDeploying showed a two-card Vast.ai listing in South Africa at $0.67 per GPU on 15 October, and Beam's own H100 article puts Vast.ai's interruptible offers 'around $0.35 per GPU-hour'⁵¹⁷. Lium, whose price includes the node's local disk, was tracked at $1.40 in Almaty on 29 September and at $2.15 in Boardman in the United States, with 26 vCPU and 221 GB of RAM, on 7 October¹⁸. None of these hours comes with a fixed machine or a credit when the host goes away, and on Vast.ai the disk and bandwidth are billed on top of the ask¹¹.

Massed Compute sells an 'H100 (80GB)' at $2.73 an hour with 20 vCPU, 128 GB of RAM and 1,250 GB of storage, next to separate rows for the H100 NVL and the H100 SXM5, so by elimination it is the PCIe card even though the page never says so¹⁹. On the marketplace side, Lium had no H100 PCIe free when we looked, only SXM listings, and OpenRelay was tracked at $2.20²⁰⁸. None of these is a reason to skip the providers above, but QuantaCloud and Massed Compute are worth a look if you need a US machine with a large local disk and the cheaper clouds are sold out.

Storage is the first line that changes the bill. GPUBank charges $0.03 per GB-month for a persistent disk and nothing else while a machine is stopped¹. Runpod charges $0.07 per GB-month for a network volume under 1 TB and $0.20 for the volume disk of an idle Pod²⁹, and Hyperstack's shared volumes work out to about $0.07⁴. Keeping 500 GB of checkpoints for a month therefore costs $15 at GPUBank, about $35 on a Runpod or Hyperstack volume and $100 on an idle Runpod Pod disk, and on a light month that can rival the compute. Hyperstack also bills a public IP at about $0.0067 an hour, $4.91 for a full month, where GPUBank, Beam and Lambda include traffic and GPUBank includes the address⁴²²⁶.

Billing granularity decides what a short job costs. A run of 61 minutes costs $1.37 on GPUBank's per-second meter, $2.54 on Hyperstack and $3.35 on Lambda, both billed by the minute, and $4.98 on Civo, whose billing rounds each service up to the next full hour, so the 61 minutes, worth $2.53 at its rate, are billed as two³⁰. Scaleway also bills by the hour with a one-hour minimum per resource, according to GetDeploying⁷. On a month of evaluations, restarts and quick tests, that rounding can be worth more than the gap between two hourly rates, so a team that launches many short jobs should weigh the meter before the headline price.

The last cost is time. A marketplace host that disappears mid-run, a spot instance reclaimed without warning or a listing slower than its card suggests costs the GPU-hours already spent and the hours to redo them, and on a marketplace the renter pays for both. GPUBank credits a session interrupted by a host failure in full, gives five minutes of notice before reclaiming interruptible capacity and credits the last 30 minutes if the notice is missed¹. Prepaid balances add a smaller risk: GPUBank's minimum top-up is $25 and the credit never expires, and its bonus lowers the effective rate only if the credit is actually spent. Running dry is treated very differently from one cloud to the next: Hyperstack's FAQ warns that resources on a drained balance are scheduled for termination after a short grace period and may be deleted for good after a week⁴.

Every measurement we found puts the H100 SXM about half again as fast as the PCIe card. Lambda measured a 49 to 51 percent premium for the SXM on a geomean of PyTorch FP16 and TF32 training throughput across SSD, BERT, GNMT, ResNet-50, Tacotron 2 and WaveGlow²⁷, and Runpod's vLLM test on Qwen2.5-7B-Instruct at one request per second measured 165 tokens a second on the SXM against 109 on the PCIe card, a ratio of 1.51, with a median time to first token of 43 ms on the PCIe card²⁸. The gap follows the hardware: 3.35 against 2.0 TB/s of memory bandwidth and 700 against 350 W of power to sustain clocks³¹⁹.

Against the A100 the picture is closer. In llama.cpp on Llama 3, the H100 PCIe generated 25.01 tokens a second on 70B Q4_K_M against 24.33 for the A100 SXM, almost the same, but processed prompts much faster, 7,760 against 5,864 tokens a second on 8B Q4_K_M, and 984 against 797 on 70B Q4_K_M, where the L40S managed 649³². On the small 8B model at Q4_K_M it generated 144.49 tokens a second, against 138.31 on the A100 PCIe, 133.38 on the A100 SXM and 113.60 on the L40S. llama.cpp does not use Hopper's FP8 path, so these numbers understate what the H100 gains in vLLM or TensorRT-LLM with FP8 weights. Runpod's vLLM test, which does run on a modern serving stack, put the PCIe card 14 percent ahead of the A100 SXM at 109 against 96 tokens a second²⁸.

Turned into cost at GPUBank's on-demand prices, a million single-stream tokens in Runpod's test costs $3.44 on the H100 PCIe, $3.11 on the H100 SXM, $2.75 on the A100 SXM, $3.18 on the L40S and $3.59 on the RTX PRO 6000, and in the llama.cpp 70B test a million generated tokens costs $14.99 on the H100 PCIe, $10.85 on the A100 SXM and $9.98 on the L40S. These are single-stream figures, and a batched server makes every token far cheaper on all of these cards, but the ranking carries over. A training run that takes 10 hours on the PCIe card ($13.50) takes about 6.7 hours on the SXM ($12.33). The PCIe card earns its hour when the GPU is partly idle, as in notebooks, development sessions, low-traffic endpoints and pipelines held back by CPU or storage, where wall-clock time rather than throughput sets the bill, and when a model needs more than the L40S's 48 GB.

For single-stream decoding of a model that fits in 48 GB, the L40S at $0.55 is cheaper per token than the H100 PCIe even though it is less than half as fast: 48 against 109 tokens a second in Runpod's test, and 15.31 against 25.01 on Llama 3 70B at 4-bit in llama.cpp²⁸³². The A100 SXM at $0.95 decodes at nearly the PCIe card's speed for 30 percent less, which makes it the cheaper choice for plain inference and fine-tunes that do not use FP8. The RTX PRO 6000 at $1.15 brings 96 GB, enough for a 70B model at 8-bit on one card, and CloudRift measured it slightly ahead of an H100 SXM under heavy concurrent load, 3,140 against 2,987 tokens a second serving GLM-4.5-Air at 4-bit with 256 to 512 requests in flight, though it trails the PCIe card at one request per second, 89 against 109³³²⁸.

That leaves the H100 PCIe a specific place: 80 GB of HBM, FP8 and fast prefill on one card, at a lower hour than the SXM, for work that does not keep the GPU saturated. If the job runs flat out for hours, rent the H100 SXM; if it fits in 48 GB, rent the L40S; if it needs more than 80 GB on one card, rent the RTX PRO 6000. On GPUBank all four sit on the same terms, per-second billing with the machine included, so switching cards between jobs costs nothing but a new launch¹.

The realistic purchase is one card in a standard 2U server. GPUDojo's tracker recorded a Newegg listing at $29,500 on 9 October 2026, a 90-day low after $34,989 a week earlier, and a used eBay offer at $27,999.99 in September³⁴, and Caladan Semiconductor recorded three used US asks between $28,000 and $32,399, with a median of $32,000, on 9 September³⁵. We take $29,500 for the card and assume $10,000 for the server, a resale value after two years of 35 percent of the card price and $2,000 for the server, 0.65 kW at the wall under load and $0.15 per kWh. The fixed cost is then $39,500 less $12,325 of resale over 24 months, $1,132.29 a month, plus $0.0975 of power for every hour of load.

Against GPUBank's $1.35 on demand, ownership breaks even only at 1,132.29 divided by 1.2525, about 904 hours a month, and a month has 730. At full use owning costs $1,203.47 a month against $985.50 on demand, $788.40 reserved and $615.94 at the effective rate from a $1,000 top-up, before any colocation fee; a rack slot at $150 a month would lift the fixed cost to $1,282.29 and the break-even against on demand to about 1,024 hours. With a used card at $25,000, an $8,000 server and 50 percent card resale, ownership falls to about $842 a month at full use, below the on-demand month but still above the reserved and top-up prices. Buy only if your data cannot leave your premises, the rack and power are already paid for and the card will be busy every hour for two years.