NVIDIA H200 cloud prices

On 16 October 2026 a single H200 SXM cost from $2.09 to about $5.44 per GPU-hour on demand at the specialised clouds that publish a price, and $7.91 to $10.60 per GPU on the large clouds, which sell only eight-GPU instances 6 16 23 21. GetDeploying put the median on-demand price at $4.94 across 34 providers 20. The cheapest managed offers on the providers' own pages were Beam at $2.09, with stock not reported, GPUBank at $2.49 with the IP and all traffic included, and GMI Cloud from $3.20 10 1. Marketplaces started at $2.80 on Vast.ai before storage and bandwidth, and spot capacity went lower, from Nebius's dynamic floor of $0.79 to fixed prices of $1.49 at GPUBank and $2.29 at Lyceum 4 3 15. Before comparing two figures, check what the hour includes and whether it comes only as a whole server.

It depends on what you are willing to give up. If the job can stop at any moment and tolerate a moving price, Nebius preemptible from $0.79 per GPU-hour is the lowest figure on any provider's page, with the price allowed to rise to $5.39 3. If you want a fixed interruptible price with notice, GPUBank charges $1.49 with five minutes of warning and the disk kept 1. For uninterrupted single-GPU use, Beam lists $2.09 for a machine whose stock was not confirmed, and GPUBank $2.49 with traffic and IPv4 included, about $1.56 effective when paid from a $1,000 top-up 6 1. For a month of dedicated capacity, GPUBank's reservation at $1.99 was the lowest fixed reserved price we could read for so short a term. Vast.ai hosts start at $2.80 but add storage and bandwidth 4.

The lowest spot figure on a provider's own page was Nebius's "from $0.79" per GPU-hour, listed since 8 October 2026, when its preemptible tier moved from a fixed $2.45 to dynamic pricing 3. Nebius says the price follows capacity and demand, can change as often as every 15 minutes during use and is capped one cent below the $5.40 on-demand rate, so $0.79 is a floor rather than a quote. Fixed spot prices on 16 October were $1.49 at GPUBank, with five minutes of notice and the machine stopped rather than deleted, $2.29 at Lyceum, $2.58 per GPU for a CoreWeave eight-GPU node in Europe and $2.72 at Verda 1 15 17 16. Google Cloud's A3 Ultra Spot costs $6.35 per GPU, sold only by the node 21. For checkpointed jobs, a fixed price with notice is easier to budget than a floor that can rise several times over mid-run.

Yes, at the specialised clouds and marketplaces, but not at the large clouds. AWS, Google Cloud, Azure and Oracle rent the H200 only as eight-GPU instances, so the smallest bill there is $37 to $85 an hour for the node depending on the pricing model 21 23 24. GPUBank, Beam, Civo, Hyperstack, DigitalOcean, Lyceum, Jarvislabs, Runpod, Nebius, Verda and Vast.ai all listed one-GPU H200 configurations in October 2026, from $2.09 to $5.44 on demand 1 6 12 7 16. Some low headline prices, such as Sesterce at $2.48 per GPU, come only as an eight-GPU server, which costs about $20 an hour to start 20. A single card suits serving one 70B-class model, fine-tuning and experiments; a full node is needed when a model exceeds 141 GB or a job trains across GPUs.

Both have 141 GB of HBM3e at 4.8 TB/s, but they are different products 26. The SXM version mounts on an HGX baseboard, runs at up to 700 W and connects eight GPUs through NVLink and NVSwitch at 900 GB/s per GPU. The NVL version is a dual-slot, air-cooled PCIe card with a 600 W ceiling that links two or four cards through NVLink bridges, and its FP8 rating is 3,341 teraFLOPS with sparsity against 3,958 for the SXM. Marketplaces often list both under "H200": on 16 October Vast.ai's NVL listings started at $3.80 next to SXM listings from $2.80, and Lium's NVL page started at $3.85 4 8. For a single card running one model the gap is modest, but for a model split across GPUs the SXM board's full fabric matters, so compare like with like.

The H200 is worth it when memory is your constraint, and usually not otherwise. It has the H100's compute but 76 percent more memory and 43 percent more bandwidth 26 27. In MLPerf v4.1, with both cards at 700 W, eight H200s produced 13 percent more Mixtral 8x7B tokens than eight H100s, while GPUBank charges 35 percent more for the H200, $2.49 against $1.85 29 1. On a model that fits both cards, that makes the H100 about 16 percent cheaper per token. The H200 pays for itself when its 141 GB removes a card: Llama 3.3 70B in FP8 on one H200 instead of two H100s, Llama 3.1 405B on four H200s instead of eight H100s, or DeepSeek V3 on one node instead of two. It also holds about twelve times more KV cache for a 70B model, which matters for long contexts and many users.

Yes, in FP8, with plenty of room for the cache. The FP8 checkpoint takes about 73 GB, which leaves roughly 63 GiB of an H200's 141 GB for KV cache once a working reserve is set aside, or about 410,000 tokens with an FP8 cache by our calculation 32. On an 80 GB H100 the same weights leave only about 5 GiB, around 32,500 tokens, less than one full 128K-token conversation, so in practice the model is served on two H100s with tensor parallelism. In BF16 the weights alone take about 141 GB, so one H200 cannot hold them and two are needed. One H200 at $2.49 an hour against two H100s at $3.70 is the clearest case on this page where the more expensive card is the cheaper deployment 1.

One eight-GPU H200 node holds either model. Its 1,128 GB of combined memory holds Llama 3.1 405B with room for the cache, and the vLLM team noted at launch that the model runs on a single 8xH200 machine, while H100s need FP8 on a node or sixteen GPUs in BF16 31. In FP8 the 405B weights take about 406 GB, so four H200s can hold them with a modest cache and eight give room for many users. DeepSeek V3 and R1 are published in FP8 at 671B parameters, which needs two H100 nodes but one H200 node 34. Verda measured 2,378 output tokens per second on DeepSeek V3 in BF16 and 2,864 in FP8 on such a node with untuned SGLang 30. At GPUBank the node costs $19.92 an hour on demand and $15.92 reserved 1.

AWS sells the H200 in the p5e and p5en instances, each with eight GPUs. The p5en.48xlarge costs $63.296 an hour on demand in N. Virginia, $7.91 per GPU, before EBS storage and data transfer 23. Its EC2 Capacity Blocks, which reserve a node for a fixed window paid up front, list the p5e.48xlarge at $54.924 an hour, $6.866 per GPU, and the p5en at $63.158, $7.895 per GPU, in about a dozen regions 22. The page says the next price revision is scheduled for January 2027 and that each block is charged at the rate in force when it is bought, even if it starts after a change. A month of a p5en node on demand comes to about $46,206 before storage and egress, against $11,621.60 for a month of GPUBank's reserved node 1.

Yes. Google Cloud's accelerator-optimised pricing page lists the a3-ultragpu-8g, A3 Ultra with eight H200s, 224 vCPU, 2,952 GB of RAM and 12,000 GiB of local SSD, at $84.81 an hour on its default on-demand model, $10.60 per GPU 21. The same row lists Dynamic Workload Scheduler Flex-start at $42.40 an hour, $5.30 per GPU, Calendar mode at $59.36, Spot at $50.80, $6.35 per GPU, a one-year commitment at $58.47 and a three-year commitment at $37.21, $4.65 per GPU. Flex-start is the cheapest Google option that does not need a commitment, but capacity is allocated when it is available rather than on request, which suits batch jobs better than an endpoint. Like the other large clouds, Google sells the H200 only by the eight-GPU machine.

Because the products differ more than the chip does. A price can include the CPU, memory, NVMe, public IP and traffic, or only the GPU; it can be for one GPU or an eight-GPU server; it can be fixed, reserved, set by a marketplace host or set by a spot market that moves every 15 minutes 3. Large clouds add enterprise networking, compliance and integration with their other services, which is why Oracle, Google and Azure sit at $10.00 to $10.60 per GPU 25 21 24. Marketplace hosts set their own asks, which move from week to week: Lium's cheapest H200 rose from $3.40 to $3.73 in a fortnight 8. Stock matters too, since several of the lowest aggregator figures were marked out of stock 20. Compare what a month of your workload will cost, not the headline rate.

Not over the past year. GetDeploying measured the median on-demand H200 price up about 27 percent over twelve months and 3 percent over the last 90 days, at $4.94 on 16 October 2026 20. Several providers raised their rates recently: Nebius moved its on-demand H200 from $4.50 to $5.40 on 1 October, and Lium's cheapest ask rose from $3.40 on 29 September to $3.73 3 8. At the same time the lowest prices moved down, as Nebius's spot floor fell from a fixed $2.45 to a dynamic $0.79, while AWS has scheduled its next Capacity Block revision for January 2027 22. The picture is a market where typical prices rise while the cheapest interruptible hours fall, as Blackwell supply grows and Hopper capacity is resold more aggressively at the bottom.

An eight-GPU HGX H200 server cost $320,000 to $420,000 in July 2026 according to Mercatus, which gives $370,000 as the typical integrated price, while Thunder Compute gives $350,000 to $450,000 and Ace Computers lists a Matrix HGX H200 from $478,992; the SXM module alone sells for $32,000 to $40,000 but needs an HGX baseboard 37 47 48. Add colocation, power for a system that draws about 10 kW at the wall, support and the cost of the money. Under our assumptions, $370,000 of hardware, 40 percent resale after two years, $1,500 a month of colocation, 8 percent cost of capital, $500 a month of support and power at the US commercial average of 13.97 cents per kWh, owning costs about $13,717 a month plus $1.42 per hour of use 38. That never beats GPUBank's on-demand node at $19.92 an hour within a 730-hour month, and it beats the market median only above about 360 hours 1 20. Buying makes sense mainly for teams that would otherwise pay large-cloud rates or must keep data on their own hardware.

Yes, at a few providers. GPUBank is paid only from a prepaid US dollar balance topped up in crypto, with Bitcoin, Ethereum, USDT, USDC, Monero and more than a dozen other coins accepted for top-ups of $25 to $5,000, and a bonus of 20 to 60 percent depending on the amount brings the effective H200 hour to about $2.08 with $25, $1.84 with $100 and $1.56 with $1,000; the balance never expires, and cards are not accepted for now 1. Lium accepts top-ups by card or by crypto 8. Most other clouds in this comparison, and all four large clouds, bill a card or an invoice in the usual way. A top-up bonus is a real discount only on money you will in fact spend, so size a top-up to the hours you expect to use rather than to the largest bonus tier.

That depends entirely on the provider's reclaim terms, so read them before starting a long job. On GPUBank's interruptible tier you get five minutes of notice through the metadata endpoint, a webhook and the console; your checkpoint hook runs, the machine is stopped rather than deleted, the disk and public IP survive, and if the notice is missed the last 30 minutes are credited 1. On Nebius, preemptible VMs can be stopped whenever capacity is needed and the price can change every 15 minutes, while stopped VMs keep billing only for their volumes 3. On Vast.ai, an interruptible instance is paused when a higher bid arrives and its storage keeps billing until you destroy it 5. Whatever the provider, write checkpoints to a persistent disk frequently and budget for redoing some work, which our fine-tuning scenario sets at 5 percent.