Where to rent an H200 SXM at the cheapest price

The table counts the cards needed to hold each model's weights at FP8 while leaving about 15 percent of memory for the KV cache and activations, rounded up to one, two, four or eight cards, with sixteen meaning two eight-GPU nodes. The H200 halves the card count for most of the models people actually serve. Llama 3.3 70B in FP8 takes about 71 GB at one byte per parameter, nearly 73 GB as published with its embeddings in BF16, which leaves an H100 no useful room for the cache, so it needs two H100s with tensor parallelism, or a single H200 with 70 GB to spare⁴⁴. Qwen3 235B-A22B moves from four H100s to two H200s, Llama 3.1 405B from a full H100 node to four H200s, and DeepSeek V3 or R1, published in FP8 at 671B parameters, from two H100 nodes linked by InfiniBand to one H200 node⁴⁵⁴⁶. The vLLM team made the same point at the 405B launch: one 8xH200 machine holds the model, where H100s need FP8 on a node or sixteen cards in BF16⁴⁷.

Where the card count does not change, the extra memory goes into the cache, which sets how many requests and how much context a GPU serves at once. With Llama 3.3 70B in FP8, an FP8 KV cache costs 160 KiB per token (80 layers, 8 key-value heads of 128 dimensions, keys and values at one byte). On an H100, after 67.7 GiB of weights and a working reserve, about 5 GiB are left, roughly 32,500 tokens of cache in total, which is less than one full 128K-token conversation; on an H200 about 63 GiB remain, roughly 410,000 tokens, more than twelve times as much⁴⁴. That is the difference between a handful of concurrent users and dozens, and it is why NVIDIA's own H200 comparisons ran larger batches than the H100 could hold. Models that fit an H100 with room to spare, such as Llama 3.1 8B, Qwen3 32B or gpt-oss-120b in its native MXFP4 form, gain batch and context from an H200 but not a card, and are usually cheaper on the H100⁴⁸.

A managed cloud sells a fixed machine, the GPU with its vCPUs, memory and local disk, at one hourly price, and for the H200 the machine varies a great deal: GPUBank and Runpod attach 24 vCPU and 276 GB of RAM to each GPU, Beam 24 vCPU and 192 GB, Verda 44 vCPU and 170 GB, Civo 23 vCPU and 169 GB^4135714. Some of these prices also include the public IP and all traffic, others bill them separately, and several of the lowest are only sold by the eight-GPU server, which turns a $2.48 rate into a $19.84 hour before anything runs⁶. A peer-to-peer marketplace lets independent hosts post their own GPU price and adds storage and bandwidth as separate lines, on hardware that varies from listing to listing: on Vast.ai the cheapest H200 ask on 16 October was $2.80 on a host with 28 vCPU and about 258 GB of RAM in Japan, with a reliability score of 94.8 percent².

The other three kinds trade something away for the price. Spot or interruptible capacity can be taken back, and what matters is whether its price is fixed and how much notice you get: Nebius's preemptible floor of $0.79 floats with demand up to $5.39, while GPUBank's $1.49 is fixed with five minutes of notice¹⁴. Reserved capacity is paid whether the card is busy or not, usually for months or years, and the lowest reserved rates on this page come with terms from a single month at GPUBank to 36 months at Civo⁴¹⁴. Serverless platforms bill only the seconds a container works, at a higher hourly equivalent, which is cheaper than an idle instance only when traffic is bursty. Putting a $0.79 spot hour next to a $2.49 managed hour without saying so is how most H200 price lists end up comparing things that cannot be compared.

H200 prices have risen over the past year rather than fallen. GetDeploying measured the median on-demand price up about 27 percent in twelve months, Nebius raised its on-demand rate from $4.50 to $5.40 on 1 October, Runpod's Secure Cloud pod now lists $5.29 where its own articles still quote $4.59, and Hyperbolic, which launched single H200s at $2.65 in September 2025, lists $3.99 today^6113151617. AWS's Capacity Block for the p5e went from $34.61 an instance-hour in December 2025 to $54.924 now, about 59 percent more per GPU¹⁸¹⁹. The cheapest interruptible hours moved the other way, as Nebius's spot floor fell from a fixed $2.45 to a dynamic $0.79, so the market is widening at both ends rather than softening, whatever some October guides forecast.

The captures below were taken on 16 Oct 2026 from each provider's own H200 or pricing page, so every figure can be read next to the conditions printed beside it.

Read against the card above, the terms behind GPUBank's $2.49 are what set it apart. The meter starts only once a machine has passed its health check, which GPUBank gives as a median of 48 seconds from launch, a failed launch is not billed, and if a host fails during a job the machine restarts on a healthy host with the same disk and the interrupted session is credited in full³⁰. Machines of 1, 2, 4 or 8 GPUs scale the vCPU, RAM and NVMe by the same multiple, which puts a whole 8-GPU node at $19.92 an hour, and templates for vLLM, SGLang, Ollama, llama.cpp, Unsloth and Axolotl, or any Docker image, cover the usual H200 work. Among single-GPU offers that include the IP and the traffic, $2.49 was the lowest price we found on 16 October, and the interruptible tier at a fixed $1.49 and the one-month reservation at $1.99 were the lowest fixed prices in their categories on a provider's own page⁴.

GPUBank loses in several places, and they matter. Nebius's dynamic spot price is far below $1.49 whenever its floor holds, Beam lists a single H200 machine at $2.09, 40 cents under GPUBank, and a marketplace host occasionally undercuts both for a short job¹⁵. GPUBank does not sell multi-node clusters on an InfiniBand fabric, which Runpod's Instant Clusters, CoreWeave and the large clouds do, so a training run spanning several nodes belongs elsewhere¹³²⁵. Payment is a prepaid US dollar balance topped up in crypto from $25 to $5,000, with no card payments for now; the balance never expires and top-ups carry a bonus of 20 to 60 percent, which brings the effective on-demand hour to about $2.08 with a $25 top-up, $1.84 with $100 and $1.56 with $1,000, a real discount only on money you will in fact spend⁴.

Several specialised clouds sit between the cheap single-GPU offers and the large clouds. GMI Cloud's pricing page lists the H200 "from $3.20" per GPU-hour, marked available now, in its own data centres, without itemising what comes with it²⁰. Civo charges $3.49 for one H200 with 23 vCPU, 169 GB of RAM and 400 GB of NVMe, with free transfer but usage rounded up to the next full hour, and $2.99 on a 36-month commitment¹⁴. Hyperstack lists $3.99 on demand and reserved capacity from $2.79, billed per minute with free traffic but public IPs charged separately²¹. Hyperbolic shows $3.99, refreshed weekly from its suppliers' rates, and does not bill instances that fail to come online¹⁷. DigitalOcean's H200 droplet costs $4.47 with 24 vCPU, 240 GiB of RAM, 5 TiB of scratch NVMe and 15,000 GiB of transfer, or $3.40 on a 12-month contract²². GetDeploying listed further offers we could not read on the providers' own pages: Gcore at $3.56 per GPU for an 8x bare-metal node, Canopy Wave at $4.00 for two GPUs, Runcrate at $4.24 for one, and Crusoe at $5.40 and OVHcloud at $6.20 for 8-GPU machines; Together AI could not be resolved, with $5.99 for clusters in our September reading against a $2.99 row now labelled on demand, which matches its former spot price⁶.

Higher up, Lyceum charges $4.49 on demand and $2.29 spot with per-second billing, Jarvislabs $4.59 with 28 vCPU and 300 GB of RAM billed per minute, and Verda $5.44 with 44 vCPU and 170 GB, or $2.72 as spot²³²⁴⁷. CoreWeave sells only the eight-GPU node, at $50.44 an hour on demand, $6.31 per GPU, and $20.64 to $20.93 as spot²⁵. Sesterce appeared on GetDeploying at $2.48 per GPU for an eight-GPU server marked out of stock, which we could not confirm on its own site, and the same tracker showed spot rates of $2.29 at Jarvislabs, $2.61 for a GPU-only slice at Daytona and $3.32 at Spheron⁶. Its reserved rows ran from fal's $2.10 for a custom deployment and Utho's $2.59 on 36 months to Contabo's $2.86 for one month, Impossible Cloud's $3.16 for six months on an 8x node, Cirrascale's $3.68 on twelve months, Together AI's $3.99 for six months, Verda's $4.08 on 24 months and Oblivus's $4.47 on 36, several last checked by the tracker on 31 August⁶. Among serverless platforms, Koyeb lists an H200 instance at $3.00 an hour with 15 vCPU and 240 GB of RAM, Modal bills $0.001261 per second, $4.54 an hour, with CPU and memory metered on top, and Verda's serverless containers cost $5.98²⁶²⁷⁷. A Runpod serverless worker at $5.93 that is busy 40 percent of the time costs $2.37 per wall-clock hour, which is the comparison that matters against an always-on instance.

The four large clouds rent the H200 only as eight-GPU instances, so their per-GPU figure is the instance price divided by eight and the smallest bill is an entire node. Google Cloud's a3-ultragpu-8g, with 224 vCPU, 2,952 GB of RAM and 12,000 GiB of local SSD, costs $84.81 an hour on demand in its default model, $10.60 per GPU; the same row shows $42.40 for Dynamic Workload Scheduler Flex-start ($5.30 per GPU), $50.80 for Spot ($6.35), $58.47 on a one-year commitment ($7.31) and $37.21 on a three-year one ($4.65)⁹. Microsoft's ND96isr H200 v5 costs $84.80 an hour on Linux in East US 2, West US 2 and West US 3, $10.60 per GPU, and $110.24 in regions such as Germany West Central, while one-year and three-year reservations bring it to about $5.82 and $5.28 per GPU²⁸. Oracle's price list gives the H200 at $10.00 per GPU-hour²⁹.

AWS sells the H200 in two instance types. The p5en.48xlarge costs $63.296 an hour on demand in N. Virginia, $7.91 per GPU before EBS storage and egress⁸. 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), with the next price revision scheduled for January 2027 and every block charged at the rate in force on the day it is bought¹⁹. A large-cloud H200 therefore costs $37 to $85 an hour for the node, which is why single-GPU buyers rarely use one unless credits, compliance or data already stored in that cloud decide for them; for a team that needs several nodes on one fabric with enterprise support, the three-year Google and Azure rates are the ones to compare with a specialised cloud's reservation.

The hourly rate is the largest line on an H200 bill but rarely the only one, and the other lines decide which offer is cheapest at low use. Storage comes first, and the question is what is billed while nothing runs. On GPUBank it is the persistent disk alone at $0.03 per GB-month, prorated by the hour; Nebius bills volumes but not the compute of a stopped VM; Runpod charges $0.07 per GB-month for a network volume under 1 TB and $0.20 for a volume disk while its pod is idle; Jarvislabs charges $0.10 per GB-month; and Vast.ai bills an instance's storage for as long as it exists, running or stopped, at a rate set by each host^4113243. Network is the second line: GPUBank, Civo and Hyperstack include traffic, DigitalOcean includes 15,000 GiB a month, Vast.ai hosts set their own bandwidth price, and the large clouds bill egress by the gigabyte on top of an $85 hour^41421223.

Billing granularity is the quiet tax on short jobs. GPUBank, Nebius, Runpod, Lyceum and Vast.ai bill per second, Hyperstack and Jarvislabs per minute, and Civo rounds usage up to the next full hour^411323212414. A job that runs 61 minutes costs $2.53 at $2.49 billed by the second, while the same rate rounded to the hour bills two full hours, $4.98, 97 percent more for one minute of overrun; on Civo's $3.49 that job costs $6.98. Lost time is the last line, and the terms decide who pays for it. GPUBank credits a session interrupted by a host failure and the last 30 minutes of a spot machine whose notice was missed, Hyperbolic does not bill instances that never come online, and on a marketplace a host that fails, a listing that turns out to be an NVL card or a CPU too weak for tokenisation all cost time you pay for at the full rate⁴¹⁷.

In NVIDIA's MLPerf Inference v5.0 submission, closed division, an eight-GPU H200 system produced 33,072 tokens per second on Llama 2 70B in the Server scenario and 34,988 Offline, against 98,443 and 98,858 for eight B200s, about three times more, although the Blackwell entries ran in FP4, so part of the gap is precision³¹. On Mixtral 8x7B the same table gives 61,802 tokens per second for the H200 node against 126,845 for eight B200s. In the same v5.0 round NVIDIA's DGX H100 (entry 5.0-0057) served Llama 2 70B at 31,106 tokens per second, both Hopper systems in FP8 on TensorRT 10.8, so on that model eight H200s did only 6 percent more work than eight H100s at full batch³². On Mixtral 8x7B in v4.1, with both cards at 700 W, eight H200s produced 57,177 tokens per second against 50,796 for eight H100s in Server, a gain of 13 percent³³. NVIDIA's larger H200 gains on Llama 2 70B come partly from configuring the card at 1,000 W, above a standard 700 W SXM part, so they should not be expected on a rented machine³³, and NVIDIA's datasheet figure of 1.9 times the H100 ran the H200 at batch 32 against batch 8, which measures memory headroom rather than like-for-like speed¹⁰.

At GPUBank's list price these results translate into a first estimate of the cost per output, assuming MLPerf-grade tuning that few deployments match. The H200 node at $19.92 an hour makes 119 million Llama 2 70B tokens an hour, $0.167 per million, and 206 million Mixtral tokens, $0.097 per million³¹³³⁴. On a model above 600B parameters the figures come from untuned runs: Verda served DeepSeek V3 on eight H200s with SGLang 0.4.1, 1,024 tokens in and 1,024 out, and measured 2,378 output tokens per second in BF16 at four requests per second and 2,864 in FP8 at eight, which at $19.92 an hour is $2.33 and $1.93 per million output tokens³⁴. Verda itself calls these approximate results without H200-specific kernel tuning, so a carefully configured node will do better, and tuned deployments go much further: the SGLang team's expert-parallel setup reported 22,300 output tokens per second per node on twelve 8x H100 nodes with 2,000-token inputs and put the cost at $0.20 per million output tokens³⁵. Treat Verda's figures as what a first deployment of a frontier-size open model gets from one H200 server, not as its ceiling.

The realistic purchase unit is the whole server, because the SXM module mounts on an HGX baseboard and cannot go into a workstation; the single-card option is the H200 NVL, a different product. Mercatus put an eight-GPU HGX H200 system at $320,000 to $420,000 in July 2026, with $370,000 as the typical integrated price, and the module alone at $32,000 to $40,000⁴⁰. Thunder Compute gives $350,000 to $450,000, and Ace Computers lists its Matrix HGX H200 server from $478,992, so $370,000 is a mid-market figure rather than a floor⁴¹⁴². Our base case resells the server for 40 percent, $148,000, after 24 months, which makes depreciation $9,250 a month; colocation at $1,500 a month, cost of capital at 8 percent a year on $370,000, about $2,467 a month, and $500 a month of support and spares bring the fixed cost to about $13,717 a month, and all four of those inputs are our assumptions. Power is an assumed 10.2 kW at the wall under load at the US commercial average of 13.97 cents per kWh for January to July 2026, $1.42 for each hour of real use⁴³.

Renting the same eight GPUs on GPUBank costs $19.92 per node-hour on demand, so owning would break even at 13,717 / (19.92 - 1.42) = 741 hours a month, more than a 730-hour month contains. At full use owning costs about $14,757 a month, against $14,541.60 for renting on demand, $11,621.60 for a month of GPUBank's reserved node and $9,088.50 when paid from a $1,000 top-up. Leave out the cost of capital, as a buyer paying cash that would otherwise earn nothing can, and the break-even falls to about 608 hours, 83 percent of the month. Against the market median of $4.94 per GPU owning wins above about 360 hours, and against AWS's $63.30 node-hour above about 222⁶⁸. Buy only if you would otherwise pay median or large-cloud rates for most of the month, or if data must stay on your hardware; against a specialised cloud at this price, renting costs less at every level of use.