---
title: Where to Rent an H200 SXM at the Cheapest Price (October 2026)
url: https://gpubank.io/rent-h200
description: H200 SXM prices checked 16 Oct 2026: Nebius spot from $0.79, Vast.ai from $2.80, GPUBank $2.49 all-in, AWS $7.91 and Google $10.60 per GPU-hour.
generated: 2026-11-02T00:20Z
publisher: GPUBank (GPU Compute OÜ, Tallinn, Estonia)
summary_for_ai: https://gpubank.io/llms.txt
full_text_for_ai: https://gpubank.io/llms-full.txt
---

> Markdown version of https://gpubank.io/rent-h200, generated from the live page. GPUBank is a GPU cloud billed by the second at one all-in price per GPU (vCPU, RAM, NVMe, a public IPv4 and traffic included), in Paris, Frankfurt, Ashburn and Singapore. Prices are in US dollars, excluding VAT.

Updated 16 Oct 2026 · 35 offers checked

# Where to rent an H200 SXM at the cheapest price

We read more than thirty H200 SXM offers on the providers' own pages, sorted them by what each hour really buys, and worked out the monthly bill, the speed per dollar and the point at which buying an eight-GPU server would cost less.

**GPUBank Research** 18 min read · Prices read on 16 Oct 2026

**141 GB**HBM3e

**4.8 TB/s**memory bandwidth

**900 GB/s**NVLink

**700 W**SXM5

The short answer, 16 Oct 2026

1. $0.79 /hr **Cheapest hour you can get.** Nebius preemptible H200 capacity starts at $0.79 per GPU-hour, but since 8 October it is a dynamic spot price that can move every 15 minutes up to $5.39, one cent below the on-demand rate, and the VM can be taken back at any time.
2. $2.80 /hr **Cheapest marketplace listing.** The lowest live Vast.ai ask was $2.80 for a single H200 on a host in Japan, a GPU rate to which storage and bandwidth are added at prices each host sets.
3. $2.49 /hr **Cheapest hour with IP and traffic included.** GPUBank's $2.49 includes 24 vCPU, 276 GB of RAM, 2 TB of NVMe, a public IPv4 address and all traffic, billed per second. Beam lists a lighter 1-GPU machine for less, at $2.09, without a confirmed stock level.

[H200 SXM $2.49 /hr all-in Launch an H200 SXM Spot $1.49 · reserved $1.99
Billed per second](https://console.gpubank.io/launch?gpu=h200-sxm)

The H200 SXM is rented at prices that differ by a factor of thirteen for what looks like the same hour of the same chip. On 16 October 2026 the lowest figure on a provider's own page was a $0.79 spot floor at Nebius, the cheapest marketplace ask was $2.80 on Vast.ai, a single managed GPU with its CPU, memory and disk ran from $2.09 to about $5.44 at the specialised clouds, and the four large clouds charged between $7.91 and $10.60 per GPU for an eight-GPU instance. GetDeploying, which tracks 49 providers, put the median on-demand price at $4.94 across the 34 that publish one, up about 27 percent over twelve months. The spread is not noise: each price buys a different product, from the GPU slice of a stranger's server to a whole HGX node with InfiniBand, and the cheapest line on a pricing page is often not the cheapest bill at the end of the month.

This page sorts the offers into five kinds, within which prices can fairly be compared, says what each price leaves out, works through three monthly bills and the rent-or-buy arithmetic for an eight-GPU server, and shows where the 141 GB of memory actually saves money compared with an H100. Prices were re-read on the providers' own pages on 16 October 2026; where a page could not be read, the figure is marked as coming from an aggregator, and providers that could not be checked at all were left out. The page is published by GPUBank, a GPU cloud run by GPU Compute OÜ in Tallinn, which rents the H200 SXM itself. Read our own entry with that in mind: it is measured on the same terms as every other offer, and the text says where a spot tier, a marketplace host or another cloud beats it.

01 / What 141 GB buys

## What 141 GB buys you

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 spare. 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 node. 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 BF16.

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 much. 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 H100.

Cards needed to hold the weights at FP8, with room for the cache

H100 SXM, 80 GB H200 SXM, 141 GB

**Llama 3.1 8B**

8B parameters · 8 GB at FP8

**1**

**1**

Fits either card many times over; the H200 only adds batch and context.

**Qwen3 32B**

32.8B parameters · 33 GB at FP8

**1**

**1**

One card on both; at 16-bit it is tight on an H100 and comfortable on an H200.

**Llama 3.3 70B**

70.6B parameters · 71 GB at FP8

**2**

**1**

71 GB of FP8 weights leave no cache room on 80 GB; one H200 removes tensor parallelism.

**gpt-oss-120b**

116.8B parameters · 117 GB at FP8

**2**

**1**

Shipped in MXFP4 (about 61 GiB), which OpenAI says runs on a single 80 GB GPU; the FP8 count applies only to an 8-bit conversion.

**Qwen3 235B-A22B**

235.1B parameters · 235 GB at FP8

**4**

**2**

Mixture of experts, 22B active; half the cards on H200.

**Llama 3.1 405B**

405.9B parameters · 406 GB at FP8

**8**

**4**

A full H100 node at FP8; at BF16 one 8x H200 node against sixteen H100s.

**DeepSeek V3 / R1**

671B parameters · 671 GB at FP8

**16**

**8**

Published in FP8; 16 means two H100 nodes linked by InfiniBand, against one H200 node.

Weights at FP8 = parameters × 1 byte, leaving about 15 % of memory for the KV cache and activations, rounded up to 1, 2, 4, 8 or 16 cards (16 is two 8-GPU nodes). The bar shows the weights against one H100 (80 GB) and one H200 (141 GB), on a scale to 700 GB.

Room left for context, Llama 3.3 70B FP8, one card

H100 SXM

**32,500 tokens**

H200 SXM

**410,800 tokens**

Computed, not measured: 81,559 MiB (H100) and 143,771 MiB (H200) of device memory, 95 % usable as in a vLLM gpu_memory_utilization of 0.95, minus 67.7 GiB of FP8 weights (embeddings and output layer kept in BF16) and 3 GiB of activations and runtime, divided by an FP8 KV cache of 160 KiB per token (80 layers x 8 KV heads x 128 dims x 2 x 1 byte). The model's own limit is 131,072 tokens per sequence.

02 / The card

## What the H200 SXM is, and who rents it

The H200 SXM is NVIDIA's Hopper GPU with a larger and faster memory system: 141 GB of HBM3e at 4.8 TB/s, NVLink at 900 GB/s, up to 700 W of configurable power and up to seven MIG instances of 18 GB each, mounted on four- or eight-GPU HGX boards. Its compute is the H100's, 3,958 FP8 teraFLOPS with sparsity or about 1,979 dense, so the gain is entirely in memory: against the H100 SXM's 80 GB at 3.35 TB/s it has 76 percent more capacity and 43 percent more bandwidth. That makes it faster wherever the GPU waits on memory, which is most of the time when a large language model writes tokens, and it lets a model fit on fewer cards, which removes tensor parallelism or frees room for a larger batch and a longer context. When a job is bound by arithmetic rather than memory, as large-batch training of a mid-size model often is, an H200 does little more than an H100.

The H200 NVL is a different product and is kept out of every price on this page. It is a dual-slot, air-cooled PCIe card with the same 141 GB at 4.8 TB/s, but its power ceiling is 600 W, its FP8 rating is 3,341 teraFLOPS with sparsity rather than 3,958, and it links to at most three other cards through two- or four-way NVLink bridges instead of an eight-way NVSwitch fabric. Marketplaces mix the two: on 16 October Vast.ai's search returned NVL listings from $3.80 next to SXM listings from $2.80, and Lium links an NVL page from $3.85 beside its SXM range. A listing that only says "H200" should be checked for its form factor before it is compared with an SXM price, because for a model split across GPUs the SXM board's full fabric is what you are paying for.

The card reaches two groups of customers through two kinds of seller. Large clouds and bare-metal specialists sell it in its native unit, the eight-GPU HGX server, to teams that train across a full NVLink domain or serve a model of several hundred billion parameters, and they price it by the node. Specialised clouds and marketplaces slice the same servers into one, two or four GPUs for teams serving a single 70B-class model with a long context, fine-tuning above 70B parameters, or running experiments that do not fit in 80 GB. Almost every cheap hour on this page comes from the second kind of seller, and almost every price above $6 from the first.

03 / The price ladder

## Five kinds of H200 offer, and how to compare them

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. 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 runs. 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 percent.

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 notice. 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 Civo. 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. 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 GPU. 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.

**Managed on demand** A fixed machine at one hourly price.

*Beam**$2.09***

*Sesterce**$2.48***

*GPUBank**$2.49***

*GMI Cloud**$3.20***

*Civo**$3.49***

*Hyperstack**$3.99***

*Hyperbolic**$3.99***

*DigitalOcean**$4.47***

*Jarvislabs**$4.59***

*Runpod**$5.29***

*Nebius**$5.40***

*Verda**$5.44***

*CoreWeave**$6.31***

**Marketplace** Hosts set their own price; machines vary.

*Vast.ai**$2.80***

*Runpod Community**$3.59***

*Lium**$3.73***

**Spot and interruptible** Cheaper hours that can be taken back.

*Nebius**$0.79***

*GPUBank**$1.49***

*Lyceum**$2.29***

*CoreWeave**$2.58***

*Verda**$2.72***

**Reserved** Committed for a term, paid either way.

*GPUBank**$1.99***

*Hyperstack**$2.79***

*Civo**$2.99***

*DigitalOcean**$3.40***

**Serverless** Billed by active seconds.

*Koyeb**$3.00***

*Modal**$4.54***

*Runpod Serverless**$5.93***

USD per GPU-hour, read on 16 Oct 2026. Hover a dot for the offer; crowded dots keep their label in the table below. The large clouds are listed separately.

04 / Provider by provider

## What each price buys, and what it leaves out.

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.

gpubank.io/pricing

**GPUBank** All-in

$2.49/hr

Managed cloud · 1 to 8 GPUs · Paris, Ashburn, Singapore

24 vCPU 276 GB RAM 2 TB NVMe Public IPv4 All traffic Persistent disk $0.03/GB-month

GPUBank's H200 SXM costs $2.49 per GPU-hour with the machine around it in the price: 24 vCPU, 276 GB of RAM and a 2 TB local NVMe disk per GPU, one public IPv4 address and all inbound and outbound traffic, on machines of one, two, four or eight GPUs. Billing is per second with no minimum and starts only after the machine passes its health check, so a launch that fails costs nothing, and a stopped machine costs only its persistent disk at $0.03 per GB-month. The same card is $1.49 as interruptible capacity with five minutes of notice, a checkpoint hook and a stop rather than a deletion, and $1.99 reserved for a week or a month, prepaid. Payment is a prepaid balance topped up in crypto from $25, with a bonus that lowers the effective on-demand hour to about $1.56 on a $1,000 top-up. GPUBank does not sell multi-node InfiniBand clusters.

Checked 16 Oct 2026

beam.cloud/pricing

**Beam**

$2.09/hr

Managed cloud · single-GPU machine

24 vCPU 192 GB RAM 2 TB NVMe Stock not reported No serverless H200

Beam's pricing page lists an on-demand H200 SXM5 machine from $2.09 an hour with 24 vCPU, 192 GB of RAM and a 2 TB NVMe drive, the lowest list price for a whole single-GPU H200 machine that we found on a provider's own page on 16 October. It undercuts GPUBank by 40 cents an hour and comes with the same 24 vCPU and the same local disk, but 84 GB less system memory. Two things could not be confirmed. GetDeploying lists the offer with its stock status as not reported, so we cannot say whether a card was free to launch that day, and the page does not say how the machine is billed, whether a public IP and traffic are included, or what happens to the disk when the machine stops. Beam's serverless table, which bills by the second, does not include the H200 at all. If Beam can provision the card when you need it, it is the cheapest managed hour on this page.

Checked 16 Oct 2026

vast.ai/pricing/gpu/H200

**Vast.ai**

from $2.80/hr

Marketplace · host-set prices

vCPU and RAM of the host Storage billed while stopped Bandwidth per host SXM and NVL mixed

Vast.ai is a marketplace where independent hosts set their own price, so its H200 figure is the cheapest live ask rather than a list price. On 16 October its search returned SXM listings from $2.80 per GPU-hour, the cheapest a single H200 with 28 vCPU and about 258 GB of RAM on a host in Japan with a reliability score of 94.8 percent, then $4.00 in Georgia and $4.61 in Romania. Separate H200 NVL listings started at $3.80, so check the form factor before comparing. The rate covers the GPU and the host's CPU and memory share; storage is billed for as long as the instance exists, running or stopped, and bandwidth is charged at each host's rate, which can add tens of dollars a month to a cheap hour. Billing is by the second from a prepaid balance. Vast suits short, checkpointed jobs on a budget, less so an endpoint that must stay up.

Checked 16 Oct 2026

lium.io/gpu/h200

**Lium**

$3.73 to $5.70/hr

Marketplace · pods from independent providers

Pod incl. local disk Specs vary by pod Lowest asks on 8-GPU pods

Lium is a marketplace of pods offered by independent providers, priced for the whole pod including the node's local disk, with vCPU, memory and storage that differ from one pod to the next. Its H200 page showed a live range of $3.73 to $5.70 per GPU-hour on 16 October, against a marketplace reference price of $3.65, and links a separate H200 NVL page from $3.85. On our 29 September reading the cheapest ask had been $3.40, so the floor moved up by a third of a dollar in two weeks, which is typical of a market where a handful of hosts set the price. The lowest asks tend to be on multi-GPU pods, so check the pod size before assuming a single card is available at the headline figure. Lium bills by the second and suits users who want a marketplace with per-pod pricing rather than separately billed disk.

Checked 16 Oct 2026

docs.nebius.com/compute/resources/pricing

**Nebius**

$5.40/hr · spot from $0.79

Managed cloud · presets of 1 or 8 GPUs

16 vCPU 200 GB RAM Volumes extra Spot price changes every 15 min

Nebius raised its on-demand H200 price from $4.50 to $5.40 per GPU-hour on 1 October 2026 and on 8 October switched its preemptible tier from a fixed $2.45 to dynamic spot pricing listed "from $0.79". That floor is the lowest H200 figure on any provider's own page, but Nebius states that spot prices follow capacity and demand, can change during use as often as every 15 minutes, and can rise to one cent below the on-demand rate, so a job budgeted at $0.79 may pay up to $5.39. GPUs on running VMs are billed per second, and stopped VMs are charged only for their volumes. The on-demand presets come with one or eight GPUs in its European and US regions. For checkpointed work that can stop at any moment and tolerate a moving price, Nebius spot is the cheapest place to run an H200; for a fixed budget it is a bet rather than a quote.

Checked 16 Oct 2026

runpod.io/pricing

**Runpod**

$5.29/hr

Managed cloud and community hosts · per second

24 vCPU 276 GB RAM Network volume $0.07/GB-month Idle volume disk $0.20/GB-month

Runpod's pricing page, updated 8 October 2026, lists the H200 SXM pod at $5.29 an hour on Secure Cloud with 24 vCPU and 276 GB of RAM, billed per second, while the page's structured data still carries a Community Cloud rate of $3.59 that the visible card no longer shows and GetDeploying marked out of stock. The same page gives $5.93 for a serverless H200 worker and $6.35 per GPU for Instant Clusters spanning several nodes, which is one of the few ways to rent a multi-node H200 fabric by the hour. Storage is extra: a network volume costs $0.07 per GB-month under 1 TB and $0.05 above, and a volume disk costs $0.10 while its pod runs and $0.20 while it is idle. Runpod is a mature platform with templates and a large community, but on 16 October its single-GPU H200 cost more than twice the cheapest managed offers.

Checked 16 Oct 2026

### Where GPUBank wins, and where it does not

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 full. 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 page.

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 job. 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 elsewhere. 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 spend.

### The rest of the market

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 it. 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 commitment. Hyperstack lists $3.99 on demand and reserved capacity from $2.79, billed per minute with free traffic but public IPs charged separately. Hyperbolic shows $3.99, refreshed weekly from its suppliers' rates, and does not bill instances that fail to come online. 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 contract. 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 price.

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 spot. 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 spot. 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 Spheron. 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 August. 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 $5.98. 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 large clouds, per GPU inside an eight-GPU instance

Google Cloud (3-year CUD)

$4.65 /GPU-hr

8-GPU node · 224 vCPU, 2,952 GB RAM, 12,000 GiB SSD

Azure (3-year reservation)

$5.28 /GPU-hr

8-GPU node · 96 vCPU, 1,850 GiB RAM

AWS (Capacity Block p5e)

$6.87 /GPU-hr

8-GPU node · Instance; EBS and egress extra

AWS (p5en on demand)

$7.91 /GPU-hr

8-GPU node · Instance; EBS and egress extra

Oracle Cloud

$10.00 /GPU-hr

8-GPU node · Bare-metal BM.GPU.H200.8

Google Cloud

$10.60 /GPU-hr

8-GPU node · 224 vCPU, 2,952 GB RAM, 12,000 GiB SSD

Microsoft Azure

$10.60 /GPU-hr

8-GPU node · 96 vCPU, 1,850 GiB RAM (East US 2)

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). 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 GPU. Oracle's price list gives the H200 at $10.00 per GPU-hour.

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 egress. 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 bought. 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.

05 / Every offer

## The whole comparison, sorted by kind and price.

| Provider | USD / GPU-hr | Kind | Unit | What the price includes | Billing | Stock | Basis |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Beam | $2.09 | Managed on demand | 1 GPU | 24 vCPU, 192 GB RAM, 2 TB NVMe | Not stated | Not stated | Provider page |
| Sesterce | $2.48 | Managed on demand | 8-GPU node | 96 vCPU, 2 TB RAM | Not stated | Sold out | Aggregator |
| GPUBank | $2.49 | Managed on demand | 1 GPU | 24 vCPU, 276 GB RAM, 2 TB NVMe, IPv4, traffic | Per second | In stock | Provider page |
| GMI Cloud | $3.20 | Managed on demand | 1 GPU | Not itemised | Not stated | In stock | Provider page |
| Civo | $3.49 | Managed on demand | 1 GPU | 23 vCPU, 169 GB RAM, 400 GB NVMe, transfer | Hourly, rounded up | Not stated | Provider page |
| Hyperstack | $3.99 | Managed on demand | 1 GPU | Up to 22 pCPU, 225 GB RAM, NVMe, traffic; IP extra | Per minute | Not stated | Provider page |
| Hyperbolic | $3.99 | Managed on demand | 1 GPU | Varies by supplier node | Metered | Not stated | Provider page |
| DigitalOcean | $4.47 | Managed on demand | 1 GPU | 24 vCPU, 240 GiB RAM, 5 TiB scratch, 15,000 GiB transfer | Not stated | Not stated | Provider page |
| Jarvislabs | $4.59 | Managed on demand | 1 GPU | 28 vCPU, 300 GB RAM; storage extra | Per minute | Not stated | Provider page |
| Runpod | $5.29 | Managed on demand | 1 GPU | 24 vCPU, 276 GB RAM; volumes extra | Per second | Not stated | Provider page |
| Nebius | $5.40 | Managed on demand | 1 GPU | Preset vCPU and RAM; volumes extra | Per second | Not stated | Provider page |
| Verda | $5.44 | Managed on demand | 1 GPU | 44 vCPU, 170 GB RAM | Not stated | Not stated | Provider page |
| CoreWeave | $6.31 | Managed on demand | 8-GPU node | 128 vCPU, 2,048 GB RAM, 61.44 TB NVMe | Not stated | Not stated | Provider page |
| Vast.ai | $2.80 | Marketplace | 1 GPU | Host vCPU and RAM; disk and bandwidth extra | Per second | In stock | Provider page |
| Runpod Community | $3.59 | Marketplace | 1 GPU | 24 vCPU, 276 GB RAM | Per second | Sold out | Provider page |
| Lium | $3.73 | Marketplace | 8-GPU pod | Whole pod incl. local disk; $5.70 on smaller pods | Per second | In stock | Provider page |
| Nebius | $0.79 | Spot and interruptible | 1 GPU | Dynamic floor, up to $5.39 | Per second | Not stated | Provider page |
| GPUBank | $1.49 | Spot and interruptible | 1 GPU | As on demand; 5 min notice, disk kept | Per second | In stock | Provider page |
| LLyceum | $2.29 | Spot and interruptible | 1 GPU | Not itemised | Per second | In stock | Provider page |
| CoreWeave | $2.58 | Spot and interruptible | 8-GPU node | 128 vCPU, 2,048 GB RAM (Europe) | Not stated | Not stated | Provider page |
| Verda | $2.72 | Spot and interruptible | 1 GPU | 44 vCPU, 170 GB RAM | Not stated | Not stated | Provider page |
| GPUBank | $1.99 | Reserved | 1 GPU | As on demand; week or month, prepaid | Per term | In stock | Provider page |
| Hyperstack | $2.79 | Reserved | 1 GPU | From price; term on request | Monthly invoice | On request | Provider page |
| Civo | $2.99 | Reserved | 1 GPU | 36-month commitment | Per term | Not stated | Provider page |
| DigitalOcean | $3.40 | Reserved | 1 GPU | 12-month contract | Per term | On request | Provider page |
| Google Cloud (3-year CUD) | $4.65 | Large clouds | 8-GPU node | 224 vCPU, 2,952 GB RAM, 12,000 GiB SSD | Committed | Not stated | Provider page |
| Azure (3-year reservation) | $5.28 | Large clouds | 8-GPU node | 96 vCPU, 1,850 GiB RAM | Committed | Not stated | Provider page |
| AWS (Capacity Block p5e) | $6.87 | Large clouds | 8-GPU node | Instance; EBS and egress extra | Prepaid block | Not stated | Provider page |
| AWS (p5en on demand) | $7.91 | Large clouds | 8-GPU node | Instance; EBS and egress extra | Per second | Not stated | Provider page |
| Oracle Cloud | $10.00 | Large clouds | 8-GPU node | Bare-metal BM.GPU.H200.8 | Per hour | Not stated | Provider page |
| Google Cloud | $10.60 | Large clouds | 8-GPU node | 224 vCPU, 2,952 GB RAM, 12,000 GiB SSD | Per second | Not stated | Provider page |
| Microsoft Azure | $10.60 | Large clouds | 8-GPU node | 96 vCPU, 1,850 GiB RAM (East US 2) | Per hour | Not stated | Provider page |
| Koyeb | $3.00 | Serverless | 1 GPU | 15 vCPU, 240 GB RAM, 320 GB disk | Per second | Not stated | Provider page |
| Modal | $4.54 | Serverless | 1 GPU | GPU only; CPU and RAM metered | Per second | Not stated | Provider page |
| Runpod Serverless | $5.93 | Serverless | 1 GPU | Worker; storage extra | Per second | Not stated | Provider page |

"Aggregator" rows could not be read on the provider's own page that day; "Provider snapshot" rows come from a page dated by the provider.

06 / The real monthly bill

## What an H200 hour really costs

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. 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.

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. 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 rate.

### Three workloads, priced for a month

**GPUBank**

264 h GPU $657.36

Storage $9.00

Other None

Month $666.36

$2.49 all-in, traffic and IPv4 included

All-in

**Beam**

264 h GPU $551.76

Storage

Other Not stated

Month $551.76

$2.09 machine with 2 TB NVMe; stock and disk terms not confirmed

**Vast.ai**

264 h GPU $739.20

Storage $45.00

Other Bandwidth

Month $784.20

Cheapest live ask; storage assumed at $0.15/GB-month, bandwidth not priced

**Civo**

264 h GPU $921.36

Storage

Other None

Month $921.36

$3.49, hourly rounding (whole hours here), transfer free

**Runpod**

264 h GPU $1,396.56

Storage $21.00

Other None

Month $1,417.56

Secure Cloud pod, network volume at $0.07/GB-month

One H200 serving a 70B model in FP8 with a long context during working hours runs 264 hours a month. On GPUBank that is 264 times $2.49, $657.36, plus $9.00 for 300 GB of persistent disk, $666.36 with the traffic included, or about $416 of cash when paid from a $1,000 top-up credited at $1,600. Beam's $2.09 machine costs $551.76 for the same hours and wins on list price if a card is free and its disk terms suit you, which we could not confirm. Vast.ai's cheapest ask comes to $739.20 plus about $45 of storage at an assumed $0.15 per GB-month, because an instance that sleeps at night still pays for its disk, before any bandwidth. Civo's whole-hour rounding costs nothing here since the endpoint runs in whole hours, but its $3.49 rate makes $921.36, and Runpod's Secure pod with a network volume comes to $1,417.56.

07 / Measured speed

## Performance, measured by others

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 precision. 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 batch. 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 percent. 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 machine, 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 speed.

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 million. 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 tokens. 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 tokens. Treat Verda's figures as what a first deployment of a frontier-size open model gets from one H200 server, not as its ceiling.

**Llama 2 70B, MLPerf Inference v5.0** Closed division, Server scenario, eight-GPU systems, TensorRT 10.8; Hopper in FP8, B200 in FP4

8x H100 SXM

31,106 tokens/s

8x H200 SXM

33,072 tokens/s

8x B200

98,443 tokens/s

MLCommons entries 5.0-0057 (DGX H100), 5.0-0060 (H200) and 5.0-0056 (DGX B200), all NVIDIA. Offline: 31,307, 34,988 and 98,858.

**Mixtral 8x7B, MLPerf Inference v4.1** Closed division, Server scenario, eight-GPU systems, both at 700 W

8x H100 SXM

50,796 tokens/s

8x H200 SXM

57,177 tokens/s

NVIDIA Technical Blog, MLPerf Inference v4.1. Offline: 52,416 and 59,022.

**DeepSeek V3 on one 8x H200 node** SGLang 0.4.1, tensor parallel 8, 1,024 tokens in and out, output tokens per second by request rate, untuned

BF16, 1 req/s

640 tok/s

BF16, 4 req/s

2,378 tok/s

FP8, 4 req/s

2,215 tok/s

FP8, 8 req/s

2,864 tok/s

Verda, published 8 January 2025; FP8 runs used an fp8_e5m2 KV cache; median TTFT 767 ms at BF16, 4 req/s.

SPOT FLOOR

$0.79

Nebius preemptible, dynamic, up to $5.39

MARKET MEDIAN

$4.94

On demand, 34 providers, GetDeploying

GPUBANK ALL-IN

$2.49

24 vCPU, 276 GB RAM, 2 TB NVMe, IPv4, traffic

08 / Per job, not per hour

## H200 or another card, per job rather than per hour

Against the H100 the arithmetic is close. At GPUBank the H200 costs $2.49 and the H100 SXM $1.85, 35 percent more, while the measured like-for-like gain is 13 percent on Mixtral 8x7B in MLPerf v4.1 and 6 percent on Llama 2 70B in v5.0, so eight H100s make a million Mixtral tokens for about $0.081 against $0.097 on eight H200s, and a million Llama 2 70B tokens for $0.132 against $0.167. For any model whose weights and working cache fit comfortably in 80 GB, the H100 is the cheaper rental per job. The H200 wins when its 141 GB changes the shape of the deployment by removing cards, as the card-count table on this page shows model by model, and each card removed recovers the 35 percent premium several times over.

The B200 at $4.15 makes Llama 2 70B tokens for about $0.094 per million against $0.167 on the H200 in the MLPerf v5.0 results, 44 percent less, so a team that can serve in FP4 or FP8 on Blackwell-optimised software gets more output per dollar, and its 180 GB holds more than the H200's 141 GB. The MI300X at $1.65 offers 192 GB at 5.3 TB/s for two-thirds of the H200's price, but the evidence on its speed disagrees. Tom's Hardware read AMD's first MLPerf Llama 2 70B results as roughly in line with the H100, while The Next Platform put the same MI300X result at about 60 percent of an H200's work, which at 66 percent of the price would make it about 10 percent dearer per token; an October 2025 arXiv study measured 37 to 66 percent of H100 and H200 throughput in realised inference, and SemiAnalysis found it on par with the H200 on vLLM for Llama 3.1 405B but behind on TensorRT-LLM. It wins per job only where your model and engine run well on ROCm. The A100 SXM at $0.95 is the cheap choice for models that fit in 80 GB and do not need FP8, which Ampere lacks. The H200 remains the default when you want CUDA, FP8 and 141 GB on one card without changing your stack.

MI300X

**192 GB** HBM3 · 5.3 TB/s

$1.65 /hr

Very large models on a single card, on ROCm

H100 SXM

**80 GB** HBM3 · 3.35 TB/s

$1.85 /hr

AI training, fine-tuning and production LLM inference

H200 SXM

**141 GB** HBM3e · 4.8 TB/s

$2.49 /hr

Long-context LLM inference and fine-tuning above 70B

B200 SXM

**180 GB** HBM3e · 8 TB/s

$4.15 /hr

Large-scale training and frontier model inference

GPUBank on-demand prices, all-in. Memory bar scaled to 192 GB.

09 / Rent or buy

## Rent or buy an eight-GPU H200 server

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 $40,000. 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 floor. 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 use.

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 222. 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.

Monthly cost of one 8-GPU HGX H200 server (8 GPUs) against hours of use a month, 0 to 730. Owning: price + $71,200 host + $1,500.00 a month of colocation − $148,000 resale over 24 months, plus 10.2 kW at the wall under load. Rent: per GPU-hour × 8.

| Hours a month | Rent at $2.49 | Rent at $1.56 effective | Own | Cheaper |
| --- | --- | --- | --- | --- |
| 100 | $1,992.00 | $1,245.00 | $13,859.16 | Rent |
| 200 | $3,984.00 | $2,490.00 | $14,001.66 | Rent |
| 300 | $5,976.00 | $3,735.00 | $14,144.15 | Rent |
| 400 | $7,968.00 | $4,980.00 | $14,286.65 | Rent |
| 500 | $9,960.00 | $6,225.00 | $14,429.14 | Rent |
| 600 | $11,952.00 | $7,470.00 | $14,571.63 | Rent |
| 730 | $14,541.60 | $9,088.50 | $14,756.88 | Rent, by about $215 on demand |

Server price: Mercatus typical integrated price, 8 July 2026 (range $320,000 to $420,000). The $71,200 shown as host is our cost of capital ($59,200 at 8 percent a year) and support ($12,000) over 24 months; colocation at $1,500 a month, 40 percent resale and the 10.2 kW draw are our assumptions; power at the EIA US commercial average.

10 / What to choose

## The cheapest H200 depends on the job.

Serving a 70B model with long contexts

One H200 on demand or reserved

The FP8 weights fit on one card with about 410,000 tokens of cache, where an H100 needs a second card to serve the same model.

Running a 400B to 671B open model

One 8x H200 node, reserved for the month

Llama 3.1 405B and DeepSeek V3 fit one H200 node in FP8; on H100s DeepSeek needs two nodes and InfiniBand.

Fine-tuning above 70B with checkpoints

Interruptible H200 at a fixed price

A fixed spot price with five minutes of notice keeps the bill predictable and limits rework to the steps since the last checkpoint; Nebius spot is cheaper only while its floor holds.

Models that fit in 80 GB

An H100 SXM instead

Same compute, and the H200 costs 35 percent more at GPUBank while it served only 6 to 13 percent more tokens in MLPerf, so the H100 is cheaper per token.

11 / Questions

## What people ask before renting one.

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. GetDeploying put the median on-demand price at $4.94 across 34 providers. 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. 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. Before comparing two figures, check what the hour includes and whether it comes only as a whole server.

Sources

## Sources

1. 01 [Nebius AI Cloud documentation, Compute pricing](https://docs.nebius.com/compute/resources/pricing) H200 $5.40 from 1 October, preemptible from $0.79 from 8 October, spot rules, per-second billing · read 16 Oct 2026
2. 02 [Vast.ai, live H200 offers (public search API, same data as vast.ai/pricing/gpu/H200)](https://vast.ai/pricing/gpu/H200) Cheapest SXM ask $2.80 (Japan), NVL listings from $3.80 · read 16 Oct 2026
3. 03 [Vast.ai documentation, Pricing](https://docs.vast.ai/documentation/instances/pricing) Storage billed while stopped, bandwidth per host, interruptible pausing · read 15 Oct 2026
4. 04 [GPUBank, pricing](https://gpubank.io/pricing) H200 SXM on-demand, interruptible and reserved prices, inclusions, billing, top-up bonuses, regions · read 16 Oct 2026
5. 05 [Beam, pricing](https://www.beam.cloud/pricing) H200 SXM5 machine from $2.09/hr, 24 vCPU, 192 GB RAM, 2 TB NVMe · read 16 Oct 2026
6. 06 [GetDeploying, Nvidia H200 rental prices (aggregator)](https://getdeploying.com/gpus/nvidia-h200) Median $4.94 across 34 providers, 12-month trend, Sesterce and Jarvislabs spot rows, Beam stock status · read 16 Oct 2026
7. 07 [Verda, pricing](https://verda.com/pricing) H200 SXM5 $5.44 on demand, $2.72 spot, serverless $5.98 · read 16 Oct 2026
8. 08 [AWS, EC2 On-Demand pricing (public price list, US East N. Virginia, Linux)](https://aws.amazon.com/ec2/pricing/on-demand/) p5en.48xlarge $63.296 per hour · read 16 Oct 2026
9. 09 [Google Cloud, accelerator-optimized VM pricing](https://cloud.google.com/products/compute/pricing/accelerator-optimized) a3-ultragpu-8g on-demand, Flex-start, Calendar, Spot and CUD prices · read 16 Oct 2026
10. 10 [NVIDIA H200 Tensor Core GPU](https://www.nvidia.com/en-us/data-center/h200/) SXM and NVL specifications · read 16 Oct 2026
11. 11 [NVIDIA H100 Tensor Core GPU](https://www.nvidia.com/en-us/data-center/h100/) H100 SXM 80 GB at 3.35 TB/s · read 15 Oct 2026
12. 12 [Lium, H200](https://lium.io/gpu/h200) Live range $3.73 to $5.70, lowest on 8-GPU pods, reference $3.65, NVL from $3.85, card or crypto top-ups · read 16 Oct 2026
13. 13 [Runpod, pricing (updated 8 October 2026)](https://www.runpod.io/pricing) H200 pod $5.29, Community $3.59, serverless $5.93, Instant Clusters $6.35, storage prices · read 16 Oct 2026
14. 14 [Civo, pricing](https://www.civo.com/pricing) H200 $3.49 on demand, $3.29 to $2.99 committed, hourly rounding, free transfer · read 16 Oct 2026
15. 15 [Runpod, NVIDIA H200 guide](https://www.runpod.io/articles/guides/nvidia-h200-gpu) Older Secure Cloud figure of $4.59 and Community $3.59 · read 15 Oct 2026
16. 16 [Hyperbolic blog, Introducing H200s on Hyperbolic](https://www.hyperbolic.ai/blog/introducing-h200s-on-hyperbolic) H200 launched at $2.65 per hour in September 2025 · read 15 Oct 2026
17. 17 [Hyperbolic, marketplace](https://www.hyperbolic.ai/marketplace) H200 $3.99, weekly refresh, no charge for failed instances · read 16 Oct 2026
18. 18 [The Register, AWS raises EC2 Capacity Block prices (5 January 2026)](https://www.theregister.com/2026/01/05/aws_price_increase/) p5e Capacity Block from $34.61 to $39.80 per instance-hour · read 15 Oct 2026
19. 19 [AWS, Amazon EC2 Capacity Blocks for ML pricing](https://aws.amazon.com/ec2/capacityblocks/pricing/) p5e $54.924 and p5en $63.158 per instance-hour, next update January 2027 · read 16 Oct 2026
20. 20 [GMI Cloud, pricing](https://www.gmicloud.ai/pricing) H200 from $3.20 per GPU-hour, available now · read 16 Oct 2026
21. 21 [Hyperstack, GPU pricing](https://www.hyperstack.cloud/gpu-pricing) H200 SXM $3.99 on demand, reserved from $2.79, per-minute billing, free traffic, IP extra · read 16 Oct 2026
22. 22 [DigitalOcean, GPU Droplets pricing](https://www.digitalocean.com/pricing/gpu-droplets) H200 $4.47, 12-month $3.40, droplet specification and transfer · read 16 Oct 2026
23. 23 [Lyceum, pricing](https://lyceum.technology/pricing) H200 $4.49 on demand, $2.29 spot, per-second billing · read 16 Oct 2026
24. 24 [Jarvislabs, pricing](https://jarvislabs.ai/pricing) H200 SXM $4.59, 28 vCPU, 300 GB RAM, per-minute billing, storage $0.10/GB-month · read 16 Oct 2026
25. 25 [CoreWeave, pricing](https://www.coreweave.com/pricing) HGX H200 node $50.44/hr, spot $20.93 (NA) and $20.64 (Europe) · read 16 Oct 2026
26. 26 [Koyeb, pricing](https://www.koyeb.com/pricing) H200 serverless instance $3.00/hr, 15 vCPU, 240 GB RAM · read 16 Oct 2026
27. 27 [Modal, pricing](https://modal.com/pricing) H200 $0.001261 per second, CPU and memory rates · read 16 Oct 2026
28. 28 [Microsoft Azure, Linux virtual machines pricing (Retail Prices API, Standard_ND96isr_H200_v5)](https://azure.microsoft.com/en-us/pricing/details/virtual-machines/linux/) $84.80/hr in East US 2, $110.24 in other regions, 1-year and 3-year reservations · read 16 Oct 2026
29. 29 [Oracle Cloud price list (Compute, GPU, H200)](https://www.oracle.com/cloud/price-list/) H200 $10.00 per GPU-hour · read 16 Oct 2026
30. 30 [GPUBank, home page](https://gpubank.io/) Health check before billing, host failure credit, 48 s median launch · read 15 Oct 2026
31. 31 [NVIDIA Technical Blog, Blackwell in MLPerf Inference v5.0](https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0) 8x H200 and 8x B200 on Llama 2 70B, Mixtral 8x7B, SDXL · read 16 Oct 2026
32. 32 [MLCommons, MLPerf Inference v5.0 results (summary_results.json)](https://github.com/mlcommons/inference_results_v5.0) Closed datacenter entries 5.0-0056 (DGX B200), 5.0-0057 (DGX H100), 5.0-0060 (H200): Llama 2 70B, FP8 on Hopper, TensorRT 10.8 · read 16 Oct 2026
33. 33 [NVIDIA Technical Blog, Blackwell in MLPerf Inference v4.1](https://developer.nvidia.com/blog/nvidia-blackwell-platform-sets-new-llm-inference-records-in-mlperf-inference-v4-1) 8x H200 and 8x H100 on Mixtral 8x7B at 700 W, H200 at 1,000 W on Llama 2 70B · read 16 Oct 2026
34. 34 [Verda, DeepSeek V3 inference benchmark on NVIDIA H200](https://verda.com/blog/deepseek-v3-llm-nvidia-h200-gpu-inference-benchmarking) SGLang 0.4.1 throughput on 8x H200, BF16 and FP8 · read 16 Oct 2026
35. 35 [LMSYS blog, large-scale expert parallelism for DeepSeek with SGLang (5 May 2025)](https://lmsys.org/blog/2025-05-05-large-scale-ep/) 52.3k input and 22.3k output tokens/s per node on 12 nodes of 8x H100; $0.20 per million output tokens · read 16 Oct 2026
36. 36 [Tom's Hardware, AMD posts first Instinct MI300X MLPerf results](https://www.tomshardware.com/tech-industry/artificial-intelligence/amd-posts-first-instinct-mi300x-mlperf-benchmark-results-roughly-in-line-with-nvidia-h100-performance) MI300X roughly in line with H100 on MLPerf Llama 2 70B · read 16 Oct 2026
37. 37 [The Next Platform, the first AI benchmarks pitting AMD against Nvidia](https://www.nextplatform.com/2024/09/03/the-first-ai-benchmarks-pitting-amd-against-nvidia/) MI300X about 60 percent of the H200 Llama 2 70B inference load in MLPerf v4.1 · read 15 Oct 2026
38. 38 [Ambati and Diep, MI300X inference performance analysis (arXiv 2510.27583)](https://arxiv.org/pdf/2510.27583) MI300X at 37 to 66 percent of H100 and H200 realised LLM inference · read 15 Oct 2026
39. 39 [SemiAnalysis, AMD vs NVIDIA inference benchmark (23 May 2025)](https://newsletter.semianalysis.com/p/amd-vs-nvidia-inference-benchmark-who-wins-performance-cost-per-million-tokens) MI300X on par with H200 on vLLM for Llama 3.1 405B; H200 ahead on TensorRT-LLM · read 15 Oct 2026
40. 40 [Mercatus, NVIDIA H200 server price 2026](https://www.mercatus-ai.com/blog/h200-server-price) $320,000 to $420,000, typical $370,000, module $32,000 to $40,000 (8 July 2026) · read 16 Oct 2026
41. 41 [Thunder Compute, NVIDIA H200 pricing (October 2026)](https://www.thundercompute.com/blog/nvidia-h200-pricing) HGX H200 server $350,000 to $450,000 · read 15 Oct 2026
42. 42 [Ace Computers, Matrix HGX H200 server](https://acecomputers.com/product/pmh-f6rarh/) Listed from $478,992 · read 15 Oct 2026
43. 43 [EIA, Electric Power Monthly, table 5.3](https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_3) US commercial average 13.97 cents per kWh, January to July 2026 (preliminary) · read 16 Oct 2026
44. 44 [Meta, Llama 3.3 70B Instruct model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) 70B parameters, 128K context, architecture used for the KV cache arithmetic · read 16 Oct 2026
45. 45 [Qwen, Qwen3-235B-A22B model card](https://huggingface.co/Qwen/Qwen3-235B-A22B) 235B total, 22B active parameters · read 16 Oct 2026
46. 46 [DeepSeek, DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3) 671B parameters, FP8 weights · read 16 Oct 2026
47. 47 [vLLM blog, Announcing Llama 3.1 support in vLLM](https://vllm.ai/blog/2024-07-23-llama31) 405B on a single 8xH200 machine, FP8 on 8xH100, 16 GPUs in BF16 · read 16 Oct 2026
48. 48 [OpenAI, gpt-oss-120b model card](https://huggingface.co/openai/gpt-oss-120b) 117B parameters, MXFP4, single 80 GB GPU · read 16 Oct 2026

What we could not confirm

Beam's $2.09 machine was read on Beam's own pricing page, but no source confirmed that a card was in stock and the page does not state its billing granularity, IP or traffic terms. Vast.ai's H200 page headline read $2.80, the same as the cheapest SXM ask in Vast's public search API that day, but the page does not separate SXM from NVL listings. Sesterce's $2.48 and Jarvislabs' $2.29 spot rate come from GetDeploying only, and Runpod's $3.59 Community rate appears only in its page's structured data, not on the visible card. GMI Cloud's page lists $3.20 where a 29 September reading and GMI's own blog had $2.60; we use the pricing page. Providers we could not re-read on 16 October (io.net, Shadeform, Seeweb, Gcore, E2E Networks, AceCloud, Clore.ai, Taiga Cloud, Northflank, Akash, Massed Compute, which is now labelled NVL) were left out rather than shown with stale prices. Storage prices for Nebius, Lyceum and CoreWeave were not read, Vast.ai storage at $0.15 per GB-month is an assumption, and the AWS egress estimate uses its published $0.09 per GB tier without a re-read. The KV cache capacities are our arithmetic, not measurements, and the colocation cost, support, cost of capital, 10.2 kW draw and resale value in the ownership model are assumptions. The server prices from Thunder Compute and Ace Computers and the AWS Capacity Block history come from third-party pages, and the GetDeploying rows for Gcore, Canopy Wave, Runcrate, Crusoe, OVHcloud and the reserved offers were not read on the providers' own pages.

H200 SXM on GPUBank

## A 70B model on one card, billed by the second.

- 24 vCPU, 276 GB RAM, 2 TB NVMe, IPv4 and traffic in $2.49/hr
- Billed by the second after the health check
- Interruptible $1.49, reserved $1.99

[Launch an H200 SXM · $2.49/hr](https://console.gpubank.io/launch?gpu=h200-sxm)

[Compare 63 GPU clouds](https://gpubank.io/compare)

More GPU price guides

## Where to rent the other cards at the cheapest price.

[B200 SXM 180 GB HBM3e From $2.49 spot · $4.15 all-in](https://gpubank.io/rent-b200)

[H100 SXM 80 GB HBM3 From $1.11 spot · $1.85 all-in](https://gpubank.io/rent-h100)

[H100 PCIe 80 GB HBM2e From $0.81 spot · $1.35 all-in](https://gpubank.io/rent-h100-pcie)

[MI300X 192 GB HBM3 From $0.99 spot · $1.65 all-in](https://gpubank.io/rent-mi300x)

[RTX PRO 6000 96 GB GDDR7 ECC From $0.69 spot · $1.15 all-in](https://gpubank.io/rent-rtx-pro-6000)

[A100 SXM 80 GB HBM2e From $0.57 spot · $0.95 all-in](https://gpubank.io/rent-a100)

[L40S 48 GB GDDR6 ECC From $0.33 spot · $0.55 all-in](https://gpubank.io/rent-l40s)

[RTX 5090 32 GB GDDR7 From $0.29 spot · $0.48 all-in](https://gpubank.io/rent-rtx-5090)

[RTX 4090 24 GB GDDR6X From $0.14 spot · $0.23 all-in](https://gpubank.io/rent-rtx-4090)