---
title: Where to Rent an H100 SXM at the Cheapest Price (October 2026)
url: https://gpubank.io/rent-h100
description: H100 SXM rental prices checked on 16 October 2026, from $0.69/hr spot to $12.29/hr on Azure, with what each hour includes, real bills and rent vs buy.
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-h100, 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 H100 SXM at the cheapest price

We read the H100 SXM offers on each provider's own page, priced what every hour really includes, and worked out the monthly bill, the cost per job and the point where buying an 8-GPU server would cost less than renting one.

**GPUBank Research** 19 min read · Prices read on 16 Oct 2026

Spot floor

$0.69

Vast.ai interruptible bid

Marketplace

$1.39

Lium, one GPU

GPUBank

$1.85

All-in, per second

Market median

$3.83

39 providers

AWS p5

$6.88

Per GPU of 8

Azure ND

$12.29

Per GPU of 8

The short answer, 16 Oct 2026

1. $0.69 /hr **Cheapest hour you can get.** Vast.ai interruptible capacity on a single unverified host in California, at a minimum bid of $0.687 per GPU-hour; the machine pauses when someone outbids you, and disk and bandwidth are billed on top. Nebius spot comes next, from $0.79 in eu-north1.
2. $1.39 /hr **Cheapest marketplace listing.** A single H100 80GB HBM3 pod on the Lium marketplace in Secaucus, with 8-GPU nodes in Helsinki from $1.69 per GPU; listings appear and disappear as hosts come and go, and no persistent volume is included.
3. $1.85 /hr **Cheapest managed price that includes the whole machine.** GPUBank, with 20 vCPU, 125 GB of RAM, a 2 TB local NVMe disk, a public IPv4 and all traffic in the rate, billed per second, or about $1.37 effective after a $100 top-up and $1.16 after a $1,000 one.

[H100 SXM $1.85 /hr all-in Launch an H100 SXM Spot $1.11 · reserved $1.48
Billed per second](https://console.gpubank.io/launch?gpu=h100-sxm)

The H100 SXM has the longest price list of any GPU you can rent. GetDeploying tracked 57 providers and 445 price points for the H100 on 15 October 2026, with a median on-demand price of $3.83 per GPU-hour across the 39 providers that publish one, and the spread around that median is enormous: the same hour of the same Hopper silicon cost $0.69 as an interruptible bid on Vast.ai and $12.29 per GPU on a Microsoft Azure ND H100 v5 virtual machine in East US. Neither figure is wrong. Each one buys a different machine, a different meter and a different promise about what happens when a host fails, which is why the cheapest line on a pricing page so often turns into a bill that is not the cheapest at the end of the month. We read each offer on the provider's own page on 15 and 16 October 2026, and where a page could not be read we name the aggregator and its date.

A good part of the market moved in the three weeks before this page was written, and almost all of it moved up. Nebius raised its on-demand H100 from $3.85 to $4.50 on 1 October and switched its preemptible tier to a dynamic spot price from 8 October; AWS now charges $5.97 per accelerator-hour for P5 Capacity Blocks in every region it lists; Verda lists $4.17, Crusoe $4.50 and JarvisLabs $3.49, all above the figures of late September, and CUDO Compute, which trackers still show at $1.79, now answers 'Quote on request' on its own pricing page. GPUBank, which publishes this page, is one of the providers compared, on the same terms as the others: we say where it is cheapest and where a marketplace listing, a spot tier or a reservation elsewhere beats it.

01 / How the price fell

## How the H100 hour fell, and why it stopped falling

The H100 entered the cloud at hyperscaler prices. Azure's ND H100 v5 has listed at $98.32 an hour for eight GPUs, $12.29 per GPU, since December 2023, and Silicon Data's series puts the hyperscaler median at about $7.76 per GPU-hour from August to December 2023, rising to $9.34 in the second half of 2024. The fall came first from below. The neocloud wave of 2024 put a median of $2.99 on specialised clouds and $2.58 on marketplaces between June and December 2024, and marketplace rates slid to $2.29 in the first five months of 2025 and $1.95 in the second half of the year. SemiAnalysis' index of one-year H100 contracts, compiled from monthly surveys of more than 100 market participants, bottomed at about $1.70 per GPU-hour in October 2025.

The hyperscalers followed in June 2025, when AWS cut P5 on-demand prices by 44 percent from 1 June, which brought the p5.48xlarge to $55.04 an hour, $6.88 per GPU; Silicon Data's hyperscaler median dropped to $6.94 that month and to $6.26 in the second half of 2025. That was the bottom. By March 2026 the SemiAnalysis contract index had climbed to $2.35, a rise of almost 40 percent in six months, Lambda raised its 8x H100 SXM rate from $2.99 to $3.99 on 9 June 2026, and in October Nebius moved from $3.85 to $4.50 and AWS Capacity Blocks reached $5.97.

One H100 SXM hour, 2023 to October 2026 Large clouds Specialised clouds Low end of the market

USD per GPU-hour. Each point is a dated reading, listed below by segment; the indices use their own methods and are not strictly comparable.

Large clouds

- Oct 2023 **$7.76** Silicon Data hyperscaler median, August to December 2023
- Dec 2023 **$12.29** Azure ND96isr H100 v5 list price, $98.32 per 8-GPU VM, in force since 1 December 2023
- Sep 2024 **$9.34** Silicon Data hyperscaler median, June to December 2024
- Jun 2025 **$6.88** AWS p5.48xlarge on demand after the 44 percent cut of 1 June 2025, $55.04 per node
- Oct 2026 **$6.88** AWS p5.48xlarge on demand, us-east-1, $55.04 per node

Specialised clouds

- Sep 2024 **$2.99** Silicon Data neocloud median, June to December 2024
- Oct 2025 **$3.33** Silicon Data neocloud median, July to December 2025
- Jun 2026 **$3.99** Lambda 8x H100 SXM on demand, raised from $2.99 on 9 June 2026
- Oct 2026 **$3.83** GetDeploying median on-demand price, 39 providers, 15 October 2026

Low end

- Sep 2024 **$2.58** Silicon Data marketplace median, June to December 2024
- Mar 2025 **$2.29** Silicon Data marketplace median, January to May 2025
- Oct 2025 **$1.70** SemiAnalysis one-year H100 contract index at its low
- Mar 2026 **$2.35** SemiAnalysis one-year H100 contract index, up almost 40 percent in six months
- Oct 2026 **$1.39** Cheapest single-GPU marketplace listing, Lium, 16 October 2026

Today the typical on-demand hour is far below the 2023 hyperscaler price but above the 2025 trough: GetDeploying's median was $3.83 across 39 providers on 15 October 2026, and Ornn's daily OCPI-H100 index, which moved between $2.49 and $3.21 over the previous month, settled at $3.13 that day. The cheapest hours did not rise with the median. A single Lium pod at $1.39 and Vast.ai bids from $0.69 sit below even the 2025 marketplace median, which is the practical lesson of the chart: a cheap H100 hour is no longer a matter of waiting for the market to fall, but of choosing the seller, the terms and the risk you are willing to carry.

### Why the large clouds look so expensive: one hour buys eight cards

AWS

p5.48xlarge

$55.04 /hr *÷ 8* **$6.88 /GPU-hr**

Minimum rental: all 8 cards

Google Cloud

a3-highgpu-8g

$88.49 /hr *÷ 8* **$11.06 /GPU-hr**

Minimum rental: all 8 cards

Microsoft Azure

Standard_ND96isr_H100_v5

$98.32 /hr *÷ 8* **$12.29 /GPU-hr**

Minimum rental: all 8 cards

Oracle Cloud

BM.GPU.H100.8

$80.00 /hr *÷ 8* **$10.00 /GPU-hr**

Minimum rental: all 8 cards

02 / The card

## What the H100 SXM is, and why one hour costs anything from $0.69 to $12.29

The H100 SXM is NVIDIA's Hopper datacenter GPU in its mezzanine form: 80 GB of HBM3 at 3.35 TB/s, NVLink at 900 GB/s between the GPUs of one server against 128 GB/s over PCIe Gen5, the FP8 Transformer Engine, 3,958 teraFLOPS of FP8 and 1,979 teraFLOPS of FP16 with sparsity, and a configurable power limit of up to 700 W. The module does not plug into a PCIe slot. It sits on an HGX baseboard of eight GPUs inside a complete server, which is why the 8-GPU node, not the single card, is the unit in which the H100 SXM is built, bought and very often rented. Two other products carry the same name on price lists and should be kept apart from it: the H100 PCIe, a slower add-in card with HBM2e, and the H100 NVL, a 94 GB PCIe variant with 3.9 TB/s of bandwidth, a 350 to 400 W limit and NVLink at 600 GB/s between pairs. GetDeploying notes that only about half of the providers it tracks label SXM explicitly, so a cheap 'H100 80GB' row may well be a PCIe card.

The people who rent it are mostly teams fine-tuning or serving open models between 8B and 70B parameters, research groups that need a full NVLink node for days at a time, and inference operators who want FP8 and a mature software stack at a lower price than an H200 or a B200. That demand explains the crowd of sellers. The range from $0.69 to $12.29 comes from four differences that most comparisons blur: the unit sold (a single-GPU slice with a share of CPU and RAM, or an 8-GPU node with two server CPUs, up to 2 TB of memory and tens of terabytes of NVMe), the fabric between servers (100 Gbps Ethernet, or 3,200 Gbps of InfiniBand that only matters for jobs spanning several nodes), the commitment (on demand, interruptible, a week, a year or three), and the seller (a hyperscaler list price, a specialised cloud, or an independent host on a marketplace whose low prices often come with a short maximum rental and a modest reliability record).

03 / The price ladder

## How to read an H100 SXM price list without comparing the wrong things

Six kinds of offer sit on the H100 SXM list, and each answers a different question. A managed on-demand cloud sells an instance at a fixed hourly price from its own fleet, with a stated amount of CPU, RAM and disk, and the thing to check is what the price leaves out: Runpod's $3.99 Secure Cloud pod includes 20 vCPU and 125 GB of RAM but not its volumes, whereas Lambda's $3.99 per GPU on its 8x instance comes with 208 vCPU, 1,800 GiB of RAM and 22 TiB of SSD. A peer-to-peer marketplace such as Vast.ai or Lium passes through a price set by an independent host, so two listings at the same price can be different machines in different countries with different disk and bandwidth charges. Spot or interruptible capacity is the same hardware sold on the understanding that it can be taken back, at anything from $0.69 on Vast.ai to $2.08 at Verda.

Reserved capacity is where the cheapest dependable hours are, but it is sold by the week, the month or the year, often by the whole node and often prepaid: Clore.ai's own dated snapshot quotes 8-GPU bare-metal contracts at $1.81 to $2.62 per GPU-hour with a 30-day minimum, Together AI charges $3.19 for 91 to 180 days and DigitalOcean $3.26 on a 12-month contract. Serverless platforms bill only while code runs, so Modal's $0.001097 per second, about $3.95 an hour for the GPU alone, can beat any always-on instance for a service that is busy a fraction of the day. The large clouds, finally, sell the H100 SXM almost exclusively as 8-GPU instances, so their per-GPU price is the instance price divided by eight and one hour of it commits you to eight cards. The fair comparison is always like with like: a marketplace headline against another marketplace headline, and a managed all-in price against what the managed competitor charges once CPU, RAM, disk, IPv4 and traffic are added.

*GetDeploying median, 39 providers $3.83*

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

*GPUBank**$1.85***

*Voltage Park**$1.99***

*Hyperstack**$3.20***

*JarvisLabs**$3.49***

*Runpod Secure Cloud**$3.99***

*Lambda**$3.99***

*Together AI**$3.99***

*Verda**$4.17***

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

*Lium**$1.39***

*Lium**$1.69***

*Vast.ai**$2.00***

*Runpod Community**$2.69***

*Akash**$2.71***

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

*Vast.ai**$0.687***

*Nebius**$0.79***

*GPUBank**$1.11***

*Together AI**$1.99***

*Verda**$2.08***

*Microsoft Azure**$2.27***

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

*GPUBank**$1.48***

*Clore.ai**$1.81***

*Hyperstack**$2.72***

*Together AI**$3.19***

*DigitalOcean**$3.26***

*Lambda 1-Click Clusters**$5.54***

*AWS Capacity Blocks**$5.97***

**Serverless** Billed by active seconds.

*Modal**$3.95***

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 H100 or pricing page, so every figure can be read next to the conditions printed beside it.

gpubank.io/pricing

**GPUBank** All-in

$1.85/hr

Managed cloud · 1 to 8 GPUs, NVLink

20 vCPU 125 GB RAM 2 TB NVMe IPv4 and all traffic Spot $1.11 Reserved $1.48 Persistent disk $0.03/GB-month

GPUBank sells the H100 SXM at $1.85 per GPU-hour with the whole machine in the price: 20 vCPU, 125 GB of RAM and a 2 TB local NVMe disk per GPU, a public IPv4 address and all inbound and outbound traffic, in machines of one to eight GPUs joined by NVLink at 900 GB/s. Billing is per second with no minimum, the meter starts only after the machine passes its health check, and a stopped machine costs nothing but its persistent disk at $0.03 per GB-month. Interruptible capacity costs $1.11 with five minutes of notice and a stop rather than a deletion, and a week or a month reserved costs $1.48. It is offered in Paris, Frankfurt, Ashburn and Singapore, with a median of 48 seconds from launch to a healthy machine. Payment is a prepaid balance topped up in crypto from $25, with bonuses that bring the effective hour to about $1.16 at $1,000; machines stop at eight GPUs, with no fabric between servers on offer.

Checked 16 Oct 2026

lium.io/gpu/h100

**Lium**

$1.39/hr

Marketplace · single pods and 8-GPU nodes

Host CPU and RAM share Local disk of the node 8x node from $1.69/GPU Stock varies by the hour

Lium is the cheapest H100 marketplace we could read on the provider's own page. On 16 October its H100 page listed a single-GPU pod in Secaucus at $1.39 per GPU-hour and 8-GPU nodes in Helsinki at $1.69 and $1.89 per GPU, with another 8-GPU node in Bellevue at $3.75. The cards are listed as H100 80GB HBM3, the SXM memory, and the Helsinki nodes come with 128 Xeon Platinum 8468 cores and 1,575 GB of RAM. Each listing is an independent provider's machine, so the CPU, the disk, the uplink and the reliability score change from one to the next; one 3-GPU host in Russia is marked with a reduced power limit, which means less throughput for the same price. Stock is the catch: the listings changed between two readings on the same day, and GPU Finder records limited availability for Lium's single-GPU offers over 30 days. We found no persistent volume in the price, so data has to be reloaded when a pod ends.

Checked 16 Oct 2026

vast.ai/pricing/gpu/H100-SXM

**Vast.ai**

$2.00/hr

Marketplace · host-set prices

26 vCPU 226 GB RAM Unverified host, California Interruptible from $0.69 Storage $0.20/GB-month Bandwidth per GB

Vast.ai is the deepest marketplace and the source of the cheapest H100 SXM hour on the market. Its public offer search returned 18 rentable H100 SXM offers on 16 October, the cheapest on demand at $2.00 per GPU-hour for one GPU on an unverified host in California with 26 vCPU and 226 GB of RAM, and the cheapest verified one at $2.13 in Germany. The same California host accepts interruptible bids from $0.687, the lowest H100 SXM price we found anywhere. Vast's own H100 SXM page advertised 'from $1.38/hr' when we captured it, a headline it recalculates every hour from live offers without saying which offer it is, so it can point to a multi-GPU host or one that is already rented. Vast's model is a bid: an interruptible instance pauses when someone outbids you or an on-demand renter takes the machine, and storage keeps billing while an instance is stopped. Each host sets its own storage price, $0.20 per GB-month on the cheapest one, and its own bandwidth price, about $0.004 per GB uploaded, so the headline is a floor rather than an all-in rate. Billing is per second from prepaid credit, and the reliability score shown on each listing is the first thing to filter on.

Checked 16 Oct 2026

runpod.io/pricing

**Runpod**

$2.69/hr

Community and Secure Cloud · 1 GPU

20 vCPU 125 GB RAM Secure Cloud $3.99 Network volume $0.07/GB-month

Runpod rents the H100 SXM in two clouds at two prices: $2.69 per hour on Community Cloud, where the machines belong to third-party hosts, and $3.99 on Secure Cloud. Both pods come with 20 vCPU and 125 GB of RAM per GPU, and both are billed per second from a prepaid balance. The page keeps the variants apart, which many lists do not: the H100 PCIe costs $1.99 and $2.89, and the H100 NVL $2.59 and $3.19, so a cheap Runpod 'H100' in a tracker is often the PCIe card. Storage is billed separately, with network volumes at $0.07 per GB-month under 1 TB and a stopped pod's volume disk at $0.20, which keeps datasets and checkpoints between sessions without paying for a running GPU. Secure Cloud rose in September; trackers that still show $3.49 for it are out of date. For a single GPU, Runpod Community is the cheapest well-known managed-style option after the marketplaces, but it costs $0.84 an hour more than GPUBank's all-in rate before storage.

Checked 16 Oct 2026

docs.nebius.com/compute/resources/pricing

**Nebius**

$4.50/hr

Managed cloud · eu-north1

vCPU and RAM per preset Per-second billing Spot from $0.79, dynamic Disks billed while stopped

Nebius shows how fast H100 prices are moving in October 2026. Its documentation lists the H100 NVLink platform, available only in eu-north1, at $3.85 per GPU-hour before 1 October, $4.50 from that date, and a preemptible price that became dynamic on 8 October, quoted as 'from $0.79'. The spot price can change as often as every 15 minutes, and the documented figure is the minimum, so the hour you actually pay depends on demand at the moment you start. Usage is billed per second against an hourly unit, and compute is not charged while a VM is stopped, although its disks are. For checkpointed batch work that can wait for a low spot price, Nebius is one of the cheapest H100 SXM hours in Europe; for steady on-demand use it is now among the more expensive specialised clouds, $2.65 an hour above GPUBank and level with Crusoe.

Checked 16 Oct 2026

lambda.ai/pricing

**Lambda**

$3.99/hr

Managed cloud · 1x to 8x instances

208 vCPU per 8x 1,800 GiB RAM 22 TiB SSD 1x at $4.29 1-Click Clusters $5.54 to $6.16

Lambda publishes one of the clearest H100 SXM price lists, with the per-GPU rate falling as the instance grows: $4.29 on a 1x instance with 26 vCPU, 225 GiB of RAM and 2.75 TiB of SSD, $4.19 on 2x, $4.09 on 4x and $3.99 on 8x, where the instance carries 208 vCPU, 1,800 GiB of RAM and 22 TiB of SSD, all plus applicable sales tax or VAT. That $3.99 was $2.99 until June 2026, one of the clearest signs that the H100 market turned upward this year. GetDeploying reported the SXM configurations out of stock on 15 October. Lambda's strength is the step beyond one node: its 1-Click Clusters of 16 to 256 H100s on InfiniBand cost $6.16 to $5.54 per GPU-hour for terms of two weeks to a year, which is the kind of offer a multi-node training run needs and that GPUBank does not sell.

Checked 16 Oct 2026

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

On a managed, all-inclusive, per-second basis, GPUBank's $1.85 is the lowest H100 SXM price we could confirm on a provider's own page. The hour includes 20 vCPU, 125 GB of RAM, a 2 TB local NVMe disk, one public IPv4 and all inbound and outbound traffic, in machines of one to eight GPUs with NVLink at 900 GB/s, and the only other line a bill can carry is a persistent disk at $0.03 per GB-month. The meter starts after the machine passes its health check, a failed launch is not billed, and a host failure during a job restarts the machine on a healthy host with the same disk and credits the interrupted session in full. Interruptible capacity costs $1.11 with five minutes of notice and a stop rather than a deletion, a week or a month reserved costs $1.48, and top-up bonuses bring the effective on-demand hour to about $1.54 with a $25 top-up, $1.37 with $100 and $1.16 with $1,000. The card is offered in Paris, Frankfurt, Ashburn and Singapore, with a median of 48 seconds from launch to a machine that has passed its health check, and any Docker image or a ready template for vLLM, SGLang, Unsloth, Axolotl, ComfyUI or JupyterLab can be started on it.

It loses on headline price to every spot tier below $1.11, Vast.ai's $0.69 bid and Nebius' dynamic price from $0.79 among them, and to Lium's marketplace pods at $1.39 for one GPU or $1.69 per GPU on an 8-GPU node, as long as such a listing is there when you need it. Its machines stop at eight GPUs and it lists no fabric between servers, so a job that must span 16 or 64 GPUs belongs on Lambda's 1-Click Clusters, Crusoe, Together AI or a hyperscaler, at two to four times the price per GPU-hour. Payment is a prepaid US dollar balance topped up in crypto from $25 to $5,000; card payments are not accepted for now, which rules GPUBank out for a buyer whose finance team can only pay by card or invoice, and the bonus only lowers the real price if the credit is actually spent, although the balance never expires.

### The rest of the market

Beyond the six providers profiled in detail, the managed tier is crowded and has mostly moved up. Voltage Park says on its own pricing page that one H100 can be rented for an hour from $1.99 without a contract, with no ingress, egress or support fees, while its Ethernet and InfiniBand configurations are priced on request. GetDeploying shows UpCloud at $1.89 an hour and $1.88 spot, both out of stock and without a stated variant. Hyperstack's H100 SXM page gives $3.20 on demand and reservations from $2.72, under a banner saying the card is only available through its Supercloud reservation, so treat the on-demand figure as a list price rather than a machine you can start today. JarvisLabs charges $3.49 per minute-billed hour with 16 vCPU and 200 GB of RAM, Verda $4.17 with 30 vCPU and 120 GB, DigitalOcean $4.41 with 20 vCPU, 240 GiB, 5 TiB of scratch NVMe and 15,000 GiB of transfer, and Crusoe and Nebius both $4.50.

Together AI publishes $3.99 on demand and $1.99 preemptible for its HGX H100 clusters, with reservations from $3.69 for 3 to 30 days down to $3.19 for 91 to 180 days. CoreWeave sells by the 8-GPU node, at $6.16 per GPU on demand on GetDeploying's reading, with spot from $2.46. HyperAI shows an H100 at $1.80 an hour on its own page, which would undercut most of the list, but the page does not say SXM or PCIe and links to a PCIe benchmark, so we leave it out of the SXM comparison. Spheron lists $2.65 on demand and $2.20 spot per minute after a 20-minute minimum, again without naming the form factor. CUDO Compute, often quoted at $1.79 for the GPU alone, now shows 'Quote on request' for the H100 SXM.

Among marketplaces, Clore.ai's own page says per-minute listings start from $0.99 per GPU-hour at host-set prices, as of 21 September 2026, without stating whether the cheapest listing is an SXM or a PCIe card; its bare-metal contracts are the clearer offer, from $1.81 per GPU-hour on terms of 180 days or more in the USA, Slovenia and Japan. Akash's own snapshot of 12 September put its live weighted-average H100 rate at about $2.71, for the GPU only, with CPU, memory, storage and IP priced in each provider's bid. GetDeploying also records Novita at $1.70 spot and $3.39 on demand, TensorDock at $2.36, Denvr at $2.45 for 8-GPU nodes, Civo at $2.99 on demand and $2.49 on a 36-month term, and Thunder Compute at $3.20 but out of stock, none of which we could read on the provider's own page today. The same tracker lists the cheapest reservations below GPUBank's on-demand rate, HyperAI at $1.53 for one month and a 48-month Vast.ai reservation at $1.62, both close to GPUBank's $1.48 reserved hour but on longer terms or an unstated form factor.

### The large clouds

The four large clouds sell the H100 SXM by the 8-GPU instance, and their list prices are three to seven times those of the specialised clouds. AWS's p5.48xlarge costs $55.04 an hour on demand in us-east-1, $6.88 per GPU, with 192 vCPU and 2,048 GiB of RAM, and EBS storage and egress on top; that price dates from June 2025, when AWS cut P5 on-demand rates by 44 percent. Google Cloud's a3-highgpu-8g lists at $88.49 an hour in us-central1, $11.06 per GPU. Azure's Standard_ND96isr_H100_v5 costs $98.32 an hour for Linux in East US, $12.29 per GPU, a rate unchanged since December 2023. Oracle's price list charges $10.00 per GPU-hour for the H100 shape, $80 per node, the same in every public region.

Commitments and spot change those figures sharply, which is how large customers actually pay. Azure's own price list puts its ND H100 v5 spot rate at $18.17 per VM-hour, $2.27 per GPU, and its one-year and three-year reservations at about $7.87 and $5.40 per GPU. Google's a3-highgpu-8g spot price works out to $6.29 per GPU and its three-year committed-use price to about $4.86, AWS's three-year reserved p5.48xlarge to about $2.97 per GPU, and AWS Capacity Blocks, which reserve P5 capacity for a fixed window, cost $47.757 per instance-hour, $5.97 per accelerator, in every listed region, up from about $5.19 in US regions before the October revision, so a guaranteed P5 window now costs almost as much as the on-demand hour. What these prices buy is integration with the rest of the cloud, compliance coverage, and large EFA, InfiniBand or RDMA clusters; for raw GPU-hours on one node they are rarely the cheap choice.

05 / Every offer

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

| Provider | USD / GPU-hr | Kind | Unit | What the price includes | Billing | Stock | Basis |
| --- | --- | --- | --- | --- | --- | --- | --- |
| GPUBank | $1.85 | Managed on demand | 1 GPU | 20 vCPU, 125 GB RAM, 2 TB NVMe, IPv4, traffic | Per second | In stock | Provider page |
| Voltage Park | $1.99 | Managed on demand | 1 GPU | Starting price; no egress fees | Not stated | Not stated | Provider page |
| Hyperstack | $3.20 | Managed on demand | 1 GPU | List price; page says reservation only | Not stated | On request | Provider page |
| JarvisLabs | $3.49 | Managed on demand | 1 GPU | 16 vCPU, 200 GB RAM; storage extra | Per minute | Not stated | Provider page |
| Runpod Secure Cloud | $3.99 | Managed on demand | 1 GPU | 20 vCPU, 125 GB RAM; volumes extra | Per second | In stock | Provider page |
| Lambda | $3.99 | Managed on demand | 8-GPU node | 208 vCPU, 1,800 GiB RAM, 22 TiB SSD | Not stated | Sold out | Provider page |
| Together AI | $3.99 | Managed on demand | cluster | HGX H100 cluster; node not itemised | Not stated | Not stated | Provider page |
| Verda | $4.17 | Managed on demand | 1 GPU | 30 vCPU, 120 GB RAM | Not stated | Not stated | Provider page |
| Lambda | $4.29 | Managed on demand | 1 GPU | 26 vCPU, 225 GiB RAM, 2.75 TiB SSD | Not stated | Not stated | Provider page |
| DigitalOcean | $4.41 | Managed on demand | 1 GPU | 20 vCPU, 240 GiB RAM, 5 TiB scratch, 15,000 GiB transfer | Not stated | Not stated | Provider page |
| Nebius | $4.50 | Managed on demand | 1 GPU | vCPU and RAM per preset; eu-north1 | Per second | Not stated | Provider page |
| Crusoe | $4.50 | Managed on demand | 1 GPU | HGX H100; machine not detailed on the page | Not stated | Sold out | Provider page |
| Lium | $1.39 | Marketplace | 1 GPU | Host CPU, RAM and local disk | Per second | In stock | Provider page |
| Lium | $1.69 | Marketplace | 8-GPU node | 128 cores, 1,575 GB RAM, 895 GB disk | Per second | In stock | Provider page |
| Vast.ai | $2.00 | Marketplace | 1 GPU | 26 vCPU, 226 GB RAM; disk and bandwidth extra | Per second | In stock | Provider page |
| Runpod Community | $2.69 | Marketplace | 1 GPU | 20 vCPU, 125 GB RAM; volumes extra | Per second | Not stated | Provider page |
| Akash | $2.71 | Marketplace | 1 GPU | GPU only; rest in the provider bid | Lease | Not stated | Provider snapshot |
| Vast.ai | $0.687 | Spot and interruptible | 1 GPU | Minimum bid; pauses if outbid | Per second | In stock | Provider page |
| Nebius | $0.79 | Spot and interruptible | 1 GPU | Dynamic minimum, changes every 15 min | Per second | Not stated | Provider page |
| GPUBank | $1.11 | Spot and interruptible | 1 GPU | As on demand; 5 min notice, disk kept | Per second | In stock | Provider page |
| Together AI | $1.99 | Spot and interruptible | cluster | Preemptible HGX H100 | Not stated | Not stated | Provider page |
| Verda | $2.08 | Spot and interruptible | 1 GPU | 30 vCPU, 120 GB RAM | Not stated | Not stated | Provider page |
| Microsoft Azure | $2.27 | Spot and interruptible | 8-GPU node | ND96isr H100 v5 spot, East US | Not stated | Not stated | Provider page |
| GPUBank | $1.48 | Reserved | 1 GPU | Week or month, prepaid | Per term | In stock | Provider page |
| Clore.ai | $1.81 | Reserved | 8-GPU node | Bare metal, 180 days or more | Per term | On request | Provider snapshot |
| Hyperstack | $2.72 | Reserved | 1 GPU | Starting reservation price | Per term | On request | Provider page |
| Together AI | $3.19 | Reserved | cluster | 91 to 180 days | Per term | On request | Provider page |
| DigitalOcean | $3.26 | Reserved | 1 GPU | 12-month contract | Per term | On request | Provider page |
| Lambda 1-Click Clusters | $5.54 | Reserved | cluster | 256 GPUs on InfiniBand, 2 weeks to 1 year | Per term | On request | Provider page |
| AWS Capacity Blocks | $5.97 | Reserved | 8-GPU node | p5.48xlarge for a fixed window; EBS extra | Per term | Not stated | Provider page |
| AWS | $6.88 | Large clouds | 8-GPU node | p5.48xlarge, 192 vCPU, 2,048 GiB; EBS extra | Per second | Not stated | Aggregator |
| Oracle Cloud | $10.00 | Large clouds | 8-GPU node | BM.GPU.H100.8 bare metal | Not stated | Not stated | Provider page |
| Google Cloud | $11.06 | Large clouds | 8-GPU node | a3-highgpu-8g, us-central1 | Per second | Not stated | Aggregator |
| Microsoft Azure | $12.29 | Large clouds | 8-GPU node | ND96isr H100 v5, 96 vCPU, 1,900 GiB | Not stated | Not stated | Provider page |
| Modal | $3.95 | Serverless | 1 GPU | GPU only; CPU and RAM metered | 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 H100 SXM hour really costs once storage, traffic and lost time are added

Storage is the most common line that the hourly rate hides. On Vast.ai each host sets a storage price that keeps running for as long as the instance exists, stopped or not; the cheapest H100 SXM host on 16 October charged $0.20 per GB-month, others up to $0.66. Runpod bills network volumes at $0.07 per GB-month and a stopped pod's volume disk at $0.20, Nebius stops charging for compute when a VM is stopped but keeps billing its disks, and GPUBank charges only a persistent disk at $0.03 per GB-month, so 500 GB kept all month costs $15. Traffic is the second line. Vast.ai hosts bill bandwidth per gigabyte, the cheapest host at about $0.004 per GB of upload, AWS and Google Cloud bill egress separately from their instances, and DigitalOcean includes 15,000 GiB of transfer before charging. For an inference service sending 5 TB a month, these terms move the bill by anything from nothing to several hundred dollars.

Billing granularity decides what a short job costs. A job that runs 61 minutes costs 61 sixtieths of $1.85, or $1.88, on GPUBank, which bills per second with no minimum; Nebius and Vast.ai also bill by the second, JarvisLabs and Spheron by the minute, Spheron after a 20-minute minimum, and a cloud that bills each started hour charges the same job two hours, which doubles it. Prepaid balances add their own terms: Vast.ai, Runpod, Lium, Clore.ai and GPUBank all draw from credit bought in advance, so the questions are whether unused credit expires and what happens when the balance reaches zero. A reservation is the other prepaid form. At GPUBank, $1.48 divided by $1.85 is 0.80, so a reservation only beats on demand if the cards would otherwise run at least 80 percent of the reserved hours.

The last hidden cost is time you pay for and cannot use, and the question is who carries it. On a marketplace, a host that goes offline ends the session, everything since the last checkpoint is lost, and the renter pays for those hours; a reclaimed spot instance costs the work since the last checkpoint plus the time to relaunch elsewhere. A listing labelled 'H100' may also be slower than expected: Lium shows one 3x H100 host with a reduced power limit, and in a public TensorRT-LLM issue a user who could not reproduce NVIDIA's documented throughput turned out to be running an H100 PCIe rather than the SXM the documentation used. A short throughput test on any new listing, before committing a long job to it, is cheap insurance. GPUBank's terms put the cost of a host failure on the provider by crediting the interrupted session, and credit the last 30 minutes if a spot notice is missed.

### Three workloads, priced for a month

**GPUBank spot**

120 h GPU $133.20

Storage $15.00

Other Rework after a reclaim

Month $148.20

Checkpointed sessions at $1.11

All-in

**Lium**

120 h GPU $166.80

Storage $0.00

Other Data reloaded each session

Month $166.80

A $1.39 pod free when needed

**GPUBank**

120 h GPU $222.00

Storage $15.00

Other None

Month $237.00

On demand at $1.85, all-in

All-in

**Vast.ai**

120 h GPU $240.00

Storage $100.00

Other Bandwidth per GB

Month $340.00

$2.00 host, disk at $0.20/GB-month

**Runpod Community**

120 h GPU $322.80

Storage $35.00

Other None

Month $357.80

$2.69 pod, network volume at $0.07

GPUBank on demand costs 120 times $1.85, or $222, plus $15 for 500 GB of persistent disk, $237 in all, and paid from $1,000 top-ups credited as $1,600 it comes to about $148 of cash. Its interruptible tier brings the list price to $148.20 for checkpointed sessions. Lium's $1.39 pod is cheaper on compute, $166.80, but only when a listing is free at the moment you sit down, and with no persistent volume in the price the datasets are reloaded each time. Vast.ai's cheapest host costs $240 of GPU time, and because its storage at $0.20 per GB-month keeps billing for as long as the instance exists, 500 GB adds $100. Runpod Community at $2.69 with a $35 network volume comes to $357.80. At list prices the order is GPUBank interruptible, then Lium, then GPUBank on demand; Lium beats on-demand GPUBank only for a user who can work around marketplace stock and reload data each session, and loses that lead as soon as the GPUBank balance is bought in large top-ups.

07 / Measured per job

## What the H100 SXM delivers per job, measured rather than claimed

The H100 SXM is the right card for LoRA and full fine-tuning of models up to about 70B parameters on one 8-GPU node, for FP8 inference of 8B to 70B models, and for training that needs NVLink bandwidth between eight GPUs. It stops being the right card when the model and its KV cache no longer fit in 80 GB per GPU, which is where the 141 GB H200 or the 180 GB B200 saves both GPUs and time; when the job is small enough that HBM3 and NVLink sit idle, as with a 7B LoRA or speech transcription, which an A100 does for about half the price; and when a job spans many nodes, where the fabric between servers matters more than the card.

The cleanest training measurement is MLPerf Training's Llama 2 70B LoRA test, a fine-tune on the GovReport dataset at a sequence length of 8,192, timed to a fixed quality target. On one 8-GPU H100 node, NVIDIA's verified entry 4.1-0050 in MLPerf Training v4.1 finished in 27.93 minutes, which is 3.72 GPU-hours, or $6.89 at GPUBank's on-demand rate. In v4.0 an 8-GPU H200 system finished in 24.7 minutes, 3.29 GPU-hours or $8.20 at GPUBank's $2.49, and in v5.1 an 8-GPU HGX B200 system finished in 9.55 minutes, 1.27 GPU-hours or $5.28 at $4.15. The B200 is the cheapest per job despite costing more than twice as much per hour, and the H200 is the most expensive per job on this test, because its gain in speed is smaller than its gain in price. A real project takes longer than the benchmark, since MLPerf uses a tuned NVIDIA stack, a modest dataset and a single successful run.

For serving, the cleanest cross-card measurement is MLPerf Inference v5.0, where NVIDIA's DGX H100 (entry 5.0-0057) served Llama 2 70B at 31,106 tokens per second in the server scenario and 31,307 offline, against 33,072 and 34,988 for NVIDIA's 8-GPU H200 system and 98,443 and 98,858 for the DGX B200, with both Hopper systems in FP8, the B200 in FP4, and TensorRT 10.8 and CUDA 12.8 throughout. Divided by eight, the H100 makes about 14 million tokens an hour per GPU, $0.132 per million at GPUBank's $1.85, against $0.167 on the H200 at $2.49 and $0.094 on the B200 at $4.15. On a 70B model that fits on both Hopper cards, the H200's 6 percent more throughput does not pay for its 35 percent higher price, and the B200 is cheaper per token only when the stack runs in FP4.

For inference, NVIDIA's TensorRT-LLM performance overview, generated with trtllm-bench on the PyTorch backend at maximum load on synthetic data, reports 14,992 output tokens per second for Llama 3.1 8B in FP8 on a single H100 SXM at 1,000 input and 1,000 output tokens, 26,401 at 128 and 128, and 9,462 at 2,048 and 2,048. At $1.85 an hour, 14,992 tokens per second is about 54 million tokens an hour, or $0.034 per million output tokens. Llama 3.3 70B in FP8 on two H100s with tensor parallelism of 2 reaches 4,181 tokens per second at 1,000 and 1,000, or about $0.25 per million on two GPUs at $3.70 an hour. These are reference points under full batching, not what a single user sees, and a user who measured 15,897 tokens per second where the documentation for v0.20 expected 27,688 at 128 and 128 was told the gap came mainly from running a PCIe card.

**Llama 2 70B LoRA fine-tune, time to train** MLPerf Training, closed division, one 8-GPU node, GovReport at sequence length 8,192 to a fixed quality target

8x H100 SXM, v4.1

27.9 minutes

8x H200 SXM, v4.0

24.7 minutes

8x B200 in GB200, v5.0

11.1 minutes

8x HGX B200, v5.1

9.55 minutes

Verified MLPerf results: H100 entry 4.1-0050 (v4.1) and H200 (v4.0) from NVIDIA, GB200 entry 5.0-0071, HGX B200 from Nebius (v5.1). The H100 and H200 rows come from different rounds; NVIDIA's unverified June 2025 H200 figure of 21.84 minutes is not shown.

**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), 5.0-0056 (DGX B200), all NVIDIA. Offline: 31,307, 34,988, 98,858.

**Llama 3.1 8B FP8 inference on one H100 SXM** TensorRT-LLM performance overview, trtllm-bench throughput mode, PyTorch backend, synthetic data; input / output tokens per request

128 / 128 tokens

26,401 output tok/s

1k / 1k tokens

14,992 output tok/s

2k / 2k tokens

9,462 output tok/s

NVIDIA reference figures at maximum load on an H100 SXM 80GB, page updated 15 September 2025 (commit 0c9430e). Not single-user speeds.

MARKET MEDIAN

$3.83

On demand per GPU-hour, 39 providers, 15 Oct 2026

AWS P5 CUT

44 %

On-demand price reduction from 1 June 2025

CONTRACT REBOUND

+38 %

One-year H100 contracts, Oct 2025 to Mar 2026

08 / H200, B200 or A100

## When an H200, a B200 or an A100 is the better rental per job

Per hour the H100 SXM sits in the middle of GPUBank's datacenter cards, at $1.85 against $0.95 for the A100 SXM, $1.35 for the H100 PCIe, $1.65 for the MI300X, $2.49 for the H200 SXM and $4.15 for the B200 SXM. Per job the order changes: on the MLPerf Llama 2 70B LoRA fine-tune above, the B200 is the cheapest run and the H200 the dearest, so a team running large fine-tunes all month pays less on the more expensive card, and the MLPerf Inference v5.0 results give the same order for serving a 70B model: $0.094 per million tokens on the B200, $0.132 on the H100 and $0.167 on the H200. The H200 earns its premium in memory rather than speed: with 141 GB of HBM3e at 4.8 TB/s, a 70B model in FP8 fits on one GPU with room for a long KV cache, where the H100 needs two GPUs or a shorter context.

At the other end, the A100 SXM at $0.95 has the same 80 GB but HBM2e at 2.0 TB/s and no FP8, so it remains the cheaper tool for small LoRA runs, embedding jobs and inference of models that run in 16-bit anyway, and the H100 PCIe at $1.35 suits single-GPU inference that never uses NVLink. The MI300X at $1.65 offers 192 GB of HBM3 at 5.3 TB/s, more memory per dollar than any NVIDIA card here, for teams whose stack runs well on ROCm. We did not find measured results for the A100, the H100 PCIe or the MI300X on the same benchmarks and settings, so we do not compute a cost per job for them.

A100 SXM

**80 GB** HBM2e · 2.0 TB/s

$0.95 /hr

LoRA and QLoRA fine-tuning, batch inference

H100 PCIe

**80 GB** HBM2e · 2.0 TB/s

$1.35 /hr

Single-GPU training and inference on a budget

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 180 GB.

09 / Rent or buy

## Rent or buy an 8-GPU HGX H100 server

The realistic purchase unit for the H100 SXM is a whole 8-GPU HGX server, because the module only runs on an HGX baseboard. Mercatus, writing in July 2026, gives $250,000 to $320,000 for a new system with about $285,000 typical from the major OEMs, and CloudZero gives the same range in its October 2026 update. Used machines are cheaper and thinly traded: Rack Rate's index of 18 July 2026 showed a median of $250,000 for used, tested 8-GPU systems, in a range of $225,000 to $275,000, and $265,000 for refurbished ones with a 90-day warranty. Mercatus' own resale estimates for a system after 36 months run from $62,000 to $146,000 counting the GPUs alone, which shows how fast this value is falling as Blackwell ships.

For the arithmetic we take a new system at $285,000, worth 45 percent of that after two years, or $128,250, and add $2,000 a month for colocation, network, remote hands and spares, $48,000 over 24 months; all three are our assumptions, not quotes. The fixed cost is then $204,750 over 24 months, about $8,531 a month, and the money tied up in the server, at 8 percent a year on $285,000, adds about $1,900, for $10,431 a month. For power we take the roughly 10.2 kW peak per system in NVIDIA's DGX H100 specifications and SuperPOD design guide at the US commercial average of 13.97 cents per kWh for January to July 2026, about $1.42 per node-hour under load. Renting the same eight GPUs at GPUBank costs $14.80 per node-hour on demand, so owning would need 10,431 / (14.80 - 1.42) = 780 node-hours a month to break even, more than the 730 a month contains.

Leave the cost of capital out, as a buyer paying cash that would otherwise earn nothing may, and the break-even with on demand falls to about 638 node-hours, 87 percent of the month; against GPUBank's reserved rate, $11.84 per node-hour, it is about 819 hours and against the effective $9.25 of $1,000 top-ups about 1,090, so a new server never wins against either. A used system at $225,000 that resells for $90,000 lowers the cash-only threshold to about 570 node-hours, but the resale assumption moves everything: every 10 points of resale value on a $285,000 system changes the monthly cost by about $1,190. Buy only if a team will keep an 8-GPU node busy more than nine hours in ten for two years against on-demand rates, already has power and rack space, and can carry the resale risk; everyone else pays less by renting, and least by renting reserved or prepaid capacity.

Monthly cost of one 8-GPU HGX H100 server (8 GPUs) against hours of use a month, 0 to 730. Owning: price + $93,600 host − $128,250 resale over 24 months, plus 10.2 kW at the wall under load. Rent: per GPU-hour × 8.

| Hours a month | Rent at $1.85 | Rent at $1.16 effective | Own | Cheaper |
| --- | --- | --- | --- | --- |
| 100 | $1,480.00 | $925.00 | $10,573.74 | Rent |
| 200 | $2,960.00 | $1,850.00 | $10,716.24 | Rent |
| 300 | $4,440.00 | $2,775.00 | $10,858.73 | Rent |
| 400 | $5,920.00 | $3,700.00 | $11,001.23 | Rent |
| 500 | $7,400.00 | $4,625.00 | $11,143.72 | Rent |
| 600 | $8,880.00 | $5,550.00 | $11,286.21 | Rent |
| 730 | $10,804.00 | $6,752.50 | $11,471.46 | Rent |

Typical new OEM price, Mercatus (14 July 2026) and CloudZero (updated 9 October 2026). The $93,600 shown as host is 24 months of colocation and maintenance at $2,000 a month plus cost of capital at 8 percent a year ($45,600); resale at 45 percent is our assumption; power at the EIA US commercial average.

10 / What to choose

## The cheapest H100 depends on the job.

You serve an 8B to 70B model around the clock

A reserved month, or a managed on-demand GPU

A month reserved at GPUBank costs $1.48 an hour with traffic included, and a production endpoint should not sit on a marketplace host that can disappear overnight.

You fine-tune in working sessions

An all-in on-demand GPU, or spot if you checkpoint

Per-second billing and a cheap persistent disk matter more than the headline rate when the GPU runs 120 hours a month and the data stays all month.

You train across more than one 8-GPU node

An InfiniBand cluster from Lambda, Together AI or a hyperscaler

NVLink stops at the edge of the server; beyond eight GPUs the fabric decides the speed, and only clusters sell it, at $5.54 to $6.16 per GPU-hour on Lambda.

You run short, restartable experiments

Spot or a marketplace listing

Vast.ai bids from $0.69 and Lium pods from $1.39 are the cheapest H100 hours there are, as long as losing a host costs you minutes rather than days.

11 / Questions

## What people ask before renting one.

The cheapest H100 SXM hour on 16 October 2026 was an interruptible bid on Vast.ai at $0.687 per GPU-hour, on an unverified host in California that pauses your instance if someone outbids you. Nebius' dynamic spot price starts at $0.79 in eu-north1. The cheapest hour that cannot be taken back was a single-GPU Lium marketplace pod at $1.39, with 8-GPU nodes from $1.69 per GPU, although listings come and go during the day. Among managed clouds that fix the machine and put CPU, RAM, local disk, an IPv4 address and traffic in the price, the lowest we confirmed on a provider's own page is GPUBank at $1.85, billed by the second, or about $1.16 effective with a $1,000 top-up. CUDO Compute, still shown at $1.79 by trackers, now quotes on request.

Sources

## Sources

1. 01 [Vast.ai, public offer search API (H100 SXM, rentable offers)](https://console.vast.ai/api/v0/bundles/) Cheapest on-demand and interruptible H100 SXM offers, host storage and bandwidth rates · read 16 Oct 2026
2. 02 [Nebius documentation, Compute pricing](https://docs.nebius.com/compute/resources/pricing) H100 NVLink on demand and spot prices by effective date, billing · read 16 Oct 2026
3. 03 [Lium, H100 GPUs](https://lium.io/gpu/h100) Live H100 pods and nodes, prices per GPU-hour · read 16 Oct 2026
4. 04 [GPUBank, pricing (updated 15 October 2026)](https://gpubank.io/pricing) GPUBank rates for all cards, inclusions, billing, top-up bonuses · read 16 Oct 2026
5. 05 [GetDeploying, Nvidia H100 rental prices (data of 15 October 2026)](https://getdeploying.com/gpus/nvidia-h100) Median on-demand price, provider count, aggregator rows for UpCloud, CoreWeave, Novita, Civo, Denvr, TensorDock, Thunder Compute · read 16 Oct 2026
6. 06 [Azure Retail Prices API, Standard_ND96isr_H100_v5, East US](https://prices.azure.com/api/retail/prices?$filter=armSkuName%20eq%20%27Standard_ND96isr_H100_v5%27%20and%20armRegionName%20eq%20%27eastus%27) On-demand, spot and reservation prices with effective dates · read 16 Oct 2026
7. 07 [AWS, EC2 Capacity Blocks for ML pricing](https://aws.amazon.com/ec2/capacityblocks/pricing/) $47.757 per instance-hour, $5.970 per accelerator-hour for p5.48xlarge in every listed region · read 16 Oct 2026
8. 08 [Verda, pricing](https://verda.com/pricing) H100 SXM5 on demand and spot rates · read 16 Oct 2026
9. 09 [Crusoe Cloud, pricing](https://www.crusoe.ai/cloud/pricing) H100 on-demand rate · read 16 Oct 2026
10. 10 [JarvisLabs, pricing](https://jarvislabs.ai/pricing) H100 SXM rate, machine, per-minute billing · read 16 Oct 2026
11. 11 [CUDO Compute, pricing](https://www.cudocompute.com/pricing) H100 SXM quote on request · read 16 Oct 2026
12. 12 [NVIDIA H100 Tensor Core GPU](https://www.nvidia.com/en-us/data-center/h100/) H100 SXM and NVL specifications · read 16 Oct 2026
13. 13 [Runpod, pricing](https://www.runpod.io/pricing) H100 SXM, PCIe and NVL rates on Community and Secure Cloud · read 16 Oct 2026
14. 14 [Lambda, GPU cloud pricing](https://lambda.ai/pricing) H100 SXM 1x to 8x instance rates, 1-Click Cluster rates · read 16 Oct 2026
15. 15 [Clore.ai, Rent H100 (snapshot of 21 September 2026)](https://clore.ai/rent-h100.html) Marketplace starting price, bare-metal contract rates and terms · read 16 Oct 2026
16. 16 [Together AI, pricing](https://www.together.ai/pricing) HGX H100 on demand, preemptible and reserved rates · read 16 Oct 2026
17. 17 [DigitalOcean, GPU Droplets pricing](https://www.digitalocean.com/pricing/gpu-droplets) H100 on demand and 12-month rates, machine, transfer · read 16 Oct 2026
18. 18 [Modal, pricing](https://modal.com/pricing) H100 per-second rate, CPU and memory rates · read 16 Oct 2026
19. 19 [Voltage Park, pricing](https://www.voltagepark.com/pricing) H100 starting price, no egress fees · read 16 Oct 2026
20. 20 [Hyperstack, NVIDIA H100 SXM](https://www.hyperstack.cloud/nvidia-h100-sxm) On-demand and reservation prices, reservation-only banner · read 16 Oct 2026
21. 21 [HyperAI, pricing](https://hyper.ai/en/pricing) H100 at $1.80, variant not stated · read 16 Oct 2026
22. 22 [Spheron, GPU rental](https://www.spheron.network/gpu-rental/) H100 on demand and spot rates, 20-minute minimum · read 16 Oct 2026
23. 23 [Akash Network, The Bid: H100 rental price 2026 (checked 12 September 2026)](https://akash.network/the-bid/h100-rental-price-2026-cost-per-hour/) Akash weighted-average H100 rate · read 16 Oct 2026
24. 24 [Vantage, p5.48xlarge pricing (from AWS price data)](https://instances.vantage.sh/aws/ec2/p5.48xlarge) On-demand, spot and three-year reserved price, specifications · read 16 Oct 2026
25. 25 [AWS News Blog, up to 45% price reduction for EC2 NVIDIA GPU instances (5 June 2025)](https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/) 44 percent P5 on-demand cut from 1 June 2025 · read 16 Oct 2026
26. 26 [gcloud-compute.com, a3-highgpu-8g (from Google Cloud SKU data, updated 11 October 2026)](https://gcloud-compute.com/a3-highgpu-8g.html) On-demand, spot and committed-use prices in us-central1 · read 16 Oct 2026
27. 27 [Oracle Cloud price list API, Compute GPU H100 (part B98415)](https://apexapps.oracle.com/pls/apex/cetools/api/v1/products/?currencyCode=USD) $10.00 per H100 GPU-hour, pay as you go · read 16 Oct 2026
28. 28 [GPUBank, home page](https://gpubank.io/) Regions, templates, median launch time, host failure credit · read 16 Oct 2026
29. 29 [GetDeploying, Vast.ai review and pricing](https://getdeploying.com/vast-ai) Vast.ai billing model, storage billed while stopped · read 16 Oct 2026
30. 30 [GitHub, NVIDIA/TensorRT-LLM issue 6294](https://github.com/NVIDIA/TensorRT-LLM/issues/6294) Reproduction gap attributed to an H100 PCIe · read 16 Oct 2026
31. 31 [NVIDIA Technical Blog, Blackwell in MLPerf Training v5.0](https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/) Llama 2 70B LoRA, 8x H100 27.93 min (4.1-0050), GB200 11.14 min (5.0-0071) · read 16 Oct 2026
32. 32 [NVIDIA Technical Blog, MLPerf Training v4.0 records](https://developer.nvidia.com/blog/nvidia-sets-new-generative-ai-performance-and-scale-records-in-mlperf-training-v4-0) Llama 2 70B LoRA, 8x H200 24.7 min · read 16 Oct 2026
33. 33 [Nebius, MLPerf Training v5.1 results](https://nebius.com/blog/posts/mlperf-training-v5-1-results) Llama 2 70B LoRA, 8x HGX B200 9.55 min · read 16 Oct 2026
34. 34 [MLCommons, MLPerf Inference v5.0 results (summary_results.json)](https://github.com/mlcommons/inference_results_v5.0) Closed datacenter entries 5.0-0057 (DGX H100), 5.0-0060 (H200), 5.0-0056 (DGX B200): Llama 2 70B, TensorRT 10.8 · read 16 Oct 2026
35. 35 [NVIDIA TensorRT-LLM, performance overview](https://nvidia.github.io/TensorRT-LLM/performance/perf-overview.html) Llama 3.1 8B and Llama 3.3 70B FP8 throughput on H100 SXM · read 16 Oct 2026
36. 36 [Mercatus, NVIDIA H100 server price 2026 (14 July 2026)](https://www.mercatus-ai.com/blog/h100-server-price) New 8-GPU HGX H100 price range and typical price, resale estimates · read 16 Oct 2026
37. 37 [CloudZero, Nvidia H100 price in 2026 (updated 9 October 2026)](https://www.cloudzero.com/blog/h100-gpu-cost/) New 8-GPU HGX H100 price range · read 16 Oct 2026
38. 38 [Rack Rate, HGX H100 SXM5 server, used and refurbished (18 July 2026)](https://rackrate.ai/hardware/hgx-h100) Used and refurbished 8-GPU system prices · read 16 Oct 2026
39. 39 [NVIDIA DGX SuperPOD data center design (H100), electrical specifications](https://docs.nvidia.com/dgx-superpod/design-guides/dgx-superpod-data-center-design-h100/latest/electrical.html) Peak demand of 20.4 kW per circuit for four DGX H100, about 10.2 kW per system · read 16 Oct 2026
40. 40 [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 15 Oct 2026
41. 41 [Silicon Data, H100 rental price over time (2023 to 2025)](https://www.silicondata.com/blog/h100-rental-price-over-time) Hyperscaler, neocloud and marketplace medians by period · read 16 Oct 2026
42. 42 [Tiger Newspress, Nvidia H100 rental costs jump nearly 40% over six months: SemiAnalysis (3 April 2026)](https://www.itiger.com/news/1185188896) One-year contract index, $1.70 in October 2025 and $2.35 in March 2026 · read 16 Oct 2026
43. 43 [UsagePricing, Lambda pricing history](https://www.usagepricing.com/blueprint/lambda-labs) 8x H100 SXM raised from $2.99 to $3.99 on 9 June 2026 · read 16 Oct 2026
44. 44 [Ornn, OCPI-H100 index](https://data.ornn.com/markets/h100-sxm) Daily H100 SXM index, $3.13 on 15 October 2026 · read 16 Oct 2026
45. 45 [GPU Finder, H100 cloud pricing](https://gpufinder.dev/gpu/h100) Availability of Lium's H100 listings over 30 days · read 16 Oct 2026

What we could not confirm

Several figures rest on something other than the provider's own pricing page, and are labelled so. Vast.ai's pricing page renders its prices in the browser and showed no figure to our reader, so its offers come from its public search API on 16 October, which returns live listings that can change within minutes. The AWS on-demand, spot and reserved figures come from Vantage and the Google Cloud figures from gcloud-compute.com, both built on the providers' price data, because the AWS and Google pricing pages load their tables by script; the Azure and Oracle prices come from their official price APIs. UpCloud, CoreWeave, Novita, Civo, Denvr, TensorDock and Thunder Compute appear only through GetDeploying. Lium's listings changed between two readings on the same day, so its $1.39 and $1.69 are moments rather than standing prices. HyperAI's $1.80 and Spheron's $2.65 do not state the form factor and were left out of the SXM comparison, and CUDO Compute now quotes on request. Stock on the day is confirmed only where the table says so; most providers do not show live availability without an account. The Silicon Data, SemiAnalysis and Ornn history figures are their own index methods and are not directly comparable with one another. In the rent or buy sum, resale value, colocation and maintenance are our assumptions, the 10.2 kW power figure is derived from NVIDIA's rack design guide rather than measured, and the electricity price is the EIA US commercial average for January to July 2026. The H100 and H200 LoRA times come from different MLPerf rounds, v4.1 and v4.0, so their cost-per-job comparison carries that caveat. No measured results on the same benchmarks were found for the A100, H100 PCIe or MI300X.

H100 SXM on GPUBank

## One H100, not eight, billed by the second.

- 20 vCPU, 125 GB RAM, 2 TB NVMe, IPv4 and traffic in $1.85/hr
- Billed by the second after the health check
- Interruptible $1.11, reserved $1.48

[Launch an H100 SXM · $1.85/hr](https://console.gpubank.io/launch?gpu=h100-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)

[H200 SXM 141 GB HBM3e From $1.49 spot · $2.49 all-in](https://gpubank.io/rent-h200)

[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)