In October 2026, published on-demand list prices for an H100 ranged from about $2 to about $12 per GPU-hour, depending on the provider, the billing model and what is bundled. Specialist clouds mostly sit between $3 and $6. Spot and long commitments can cut the price by half or more. B200 list prices run from about $6 to $14.
This article sets out the published numbers, says where each came from and when, and explains the forces behind price changes. Prices change often. Check the provider's own page before you budget, and treat everything below as a dated snapshot. Where a provider's own pricing page shows the figure, that page is the source; where only a third party survey reports it, the survey is named.
Key numbers at a glance
- AWS Capacity Blocks, p5.48xlarge (8 H100s), N. Virginia, Ohio and Oregon: $47.757 per instance-hour, $5.970 per GPU-hour. AWS Capacity Blocks pricing page, 19 October 2026.
- AWS p5.48xlarge on-demand (8 H100s), N. Virginia, Ohio, Oregon: $6.88 per GPU-hour, as checked by IntuitionLabs on 10 October 2026.
- Azure NC40ads H100 v5 (1 GPU), East US, Linux pay-as-you-go: $6.98 per hour. Azure ND96isr H100 v5 (8 GPUs): $98.32 per VM-hour, $12.29 per GPU-hour. Azure Retail Prices API, 19 October 2026.
- Google Cloud a3-highgpu-8g (8 H100s), Iowa, on-demand: $88.49 per VM-hour, which is $11.06 per GPU-hour. IntuitionLabs, 10 October 2026.
- Oracle Cloud BM.GPU.H100.8: $10.00 per GPU-hour list. IntuitionLabs, 10 October 2026.
- CoreWeave HGX H100, 8 GPUs, on-demand: $49.24 per instance-hour, about $6.16 per GPU. CoreWeave pricing page, 19 October 2026.
- Lambda H100 SXM, 8-GPU instance: $3.99 per GPU-hour. Lambda pricing page, 19 October 2026. IntuitionLabs records $2.99 in a 2025 snapshot.
- Runpod H100 SXM: $3.99 Secure Cloud and $2.69 Community Cloud. H100 PCIe: $1.99 Community Cloud. Runpod pricing page, 19 October 2026.
- Nebius HGX H100 on demand: $4.50 per GPU-hour. Nebius prices page, 19 October 2026. IntuitionLabs says this took effect on 1 October 2026, up from $3.85.
- H200 on-demand list prices: $5.29 (Runpod Secure Cloud), $5.40 (Nebius), about $6.31 (CoreWeave, $50.44 per 8-GPU instance). Provider pages, 19 October 2026.
- B200 on-demand list prices: $6.69 (Lambda, 8 GPUs), $7.99 (Runpod Secure Cloud), $8.50 (Nebius), $8.60 (CoreWeave, $68.80 per 8-GPU instance). AWS Capacity Blocks: $14.208. Provider pages, 19 October 2026.
- B200 index price: $2.31 per hour in early March 2026 and $4.95 in late April 2026, per the Ornn Compute Price Index as reported on tomtunguz.com on 27 April 2026.
First, what "per GPU-hour" hides
A GPU-hour is one GPU rented for one hour. It is the standard unit, but it hides a lot.
Price per GPU-hour tells you nothing about what is around the GPU. A cheap hourly rate on a machine with slow networking, little storage, or a long queue may cost more per finished job than a pricier rate on a well connected cluster. A few things differ between listings and are worth checking:
- The chip variant. The H100 comes in a PCIe version and an SXM version. SXM is the faster, more tightly connected one used in 8-GPU servers. Runpod's own page lists its H100 PCIe and H100 SXM at different prices for this reason ($1.99 and $2.69 in Community Cloud).
- Whether the price is per GPU, per node or per cluster. CoreWeave quotes a whole 8-GPU instance, while Lambda and Nebius quote per GPU. The IntuitionLabs survey divided node prices by eight to compare and warns that node pricing distorts per-GPU comparisons.
- The billing model. On-demand means you pay by the hour with no commitment. Spot or preemptible means cheaper capacity that the provider can take back. Committed means you promise a term. Prepaid capacity blocks mean you reserve a window in advance.
- Region and operating system. AWS and Azure prices vary by region, and Azure's Windows rates are higher than Linux for the same VM ($8.82 against $6.98 for the NC40ads H100 v5 in East US, per the Azure Retail Prices API on 19 October 2026).
- Memory. AWS's P5 page lists 80 GB of HBM3 per H100 and 1,128 GB of HBM3e per 8-GPU H200 instance, about 141 GB per GPU. More memory lets a larger model or a longer context fit on fewer GPUs, which is part of why the H200 costs more.
The IntuitionLabs survey says plainly that its entries mix prepaid capacity, on-demand VMs, pods, and cluster and node rates, and are not interchangeable. Its inclusion rule is publicly displayed hourly rates for a matching product, taken from first-party pages, with promotional rates excluded. It is a survey by one publisher, not an official price index, and its method is as the publisher describes it.
H100: the 2026 snapshot
The H100 is the older workhorse. Here is the picture from the 10 October 2026 survey and provider pages read on 19 October 2026.
The large clouds charge the most on-demand. IntuitionLabs lists AWS at $6.88 per GPU-hour for the 8-GPU instance in three US regions and $8.60 in N. California. Azure's own price feed shows $6.98 for a single-GPU VM and $98.32 for the 8-GPU VM in East US, which is $12.29 per GPU. IntuitionLabs lists Google at $11.06 per GPU in Iowa and Oracle at $10.00.
Specialist clouds list lower. Lambda's pricing page shows the 8-GPU H100 SXM at $3.99 per GPU, with smaller instances at $4.09, $4.19 and $4.29 for 4, 2 and 1 GPUs. Runpod's page shows its SXM pod at $3.99 in Secure Cloud and $2.69 in Community Cloud. Lambda's 1-Click Cluster product lists H100 at $6.16 per GPU for 16 GPUs, $5.85 for 64 and $5.54 for 256, for terms of two weeks to a year. CoreWeave's on-demand HGX H100 instance works out at about $6.16 per GPU. IntuitionLabs also reports one marketplace query on 10 October that returned single-GPU H100 SXM offers from about $1.68 per hour on an unverified host, and is clear that a marketplace query is not a list price.
Several of these went up. According to the survey, AWS Capacity Blocks rose from $5.19 to $5.97, Runpod Secure Cloud H100 SXM from $3.29 to $3.99, Lambda's 8-GPU SXM from $2.99 (a 2025 snapshot) to $3.99, and Nebius's HGX H100 from $3.85 to $4.50 effective 1 October 2026. The current figures match the providers' own pages on 19 October 2026. The survey says AWS on-demand and Azure list prices were not re-verified as lower than the 2025 third-party figures.
How much cheaper are spot and committed rates?
The IntuitionLabs survey lists discount programs at Google Cloud for the 8-GPU VM, per VM-hour: Spot $50.32, Dynamic Workload Scheduler Flex-start $38.32, Calendar $41.60, one-year committed use $61.38 and three-year committed use $38.86. Against the $88.49 on-demand price, the discounts work out at about 43% for Spot, about 31% for one year and about 56% for three years (calculated from the listed prices). Per GPU, that is $6.29 for Spot, $7.67 for one year and $4.86 for three years.
Azure's own price feed shows spot capacity at about $1.29 per hour for the single-GPU NC40ads H100 v5 against $6.98 pay-as-you-go, roughly 82% lower, and about $18.17 per hour for the 8-GPU ND96isr against $98.32. Nebius's prices page lists preemptible H100 and H200 capacity from $0.79 per GPU-hour and B200 from $0.99, and says spot prices are dynamic and move with available capacity. Nebius also says customers who reserve large clusters for multiple months can pay up to 35% less than on-demand rates, without listing fixed committed prices.
A low spot price comes with a catch: the provider can reclaim the machine, so the work must tolerate interruptions.
The pattern is simple. The longer you commit and the more interruption you accept, the lower the hourly price. The right choice depends on how steady your demand is, which ties into the companion explainer on GPU utilization.
What an hourly price means for a year
A year has 8,760 hours. Here is a simple calculation, using three published prices for one H100, billed around the clock with no discounts:
- At $2.69 per hour (Runpod Community Cloud SXM): about $23,560 a year.
- At $3.99 (Lambda or Runpod Secure Cloud): about $34,950 a year.
- At $6.16 (CoreWeave on-demand HGX instance, per GPU): about $53,960 a year.
For a node of eight GPUs, multiply by eight. These are multiplications of published list prices, not quotes. They show why a buyer with steady demand pushes for a commitment: a few dollars an hour per GPU is tens of thousands of dollars a year per chip, and thousands of chips make it hundreds of millions.
H200 and B200
The H200 is a later Hopper chip with more and faster memory, about 141 GB of HBM3e per GPU on AWS's P5e and P5en instances. The B200 is NVIDIA's Blackwell generation GPU; Lambda's page lists it with 180 GB of memory.
First-party list prices on 19 October 2026 give a clearer picture than earlier in the year:
- H200 on-demand: Runpod Secure Cloud $5.29 and Community Cloud $3.59; Nebius $5.40; CoreWeave $50.44 per 8-GPU instance, about $6.31 per GPU.
- H200 Capacity Blocks on AWS: $6.866 per GPU-hour on p5e.48xlarge and $7.895 on p5en.48xlarge, in the US regions listed.
- B200 on-demand: Lambda $6.69 per GPU on an 8-GPU instance and $6.99 for one GPU; Runpod Secure Cloud $7.99 and Community Cloud $5.98; Nebius $8.50; CoreWeave $68.80 per 8-GPU instance, about $8.60 per GPU.
- B200 Capacity Blocks on AWS: $113.666 per instance-hour, $14.208 per GPU-hour, on p6-b200.48xlarge.
- B200 clusters on Lambda: $9.86 per GPU for 16 GPUs, $9.36 for 64 and $8.87 for 256 or more, for terms of two weeks to a year.
AWS's Capacity Blocks page says its rates are updated regularly, with the next update scheduled for January 2027, and that a block is charged at the rate in effect when it is bought.
The index view is different. The Ornn Compute Price Index, reported in a post dated 27 April 2026 on tomtunguz.com, put the B200 at $4.95 per hour that week, up from $2.31 in early March, a 114% rise in six weeks. The author also reports that the B200 premium over H200 fell to $0.28 per hour in November 2025 and widened to $1.80 by late April 2026. That is one index, from one author, describing a spot style market, so its levels sit below most on-demand list prices. The takeaway is directional: Blackwell prices were rising in spring 2026, and the gap to the previous generation was widening.
Why prices move
Several forces act at once, and sources disagree about their weight.
Supply. When new chips ship in volume, older chips tend to fall in price. IntuitionLabs records that AWS announced an up to 45% price reduction on 5 June 2025, which the survey says included about 44% off H100 on-demand, and that providers cut H100 rates by roughly 20 to 25% in single moves in June and July 2025. That is the historical context for 2025.
Demand. The tomtunguz.com author argues that frontier model releases coincide with jumps in demand for the newest chips. The author links the B200 spread widening to a model launch in February 2026 and a larger context window in April 2026 that needs more memory. The author admits the correlation is not perfect and that supply shocks matter. This is an author's interpretation, not a measured fact.
Capacity running out. The October 2026 increases at Lambda, Nebius, Runpod and AWS Capacity Blocks suggest H100 capacity is tighter than in mid 2025. The survey reports the increases but does not give reasons, and none of the providers states a cause on its pricing page.
Contract structure. As the companion explainer on GPU as a service economics sets out, large providers sign multi-year take-or-pay contracts. Capacity already promised to big customers is not available to walk-up renters, so list prices can rise while some marketplace offers stay low.
How to compare quotes sensibly
Ask for the price per GPU-hour, and then ask what is included. A short checklist:
- Which chip and which variant, and what memory per GPU.
- What network connects the GPUs, if you need more than one node.
- Whether the quote is on-demand, spot, committed or prepaid, and for how long.
- What storage, data transfer and support cost on top.
- What happens if the provider needs the machine back.
- Whether the machine is bare metal or a virtual machine.
Then compare on cost per useful result, such as dollars per million tokens served or per training run, not on the hourly rate alone. A cheaper GPU that sits idle half the time costs more per result than a dearer one kept busy.
Matching the pricing model to the workload
The same list price means different things to different buyers. The published price structures point to a few general matches.
Short, bounded jobs such as a fine-tuning run fit on-demand or a reserved block. On AWS, a Capacity Block for 8 H100s costs $5.97 per GPU-hour in the listed US regions, against $6.88 on-demand in the IntuitionLabs survey, and a specialist cloud's on-demand 8-GPU H100 lists at $3.99. For a run across several GPUs, the networking between them matters as much as the rate.
Steady serving fits a commitment. Google's three-year committed rate for the 8-GPU H100 VM is $38.86 per VM-hour against $88.49 on-demand, about 56% lower, and a one-year term at $61.38 gives a smaller discount with less lock-in, per the survey. A three-year promise is a bet that the workload and the chip will still be right at the end of the term, when a newer generation may be cheaper per unit of work.
Interruptible work such as batch experiments fits spot or preemptible capacity. Azure's spot rate of about $1.29 for a single H100 VM is a fraction of the $6.98 pay-as-you-go rate in its own price feed. The work has to save progress often, because the provider can take the machine back.
In all three cases, the hourly rate is only the start. The better question is what it costs to finish the job, and that depends on how busy the GPUs are kept. The companion explainer on GPU utilization covers why a cheap rate on an idle machine is expensive.
Why list prices are not the whole market
List prices are the starting point of a negotiation for large buyers. The IntuitionLabs survey says Oracle's price, for example, is a pay-as-you-go list price in US dollars before negotiated discounts. Lambda's page asks buyers of clusters for a year or longer to contact sales, and Nebius lists GB200 and GB300 NVL72 systems as contact sales. A company buying thousands of GPUs on a multi-year contract will not pay list. Those deals are private and are not reported in public price lists or the surveys cited here.
That means the public numbers are best read as a ceiling for walk-up buyers and as a signal of direction. If list prices at several providers rise in the same month, as they did around October 2026 according to the survey, it suggests capacity is tighter, but it does not tell you what a large contract costs.
What is firm and what is uncertain
Firm: the list prices shown on the AWS Capacity Blocks, CoreWeave, Lambda, Nebius and Runpod pricing pages and in the Azure Retail Prices API on 19 October 2026. Also firm, as a dated survey: the IntuitionLabs figures checked on 10 October 2026 from first-party pages, including the AWS on-demand, Google and Oracle prices not confirmed here directly. Also firm as reports: the Ornn index figures as quoted on tomtunguz.com.
Uncertain: the underlying cause of the 2026 H100 increases; how the index prices for H200 and B200 relate to the higher on-demand list prices; negotiated enterprise prices, which are usually well below list; and whether any figure still holds when you read this. Independent rental indexes from other publishers report their own numbers, but this post uses none of them, so treat the index view here as one source.
What to watch next
- Provider price pages, monthly. Lists have been changing every few weeks.
- AWS's next scheduled Capacity Blocks price update, which its page says is due in January 2027.
- Whether the B200 premium over H200 keeps widening or collapses again, as it did by November 2025.
- Whether AWS and Azure on-demand H100 prices move, since both were described as not reflecting lower 2025 third-party figures.
- Whether providers publish fixed multi-year commitment prices as Blackwell supply grows, rather than "contact sales".
See also GPU as a service economics, GPU utilization explained and the AI chip market in 2026.
Teams that rent GPUs also run a lot of open source code on them; Safeguard offers SBOM and dependency scanning and a free scan at gold.safeguard.sh/scan.
Quick check
Is $2 per hour a realistic price for an H100 in October 2026?
Only at the very bottom. The lowest published on-demand list price was $1.99 for an H100 PCIe on Runpod Community Cloud, and most specialist list prices for H100 SXM were $2.69 to $4.50. Large clouds listed $6.88 to $12.29 per GPU.
Why is spot capacity so much cheaper?
The provider can take it back. Azure's single-GPU H100 spot rate of about $1.29 compares with $6.98 pay-as-you-go in its own price feed, so the discount is payment for accepting interruptions. It suits work that can pause and resume.
Do GPU prices only go down over time?
No. Prices fell in 2025, but the October 2026 survey records increases at several providers, and the Ornn index shows B200 more than doubling from March to April 2026.
Sources
- H100 rental prices compared, IntuitionLabs, updated 11 October 2026, first-party prices checked 10 October 2026
- B200 GPU pricing, spot market and model releases, tomtunguz.com, 27 April 2026
- Amazon EC2 Capacity Blocks for ML pricing, AWS, read 19 October 2026
- Amazon EC2 P5 instances, AWS, read 19 October 2026
- Azure Retail Prices API, Microsoft, East US Linux and Windows consumption prices, queried 19 October 2026
- CoreWeave pricing, read 19 October 2026
- Lambda pricing, read 19 October 2026
- Nebius prices, read 19 October 2026
- Runpod pricing, read 19 October 2026
Prices are list or index figures as published or reported on the dates shown. Percent differences and per-GPU figures derived from per-instance prices are calculated from the listed prices. This is not purchasing advice.
Last reviewed: 19 October 2026.
