H100 pricing in October 2026: every major cloud provider compared

On October 1, 2026, the on-demand price of H100 on Nebius rose from $3.85 to $4.50 per GPU hour. This is a 16.9% increase for a Hopper family GPU.

The gap between vendors is much bigger than the increase. The listed on-demand price for H100 ranges between $3.99 per GPU hour on Lambda and $12.29 on Azure.

AWS is next on October 7. The reserved price of H100 in AWS Capacity Block increased by 15% to $5.970 per GPU hour.

In this guide, we compare H100 pricing among seven vendors. It includes the source and the disclaimers for each. We also clarify why two websites may give different prices for the same vendor. Prices are based on the sources below as of October 1 through October 6, 2026.

Updated: October 6, 2026.

Key takeaways

• Nebius increased its rate for its H100 on-demand to 16.9% on October 1 from $3.85 to $4.50 per GPU-hour. Its preemptible price remains at $2.15.
• On-demand H100 prices currently range from $3.99 (Lambda) to $12.29 (Azure) per GPU-hour.
• The price of AWS’ H100 reserved Capacity Block increases by 15% on October 7 to $5.970 per GPU-hour.
• CoreWeave, Oracle, Google Cloud, and Azure offer H100s as eight-GPU nodes, so you calculate the price per GPU by dividing the cost of one node by eight.
• The price differs depending on the region, GPU configuration, and purchase type, so do a like-to-like comparison.
• Outdated figures include $2.99 for Lambda and $98.32 for AWS p5.48xlarge.

What changed on October 1

As of October 1, 2026, Nebius revised the pricing of compute for four GPUs. You can find the old and new pricing in Nebius’s documentation.

They have not adjusted the preemptible pricing. Therefore, the H100 preemptible price remains at $2.15. This widens the gap with the on-demand option.

This change also affected the positions of two Neoclouds. Before October 1, the $3.85 was lower than the $3.99 from Lambda. However, now the price has increased to $4.50 and has become higher than Lambda’s.

The same revision resulted in price adjustments for non-GPU services. AMD Epyc Genoa vCPU cost went up from $0.012 to $0.015 per hour, and memory–from $0.0032 to $0.0045 per GiB-hour.

If one 8-GPU H100 instance works for 720 hours, the price will increase from $22,176 to $25,920, i.e., $3,744 more each month without counting storage and networking.

FinanceFeeds’s check on October 2 revealed that the page still displayed both price options. A customer could not see what pricing the console would charge him or her.

What changes on October 7

There is a change in the EC2 Capacity Block rates that AWS will implement on October 7. EC2 Capacity Block refers to a buyer reserving GPU capacity ahead of time. The above rate is a reserved rate rather than an on-demand rate, as indicated in the table.

The rate for the P5 instance family, which is the H100 family, increases to $5.970 per hour of GPU in all regions. According to FourWeekMBA, AWS does not provide the old and new numbers when providing the new rate on its webpage.

Axis Intelligence indicates the increase to be about 15% from $5.191 to $14.208 for B200 reservations.

The increase only applies to Capacity Blocks, yet the new rate remains lower than the $6.88 on-demand rate.

How to read H100 GPU pricing

While there appears to be a single pricing value for H100 per hour, this is not the case, as it is made up of five distinct variables. Providers vary in terms of form factor, billing unit, region, purchase choice, and price inclusion.

The form factor matters because they offer the H100 in SXM, NVLink, and PCIe formats. The price for each is distinct.

The billing unit is next on the list, as CoreWeave, Oracle, Google Cloud, and Azure all offer H100 as eight-GPU nodes at a price per GPU equal to the node price/8.

For instance, the price per node at Azure is $98.32, and this translates to $12.29 per GPU. Nebius bills by the second and shows the prices as hourly figures. Thus, in a case where you use the service for 30 minutes, you pay half of the hourly fee. The term may be new to you, but you should find out the meaning of GPU-hour.

The location also dictates the pricing. According to one tracker, the difference between the cheapest and the most expensive location for Google’s eight GPU H100 is 30 percent.

There are also different options for purchasing. The on-demand option comes with the highest prices, followed by spot, preemptible, and committed.

Finally, check out the included cost of the price. There are additional costs such as egress and storage that come on top of the GPU price. Our guide to reading a GPU pricing page covers them.

NVIDIA H100 pricing by provider

The table lists on-demand rates per GPU-hour, lowest first. Each provider’s section below gives the source and the main caveat.

Lambda Labs H100 pricing

Lambda, previously known as Lambda Labs, quotes a price of $3.99 for an H100 SXM eight-GPU instance on their current pricing page. The page also mentions that there are no egress fees.

This price supersedes the previous $2.99 per hour rate, which is seen on cached pages of Lambda’s site and some trackers.

Nebius H100 pricing

Nebius charges $4.50 per GPU-hour for H100 NVLink, effective October 1st. The H100 is available only in Nebius’ eu-north1 location, according to their documentation.

CoreWeave H100 pricing

CoreWeave is currently selling the H100 as an eight-GPU HGX node for $49.24 per hour, or $6.16 per GPU-hour. CoreWeave’s pricing page mentions this information.

Their classic pricing page continues to mention the H100 HGX at $4.76 per GPU. Verify what product you are being quoted for. CoreWeave also mentions savings up to 60 percent for committed usage.

AWS H100 pricing

In US East, AWS has the p5.4xlarge priced at $6.88 per hour in the devzero.io tracker. The AWS pricing page automatically loads the pricing; hence, we could not quote directly.

The pricing increases beyond the US region. According to the Holori regional tracker, the price for the same instance is $11.56 per hour in São Paulo.

Note the $98.32 rate. Some trackers may still quote the price for the eight-GPU p5.48xlarge. A reduction of 44% from the said rate will be close to the current $55.04 rate in AWS June 2025 H100 price reduction.

Oracle Cloud H100 pricing

Oracle lists the eight-GPU BM.GPU.H100.8 bare-metal shape at $80 per hour or $10.00 per GPU-hour. Economize’s tracker, updated September 9, 2026, also lists the same rate.

The Oracle pricing page loads dynamically; hence, we could not check it directly. The Thunder Compute guide, reviewed on October 1, 2026, also lists the same $10.00.

GCP H100 pricing

Google on its pricing page for Google accelerators prices at $88.49 per hour for the eight-GPU a3-highgpu-8g. This amounts to $11.06 per GPU-hour.

According to Google’s GPU Pricing page, spot prices are between 60% and 91% lower than on-demand prices for most machine types.

Azure H100 pricing

The ND96isr H100 v5 in Azure with eight GPUs costs $98.32 per hour, or $12.29 per GPU. This is according to the Azure Retail Prices API via AzureSpeed’s pricing tool.

There is also a one-GPU NC H100 v5 VM in Azure. The price we came across is from the December 2025 edition of Thunder Compute’s pricing guide.

What eight GPUs cost over 720 hours

Multiplying each rate by eight GPUs and 720 hours shows what a month of continuous use costs. These totals exclude storage and egress.

The ranking matches the hourly table, but the dollar gap is easier to feel. Azure’s monthly cost is $47,808 more than Lambda’s.

These totals use on-demand list rates, so reserved or preemptible capacity would cost less.

Which provider fits which job

Single-GPU work points to Lambda or AWS. A tracker lists Lambda’s one-GPU H100 at $4.29, and AWS lists a single-GPU p5.4xlarge at $6.88. CoreWeave sells H100s only as eight-GPU nodes.

Interruptible jobs favour the preemptible and spot tiers. Nebius lists $2.15 per GPU-hour. Economize lists Oracle’s preemptible eight-GPU shape at $40 an hour, or $5.00 per GPU.

Long, steady runs favour commitments. Oracle’s one-year rate works out to $6.70 per GPU-hour and its three-year rate to $5.00. CoreWeave advertises up to 60% off, and Lambda sends reserved buyers to sales.

AWS’s prepaid Capacity Blocks rise to $5.970 per GPU-hour on October 7.

Region limits matter too. Nebius offers the H100 only in eu-north1, and Google’s rate varies by region. Check that the cheapest listing is available where your data sits.

What the Nebius increase does and doesn’t tell us

Nebius has not published a reason in the documentation we reviewed. Toobit’s report on the notice reads the increase as a sign that scarce capacity still carries pricing power.

That is an interpretation, not a statement from Nebius.

One increase at one provider does not set the market. AWS cut H100 on-demand prices by 44% in June 2025, per Data Centre Dynamics. CoreWeave’s H100 node price did not change in our check.

Our GPU price increase analysis covers the AWS side in more detail.

How to lower your H100 bill

Start with the lower tiers. Nebius’s preemptible H100 costs $2.15 per GPU-hour, less than half its new $4.50 on-demand rate. Providers can reclaim preemptible and spot capacity, so it suits interruptible jobs.

For one eight-GPU node over 720 hours, $2.15 per GPU costs $12,384 against $25,920 on-demand.

Commitments cut rates further. Oracle’s one-year commitment lists at $53.60 an hour for eight GPUs and its three-year commitment at $40.00 per Economize’s tracker. Over 720 hours, those come to $38,592 and $28,800 against $57,600 on-demand.

Lambda directs buyers to sales for its lowest reserved prices.

Egress can erase part of a headline discount. Lambda and CoreWeave list free data transfer. AWS charges about $0.09 per GB after a free allowance, per Cloudflare’s AWS pricing summary.

For a full 30-day cost comparison that includes egress, see our worked cost comparison.

FAQ: H100 pricing questions

Why do other sites show different H100 prices?

Older snapshots linger, and some sites mix purchase types. Lambda’s $2.99 and AWS’s $98.32 for a p5.48xlarge still appear, and some pages list AWS’s reserved rate beside on-demand rates.

Does the Nebius increase apply to committed contracts?

Not according to Toobit’s report, which says the notice covers on-demand capacity only. Check your own agreement, since terms vary.

Do I get the eight-GPU rate if I rent fewer GPUs?

Not always. A ComputePrices tracker shows Lambda at $4.29 for one GPU against $3.99 for eight.

How much more do the H200 and B200 cost than the H100?

At Nebius, the H200 is $0.90 more per GPU-hour, and the B200 is $4.00 more. Those gaps come from its post-October rate card.

Will Vera Rubin make H100 cheaper?

Not yet in the listings. Nebius’s increase took effect the day after CoreWeave announced limited Vera Rubin availability.

Conclusion

Listed H100 rates now run from $3.99 to $12.29 per GPU-hour. Nebius just raised its price by 16.9%, and AWS raises its reserved rate on October 7.

The right number depends on form factor, region, billing unit, and commitment, so compare like for like.

Rates in this market change often. Re-check each provider’s page on the day you budget, and treat third-party trackers as a starting point.

Before you commit, confirm three things. Check the form factor, check the region, and check whether the quoted price is per GPU or per node.