Global · price of compute
What it costs to rent an H100, H200 or B200 by the hour
Renting one Nvidia H100 for an hour costs a median $5.33 on demand across 8 clouds whose price lists we can read (AWS, Azure, Oracle and five neoclouds; not Google Cloud), from $3.99 to $12.29 (list prices, 2026-10-11). The big clouds ask 2.5 times what neoclouds do. Over the past year Lambda (+33%) and Nebius (+53%) raised their H100 list prices; Azure and CoreWeave held theirs. Azure charges 20% more in Dublin than in Virginia for the same H200 machine.
Updated 11 October 2026 · data build 2026-10-10 · how we know · corrections
An H100 rents for a median $5.33 an hour; the big clouds ask 2.5× the neoclouds
On-demand list price for one GPU for one hour, 2026-10-11. The median is over these 8 clouds, each at its cheapest US region: AWS, Azure, Oracle, CoreWeave, Nebius, Lambda, Together, RunPod. Google Cloud (no keyless price list) and marketplace asks such as Vast.ai are not included; adding Vast.ai's median ask, as our research note did, lowers the median to about $4.50 (11 Oct 2026). Measured
Source: AWS · Azure · Oracle · CoreWeave · Nebius · Lambda · Together · RunPod · as of 2026-10-11 · H100 SXM, on demand, cheapest US region where a cloud lists several · Compiled by StackPath
Eight clouds, six GPUs: the B200 costs a median $8.5 an hour, the GB200 $16
On-demand $/GPU-hour by cloud and chip, each cloud's main US region (Oracle and Nebius price all regions alike). A dash means the cloud lists no on-demand price for that chip. Measured
| Cloud | Type | H100 | H200 | B200 | B300 | GB200 | MI300X | A100 |
|---|---|---|---|---|---|---|---|---|
| AWS | hyperscaler | 6.88 | 7.91 | 14.24 | 17.80 | – | – | 2.74 |
| Azure | hyperscaler | 12.29 | 10.60 | – | – | 27.04 | 6.00 | – |
| Oracle | hyperscaler | 10.00 | 10.00 | 14.00 | 15.00 | 16.00 | 6.00 | 4.00 |
| CoreWeave | neocloud | 6.16 | 6.30 | 8.60 | – | 10.50 | – | 2.70 |
| Nebius | neocloud | 4.50 | 5.40 | 8.50 | 9.50 | – | – | – |
| Lambda | neocloud | 3.99 | – | 6.69 | – | – | – | 2.79 |
| Together | neocloud | 3.99 | 5.99 | 8.19 | 9.99 | – | – | – |
| RunPod | neocloud | 3.99 | 5.29 | 7.99 | 8.99 | – | 2.39 | 1.79 |
| Median | 5.33 | 6.30 | 8.50 | 9.99 | 16.00 | 6.00 | 2.74 |
Nebius is the only cloud here headquartered in Europe (Amsterdam). Nebius raised its on-demand prices on October 1, 2026: B300 $7.85 → $9.50 (+21%), B200 $7.15 → $8.50 (+19%), H200 $4.50 → $5.40 (+20%), H100 $3.85 → $4.50 (+17%). Google Cloud is missing: it has no public price list a script can read without a key.
Download CSV ↓Source: AWS · Azure · Oracle · CoreWeave · Nebius · Lambda · Together · RunPod · as of 2026-10-11 · Compiled by StackPath
List prices stopped falling: Lambda +33%, Nebius +53% on the H100 in a year
On-demand $/GPU-hour each month, from archived copies of each price page (Wayback Machine) and our weekly pulls. Gaps are months with no archived copy. Measured
| Cloud | 2025-10 | Now | Change |
|---|---|---|---|
| Azure | $12.29 | $12.29 | 0% |
| CoreWeave | $6.16 | $6.16 | 0% |
| Lambda | $2.99 | $3.99 | +33% |
| Nebius | $2.95 | $4.50 | +53% |
Same cloud, same chip, a year apart. Azure's months come from the effective date the Azure price API gives for each price.
Download CSV ↓Source: Wayback Machine (archived price pages) · AWS · Azure · Oracle · CoreWeave · as of 2026-10-11 · 71 archived pages read · Compiled by StackPath
Committing pays: three-year reservations at Azure are about half the on-demand price
Reserved, term, spot and community prices next to the same cloud's on-demand price, where the price page or API shows them. Measured
Source: Azure Retail Prices API · Together pricing · CoreWeave pricing · Nebius prices · RunPod pricing · as of 2026-10-11 · Azure reservations converted to hourly (1 year = 8,760 hours); Nebius spot is its 'from' price · Compiled by StackPath
Lambda's 1-Click Clusters (16 or more GPUs with fast networking, 2 weeks to 1 year) list above its single-machine on-demand price: $6.16 an hour for an H100 against $3.99. A cluster is a different product, so it is left out of the table.
Per unit of compute, Blackwell is cheaper: $3.78 per petaFLOP-hour on a B200 against $5.39 on an H100
Median on-demand price divided by each chip's dense BF16 peak (Epoch AI hardware database). Lower is better value. Calculated
| GPU | Released | Median $/h | Dense BF16 PFLOP/s | $ per BF16 PFLOP-h | $ per FP8 PFLOP-h | Clouds |
|---|---|---|---|---|---|---|
| H100 | 2022-09-20 | 5.33 | 0.99 | 5.39 | 2.69 | 8 |
| H200 | 2024-11-18 | 6.30 | 0.99 | 6.37 | 3.18 | 7 |
| B200 | 2024-11-15 | 8.50 | 2.25 | 3.78 | 1.89 | 7 |
| B300 | 2025-08-22 | 9.99 | 2.25 | 4.44 | 2.22 | 5 |
| GB200 | 2025-02-15 | 16.00 | 2.5 | 6.40 | 3.20 | 3 |
| MI300X | 2023-12-06 | 6.00 | 1.31 | 4.59 | – | 3 |
| A100 | 2020-11-16 | 2.74 | 0.31 | 8.78 | – | 5 |
Source: Epoch AI, Machine learning hardware (CC BY 4.0) · price panel above · as of 2026-10-11 · Compiled by StackPath
Europe costs more at the hyperscalers: Azure's Dublin region is 20% above Virginia; AWS has no H100 in Ireland
Same cloud, same machine, on-demand list price by region, against that cloud's cheapest US region. Europe here = EU-27, UK, Norway and Switzerland regions. Measured
| Cloud | Region | $/GPU-h | vs US |
|---|---|---|---|
| AWS | N. Virginia (us-east-1) | 6.88 | cheapest US |
| AWS | Ohio (us-east-2) | 6.88 | cheapest US |
| AWS | Oregon (us-west-2) | 6.88 | cheapest US |
| AWS | Stockholm (eu-north-1) | 7.36 | +7% |
| AWS | London (eu-west-2) | 8.94 | +30% |
| Azure | Arizona (westus3) | 12.29 | cheapest US |
| Azure | Virginia (eastus) | 12.29 | cheapest US |
| Azure | Washington (westus2) | 12.29 | cheapest US |
| Azure | Virginia (eastus2) | 12.29 | cheapest US |
| Azure | Texas (southcentralus) | 14.75 | +20% |
| Azure | westcentralus | 14.75 | +20% |
| Azure | Illinois (northcentralus) | 14.75 | +20% |
| Azure | Iowa (centralus) | 15.12 | +23% |
| Azure | California (westus) | 15.98 | +30% |
| Azure | London (uksouth) | 15.36 | +25% |
| Azure | Warsaw (polandcentral) | 15.98 | +30% |
| Azure | Netherlands (westeurope) | 15.98 | +30% |
| Azure | Gävle (swedencentral) | 15.98 | +30% |
| Azure | Madrid (spaincentral) | 15.98 | +30% |
| Azure | Zurich (switzerlandnorth) | 17.57 | +43% |
| CoreWeave | North America | 6.16 | cheapest US |
| CoreWeave | Europe | 6.16 | 0% |
'vs US' compares with the same cloud's cheapest US region for the same machine.
Download CSV ↓Source: AWS EC2 on-demand pricing · Azure Retail Prices API · CoreWeave pricing · as of 2026-10-11 · Compiled by StackPath
What SemiAnalysis says about rental prices, and what list prices show
Paraphrased from the free sections of SemiAnalysis posts; its index data and paywalled analysis are not used. Editorial judgement
Our reading. A GPU-hour is the price of AI's raw material, and in 2026 it stopped getting cheaper for the chips most people rent. The spread is the bigger story: the same H100 lists at $3.99 at a neocloud and $12.29 at Azure, and European regions of AWS and Azure list 7–54% above the same cloud's cheapest US region (CoreWeave lists the same price in both). Most capacity, though, is sold on contracts that never appear on a price page; the $3.99–$12.29 band is the ceiling buyers negotiate down from. How the neoclouds fund the machines behind these prices is on the cost of GPU debt; what the compute turns into is on AI token prices; Europe's own GPU clouds are on European neoclouds.
List prices, not paid prices: large buyers sign one- to five-year contracts at discounts that are not published. Per-GPU price = instance price ÷ GPUs in the instance. Instances differ in CPUs, memory, storage and networking, so the same GPU is not quite the same product at each cloud. AWS: Linux on-demand, canonical 8-GPU instance (p5.48xlarge H100, p5en H200, p6-b200, p6-b300). Azure: ND v5 8-GPU InfiniBand VMs and ND GB200 v6 (4 GPUs per VM, Azure size page); pay-as-you-go, reservations converted to an hourly rate (1 year = 8,760 h). Oracle's price list has one price for all regions. CoreWeave lists North America; Lambda and Together list US capacity; RunPod and Nebius do not split prices by region. Nebius changed its on-demand prices on 1 Oct 2026; the page shows both columns and we take the new one. Back-history comes from Wayback Machine snapshots (about one a month where archived); months with no snapshot are gaps, not zeros. Azure months are filled from each price's effectiveStartDate in the API. Price per PFLOP-hour uses Epoch AI's dense BF16 / FP8 FLOP/s per chip (spec sheet peaks, not achieved throughput). Google Cloud: no keyless price API (the Cloud Billing Catalog API needs a key; its pricing pages render in JavaScript). Vast.ai: its offers API returns HTTP 403 to scripts, so marketplace asks are not included. Nscale: no public price page. SemiAnalysis, Ornn and Silicon Data GPU indices: paywalled or keyed; at most one attributed headline each on the page.
Questions people ask
- How much does it cost to rent an H100 per hour?
- On-demand list prices on 2026-10-11: a median $5.33 per GPU-hour across 8 clouds, from $3.99 (neoclouds such as Lambda, Together and RunPod) to $12.29 (Azure). Large buyers pay less on one- to three-year contracts; those prices are not published.
- Why do AWS, Azure and Oracle charge more than neoclouds for the same GPU?
- In this panel the hyperscalers' median H100 price is $10 against $3.99 at neoclouds, 2.5 times as much. On-demand is the hyperscalers' dearest way to buy: their reserved prices are far lower (Azure's three-year reservation is 56% below its on-demand rate), and the clouds do not publish what large customers pay under negotiated contracts.
- Are GPU rental prices going up or down?
- H100 on-demand list prices, for the clouds with a price a year earlier: Azure $12.29 → $12.29 (0%); CoreWeave $6.16 → $6.16 (0%); Lambda $2.99 → $3.99 (+33%); Nebius $2.95 → $4.50 (+53%). 2 rose, 0 fell, 2 unchanged.
- Is renting GPUs more expensive in Europe or Ireland?
- At the hyperscalers, yes. Azure lists its 8-GPU H200 machine at $12.72 per GPU-hour in North Europe (Dublin), 20% above its cheapest US region. AWS offers no on-demand H100, H200, B200 or B300 instances in its Ireland region at all, only the older A100. CoreWeave lists the same on-demand prices in Europe and North America.
- Is a B200 better value than an H100?
- Per unit of compute, yes, at list prices. The median B200 rents for $8.5 an hour and the median H100 for $5.33, but the B200 does about 2.3 times the dense BF16 maths, so a petaFLOP-hour costs $3.78 on a B200 against $5.39 on an H100 (Epoch AI chip specs).