A field guide to the companies renting GPUs and other AI accelerators. Trace who supplies whom, compare GPU prices, and browse sourced profiles of 97 providers.
Companies that rent, host, or resell AI compute. Select one for its full profile and sources.
Events from the first GPU miners to October 2026. Years before 2025 are compressed because most of the activity is recent. Click a dot to read it, or step through with the arrow keys.
Show every event as a list
Published on-demand rates for an H100 run from $3.20 per GPU-hour at Hyperstack to $12.29 at Microsoft Azure. Every price on this page comes from the provider's own rate card, read on October 14, 2026.
H100 on-demand prices
US dollars per GPU-hour for an H100 SXM, from each provider's rate card. Hover over or focus a dot for its terms and source, or open the table below.
Show prices and sources as a table
Rental terms
Reserved pricing trades a spending or time commitment for a lower hourly rate than on-demand, and spot capacity costs less again because the provider can interrupt it. Compare offers within the same rental model before comparing providers.
AWS lists its H100 p5 instance at $6.88 per GPU-hour on demand, $5.42 under a one-year Compute Savings Plan and $2.97 under a three-year EC2 Instance Savings Plan. Microsoft lists its ND H100 v5 at $12.29, falling to $7.87 with a one-year reservation and $5.40 with a three-year one. The lower rates come with longer commitments, so neither figure alone shows which offer costs less for a given workload.
The neocloud and AI platform rate cards list H100s at $3.20 to $6.16 per GPU-hour on demand. Together AI lists $3.19 for a 91 to 180 day reservation, and Verda's published discounts bring a two-year term to about $3.07. CoreWeave advertises committed-use discounts of up to 60%, and Nebius up to 35%. Advertised discounts do not establish the prices customers negotiate in private contracts.
| Offer | Per GPU-hour | Range and examples |
|---|---|---|
| H100 on-demand | $3.20 to $12.29 | Hyperstack to Microsoft Azure. Neoclouds list $3.20 to $6.16 and hyperscalers $6.88 to $12.29. |
| H100 reserved | $2.97 to $7.87 | A three-year AWS EC2 Instance Savings Plan to a one-year Azure reservation. DigitalOcean lists $3.26 on a 12-month contract. |
| H100 spot | from $0.79 | Nebius sets spot prices dynamically from $0.79. Together AI lists preemptible capacity at $1.99, Azure spot at $2.27 and CoreWeave spot at $2.46. |
| B200 on-demand | $6.00 to $14.24 | Hyperstack to AWS P6. Lambda lists $6.69, Nebius $8.50, CoreWeave $8.60 and Oracle $14.00. |
| GB200 NVL72 | $10.50 to $27.04 | CoreWeave on-demand to Azure pay-as-you-go. Oracle lists $16.00, and a three-year Azure reservation brings Azure to $11.90. |
Availability and price changes
Lambda's rate card lists on-demand H100s without showing live availability, so a published rate does not guarantee that capacity is free to rent. The rates also price different units. Lambda, CoreWeave, AWS, Google Cloud and Azure quote a share of an eight-GPU instance, while Verda and Runpod price a single GPU. Check the instance size as well as the hourly rate.
SemiAnalysis reports that one-year H100 rental contracts rose from $1.70 per GPU-hour in October 2025 to $2.35 in March 2026, an increase of 38%. Those are year-long contracts, so the series measures a different market from the hourly rates above.
Cost per completed job
An hourly rate becomes useful when you know how many billed hours a job needs. Compare the full instance charge, runtime, and any storage or data-transfer charges. For reserved capacity, count the hours you pay for but do not use. For spot capacity, include the work lost to interruptions.
Consider a simplified example. At $2.50 per reserved GPU-hour, using only half the hours you pay for raises the compute cost to $5.00 per used GPU-hour. If failures also consume 8% of those used hours, the cost reaches about $5.43 per useful GPU-hour. These figures assume idle and failed hours remain billable and exclude other charges.