GPU cost calculator
We may earn commissions when you shop through the links below. Learn more
Hourly GPU prices turned into a monthly estimate, with the provider range, every matching price and its source.
Estimated monthly cost, October 2026
$695–$802≈ €621–€716
Lowest to highest of 3 complete price options from 3 providers for H100 SXM 80GB, 176 GPU-hours a month (median $776≈ €693). Excludes VAT and sales tax. 2 more options are missing a published storage or egress cost and listed separately.
| Provider and offer | Monthly total |
|---|---|
| Modal | $695≈ €621 per month |
| DigitalOcean | $776≈ €693 per month |
| Nebius | $802≈ €716 per month |
| Lambda | $755≈ €675 per month, storage price not published |
| Together AI | $966≈ €863 per month, no storage that attaches to this offer |
All matching prices
Monthly cost in October 2026 of each published price that matches your inputs, cheapest complete estimate first. Prices are per GPU-hour from each provider's own pricing page, at the price that applies in that month; open the source to check the current figure. The ranking uses the dataset's USD figures whatever currency is shown. This GPU also has 2 spot and 8 reserved prices; switch “Commitment” to compare them.
Not included
- DigitalOcean, on-demand at $4.41/GPU-hr: Sold as 8-GPU nodes; set GPUs per instance to 8 or more., on-demand at ≈ €3.94/GPU-hr: Sold as 8-GPU nodes; set GPUs per instance to 8 or more.
- Hyperstack, on-demand at $3.20/GPU-hr: Sold as 8-GPU nodes; set GPUs per instance to 8 or more., on-demand at ≈ €2.86/GPU-hr: Sold as 8-GPU nodes; set GPUs per instance to 8 or more.
- Lambda, on-demand at $3.99–$4.19/GPU-hr: Sold as 2-, 4- and 8-GPU nodes; set GPUs per instance to 2 or more., on-demand at ≈ €3.56–€3.74/GPU-hr: Sold as 2-, 4- and 8-GPU nodes; set GPUs per instance to 2 or more.
- Scaleway, on-demand at $3.61–$3.77/GPU-hr: Sold as 2-, 4- and 8-GPU nodes; set GPUs per instance to 2 or more., on-demand at ≈ €3.23–€3.37/GPU-hr: Sold as 2-, 4- and 8-GPU nodes; set GPUs per instance to 2 or more.
- Together AI, on-demand at $3.99/GPU-hr: Sold as 8-GPU nodes; set GPUs per instance to 8 or more., on-demand at ≈ €3.56/GPU-hr: Sold as 8-GPU nodes; set GPUs per instance to 8 or more.
How this estimate is calculated
Formula
- compute + storage + egress
- price per GPU-hour × GPUs per instance × hours per day × days per month × instances (730 hours a month in place of hours × days for reserved and monthly-billed prices; at most the monthly price per server where hourly billing stops at one)
- storage GB beyond what the instance includes × the price of the cheapest persistent storage that attaches to the offer, after its free allowance, minimum size, size steps and volume tiers
- egress GB × the provider's published price per GB
Assumptions
- Prices are USD per single GPU-hour, taken from each provider's own pricing page or API and stored with the date they were checked. VAT, sales tax and negotiated discounts are not included.
- Providers that publish prices in euros are converted to US dollars at the European Central Bank reference rate noted with each price; the provider page shows the original euro price.
- A price the provider does not publish is left out and marked “Not verified”. It is never counted as $0.
- Prices a provider's terms do not allow in a comparison service are left out of every estimate and range.
- Some GPUs are only sold as multi-GPU nodes, such as 8 × H100. Those prices apply once you ask for at least that many GPUs per instance; larger requests are rounded up to whole nodes. Smaller offers are combined when you ask for more GPUs than one instance holds.
Show all 17 assumptionsShow fewer assumptions
- Each day's hours count as one run, billed in the provider's own increment: a started hour is billed in full where billing is hourly (1 hour 1 second costs 2 hours) and a started minute where it is per minute.
- “No commitment” covers on-demand, secure and community cloud, serverless and marketplace listings; marketplace lowest and median listings stay separate. Spot and reserved prices only appear when you choose them.
- Serverless prices billed while running assume you pay only for the hours you enter; cold starts, idle windows and minimum charges come on top.
- Reserved prices and offers billed per month are charged for every hour of the month (730 hours), whatever hours you enter: a reservation is paid for its whole term, whether or not the GPUs run.
- Where a provider's hourly billing stops at a monthly price (Hetzner's dedicated GPU servers), each server costs at most that price a month. One-off setup fees are not part of the monthly estimate; the offer lists them next to its price.
- Storage uses the provider's published persistent block, volume or file storage that attaches to the offer and costs least for the size you ask for, billed for the whole month. Published free allowances (Modal's first 1 TiB), minimum sizes and size steps (Together AI's 1 TiB volumes), volume tiers and higher rates while an instance is stopped are applied where a provider publishes them. A persistent disk included with the instance is used first; a disk that only comes with the instance is never counted as free storage for more data. Object storage and temporary disks are left out.
- Storage is only priced where it can be attached: Hetzner's cloud volumes, for example, cannot be attached to its dedicated GPU servers. When none of the provider's published storage attaches to the offer, as with a dedicated model endpoint (Together AI), the estimate is marked incomplete rather than priced with storage it cannot use.
- Egress uses the provider's published price per GB, from the day the provider starts charging it. Modal's egress allowance is shared by the whole workspace per billing cycle: only what your workload adds beyond the allowance left after the workspace's other egress is charged, and with an unknown plan or unknown other usage the estimate shows a range, never a free figure. Other free allowances and volume tiers are not deducted, so estimates with heavy egress can run high.
- Charges a provider adds per instance outside the GPU price are not included: for example, an additional public or floating IP address (about €1.97 a month at OVHcloud, whose instances include one IPv4 address by default).
- Each price applies in the billing month you pick from the day it takes effect (00:00 UTC). Where a fixed price turns into a dynamic spot price inside the month (Nebius's preemptible H100, H200 and B200 from 8 Oct 2026), the days from that date have no fixed price: the estimate shows no monthly total and lists only the known costs, compute for the days before at the fixed rate (your hours spread evenly over the month's days) plus storage and egress for the whole month. A month entirely after the change has no estimate for that price.
- The range shows the lowest and highest monthly total, with the median, across estimates where every cost you asked for has a published price. Money is calculated in cents and rounded once per cost; totals show whole dollars.
- “Preferred region” uses the regions listed for each offer. When a provider operates in the region but does not say which GPUs run there, the offer stays in and is marked “not confirmed”.