Cloud GPU Pricing in 2026: What an H100 Hour Actually Costs
Per-GPU hourly rates off the vendor pages, and the thing the rate sheet gives away: the cheap-to-expensive order is not the same on Blackwell as it is on Hopper.
Here is the number you came for, per GPU per hour, taken off each vendor’s own pricing page on 22 August 2026.
| H100 SXM | H200 | B200 | B300 | |
|---|---|---|---|---|
| RunPod Community | $2.69 | $3.59 | $5.98 | $6.94 |
| RunPod Secure | $3.29 | $4.59 | $6.79 | $7.89 |
| Lambda | $3.99 | not listed | $6.69 | not listed |
| AWS Capacity Blocks | $5.19 | $5.97 | $12.36 | $14.04 |
| CoreWeave | $6.16 | $6.31 | $8.60 | contact sales |
The AWS column is Capacity Blocks for ML divided by eight, since those are whole-node reservations: p5.48xlarge at $41.528 an hour, p5e.48xlarge at $47.76, p6-b200.48xlarge at $98.84, p6-b300.48xlarge at $112.32, all US East. Lambda and CoreWeave are eight-GPU instances divided the same way. RunPod rents single GPUs, and its H100 PCIe is cheaper still at $1.99 on Community and $2.89 on Secure, which is a different chip from the SXM part in the table and not a cheaper way to buy the same thing.
Now look at the shape of it rather than the digits.
The order changes between generations
On H100, the ladder runs the way everyone says it does, except at the top. RunPod Community is the floor, CoreWeave is the ceiling, and AWS sits between Lambda and CoreWeave. A hyperscaler reservation is not the most expensive H100 hour on this table. CoreWeave is.
On B200, that inverts. AWS Capacity Blocks costs $12.36 per GPU per hour, which is 1.4 times CoreWeave and 2.1 times RunPod Community. The same provider that undercut CoreWeave on Hopper is the most expensive line on Blackwell.
I would not read that as AWS gouging Blackwell buyers specifically. It reads like allocation. Capacity Blocks pricing is set against what AWS can actually place, and the newest silicon is the part it cannot get enough of. AWS says as much on the pricing page: reservation prices are “updated regularly based on trends in supply and demand,” and it names the next repricing date. Right now that date is October 2026.
That published date is the most useful thing on the page. Reserved GPU capacity is not a fixed cost you can model out for a year. It is a rate that resets on a schedule the vendor announces in advance, and you can go read when the next reset lands before you sign anything longer than it.
Cost per hour is the wrong denominator, and you can prove it without a benchmark
The usual pitch for Blackwell is that it costs more per hour and less per token. The usual evidence is somebody else’s benchmark on somebody else’s model.
You do not need it. Take the two prices from the same vendor and you get your own break-even. On RunPod Secure, a B200 hour costs $6.79 against $3.29 for an H100 SXM, so B200 is 2.06 times the price. It pays for itself the moment it finishes your job in less than half the time, and it loses money if it does not. On Lambda the ratio is 1.68 ($6.69 against $3.99), which is a much easier bar to clear. On AWS Capacity Blocks it is 2.38.
So the answer differs by supplier for the same pair of chips, which is why a published cost-per-million-tokens figure from a third party cannot tell you what to rent. Run your own job for one hour on each and divide by the work it did. An afternoon of that is worth more than any benchmark table, including the ones in articles like this.
Where the hardware genuinely decides things is memory, not throughput. NVIDIA lists 141GB and 4.8TB/s for the H200 SXM against 80GB on the H100, and the HGX B200 system page gives 1.4TB across eight GPUs, so 180GB each. If your model fits in 141GB but not 80GB, the H200 is not a 15% upgrade over the H100, it is the difference between one GPU and a sharded setup with all the latency and complexity that brings. That threshold does not move with the price sheet.
What each tier is actually selling
The per-hour spread is only confusing if you think all five rows sell the same thing.
RunPod sells self-serve speed. Per-second billing, a GPU in a few minutes, an API that does not need a solutions architect. Community Cloud is hardware from third-party hosts and Secure Cloud is RunPod’s own datacenters, which is the entire difference between the two columns. It suits prototyping, fine-tuning runs, and bursty inference. It does not suit a 200-GPU synchronized training job.
CoreWeave charges the most per H100 hour and the money is not going into the GPU. It goes into InfiniBand fabric and large contiguous clusters, which is what stops multi-node training from choking on gradient synchronization. If your job fits on one node you are paying for a network you are not using. If it spans forty, the per-GPU rate stopped being the number you optimize a while ago.
Lambda sits in between on price and has a long-running research following. The recurring complaint is capacity: popular instance types go unavailable during demand spikes and you wait. Worth planning around if Lambda is your only supplier.
Vast.ai is a marketplace, and it deliberately publishes no rate card. Its pricing page says prices are “set by the market, not by Vast,” so the only honest way to price it is to query the live listings for the exact configuration you want. It is usually the cheapest way to get an H100 and the variance is the price of that. Host quality differs listing to listing, and a good rate from a flaky host is not a good deal.
The hyperscalers sell the thing that has nothing to do with GPUs. You stay on AWS, GCP, or Azure when your compliance boundary is already drawn there, when your data and VPC already live there and egress would cost more than the compute savings, or when you need a support contract with a name on it at three in the morning. Those are real. What they are not is a reason to skip the quote.
Two numbers to pull before you move anything
Idle hours first. If your GPUs sit cold most of the day, no reservation discount anywhere will beat simply not paying for the idle hours, and an unused reservation is the most expensive GPU there is. Reserve only what you can keep busy.
Egress second. On the hyperscalers it is the line item that quietly outgrows the compute when you are moving datasets and checkpoints around, and it is also the thing that makes a partial migration worse than either staying or leaving. Price the whole round trip, not the hour.
Then pull last month’s largest GPU line item and get a live quote for that one workload from RunPod and from your current provider. Two numbers, one afternoon. If your raw GPU decision then turns into a managed-inference decision, the inference provider comparison covers the layer above this one.
One caution about every figure above, including mine: these rates move week to week, and AWS has already told you its next move lands in October. Check the live page before you commit budget.