What an H100 costs per hour in October 2026, and what the price leaves out
We read eleven price lists on the same day. The spread is real, but a good part of it pays for things an inference server doesn't use.

H100 prices on 5 October 2026#
We read each provider's public price list or pricing API on the same day. Where a provider only sells eight-GPU machines, we divided the machine's price by eight.
| Provider | What you rent | Per GPU-hour | Note |
|---|---|---|---|
| Vast.ai | single GPUs from independent hosts | $1.80 | marketplace median; lowest $1.73 |
| RunPod Community Cloud | one GPU or more | $2.69 | |
| Hyperstack | one GPU or more | $3.20 | |
| RunPod Secure Cloud | one GPU or more | $3.49 | |
| Verda | 1, 2, 4 or 8 GPUs | $3.77 | $3.37 on 22 September |
| Lambda | 8-GPU instance | $3.99 | $4.29 for a single GPU |
| Nebius | one GPU or more | $4.50 | $3.85 until 1 October |
| CoreWeave | 8-GPU node | $6.16 | North America |
| AWS | p5.48xlarge (8 GPUs) or p5.4xlarge (1 GPU) | $6.88 | us-east-1 |
| Google Cloud | a3-highgpu-8g (8 GPUs) | $11.06 | us-central1 price table |
| Azure | ND96isr H100 v5 (8 GPUs) | $12.29 | East US |
H100 80GB SXM, on-demand list price per GPU-hour, 5 October 2026
| Provider | Bar | USD per GPU-hour |
|---|---|---|
| Vast.ai marketplace median | $1.80 | |
| RunPod Community | $2.69 | |
| Hyperstack | $3.20 | |
| RunPod Secure | $3.49 | |
| Verda | $3.77 | |
| Lambda 8-GPU instance | $3.99 | |
| Nebius | $4.50 | |
| CoreWeave 8-GPU node | $6.16 | |
| AWS us-east-1 | $6.88 | |
| Google Cloud us-central1 | $11.06 | |
| Azure East US | $12.29 |
Azure's rate is 4.6 times RunPod Community's. Even the cheapest hyperscaler, AWS, charges more than every specialist cloud we checked. Every row rents the same NVIDIA H100 SXM, so the gap is in what surrounds the chip: the machine it comes in, the network, the contract and the support. Some of that is worth paying for. For serving one model, much of it isn't.
Why the hyperscalers charge two to three times more#
AWS, Google Cloud and Azure sell their H100s mainly as eight-GPU machines designed for training across many machines at once. The price covers high-bandwidth networking between those machines, plus the rest of the cloud: identity and access controls, private networking, compliance programmes and the support contract you may already have.
A model served from one machine never uses the network between machines. Two H100s in the same server talk over NVLink inside the box.
Hyperscaler list prices also assume you will commit. One year takes 31 to 37% off; three years takes about 56%:
| Provider | On-demand | 1-year commitment | 3-year commitment |
|---|---|---|---|
| AWS (EC2 Instance Savings Plan, no upfront) | $6.88 | $4.33 | $2.97 |
| Google Cloud (resource commitment) | $11.06 | $7.67 | $4.86 |
| Azure (reserved VM instance) | $12.29 | $7.87 | $5.40 |
A three-year AWS commitment lands within 28 cents of the cheapest on-demand H100 we found. That is a good deal if your load is steady, you already spend with AWS, and you are confident an H100 will still be the right GPU for your model in 2029. AWS also cut on-demand P5 prices by up to 45% in June 2025, so $6.88 is already the reduced rate.
SXM, PCIe and NVL are different products#
"H100" covers three cards with different memory and bandwidth:
| H100 variant | Memory | Memory bandwidth | Lowest on-demand price we found |
|---|---|---|---|
| SXM | 80 GB | 3.35 TB/s | $2.69 (RunPod Community) |
| PCIe | 80 GB | 2.0 TB/s | $1.99 (RunPod Community) |
| NVL | 94 GB | 3.9 TB/s | $2.59 (RunPod Community) |
Each generated token requires reading the model's weights from memory, so for a single stream of text the speed limit is memory bandwidth rather than compute. NVIDIA's inference guide walks through why. Per unit of bandwidth, the SXM card is the better buy: $2.69 for 3.35 TB/s works out at $0.80 per TB/s-hour, against $1.00 for the PCIe card at $1.99.
The NVL's extra 14 GB matters in one common case. A 70B model at FP8 loads on an 80 GB card with almost nothing left over for users' conversations; on 94 GB it has room to work. We did the arithmetic in how much GPU memory a 70B model needs.
Costs the hourly rate leaves out#
Egress#
| Provider | Internet egress |
|---|---|
| AWS (us-east-1) | first 100 GB a month free, then $0.09 per GB |
| Azure (East US) | first 100 GB free, then $0.087 per GB |
| Google Cloud (Premium tier) | first 1 GiB free, then $0.12 per GiB ($0.11 above 1 TiB) |
| Lambda, RunPod, CoreWeave, Nebius, Hyperstack | free |
Sending 10 TB a month costs about $891 on AWS, $861 on Azure and $1,035 on Google Cloud's Premium tier, and nothing at the five specialist clouds above.
For a text endpoint, egress is usually small. Each streamed token arrives as a small JSON event of very roughly 100 to 200 bytes, so a billion streamed tokens is around 100 to 200 GB: $9 to $18 at AWS's rate. It adds up when you move other data out: embeddings, generated images or audio, logs shipped elsewhere, or traffic between an application on one cloud and GPUs on another.
Storage#
A 70B model at FP8 is about 73 GB on disk. At $0.07 to $0.10 per GB a month at most providers, and $0.20 at Lambda and Verda, keeping it costs $5 to $15 a month. It won't decide anything.
Billing increments#
- Per second: AWS (60-second minimum), Google Cloud (1-minute minimum), RunPod, Nebius and Vast.ai.
- Per minute: Azure, Lambda and Hyperstack.
- Prepaid 10-minute blocks: Verda, with unused time refunded.
This matters if you add and remove replicas through the day. For a server that runs around the clock, it doesn't.
The smallest unit you can rent#
Several offerings come only as whole eight-GPU machines: AWS's H200 and B200 instances, Google Cloud's A3 Ultra and A4, Azure's ND H100 v5 and H200 v5, and CoreWeave's on-demand nodes. If your model needs two GPUs and you rent eight at $6.88 each, you pay $55.04 an hour and six GPUs sit idle unless you pack other work onto them. AWS does sell a single H100 (p5.4xlarge) at the same per-GPU rate. Lambda, RunPod, Verda, Hyperstack and Vast.ai rent single GPUs.
Rent the smallest unit that holds the model and its cache.
Spot and preemptible prices#
| Provider | Spot or preemptible H100 | Below on-demand by |
|---|---|---|
| AWS (p5.48xlarge, us-east-1, read at 16:26 UTC) | $3.20 | 54% |
| Google Cloud (a3-highgpu-8g) | $6.29 | 43% |
| Azure (ND96isr H100 v5, East US) | $2.27 | 82% |
| CoreWeave | $2.46 | 60% |
| Verda | $1.89 | 50% |
| Together AI (GPU clusters) | $1.99 | 50% |
Spot capacity can be taken back at short notice. It suits batch work such as evaluations, embeddings and offline generation. For an endpoint people are waiting on, use it only behind a fallback that can absorb the traffic when machines disappear.
Prices rose this autumn#
- Nebius raised on-demand prices on 1 October: H100 by 17% ($3.85 to $4.50), H200 by 20% and B200 by 19%.
- AWS announced a 15% rise in EC2 Capacity Block prices from 7 October, taking a P5 H100 from $5.19 to $5.97 per GPU-hour.
- Verda's H100 went from $3.37 on 22 September to $3.77 on 5 October, up 12% in 13 days, according to the Wayback Machine's captures of its price page.
Any price survey goes stale within weeks, this one included. For today's number, live trackers such as gpufinder.dev, which polls 22 vendors every hour, and SkyPilot's GPU Compass are better sources. A survey is more useful for explaining what the number buys.
What we would rent for inference this month#
- One model on one or two GPUs, steady traffic: single or paired H100 SXM GPUs at a specialist cloud such as Hyperstack ($3.20), RunPod Secure ($3.49) or Verda ($3.77). Check capacity in your region first. A list price says nothing about whether the GPU is free to rent.
- Already on a hyperscaler with credits or compliance requirements: stay, but commit. Three years on AWS is $2.97 per GPU-hour, and the single-GPU p5.4xlarge saves you from renting eight.
- Experiments and batch jobs: Vast.ai or spot capacity, with checkpoints so a lost machine costs minutes instead of a whole run.
- A model that needs more than 80 GB per GPU: price one H200 ($3.59 to $10.60 on-demand) against a pair of H100s before reaching for an eight-GPU machine.
The lowest rate on the list is rarely the cheapest deployment. A GPU that sits idle half the day costs twice its hourly rate per useful hour, which is the subject of whether renting GPUs beats paying per token.
This comparison is what our configurator runs for each workload: price and capacity across these providers for the GPU the model needs, ranked so you can choose where it runs. After it's deployed, we tell you when a faster or cheaper option appears, and whether to move is your call. We own no GPUs and favour no provider.
How we worked this out#
Sources#
All accessed 5 October 2026.
- AWS: EC2 on-demand pricing, Savings Plans, Spot pricing, Capacity Blocks pricing, June 2025 GPU price reduction
- Google Cloud: accelerator-optimized VM pricing, network pricing
- Azure: Retail Prices API, bandwidth pricing
- Lambda pricing and billing documentation
- RunPod pricing and pricing documentation
- Nebius prices
- Hyperstack GPU pricing
- Verda pricing, billing documentation, Wayback Machine captures of 22 September, 28 September and 30 September 2026
- CoreWeave pricing
- Vast.ai pricing
- Together AI pricing
- NVIDIA: H100 specifications, H100 PCIe product brief, Mastering LLM techniques: inference optimization

