Rent an H100: on-demand, spot, reserved and serverless options by provider

H100 capacity is easy to find in single units now. The decision is less about whether you can get one and more about which form factor, which rental model and how much you have to commit to.

First, which H100

• H100 SXM is the version in 8-GPU HGX servers with NVLink between GPUs. Pick it when a job spans several GPUs in one node (multi-GPU training, large models that need tensor parallelism). Providers often rent it one GPU at a time too.
• H100 PCIe is the card version. It is cheaper wherever both are listed and fine for single-GPU work, but multi-GPU communication is slower.
• H100 NVL (94 GB) is a PCIe variant with more memory; in the index only Runpod lists it.

The index keeps them apart, so a cheap PCIe row is not mistaken for SXM capacity.

Pick by situation

• A few hours of experiments or fine-tuning on one GPU. Per-second or per-minute on-demand billing with no minimum fits best: Runpod (per second), Hyperstack (per minute), DigitalOcean (per second, 5-minute minimum). Runpod Community Cloud and Vast.ai listings are lower, but run on third-party hosts whose reliability varies.
• A multi-day training run on 8 GPUs in one node. Look at 8-GPU HGX rows. CoreWeave sells H100 as 8-GPU instances only, and DigitalOcean offers 1 or 8 GPUs per Droplet and Lambda 1, 2, 4 or 8 per instance. Check that the provider lets you keep the node for the whole run; spot capacity can be reclaimed.
• Interruptible batch work. CoreWeave’s North American spot price is far below its on-demand rate, and Hyperstack runs PCIe spot VMs. Design for checkpoints, because capacity can be taken back.
• Steady capacity for months. Reserved pricing changes the economics: DigitalOcean publishes a 12-month H100 rate, Hyperstack publishes “starting from” reservation prices, and Lambda’s 1-Click Clusters start at 16 GPUs for a minimum of two weeks. CoreWeave and Nebius quote commitments through sales.
• An inference endpoint with uneven traffic. A serverless worker that scales to zero can cost less overall despite a higher hourly rate. See serverless GPU platforms compared.

Caveats

• Lambda. Lambda’s on-demand H100 prices for today are shown only to logged-in users. The figures on lambda.com apply from 16 November 2026 and are labelled as announced.
• Availability. “Self-serve” means you can sign up and launch without talking to sales. It is not a guarantee of stock; popular configurations sell out.
• What the rate includes. Some rates bundle CPU, RAM and local NVMe, while others (Modal) bill them separately. The GPU Cloud Index shows each provider’s extras and billing unit.