"How much to rent an H100" is a constant search, and now B300 has been added to the list. We checked three providers' public pricing pages on Oct 22, 2026. Prices change often, so treat this as a snapshot and check the links before you launch anything.
All prices are in US dollars, on demand, before tax.
Per-GPU prices at a glance
| GPU | Lambda (per GPU-hour) | Runpod Secure Cloud | Runpod Community Cloud | CoreWeave (8-GPU node, per hour) |
|---|---|---|---|---|
| H100 SXM (80GB) | $3.99 (8x) to $4.29 (1x) | $3.99 | $2.69 | $49.24 (≈ $6.16/GPU) |
| H100 PCIe (80GB) | $3.29 (1x) | $2.89 | $1.99 | n/a |
| H200 (141GB) | not listed | $5.29 | $3.59 | $50.44 (≈ $6.31/GPU) |
| B200 (180GB) | $6.69 (8x) to $6.99 (1x) | $7.99 | $5.98 | $68.80 (≈ $8.60/GPU) |
| B300 (288GB on Runpod) | not listed | $8.99 | $6.94 | Contact sales; spot $35.84 (≈ $4.48/GPU) |
A few things to know when reading this table:
- Lambda's price is per GPU, and the per-GPU rate drops slightly as you take more GPUs in one instance. Lambda's 1-Click Clusters (16+ GPUs, 2 weeks to 1 year) list at $5.54–$6.16 per H100-hour and $8.87–$9.86 per B200-hour.
- Runpod has two tiers, Secure Cloud and a cheaper Community Cloud. Read Runpod's docs on the difference before putting sensitive data on either. Runpod's page says it was last updated Oct 8, 2026.
- CoreWeave sells whole nodes. We divided by 8 to get the per-GPU figures (our math, not CoreWeave's). CoreWeave also lists "single GPU" inference prices of $6.16 (H100), $6.31 (H200) and $8.60 (B200), available only to its inference-platform customers. It says reserved commitments can save up to 60% versus on-demand.
- Spot means interruptible. CoreWeave's spot H100 node was $19.71/hour in North America, but spot capacity can be taken back.
What a job actually costs
The formula is simple: GPUs × hours × hourly rate, then add storage and any data you keep around afterwards.
Example 1: a 10-hour fine-tune on 8 H100s - Lambda or Runpod Secure Cloud: 8 × 10 × $3.99 = $319.20 - Runpod Community Cloud: 8 × 10 × $2.69 = $215.20 - CoreWeave on-demand node: 10 × $49.24 = $492.40
Example 2: the same job on 8 B200s - Lambda (8x instance): 8 × 10 × $6.69 = $535.20 - Runpod Secure Cloud: 8 × 10 × $7.99 = $639.20 - CoreWeave: 10 × $68.80 = $688.00
To price your own job across every offer above, use our GPU cloud price tracker.
Faster GPUs can finish the same job in fewer hours, so the cheaper hourly rate isn't always the cheaper job. Run a short test and measure.
Don't forget storage
GPU-hours are the headline, but disks keep billing. Runpod lists network storage at $0.07/GB/month under 1TB and $0.05 above. CoreWeave lists hot object storage at $0.06/GB/month and free egress. Delete checkpoints you don't need.
Why prices are where they are
High-end GPUs carry a lot of HBM, and HBM is in short supply. TrendForce expects that to last through 2027 (see HBM explained). If you're choosing between renting and buying a card for local models, start with how much VRAM you actually need. For a model that needs a whole node, such as DeepSeek V4.1 Flash, renting is usually the only realistic option.
Before you launch
- Put a spending limit or alert on the account.
- Shut down instances when the job ends; idle GPUs bill at full price.
- Keep API keys and tokens out of images and notebooks. Our guide to running coding agents safely covers the habits.
