# Compute Cheap — The world’s cheapest GPU compute

> Affordable H100, H200, GB300, and B200 GPU compute for training, inference, and everything in between.

Compute Cheap sells limited prepaid NVIDIA H100, H200, GB300, and B200 GPU capacity for training, inference, and everything in between, at some of the lowest published prices for frontier AI compute. Capacity is offered **reserved** (held for the full block at a fixed rate) or **interruptible** (cheaper, preemptible under demand).

## Current pricing

| GPU | Interruptible (preemptible) | Reserved (non-preemptible) |
| --- | --- | --- |
| NVIDIA H100 SXM 80GB | $1.15 / GPU-hour | $1.19 / GPU-hour |
| NVIDIA H200 SXM 141GB | $1.39 / GPU-hour | $1.99 / GPU-hour |
| NVIDIA B200 SXM 180GB | $2.29 / GPU-hour | $3.10 / GPU-hour |
| NVIDIA GB300 | Sold out | Sold out |

- Prices are in USD per GPU-hour, fixed for the duration of a reservation; there is no separate platform fee.
- Minimum reservation: 2,000 GPU-hours per request (about 8 GPUs for 10 days).
- Launch sizes: 1, 2, 4, or 8 GPUs per node.
- Payment: 50% deposit only after availability is confirmed, remaining 50% billed as usage completes.
- Regions: United States and European Union.
- Purchase model: request-based prepaid reservation reviewed by Compute Cheap — not self-serve or pay-as-you-go. Submitting a request does not guarantee capacity.
- Request capacity: https://compute.cheap/request-compute
- Machine-readable prices: https://compute.cheap/api/v1/pricing (JSON, no auth)

## How it works

1. **Request capacity.** Choose H100, H200, GB300, or B200, reserved or interruptible, and how many GPU-hours you need. Minimum reservation is 2,000 GPU-hours.
2. **We confirm availability.** Capacity is limited, so we confirm it first. You prepay 50% upfront at the quoted rate, and the block is reserved.
3. **Run your workload.** Your GPUs are reserved and ready. Train, serve, and scale at the price you were quoted.

## Customers

Trusted by teams training and serving frontier models, including Runway, ElevenLabs, Midjourney, fal, Reka AI, Arcee AI.

> “We train and merge small language models on interruptible H100s. Preemptions are rare enough that normal checkpointing covers them, and the per-run cost came in well under what we paid before.”
>
> — Infra @ Arcee AI

## FAQ

### What's the difference between reserved and interruptible capacity?

Reserved capacity stays yours for the full block, at a fixed rate. Interruptible capacity costs less per GPU-hour but can be reclaimed if demand spikes, so it suits fault-tolerant training and batch jobs.

### Is there a minimum reservation?

Yes, 2,000 GPU-hours per request. That's roughly 8 GPUs for 10 days, or fewer GPUs over a longer window, split however your workload needs.

### How does the 50% deposit work?

We confirm availability for your exact request before you pay anything. Once confirmed, you prepay 50% of the block upfront to reserve it, and the remaining 50% is billed as usage completes.

### What happens if the capacity I asked for isn't available?

We only take a deposit after availability is confirmed, so you're never charged for a block we can't deliver. If exact availability is tight, we'll offer the closest match before you commit.

### Which regions is capacity available in?

United States and European Union today. Mention your preferred region in your request's notes, and we'll confirm availability there before anything is reserved.

### How is the price calculated?

Hourly rate × GPU-hours, shown before you request. There's no separate platform fee layered on top of the GPU-hour price.

## Read more

- [Cheapest compute providers leaderboard](https://compute.cheap/leaderboard): B200, H100 and H200 prices compared across providers.

- [The cheapest GPU cloud providers in 2026, ranked](https://compute.cheap/blog/cheapest-gpu-cloud-providers-2026): Every major GPU cloud’s H100, H200, B200, and GB300 price per GPU-hour, ranked from cheapest to most expensive.

- [If they can make intelligence cheap, why can’t we do the same for compute?](https://compute.cheap/blog/why-compute-should-be-cheap): The question that started Compute Cheap, and how competition, distributed supply, and a leaner stack turn it into H100s at $1.15 an hour.

---

Compute Cheap · hello@compute.cheap · HTML version: https://compute.cheap/ · Site index for agents: https://compute.cheap/llms.txt