GPU Rental Price Index: H100, H200, B200 and A100 Hourly Rates
A GPU rental price index answers one question: what does an hour of a given accelerator cost right now, across the market, rather than at one provider? This page condenses the same dataset that powers the live pricing table on the homepage into a low / median / high figure per GPU, explains how the index is built, and shows how to read the other published GPU price indices you will run into.
The index: on-demand $/GPU-hour by accelerator
Each row summarises the public on-demand rate for one GPU (or one GPU inside a multi-GPU node, expressed per GPU) across every provider in the dataset that lists it. "Median" is the middle listing — a better benchmark than the minimum, which is often a single marketplace host in one region.
| Accelerator | Listings | Low $/hr | Median $/hr | High $/hr | Cheapest in dataset | Notes |
|---|---|---|---|---|---|---|
| NVIDIA B200 192GB | 5 | $5.89 | $6.89 | $7.99 | Crusoe | Blackwell; supply still tight, mostly reservation-first |
| NVIDIA H200 141GB | 9 | $3.58 | $4.95 | $10.60 | Crusoe | Hyperscaler on-demand is 2–3× the specialist-cloud rate |
| NVIDIA H100 80GB | 14 | $2.29 | $2.97 | $4.99 | FluidStack | The most liquid SKU; the reference point for every index |
| AMD MI325X 256GB | 3 | $2.25 | $2.35 | $2.49 | Vast.ai | Priced against H100 on memory per dollar |
| AMD MI300X 192GB | 2 | $1.99 | $2.44 | $2.89 | FluidStack | Popular for large-model inference (192 GB on one card) |
| NVIDIA A100 80GB | 12 | $1.65 | $2.46 | $3.67 | FluidStack | See the cheapest A100 comparison |
| NVIDIA RTX 4090 24GB | 4 | $0.65 | $0.71 | $0.79 | FluidStack | Consumer card; marketplaces only, no NVLink |
| Google TPU v7 | 1 | $8.50 (Google Cloud only) | Google Cloud | Ironwood; per-chip, on-demand | ||
| Google TPU v6e | 1 | $3.22 (Google Cloud only) | Google Cloud | Trillium; see TPU explained | ||
| AWS Trainium3 / Trainium2 | 1 each | $3.85 / $2.15 (AWS only) | AWS | Per accelerator inside Trn instances | ||
| Intel Gaudi 3 / Gaudi 2 | 1 each | $1.99 / $1.29 (Intel Cloud) | Intel Tiber | Limited third-party availability | ||
Rates are public on-demand list prices in USD per GPU-hour recorded in the ComputingPower.org dataset; single-vendor accelerators have no spread. Spot, reserved and negotiated contract rates are lower — see the spread section below. Always confirm the current figure on the provider's pricing page before committing.
How this index is built
- Scope: 18 providers, covering the three hyperscalers (AWS, Azure, Google Cloud), specialist GPU clouds (CoreWeave, Lambda, Crusoe, FluidStack, Nebius and others), marketplaces (Vast.ai, RunPod) and single-vendor accelerator clouds (Groq, Cerebras, SambaNova, Intel).
- Unit: USD per single GPU per hour. Multi-GPU node prices (an 8×H100 node at $30/hr, say) are divided by GPU count so a node and a single-card marketplace listing are comparable.
- Pricing model: on-demand only. Mixing spot and reserved rates into one number hides the most important decision you make — see spot vs on-demand vs reserved.
- Region: the cheapest listed region per provider, which is usually US. European and Asia-Pacific listings for the same SKU typically run 10–30% higher.
- Cadence: the dataset is refreshed with each editorial review; the date at the top of this page is the last time the figures were checked. The about page describes the review process.
Spot, reserved and contract: the price spread behind the index
The on-demand rate is the ceiling, not the market. For the H100 the same dataset implies roughly this ladder:
| H100 80GB pricing tier | Typical $/GPU-hr | Relative to on-demand median |
|---|---|---|
| Hyperscaler on-demand | $4.99 | +68% |
| Specialist-cloud on-demand (index median) | $2.97 | — |
| Marketplace / cheapest on-demand | $2.29 | −23% |
| 1-year committed / reserved | ≈ $2.00–2.40 | −20% to −35% |
| Spot / interruptible | ≈ $1.00–1.80 | −40% to −65% |
| Multi-year negotiated cluster contract | below public rates | not published |
Two things follow. First, an "H100 price index" that tracks spot or a spot/contract composite will sit well below this page's on-demand median, and that is not a contradiction — it is a different tier. Second, the spread between the top and bottom of the ladder is larger than the spread between GPU generations: moving from hyperscaler on-demand to specialist-cloud spot cuts the H100 bill more than moving from H100 to A100 would.
How to read other GPU price indices
Several research firms and data vendors publish H100, H200 and B200 rental indices, and they are frequently quoted in market commentary. When you compare them with the table above, check three things:
- Which tier is being measured. Composite spot/contract indices, hyperscaler-only indices and marketplace-only indices can differ by 2× for the same GPU on the same day.
- Which unit. Per GPU-hour, per 8-GPU node-hour, or per month. A node-hour figure divided by eight is the only one comparable with this page.
- What "the H100 price" includes. Bare GPU-hour, or GPU plus the CPU, RAM, local NVMe and networking that come with the instance. Hyperscaler figures bundle the latter; marketplace figures usually do not. The extras that never appear in any index — storage, egress, idle time — are covered in hidden cloud GPU costs.
Using the index
- Budgeting a training run: multiply GPU-hours by the median (not the minimum) and see how to estimate LLM training cost for the GPU-hour side of the equation.
- Choosing a provider: the GPU cloud providers comparison lists who carries which SKU and where; GPU pricing by region splits the same listings into US, Europe and global marketplaces.
- Choosing a pricing tier: spot vs on-demand GPU pricing shows both rates per provider and when the spot discount survives interruptions.
- Choosing between generations: the H100 vs A100 vs H200 vs B200 comparison turns $/hour into $/training-run.
- Modelling a specific workload: the GPU cost calculator ranks providers for your model size, GPU count and duration.