# Cloud AI GPU Pricing Comparison

> Source: https://aiwiki.ai/wiki/cloud_gpu_pricing_comparison
> Updated: 2026-07-07
> Categories: AI Hardware, AI Infrastructure, Data Centers
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

As of July 2026, the cheapest credible on-demand NVIDIA [H100](/wiki/nvidia_h100) rental is about $2 per GPU-hour on marketplace neoclouds: [Vast.ai](/wiki/vast_ai) lists a single H100 from roughly $2.01/GPU-hr [4] and [RunPod](/wiki/runpod) Community Cloud from about $1.99/GPU-hr [1], while dedicated single-tenant neoclouds such as [Lambda Labs](/wiki/lambda_labs), [Nebius](/wiki/nebius), and [Together AI](/wiki/together_ai) sit at $3.29 to $3.99/GPU-hr [6][7][8]. The hyperscalers are the most expensive: [AWS](/wiki/amazon_web_services) charges $6.88/GPU-hr for an eight-way H100 (p5) instance [10], [Microsoft Azure](/wiki/microsoft_azure) about $6.98 [1], and [Google Cloud](/wiki/google_cloud) roughly $11.06 [4], which is three to six times the neocloud floor [13]. For the newer Blackwell [B200](/wiki/nvidia_b200), the cheapest on-demand rates are about $4.69/GPU-hr on marketplaces (Vast.ai) [3] and $5.89 to $6.99 on dedicated neoclouds (RunPod $5.89 [5], Lambda $6.69 [6]), versus $8.19 to $8.60 at Together AI and [CoreWeave](/wiki/coreweave) [7][9] and $14.24/GPU-hr on AWS [12]. Spot and interruptible tiers are cheaper still: H100 spot listings dip to about $0.36/GPU-hr [1]. All prices below are USD per GPU-hour, verified against provider pricing pages and dated aggregators in July 2026.

**Cheapest on-demand by GPU (July 2026, USD/GPU-hr):**

- H100: about $2.01 on Vast.ai [4] and $1.99 on RunPod Community Cloud [1]; roughly $3.29 to $3.99 on dedicated neoclouds (Lambda PCIe $3.29 [6], Nebius $3.85 [8], Lambda SXM and Together $3.99 [6][7]).
- H200: about $2.30 on [FluidStack](/wiki/fluidstack) [2] and $2.89 on Vast.ai [2]; $4.39 to $4.50 on RunPod and Nebius [5][8].
- B200: about $4.69 on Vast.ai [3]; $5.89 to $7.15 on RunPod, Lambda, and Nebius [5][6][8].

## Cross-provider GPU rental price table

Prices are USD per GPU-hour. "On-demand" is the published pay-as-you-go rate; "Reserved/committed" is the lowest disclosed committed rate (annual reserved, multi-month commitment, or committed-use discount). Blank cells marked n/r were not reliably published as of the verification date. Last verified: July 2026.

| Provider | GPU | On-demand $/GPU-hr | Reserved/committed $/GPU-hr | Notes |
|---|---|---|---|---|
| [Vast.ai](/wiki/vast_ai) | [H100](/wiki/nvidia_h100) SXM | 2.01 [4] | n/r [14] | Marketplace; interruptible spot from ~0.36 [1]; per-second billing |
| Vast.ai | [H200](/wiki/nvidia_h200) NVL | 2.89 [2] | n/r | Marketplace listings across 30+ hosts [2] |
| Vast.ai | [B200](/wiki/nvidia_b200) | 4.69 [3] | n/r | Spot from ~2.69 [3] |
| [RunPod](/wiki/runpod) | H100 SXM | 3.29 [5] | n/r | Community Cloud (peer hosts) from ~1.99 [1] |
| RunPod | H200 | 4.39 [5] | n/r | Community Cloud ~3.59 [2] |
| RunPod | B200 | 5.89 [5] | n/r | 180GB SXM [3] |
| [Nebius](/wiki/nebius) | H100 HGX | 3.85 [8] | 2.15 [8] | Committed clusters save up to 35% [8] |
| Nebius | H200 HGX | 4.50 [8] | 2.45 [8] | Preemptible/committed tier [8] |
| Nebius | B200 HGX | 7.15 [8] | 3.95 [8] | Preemptible/committed tier [8] |
| [Lambda Labs](/wiki/lambda_labs) | H100 SXM | 3.99 [6] | contact [6] | PCIe $3.29; 1x SXM $4.29 [6] |
| Lambda Labs | B200 SXM | 6.69 [6] | contact | 1x B200 $6.99 [3] |
| Lambda Labs | H200 | n/r | n/r | Not on public price list |
| [Together AI](/wiki/together_ai) | H100 | 3.99 [7] | 3.09 [7] | Reserved falls with term, 7-180 days [7] |
| Together AI | H200 | 5.99 [7] | 3.99 [7] | Reserved 91-180 days [7] |
| Together AI | B200 | 8.19 [7] | 6.79 [7] | Reserved 91-180 days [7] |
| [CoreWeave](/wiki/coreweave) | H100 HGX | 6.16 [9] | up to -60% [9] | Spot $2.46 [9] |
| CoreWeave | H200 HGX | 6.31 [9] | up to -60% [9] | Spot $2.62 [9] |
| CoreWeave | B200 HGX | 8.60 [9] | up to -60% [9] | Spot $4.26; GB200 NVL72 $10.50 [9] |
| [AWS](/wiki/amazon_web_services) | H100 (p5) | 6.88 [10] | 2.97 (1yr) / 2.46 (3yr) [10] | 8x H100 SXM; enterprise SLA |
| AWS | H200 (p5en) | 7.91 [11] | n/r [11] | Spot ~2.90 [11] |
| AWS | B200 (p6) | 14.24 [12] | n/r [12] | 8x B200; reserved not yet offered |
| [Microsoft Azure](/wiki/microsoft_azure) | H100 | 6.98 [1] | contact | NC H100 v5; ND H100 v5 8x SXM ~8.30 [4] |
| Azure | H200 (ND v5) | 13.78 [13] | contact | List price, ND H200 v5 |
| Azure | B200 | n/r | n/r | ND GB200 v6, reservation-based |
| [Google Cloud](/wiki/google_cloud) | H100 (A3) | 11.06 [4] | CUD (contact) | Whole-VM pricing normalized per GPU |
| Google Cloud | B200 (A4) | 16.11 [13] | reservation [13] | A4/A4X require capacity reservation |
| Google Cloud | H200 (A3 Ultra) | n/r | n/r | Priced per VM; per-GPU not published |

## Which cloud has the cheapest H100 right now?

For pure price, marketplace neoclouds win. Vast.ai matches independent hosts to renters and lists 1x H100 from about $2.01/GPU-hr on demand [4], and RunPod Community Cloud starts near $1.99/GPU-hr [1]. These tiers trade some reliability and single-tenant isolation for price. If you want a dedicated, single-tenant instance with a support contract, the cheapest tier is the dedicated neoclouds: Lambda Labs H100 PCIe at $3.29/GPU-hr [6], Nebius HGX H100 at $3.85 on demand (and $2.15 committed) [8], and Lambda SXM or Together AI at $3.99 [6][7]. Every hyperscaler is more expensive: AWS $6.88 [10], Azure $6.98 [1], and Google Cloud about $11.06/GPU-hr [4]. So the practical ranking, cheapest to priciest, is marketplace neoclouds, then dedicated neoclouds, then CoreWeave (which sits between the two groups at $6.16 on demand but discounts up to 60% on committed capacity) [9], then AWS and Azure, then Google Cloud.

## What does it cost to rent a B200 in 2026?

Blackwell B200 supply loosened through 2026, and on-demand rates now start around $4.69/GPU-hr on Vast.ai [3] and about $5.89 on RunPod [5]. Lambda Labs lists B200 SXM from $6.69 (eight-way) to $6.99 (single) [6], Nebius at $7.15 on demand or $3.95 committed [8], Together AI at $8.19 [7], and CoreWeave HGX B200 at $8.60 [9]. Hyperscalers remain the ceiling: AWS's p6-b200 works out to $14.24/GPU-hr on demand [12], and Google Cloud's A4 (B200) runs about $16.11/GPU-hr but requires a capacity reservation [13]. For rack-scale Grace-Blackwell, CoreWeave publishes GB200 NVL72 at $10.50/GPU-hr on demand [9].

## Why do neoclouds undercut the hyperscalers?

Specialist GPU clouds (often called neoclouds) are cheaper for structural reasons, not just promotions:

1. Single-purpose infrastructure. Neoclouds run GPU compute and little else, so they avoid the overhead of a broad service catalog and can pass thin margins on to renters. AIMultiple's index finds hyperscaler posted prices are typically three to six times the lowest neocloud listings for an identical GPU [13].
2. Less virtualization and bundling. Providers like Lambda, CoreWeave, and Nebius offer bare-metal or lightly virtualized instances and do not bake in the managed networking, storage, and egress premiums that inflate hyperscaler bills.
3. Cheaper and sometimes stranded power. Operators such as [Crusoe](/wiki/crusoe) historically sited capacity at low-cost or otherwise-wasted energy, lowering the largest variable cost of running an H100 fleet.
4. Marketplace and spot supply. Vast.ai and RunPod Community Cloud aggregate third-party and idle capacity, and interruptible listings drop H100 rates below $1/GPU-hr and as low as about $0.36 [1].
5. Hyperscaler price includes more than the GPU. AWS, Azure, and Google Cloud charge for enterprise SLAs, compliance certifications (HIPAA, SOC 2, FedRAMP), global regions, integrated services, and premium support, which enterprises often need [13]. That bundle, not the silicon, is much of the price gap.

## On-demand vs reserved vs spot: which tier should you pick?

The same GPU can vary 5x or more in price depending on the commitment tier, so match the tier to the workload:

- Spot / interruptible / community is cheapest and best for fault-tolerant training or batch jobs that checkpoint. Vast.ai and RunPod Community Cloud, plus CoreWeave spot ($2.46 for H100) [9] and AWS Spot (about $2.90 for H200) [11], live here.
- On-demand is the flexible default for development, inference bursts, and short runs. Neocloud on-demand ($2 to $4 for H100) is far below hyperscaler on-demand ($6.88 to $11.06) [4][10].
- Reserved / committed rewards steady, predictable demand. AWS 1-year and 3-year reservations cut the p5 H100 to $2.97 and $2.46/GPU-hr [10], Nebius committed clusters reach $2.15 for H100 [8], Together AI reserved falls to $3.09 [7], and CoreWeave discounts committed capacity up to 60% [9]. Hyperscaler committed-use and savings plans narrow but do not close the gap to neoclouds.

## Caveats: what these prices do and do not include

Published per-GPU-hour rates exclude storage, data egress, inter-node InfiniBand or NVLink fabric where charged separately, and any minimum-commitment or region surcharges. Multi-GPU instances often price slightly lower per GPU than single-GPU ones (Lambda H100 SXM is $3.99 at eight-way versus $4.29 for one GPU) [6]. Variant matters too: H100 SXM (used in [HGX](/wiki/nvidia_hgx) systems) usually costs more than H100 PCIe or NVL, and hyperscaler "single-GPU" SKUs (Azure NC v5) can look cheaper per GPU than their eight-way SXM counterparts (ND v5). GPU cloud prices move constantly as supply and demand shift, so treat every figure here as a July 2026 snapshot and reconfirm against the linked provider pages before committing spend. Where a rate could not be verified, the cell is marked n/r rather than estimated.

## References

1. GetDeploying, "H100 Cloud Pricing: Compare 47+ Providers (2026)," updated 2026-07-07. https://getdeploying.com/gpus/nvidia-h100
2. GetDeploying, "H200 Cloud Pricing: Compare Providers (2026)," updated 2026-07-07. https://getdeploying.com/gpus/nvidia-h200
3. GetDeploying, "B200 Cloud Pricing: Compare 25+ Providers (2026)," updated 2026-07-07. https://getdeploying.com/gpus/nvidia-b200
4. Thunder Compute, "NVIDIA H100 Pricing (July 2026): Cheapest Cloud GPU Rates," 2026-07-01. https://www.thundercompute.com/blog/nvidia-h100-pricing
5. RunPod, "GPU Cloud Pricing," pricing page, accessed 2026-07-07. https://www.runpod.io/pricing
6. Lambda, "GPU Cloud Pricing," accessed 2026-07-07. https://lambda.ai/service/gpu-cloud
7. Together AI, "Pricing," accessed 2026-07-07. https://www.together.ai/pricing
8. Nebius, "AI Cloud Pricing," accessed 2026-07-07. https://nebius.com/prices
9. CoreWeave, "Pricing," accessed 2026-07-07. https://www.coreweave.com/pricing
10. Vantage, "p5.48xlarge Pricing (AWS EC2, 8x H100)," accessed 2026-07-07. https://instances.vantage.sh/aws/ec2/p5.48xlarge
11. Vantage, "p5en.48xlarge Pricing (AWS EC2, 8x H200)," accessed 2026-07-07. https://instances.vantage.sh/aws/ec2/p5en.48xlarge
12. Vantage, "p6-b200.48xlarge Pricing (AWS EC2, 8x B200)," accessed 2026-07-07. https://instances.vantage.sh/aws/ec2/p6-b200.48xlarge
13. AIMultiple, "Cloud GPU Rental Price Index," July 2026. https://aimultiple.com/gpu-index
14. Vast.ai, "Pricing," accessed 2026-07-07. https://vast.ai/pricing