AI Wiki

Cloud AI GPU Pricing Comparison

AI HardwareAI InfrastructureData Centers
8 min read
Updated
Suggest editHistory
RawGraph

Last reviewed

Sources

14 citations

Review status

Source-backed

Revision

v1 · 1,693 words

As of July 2026, the cheapest credible on-demand NVIDIA H100 rental is about $2 per GPU-hour on marketplace neoclouds: Vast.ai lists a single H100 from roughly $2.01/GPU-hr [4] and RunPod Community Cloud from about $1.99/GPU-hr [1], while dedicated single-tenant neoclouds such as Lambda Labs, Nebius, and Together AI sit at $3.29 to $3.99/GPU-hr [6][7][8]. The hyperscalers are the most expensive: AWS charges $6.88/GPU-hr for an eight-way H100 (p5) instance [10], Microsoft Azure about $6.98 [1], and Google Cloud roughly $11.06 [4], which is three to six times the neocloud floor [13]. For the newer Blackwell 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 [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 [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.

ProviderGPUOn-demand $/GPU-hrReserved/committed $/GPU-hrNotes
Vast.aiH100 SXM2.01 [4]n/r [14]Marketplace; interruptible spot from ~0.36 [1]; per-second billing
Vast.aiH200 NVL2.89 [2]n/rMarketplace listings across 30+ hosts [2]
Vast.aiB2004.69 [3]n/rSpot from ~2.69 [3]
RunPodH100 SXM3.29 [5]n/rCommunity Cloud (peer hosts) from ~1.99 [1]
RunPodH2004.39 [5]n/rCommunity Cloud ~3.59 [2]
RunPodB2005.89 [5]n/r180GB SXM [3]
NebiusH100 HGX3.85 [8]2.15 [8]Committed clusters save up to 35% [8]
NebiusH200 HGX4.50 [8]2.45 [8]Preemptible/committed tier [8]
NebiusB200 HGX7.15 [8]3.95 [8]Preemptible/committed tier [8]
Lambda LabsH100 SXM3.99 [6]contact [6]PCIe $3.29; 1x SXM $4.29 [6]
Lambda LabsB200 SXM6.69 [6]contact1x B200 $6.99 [3]
Lambda LabsH200n/rn/rNot on public price list
Together AIH1003.99 [7]3.09 [7]Reserved falls with term, 7-180 days [7]
Together AIH2005.99 [7]3.99 [7]Reserved 91-180 days [7]
Together AIB2008.19 [7]6.79 [7]Reserved 91-180 days [7]
CoreWeaveH100 HGX6.16 [9]up to -60% [9]Spot $2.46 [9]
CoreWeaveH200 HGX6.31 [9]up to -60% [9]Spot $2.62 [9]
CoreWeaveB200 HGX8.60 [9]up to -60% [9]Spot $4.26; GB200 NVL72 $10.50 [9]
AWSH100 (p5)6.88 [10]2.97 (1yr) / 2.46 (3yr) [10]8x H100 SXM; enterprise SLA
AWSH200 (p5en)7.91 [11]n/r [11]Spot ~2.90 [11]
AWSB200 (p6)14.24 [12]n/r [12]8x B200; reserved not yet offered
Microsoft AzureH1006.98 [1]contactNC H100 v5; ND H100 v5 8x SXM ~8.30 [4]
AzureH200 (ND v5)13.78 [13]contactList price, ND H200 v5
AzureB200n/rn/rND GB200 v6, reservation-based
Google CloudH100 (A3)11.06 [4]CUD (contact)Whole-VM pricing normalized per GPU
Google CloudB200 (A4)16.11 [13]reservation [13]A4/A4X require capacity reservation
Google CloudH200 (A3 Ultra)n/rn/rPriced 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 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 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 ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation.

Suggest edit