GPU Cloud Provider Directory: Where to Check Real Prices
Why a directory instead of a price table? GPU cloud prices change constantly and vary by region, commitment, and availability — any static table we published would be wrong within weeks, and quoting prices that aren't ours would be misleading. So instead of pretending, this page does the honest thing: it points you to the official pricing pages of 18 GPU clouds, organized by type, so you can verify live rates yourself in minutes.
On this page
How to use this directory
A good comparison takes about 20 minutes: pick the GPU you need (see our GPU comparison guide), open 3–4 pricing pages below, and note the per-GPU hourly rate for your region and pricing model (on-demand vs spot). Then sanity-check the total with our cost calculator. Remember that the headline rate is never the whole bill — storage, egress, and minimum billing increments all matter (see how GPU cloud pricing works).
All links below were verified to load on September 29, 2026. Provider pages move sometimes; if a link breaks, please tell us and we'll fix it.
The hyperscalers
The big three clouds: the broadest GPU catalogs and deepest ecosystems (IAM, managed services, startup credits) — usually at the highest list prices.
AWS
HyperscalerEC2 GPU instances (P5/P4 families and more) plus AWS's own Trainium and Inferentia chips. Committed-use and spot options; startup credits are common.
EC2 on-demand pricing →Google Cloud
HyperscalerNVIDIA GPUs plus Google's own TPUs. Per-second billing, committed-use discounts, and spot VMs. Note that list rates here are often the highest in third-party surveys.
GPU pricing page →Microsoft Azure
HyperscalerNC/ND-series GPU VMs, reserved instances, and spot VMs. Strong choice if your team already lives in the Microsoft ecosystem.
Virtual Machines pricing →GPU neoclouds
Clouds built specifically for AI compute. They typically publish straightforward on-demand GPU rates and often undercut hyperscaler list prices — the trade-off is a smaller ecosystem of managed services.
Lambda
NeocloudAI-focused cloud (formerly Lambda Labs) with published on-demand GPU pricing, including newer Blackwell cards. Popular with researchers and startups.
Pricing page →CoreWeave
NeocloudSpecialized AI cloud repeatedly rated among the top GPU clouds in independent benchmarks. Premium list pricing; geared toward larger and enterprise workloads.
Pricing page →Nebius
NeocloudAI cloud with European roots and published on-demand GPU pricing. One of the providers listing newer-generation GPUs with public rates.
Pricing page →Hyperstack
NeocloudSeparates on-demand, reservation, and interruptible (spot) pricing on its rate card — a clear example of the three pricing models side by side.
Pricing page →DataCrunch
NeocloudEuropean GPU cloud with published rates. Worth including in any comparison if your workloads can run in EU regions.
Pricing page →TensorWave
NeocloudOne of the few clouds listing AMD Instinct MI300X instances with public rates — relevant if you want to compare beyond NVIDIA.
Pricing page →GMI Cloud
NeocloudLists on-demand rates for newer cards including H200 and GB200. Has appeared with competitive listed rates in recent third-party surveys.
Official site →DeepInfra
NeocloudServerless inference plus GPU rental. Frequently among the lowest listed on-demand rates in third-party price surveys.
Official site →Cudo Compute
NeocloudMarketplace-style cloud. Has appeared with low listed H100 rates in recent third-party surveys — a good reminder to check beyond the big names.
Official site →Crusoe
NeocloudAI cloud known for sustainable data centers powered by otherwise-wasted energy. The only provider in some surveys listing AMD MI300X/MI355X on its rate card.
Official site →On-demand & marketplaces
Developer-friendly platforms and peer-to-peer marketplaces. This is usually where the lowest listed hourly rates appear — with the caveat that availability and support vary more than at the big clouds.
RunPod
On-demandSecure Cloud (guaranteed) and Community Cloud (cheaper, peer-hosted) pods with per-second billing. Extremely popular with indie developers and researchers.
Pricing page →Vast.ai
MarketplacePeer-to-peer GPU marketplace. Frequently the lowest listed rates in third-party surveys — and the widest variance, since hosts set their own prices.
Pricing page →DigitalOcean
On-demandGPU Droplets with simple, flat pricing (Paperspace is now part of DigitalOcean). A gentle on-ramp if you already use DO for the rest of your stack.
GPU Droplets →OVHcloud
On-demandEuropean cloud with published public-cloud GPU rates. Worth checking if you want EU data residency.
Public cloud prices →Scaleway
On-demandEuropean cloud whose console reads from a public product-catalog API — pricing is unusually transparent and scriptable.
Pricing page →Salad
CommunityLets everyday PC owners share idle GPUs. A different model from the rest — interesting for small, flexible workloads, not for production guarantees.
Official site →Open price data for developers
If you'd rather query prices programmatically than click through pricing pages, these open resources exist:
- gpuhunt — an open-source Python library that queries GPU offers and prices across AWS, Azure, GCP, and other providers, including spot/interruptible instances. Useful for building your own comparisons.
- Vast.ai's public search API — the same endpoint their own search UI uses; third-party trackers use it to monitor marketplace rates.
- Community price snapshots — developers periodically publish surveyed rate tables (for example, "what it costs to rent an H100/B200/RTX 4090" roundups). Treat these as directional, and always re-verify with the provider before budgeting.
A pattern worth knowing from recent third-party surveys (September 2026): the cheapest listed on-demand rate for flagship GPUs is often 3–4× lower than the most expensive hyperscaler list rate for the same card — and the cheapest listings are frequently on marketplaces or smaller neoclouds. The gap is real, but so are the trade-offs: support, availability guarantees, networking, and managed services.
What to check on each pricing page
Before you compare two numbers, make sure you're comparing the same thing:
- Per-GPU vs per-machine. Some pages quote an 8-GPU node; divide by 8 before comparing with single-GPU rates.
- Pricing model. On-demand, spot, and reserved rates for the same GPU can differ by multiples. Match the model to your workload (see our pricing guide).
- Region. The same GPU can cost different amounts in different regions — and may be sold out in yours.
- Billing increment. Per-second vs per-minute vs per-hour minimums change the effective cost of short experiments.
- What's excluded. Storage, egress, and IP addresses are usually billed separately.
- Availability, not just price. The cheapest listed rate is worthless if there's no capacity when you need it. Have a fallback GPU in mind.
Once you have real numbers, plug them into the cost calculator alongside your estimated GPU-hours (see how to estimate them) for a budget you can defend.