Cloud Provider Database
Provider profiles, not stale price sheets.
Cloud GPU prices move weekly, so a printed price table lies within days. What stays stable is each provider's shape: how they bill, whether spot exists, what egress costs, which GPUs they stock and who they're built for. Use the profiles below to shortlist — then check the live console for today's number. Indicative H100 bands last reviewed .
AWS, Google, Azure, OCI, Alibaba. Priciest on-demand, deepest compliance, everything integrates.
CoreWeave, Lambda, Nebius, Crusoe & co. Purpose-built GPU fleets, InfiniBand, 30-70% cheaper.
Vast.ai, RunPod, TensorDock. Aggregated supply, lowest prices, reliability varies by host.
Together, Modal, Replicate, Fireworks. Per-second or per-token; zero instance management.
Quick compare
All providers, sorted by lowest indicative H100 band
| Provider | Type | H100 / GPU-hr | Billing | Egress |
|---|
Bands are on-demand list prices per GPU-hour where a provider publishes them, or the typical observed range where pricing is dynamic (marketplaces) or usage-based (serverless, converted at full utilisation). Spot, reserved and committed pricing is often 40-90% lower. Last reviewed — always verify in the provider console before committing spend.
How to choose
A 30-second decision rule
Sustained training or 24/7 inference? AI-native clouds win — reserved Nebius, Lambda or CoreWeave capacity beats everything on $/GPU-hr at high utilisation.
Bursty or experimental? Marketplaces (Vast.ai, RunPod Community) for checkpoint-tolerant jobs; serverless (Modal, RunPod Serverless) when cold-start cost beats idle cost.
Just need tokens, not GPUs? Together AI or Fireworks per-token pricing usually beats self-hosting below ~50-70% sustained utilisation.
Compliance, VPC peering, or an existing enterprise agreement? Stay hyperscaler — the premium buys audit trails and integration. OCI is the cheapest of the four.
Not sure how much VRAM you even need? Start with the GPU Memory Calculator, then come back here to price the card that fits.
Common questions
Why price bands instead of exact hourly prices?
Cloud GPU prices move weekly and vary by region, commitment and capacity. A printed price is stale the day it ships. Bands plus the provider's billing model tell you which platforms to shortlist; the provider's own console gives the live number.
What is the cheapest way to rent an H100?
Marketplaces such as Vast.ai routinely offer the lowest interruptible H100 prices, followed by AI-native clouds like Nebius, Hyperstack and RunPod on-demand. Hyperscalers cost the most on-demand but offer the deepest compliance and integration.
When does serverless beat renting a GPU?
Below roughly 50-70% sustained utilisation, per-second or per-token platforms like Modal, Together AI or Fireworks are usually cheaper than an idle rented GPU, and they remove all instance management.