GPU Cloud Price Comparison

Compare GPU cloud rental prices across the major providers for the hardware people actually train and serve models on. Pick a GPU, enter how many hours per month you expect to run it, and the table ranks providers by total cost. Prices are representative on-demand rates so you can see the spread between hyperscalers and specialist clouds at a glance. Runs in your browser.

Representative on-demand list prices per GPU-hour, gathered for general comparison and rounded. Actual rates vary by region, availability, commitment (spot/reserved), and change frequently — always confirm with the provider. 730 hours ≈ one full month of continuous use.

How to use the GPU Cloud Price Comparison

Select a GPU and enter the number of hours per month you plan to run it — 730 is a full month of continuous use, or enter fewer hours for part-time workloads. Set the GPU count if you run several in parallel. The table lists each provider's hourly rate and the resulting monthly and annual cost, sorted cheapest first, with the lowest-cost option highlighted.

Hyperscalers (AWS, Google Cloud, Azure) usually sit at the top of the price range for on-demand instances, while specialist clouds (Lambda, RunPod, Vast.ai, CoreWeave) and marketplace providers are typically much cheaper for the same hardware. The trade-off is reliability, support, networking, and data-residency guarantees, which the headline hourly rate does not capture.

Why GPU cloud prices vary so widely

The same H100 or A100 can cost three to six times more on one cloud than another, and the gap is not arbitrary. Hyperscalers bundle enterprise networking, managed services, compliance certifications, and guaranteed capacity into the hourly rate, so their on-demand prices are the highest. Specialist GPU clouds like Lambda, RunPod, and CoreWeave focus narrowly on accelerated compute and pass through lower prices, while marketplaces such as Vast.ai let individual hosts rent out idle hardware, reaching the lowest rates of all at the cost of variable reliability.

Beyond the provider, three levers move the price. Commitment matters most: spot or interruptible instances can be 50–70% cheaper than on-demand, and one-to-three-year reservations cut rates further, while on-demand is the most expensive way to buy. Region and availability shift prices as supply tightens — H100 capacity in particular has swung with demand. And form factor counts: an SXM H100 with NVLink costs more than a PCIe card, and 80GB parts command a premium over 40GB.

For a realistic budget, decide your commitment model first. If your workload runs continuously, reserved pricing or owning hardware usually wins; if it is bursty, on-demand or spot on a specialist cloud is cheaper. The figures here are on-demand list prices to give a common baseline — treat them as a starting point and confirm the current rate and commitment discounts with each provider.

Common use cases

  • Picking a provider. See which clouds offer the cheapest H100 or A100 for your hours.
  • Budgeting a project. Turn an hours-per-month estimate into a monthly and annual GPU bill.
  • Comparing form factors. Weigh an A100 against an H100 or a consumer 4090 for cost per hour.
  • Build-vs-rent decisions. Compare annual rental cost against the price of buying the card outright.

Frequently asked questions

Are these prices current?
They are representative on-demand list prices gathered for comparison and rounded. GPU pricing changes frequently with supply and demand — H100 rates especially — so use these as a baseline and confirm the live rate with each provider before committing.

Why is AWS so much more expensive than RunPod or Vast?
Hyperscalers bundle enterprise networking, managed services, compliance, and guaranteed capacity into the price. Specialist clouds and marketplaces strip that down to focus on raw GPU compute, which is why their on-demand rates are often a fraction of a hyperscaler's for the same card.

How much can I save with spot or reserved pricing?
Spot or interruptible instances are commonly 50–70% cheaper than on-demand but can be reclaimed at short notice. One-to-three-year reservations also cut rates substantially. This tool shows on-demand prices, so treat them as the upper bound for each provider.

What does 730 hours mean?
It is the average number of hours in a calendar month (365 × 24 ÷ 12), so 730 hours represents running the GPU continuously all month. Enter fewer hours for part-time or bursty workloads to see the proportionally lower cost.

Is renting cheaper than buying?
For continuous, long-term workloads, buying or reserving hardware usually beats on-demand rental. For bursty or short-term needs, on-demand and spot pricing avoid the large upfront cost and idle time. Compare the annual figure here against the card's purchase price plus power and hosting to decide.