Cheapest Cloud GPU Providers in 2026: Real Prices Compared

Every "cheapest GPU cloud" article ends the same way: a table with one provider circled and a claim that it's the winner. That claim rarely survives contact with your actual bill, because the advertised hourly rate is not the number that determines what you pay. Egress, storage, idle time, minimum billing increments, and whether a "price" is per-GPU or per-node all move the real cost by more than the gap between providers usually does.

TL;DR: As of August 2026, on-demand H100 rates range from about $1.99/GPU-hr (RunPod, PCIe, Community tier) to $6.88/GPU-hr (AWS p5, on-demand); A100 80GB ranges from $1.19 to $3.43/GPU-hr; L40S from $0.79 to $2.25/GPU-hr; RTX 4090 from $0.34/hr (RunPod) up. No single provider is cheapest across every GPU — RunPod and Lambda undercut on-demand rates most consistently, but egress fees, storage billed separately, and idle time typically add more to your bill than the difference between two providers' hourly rates. Compare total cost, not the sticker price.

What actually determines "cheapest," if not the hourly rate

The hourly rate is the number every provider puts on the homepage because it's the number that wins comparisons. It's also the smallest lever most people pull on their actual bill. Four things usually matter more:

Whether the price is per-GPU or per-node. Some providers price by the individual GPU; others sell nodes and expect you to divide. CoreWeave's H100 HGX node is $49.24/hr for 8 GPUs — that's $6.16/GPU-hr, not $49.24/GPU-hr (CoreWeave pricing). AWS's p5.48xlarge is $55.04/hr for 8 H100s, or $6.88/GPU-hr — the same instance family's 1-GPU variant, p5.4xlarge, is priced separately at $6.88/hr (AWS via Vantage). Get this backwards on any provider and you'll compare a per-node number against someone else's per-GPU number and draw a conclusion that's wrong by roughly the GPU count.

Egress. Moving data out of a provider's network — model weights, checkpoints, datasets — is frequently billed separately from compute, and that fee doesn't show up in a GPU price comparison at all. If your workflow moves large files in and out regularly, check the egress line item before the hourly rate; on some providers it's the larger cost over a month of active use.

Storage billed separately from compute. A persistent volume attached to your instance usually keeps billing after you stop the GPU, at its own per-GB rate. That's often the right tradeoff — you want your data to survive — but it means the "$X/hour" headline isn't your full bill even while the GPU itself is off.

Minimum billing increments and idle time. Some providers bill by the second, some round up to the minute or the hour. On a short job, rounding matters. But idle time dwarfs rounding: a GPU that's running but not doing useful work — between training runs, during setup, mid-debug session — bills at the full rate whether or not anything is happening on it. This is usually the single largest gap between the rate you compared and the number on your invoice.

Commitment discounts. Every provider's on-demand rate has a cheaper reserved or contract-priced sibling. SemiAnalysis's tracked one-year H100 contract pricing sat between $2.10 and $2.70/GPU-hr as of April 2026, meaningfully below spot on-demand rates in the same window (SemiAnalysis GPU index) — but a commitment only pays off if you can actually keep the GPU busy for the term, which brings you back to idle time as the real variable.

On-demand prices, provider by provider (August 2026)

All numbers below are per-GPU, on-demand (no commitment), pulled from each provider's own public pricing page on August 23, 2026. Where a provider only sells whole nodes, both the node price and the per-GPU basis are shown so you can see the math rather than take it on faith.

* Node-priced (8 GPUs); figure shown is node rate divided by GPU count. See sources at the end for each exact page.

Vast.ai deserves its own line because it isn't a rate card — it's a marketplace where independent hosts set their own prices, so the "price" is a range that changes by the hour depending on who's listing. Treat the low end as a floor you might land on, not a rate you're guaranteed. It's frequently the cheapest number on this page and also the least predictable one.

RunPod's Community tier is the cheapest H100, A100, L40S, and RTX 4090 rate we could verify on a fixed-price provider as of this snapshot — its own pricing page shows $1.99/hr for H100 PCIe and $0.34/hr for RTX 4090 (RunPod pricing). That doesn't make it the cheapest choice for every workload — Community tier trades some reliability guarantees for the lower rate, which is exactly the kind of tradeoff a headline number hides. Check current availability and terms on the live marketplace before treating any number on this page as guaranteed available today.

Total cost of ownership: why the cheapest rate isn't the cheapest job

Say you need 200 GPU-hours of H100 time to finish a training run. At RunPod's $1.99/hr Community rate, that's $398 in compute. At AWS's $6.88/hr on-demand rate, the same 200 hours costs $1,376 — a 3.5x difference that looks decisive.

Now add idle time, which almost every real workload has and almost no price comparison accounts for. If your actual session runs for 10 days and the GPU is doing useful work 8 hours a day — training between debugging, data prep, and evaluation — you're paying for 240 wall-clock hours to get 80 hours of real training, at whatever the hourly rate is. At $1.99/hr that's $477.60 for 80 hours of real work; at $6.88/hr it's $1,651.20. The provider gap didn't shrink, but it's no longer the biggest number in the equation — the 160 idle hours cost you $318.40 to $1,100.80 depending on provider, money spent on a GPU sitting there doing nothing.

This is the actual argument for stopping a GPU the moment it's idle rather than leaving a session running "just in case," and it's a bigger lever than provider choice for most workloads. Our own breakdown of what idle time costs, with the idle cost calculator, walks through the math for your specific rate and usage pattern. If you're planning a training run in advance, the training cost calculator does the same for a fixed compute budget.

The catch with stopping and restarting to avoid idle billing: most providers don't guarantee your environment survives the stop. If a terminated instance means reinstalling CUDA, rebuilding your Python environment, and re-downloading model weights, the time you spend rebuilding is its own cost, and it's easy to spend more of it than you saved on the hourly rate. That's the tradeoff a pure price table can't show you — whether you can actually act on a cheaper rate depends on whether stopping and moving is fast or expensive.

What "cheapest" doesn't capture

A lower hourly rate on an unfamiliar or thinly-reviewed provider isn't automatically a bargain. Reliability, network bandwidth, boot time, and how quickly a host resolves an outage all vary between providers and between individual machines on marketplace platforms like Vast.ai, where the same GPU model can sell for wildly different prices depending on the specific host. None of that shows up in a price column, and all of it affects whether the cheap rate actually finishes your job on time.

There's also no single provider that's cheapest across every GPU generation and every workload — the table above makes that plain. RunPod undercuts most competitors on-demand across the board in this snapshot, but a different snapshot next quarter, taken after a pricing change or a capacity shift, could look different. Compare on the GPU index and marketplace at the time you're actually renting, not against a table from a blog post.

Frequently asked questions

What is the cheapest cloud GPU provider in 2026?

There's no single answer across every GPU — it depends on the model and tier. As of August 2026, RunPod's Community tier had the lowest verified on-demand rates for H100 ($1.99/hr PCIe), A100 ($1.39/hr), L40S ($0.79/hr), and RTX 4090 ($0.34/hr) among fixed-price providers. Vast.ai's marketplace can go lower on individual listings but isn't a fixed rate.

What's the cheapest way to rent a GPU?

Combine a low on-demand rate with minimizing idle time — idle GPU-hours are usually the largest line item on a real bill, often larger than the gap between providers' hourly rates. Marketplace platforms like Vast.ai can undercut fixed-price providers on a given listing, at the cost of predictability and, on some hosts, reliability.

Is the advertised GPU price always per-GPU?

No. Some providers, including CoreWeave and AWS's multi-GPU instances, price by the node rather than the individual GPU. A CoreWeave H100 HGX node is $49.24/hr for 8 GPUs — $6.16/GPU-hr — not $49.24/GPU-hr. Always check whether a quoted rate is per-GPU or per-node before comparing it to another provider's number.

Why is my GPU cloud bill higher than the advertised hourly rate?

Usually one of: idle time (paying the hourly rate while the GPU isn't doing useful work), egress fees for moving data out, storage billed separately from compute even after you stop the GPU, or minimum billing increments that round short sessions up. All four are typically absent from the headline price on a provider's homepage.

Are commitment or reserved-price discounts worth it?

Only if you can keep the GPU utilized for the term. One-year H100 contract pricing ran $2.10–$2.70/GPU-hr as of April 2026, below on-demand rates in the same window — but that discount assumes near-continuous use. If your workload is bursty, a lower on-demand rate you can stop and start freely often beats a cheaper committed rate you're paying for during idle stretches.

Compare before you commit

Prices in the table above are a snapshot — check the GPU index and marketplace for current availability, and use /compare to line providers up side by side for your specific GPU and region. If idle time is the bigger cost in your case, which it usually is, start with the idle cost calculator instead of the price table — it'll tell you more about your real bill than another provider comparison will.

Sources

• RunPod pricing — H100, A100, L40S, RTX 4090 on-demand rates, observed August 23, 2026.
• Lambda pricing — H100, A100 on-demand and cluster rates, observed August 23, 2026.
• CoreWeave pricing — H100, A100, L40S, B200 node rates, observed August 23, 2026.
• AWS EC2 p5.4xlarge via Vantage, p4de.24xlarge — on-demand instance pricing, observed August 23, 2026.
• DataCrunch (Verda) pricing, Hyperstack pricing, Massed Compute pricing, Crusoe Cloud pricing, Nebius pricing, Paperspace/DigitalOcean pricing — observed August 23, 2026.
• Vast.ai pricing and live bundle API — marketplace rates observed August 23, 2026.
• SemiAnalysis GPU rental price index — one-year contract pricing range, April 2026.