Runpod vs Vast.ai vs Lambda: which GPU cloud is cheapest for AI?

Which GPU cloud is cheapest for AI? Vast.ai, on almost every card. On November 3, 2026 the cheapest H100 on Vast.ai was $1.79 an hour, against $2.89 for an H100 PCIe on Runpod Secure Cloud and $3.29 for one H100 PCIe on Lambda. For an RTX 4090 it was $0.37 on Vast.ai and $0.89 on Runpod; Lambda doesn't rent 4090s at all. The lower price comes with a catch: Vast.ai is a marketplace of independent hosts, so reliability and extra fees vary by offer. Runpod costs more for a managed data center and an easier start. Lambda is built for multi-GPU training; it beats Runpod Secure Cloud on the B200 ($6.99 for one GPU, $6.69 per GPU on an 8x instance, against $7.99) but costs more for an H100.

How much does each GPU cost on Runpod, Vast.ai and Lambda?

Here is every common AI card side by side, all read on November 3, 2026. The Vast.ai column comes from our own read of 512 rentable offers through Vast's API at 13:06 UTC: the cheapest on-demand offer per GPU, with the median offer in brackets. The Runpod column is the Secure Cloud price on Runpod's pricing page, which is what a new account rents. The Lambda column is from Lambda's instance price list, single-GPU (1x) unless marked.

The pattern holds across the table. Vast.ai is cheapest on every card it had offers for, often by half or more. Runpod sits in the middle at a fixed list price. Lambda is the most expensive for a single H100 or A100 (its B200 is the exception) and doesn't rent consumer cards like the RTX 4090 or 5090; its cheapest single GPUs that day were a Quadro RTX 6000 at $0.69 and an A6000 at $1.09. Use our VRAM calculator first to see which card your model actually needs, because the cheapest card that fits beats the cheapest H100.

Why is Vast.ai so much cheaper?

Because Vast.ai doesn't set the price. Its pricing page says "Prices are set by the market, not by Vast. More supply means lower prices". Independent hosts, from data centers to people with spare cards, list GPUs and compete for renters. In our November 3 read only 11.7% of the 512 offers were in data centers, 244 were marked verified, and 59% had a reliability score of 0.99 or higher. That spread is where both the low prices and the risk come from.

Vast.ai also sells an interruptible tier, which its pricing page calls "50%+ cheaper" and "Preemptible — may be reclaimed". In our read the lowest interruptible bid for an H100 SXM was $0.89 an hour. That's fine for work you can checkpoint and resume, and a bad idea for a long run you can't restart.

Is Runpod's Community Cloud cheaper than Vast.ai?

Sometimes, but it is a shrinking option. Runpod's pricing page still lists Community Cloud prices next to Secure Cloud, and on November 3, 2026 they were close to Vast.ai's: an RTX 4090 at $0.34 (Vast.ai's cheapest was $0.37), an RTX 5090 at $0.69, an H100 PCIe at $1.99, an H100 SXM at $2.69, an H200 at $3.59 and a B200 at $5.98. The catch is in Runpod's own docs: "Runpod is no longer accepting new hosts for Community Cloud. Existing Community Cloud resources remain available." So you can use those cards while they last, but don't plan around them.

What else is on the bill besides the GPU?

The hourly GPU price isn't the whole bill. Each provider adds different extras, and on Vast.ai they can quietly eat the savings:

Vast.ai's docs are blunt about the storage trap: "Storage charges continue even when instances are stopped. Delete instances completely to cease storage billing." Data transfer costs "vary by host" and are charged "per byte transferred", so a cheap GPU on a host with expensive bandwidth can cost more than it looks if you download big models. Lambda's page says "Transparent pricing with no egress fees" and "Pay by the minute". Runpod's docs say Pods "are billed by the second for compute and storage, with no fees for data ingress or egress." Runpod charges $0.10 per GB a month for a volume disk while the Pod runs and $0.20 while it's idle, so a stopped Pod with a big disk also keeps costing money. A dash means the page we read didn't state it.

What does a 10-hour H100 job cost on each?

At the list prices above, 10 hours on one H100 cost $17.88 on the cheapest Vast.ai offer ($23.35 at the median offer), $28.90 on a Runpod Secure Cloud H100 PCIe, and $32.90 on a Lambda 1x H100 PCIe (plus tax). That's our arithmetic from the November 3 prices, GPU time only. Add storage on Vast.ai and Runpod, and data transfer on Vast.ai. The Vast.ai offer is also an H100 SXM, the faster version of the card; on SXM against SXM the gap is wider still, $1.79 against $3.99 on Runpod and $4.29 on a Lambda 1x instance.

Runpod vs Vast.ai (or Lambda): which should you pick?

Pick by what you can't afford to lose: money, time or a long training run.

• Pick Vast.ai if the lowest hourly rate matters most. Filter for verified hosts with high reliability, read the bandwidth price on the offer, delete instances you're done with, and use interruptible only for work you can checkpoint. Our Vast.ai review covers the setup.
• Pick Runpod if you want it to just work at a fair, fixed price, or you'll serve a model through an API. Its serverless option bills per use; see Runpod serverless pricing explained and our Runpod review.
• Pick Lambda if you need a B200 or you're a team training on 8 GPUs. Its B200 at $6.99 for one GPU ($6.69 per GPU on 8x) and 8x H100 at $3.99 per GPU are its strongest prices, with no egress fees. For a single H100 or A100, it's the most expensive of the three. We earn nothing from Lambda, so there's no button for it here; its prices are on lambda.ai.

For one card in depth, see the cheapest H100 rental, the cheapest RTX 4090 rental, the cheapest A100 rental and what an H200 costs to rent. If you're making AI video rather than running LLMs, see what a clip costs on a rented GPU versus an app or API. And if you'll rent for months, check rent vs buy first.

The short version: Vast.ai is the cheapest GPU cloud of the three, Runpod is the easiest at a fair fixed price, and Lambda wins on the B200 and some 8-GPU training rates. Compare today's prices on our cloud-GPU compare page, and read the cloud-GPU pricing pillar for CoreWeave, Nebius and the rest of the market.