Renting GPUs in 2026: RunPod vs Lambda vs Vast.ai

The same H100 rents for somewhere between about $1.50 and $4.00 an hour depending on who you rent it from, and the difference is almost never the hardware. It is whether the machine is yours for the duration, who is on call when it stops, and what happens to your data when the job ends.

Prices in this market move weekly, so treat every figure below as a snapshot from autumn 2026 and check the current rate before you commit. The structural differences are what last.

The four shapes of GPU rental

• Marketplaces — Vast.ai is the clearest example. Independent operators list spare capacity and you bid for it. Cheapest by a distance; also the only tier where your machine can be taken back mid-run.
• Community tiers inside managed clouds — RunPod's community pool is the same idea with a single company's billing and support in front of it. Roughly half the price of its own secure tier.
• Managed GPU clouds — RunPod's secure pods, Lambda, and similar. You get a machine in a known data centre with an SLA behind it.
• Capacity providers — CoreWeave and the hyperscalers. Reserved contracts, interconnect that matters for multi-node training, and a sales conversation before you get a price.

Roughly what things cost

Autumn 2026 on-demand rates, which is to say the number you see before storage and egress:

The spread on an H100 is more than double from one end of that table to the other, for the same chip. Everything that follows is about what the extra buys.

What the cheap tiers actually cost you

Three things, and only one of them is obvious.

1. Interruption. Marketplace and community capacity can be reclaimed. If your training loop checkpoints every twenty minutes, that is an inconvenience. If it does not, a reclaim at hour nine of a ten-hour run has cost you the entire run, and the money you saved per hour is gone several times over. Checkpointing is the single thing that decides whether cheap capacity is cheap.
2. Variance in what you actually get. On a marketplace, "RTX 4090" describes the card and not the machine around it. A 4090 behind a slow CPU with four lanes of PCIe and a thin uplink will train noticeably slower than the same card in a properly specified host. Check the host's CPU, RAM and network before you judge the price.
3. Your time. Hunting for a good listing, re-provisioning after a reclaim, and debugging a host that turns out to be misconfigured is work. At one job a week it is free. At several a day it is a part-time job, and that is the point at which the managed tier is cheaper.

Storage is where the surprise bills come from

Every comparison, this one included, leads with the hourly GPU rate. Almost nobody's actual invoice is dominated by it.

Persistent volumes bill whether or not a GPU is attached. A 500 GB dataset volume you left behind after an experiment in March is still billing in September, and it does not appear on any dashboard you look at while deciding which provider is cheapest. On top of that, pulling a large container image at the start of every run is time you pay for at the GPU rate — a 20 GB image pulled before a 40-minute job is a meaningful tax on the job.

Two habits remove most of this: delete volumes when an experiment ends rather than when you remember, and keep the image slim enough that pulling it is not a measurable fraction of the run.

Serverless, and when it is the right shape

RunPod and others sell a serverless mode: your container wakes on a request, runs, and stops. The per-second rate is higher than a pod — RunPod's serverless H100 works out meaningfully above its on-demand equivalent — but you pay for seconds of inference rather than hours of idle.

The maths is straightforward. If your GPU would be busy less than roughly a third of the time, serverless is cheaper despite the higher rate. Above that, a pod you keep is cheaper. The thing to check before committing is cold start: a model that takes ninety seconds to load into VRAM makes a two-second inference request feel broken, and the usual fix — keeping a worker warm — reintroduces the idle cost you were avoiding.

When renting stops making sense

Buying an H100 outright is a five-figure purchase before the machine around it, the power, and somewhere to put something that sounds like a hairdryer. Against roughly $2–3.50 an hour rented, the break-even is well over a year of continuous use — and that is before you account for the card being a generation behind by the time you get there.

For almost everyone the answer is rent. The exception is sustained, predictable, around-the-clock load, which is also exactly the case where CoreWeave and the hyperscalers will sell you a reserved contract at well under the on-demand rate. If you are at that volume, the conversation to have is about commitment discounts, not about marketplaces.

A consumer card in a desktop under your own desk is the other honest answer for a lot of hobby work. A 4090 you already own costs nothing per hour, and for fine-tuning small models and running inference locally it is more than enough. Our comparison of laptop and desktop workstations goes into where that line sits.

Which to pick

• Experiments you can restart, and a tight budget. Vast.ai. Check the host specs, checkpoint often, and accept that some runs will die.
• Most people, most of the time. RunPod's community tier for exploratory work, secure pods for anything that matters. One account, one bill, a real support queue.
• A training run you cannot afford to lose. Lambda or RunPod secure. The premium over a marketplace is cheap insurance against repeating a long job.
• Multi-node training where interconnect matters. CoreWeave or a hyperscaler, and a reserved contract. This is the one case where the cheap options are not competing.
• Intermittent inference behind an API. Serverless, after you have measured your cold start honestly.

If the GPU is only part of the picture and you also need ordinary servers around it, our Kamatera vs DigitalOcean comparison covers the CPU side, and running a cluster on a budget host covers the architecture.