RunPod vs Vast.ai: The Discount Is Real on a 4090 and Gone on an H100

The standard version of this comparison says Vast.ai is cheaper and less reliable, RunPod is dearer and safer, and you pick according to how much you like risk.

That’s true enough to be useless, because the size of the discount varies enormously by card — and once you price it properly, the decision makes itself on most of the range without any judgement about risk at all.

On an RTX 4090, Vast.ai’s median is roughly half RunPod’s Secure Cloud rate. On an H100 SXM, it’s 7% below. Those two facts point in completely different directions.

Two different businesses

RunPod curates. It runs Secure Cloud on vetted datacentre hardware and Community Cloud on third-party capacity at lower rates, plus a serverless product that bills only while a request is processing. Prices are fixed rate cards you can plan against.

Vast.ai is a marketplace. Hosts set their own prices and compete, across 40+ datacentres and 68+ GPU types. As Vast puts it: “prices are set by the market, not by Vast. More supply means lower prices — and you always see the real rate.”

That means Vast has no rate card, which has a consequence most comparisons ignore: the number you quote depends on when you looked.

Judge Vast on the median, not the “from”

Vast publishes a floor price and a median side by side, which is more honest than most marketplaces and creates a trap for anyone skim-reading:

The RTX 4090 shows “from $0.13/hr” — and a median of $0.36. Nearly three times the headline. The gap is the marketplace working as designed: somebody somewhere has a cheap 4090 free right now, and it may not be in a country, on a host, or with an uptime record you want.

We also watched this move in real time. Checking the H100 SXM twice on the same day, the median went from $2.72 to $3.07 — a 13% swing in hours. Any price table for Vast.ai, including the one below, is a snapshot.

The comparison, by card class

Vast.ai medians against RunPod’s published rates:

Notice what happens as the cards get more serious. On consumer hardware Vast is dramatically cheaper. On an H100 SXM it’s $3.07 against RunPod Secure’s $3.29 — a 7% difference — and RunPod’s own Community tier at $2.69 is cheaper than Vast’s median.

The marketplace advantage is a consumer-GPU phenomenon. That makes sense: the supply of idle 4090s in bedrooms and small datacentres is enormous, and the supply of idle H100s is not. Where the hardware is scarce, the market clears near the professional rate.

How much slack does the discount actually buy?

Here’s the framing that decides it. If a job runs on cheaper hardware but gets interrupted, restarted, or runs on a slower host, the discount is spent on wasted time. So: how much longer can the same job take on Vast before you’ve given the saving back?

On a 4090, a job could take more than twice as long — restarts, re-uploads, lost checkpoints and all — and you’d still be ahead. That is an enormous amount of tolerance, and it’s why Vast is genuinely the right answer for experimentation and fault-tolerant batch work on consumer cards.

On an H100 SXM, you get 7%. A single interruption on a multi-hour fine-tune, one re-upload of a dataset, one restart from a checkpoint, and the discount is gone — you’ve paid more and finished later. At that point you are taking marketplace variance for nothing.

This is why “total cost of a completed job” is the only sane way to compare these two, and why the answer flips between card classes rather than between vendors.

What “unreliable” actually means here

The lazy version of this comparison says Vast has no SLA because your host is somebody’s gaming rig. That’s too crude, and it’s worth being precise because Vast sells three distinct things:

• On-demand — “Guaranteed uptime for production workloads with per-second billing and no interruptions.” This is not preemptible.
• Interruptible — advertised at 50%+ cheaper, preemptible, intended for fault-tolerant batch work. This is the tier that gets outbid and killed.
• Reserved — 1, 3 or 6 month commitments at up to 50% off.

So the interruption risk is a property of the tier you choose, not of the platform. The variance that does apply across all Vast tiers is host quality: hardware, network, disk speed, geography and operator competence differ machine to machine in a way they don’t inside a curated datacentre. Two 4090s at the same price are not necessarily the same product.

RunPod’s Community Cloud occupies similar ground — third-party hosts at lower rates — and its Secure Cloud is the vetted-datacentre tier. The structures are more alike than the marketing suggests.

The thing RunPod has that Vast doesn’t

Serverless. RunPod bills only while a request is being processed; Vast.ai has no equivalent, so an idle instance keeps costing money on either of its on-demand or reserved tiers.

For inference endpoints that aren’t continuously busy — which is most of them before real scale — that difference outweighs every hourly rate in this article. We worked the arithmetic in RunPod vs DigitalOcean GPU Droplets: serverless stays cheaper until a GPU is running roughly 60% of the month. Nothing in Vast’s catalogue competes with that shape of workload.

RunPod also bills per second, matching Vast’s granularity, so there’s no advantage either way on rounding.

The verdict

Use Vast.ai for consumer cards and fault-tolerant work. On a 4090, 3090, 5090 or L40S the discount is large enough to absorb a great deal of wasted time, and the interruptible tier at 50%+ off is the cheapest compute in this comparison by a wide margin. Experimentation, batch inference, hyperparameter sweeps, anything that checkpoints and can restart — this is the right home for it, and the saving is real rather than theoretical.

Use RunPod for H100-class work and anything with a deadline. At 7% the discount no longer pays for the variance, and RunPod’s Community Cloud undercuts Vast’s median on the H100 SXM anyway. Add Secure Cloud when there’s a customer or a compliance story attached, and Serverless when the GPU will spend most of its time idle.

Check RunPod pricing → · Read our RunPod review

For the three-way view including Lambda Labs — which sells whole eight-GPU nodes and interconnected clusters rather than single cards — see RunPod vs Vast.ai vs Lambda Labs.

How we picked

Vast.ai figures are medians from its live pricing page, read in August 2026 and visible in the screenshot above. We have used the median throughout rather than the advertised “from” price, because the two differ by up to threefold — the RTX 4090 shows a $0.13 floor against a $0.36 median — and comparing a marketplace’s best case to a competitor’s standing rate is not a comparison.

Those medians move. We observed the H100 SXM median shift from $2.72 to $3.07 within a single day while writing this, so treat every Vast figure here as a snapshot and check the live page before acting on it. RunPod’s rates are its published Secure and Community Cloud pricing.

The break-even table is our arithmetic: the ratio of RunPod’s Secure rate to Vast’s median, expressed as how much longer a job can run before the saving is consumed. It deliberately measures only time, because that’s what interruptions and slower hosts cost. It does not price re-uploading a dataset over a slow host link, engineering time spent babysitting restarts, or the risk of a run failing near completion — all of which push the real break-even lower than the table shows, and all of which hit hardest on exactly the long jobs where the discount is thinnest.

What we have not done: we have not run the same fine-tune on both platforms and counted restarts, which is what the brief for this piece called for. Neither vendor publishes interruption rates, and inventing one would decide the article’s conclusion by assumption. The break-even framing above is the honest substitute — it tells you how much disruption the discount can absorb, and you can compare that against what you observe on your own workload.

Only RunPod runs an affiliate programme we participate in, and Vast.ai has no product page on Findstack to link to. Worth knowing given Vast wins most of the price comparisons above on our own numbers.

FAQ

Is Vast.ai cheaper than RunPod?

On consumer GPUs, substantially — an RTX 4090 at a $0.36 median against RunPod Secure’s $0.74. On an H100 SXM the gap collapses to 7% ($3.07 vs $3.29), and RunPod’s Community Cloud at $2.69 is cheaper than Vast’s median outright.

Why is Vast.ai’s advertised price so much lower than what I pay?

It’s a marketplace floor, not a rate card. The RTX 4090 advertises “from $0.13/hr” against a $0.36 median — the floor is one host’s price at one moment, on hardware and in a location you may not want.

Does Vast.ai interrupt your instances?

Only on the interruptible tier, which it advertises at 50%+ below on-demand precisely because it’s preemptible. Vast’s on-demand tier is sold as “guaranteed uptime for production workloads… no interruptions.” The variance that applies across all tiers is host quality rather than preemption.

Which is better for fine-tuning?

Depends on the card. On consumer hardware Vast’s discount absorbs more than double the job duration before breaking even, so it wins comfortably. On H100-class hardware it absorbs 7%, meaning one restart erases the saving — use RunPod, and Secure Cloud if the run matters.

Which one scales to zero?

RunPod, via Serverless. Vast.ai has no equivalent, so an idle instance keeps billing. For inference that isn’t continuously busy, that’s usually the deciding factor rather than the hourly rate.

Is RunPod Community Cloud the same thing as Vast.ai?

Structurally similar — both are third-party capacity at lower rates than vetted datacentre hardware. The difference is that RunPod sets the price and vets the hosts, while Vast lets hosts compete freely, which produces both the lower floor and the wider spread.

How often should I re-check these prices?

Before every significant purchase on Vast, because the medians move daily. RunPod’s rate card is stable enough to plan against for months.