RunPod vs Vast.ai vs Lambda Labs: Three Products, Not Three Prices

The premise of most comparisons here is that one of these three is cheapest per GPU class, and your job is to find out which. That’s not quite the shape of it.

These are three different businesses selling GPU time in three different ways. Vast.ai is a marketplace where hosts compete and the price moves. Lambda sells whole eight-GPU nodes with enormous CPU, memory and storage attached. RunPod sells single GPUs across three reliability tiers, plus a serverless product neither of the others has.

Compare them on the hourly rate alone and you’ll pick wrong, because the hourly rate is buying different amounts of machine.

The H100, on like-for-like terms

Every rate below is per GPU per hour, for an 80GB H100:

Two things in that table do most of the work, and both are usually missed.

Vast.ai’s headline price is not the price

Vast.ai publishes live marketplace rates, and to its credit it shows both the floor and the median side by side:

The H100 SXM floor is $1.60/hr. The median is $2.72. That’s a 70% spread between the number that gets quoted in comparison articles and the number you’ll typically pay.

The pattern repeats across the board: B200 from $5.00 against a $6.34 median, H200 from $3.56 against $4.13. The floor is a real price paid by someone, on some host, at some moment — it is not a rate card, and any article that puts Vast’s “from” price in a comparison table next to RunPod’s and Lambda’s fixed rates is comparing a best case to a standing price.

Judged on the median, Vast.ai is still the cheapest of the three for an H100 SXM — 17% below RunPod’s Secure Cloud — which is a genuine advantage. It’s just a smaller one than the marketing implies, and it comes with the marketplace’s variance in host quality, uptime and location.

Vast’s structural advantages are real: per-second billing with “no minimum hours, no rounding up”, an interruptible tier it advertises as 50%+ cheaper for fault-tolerant batch work, reserved terms at up to 50% off, and 68+ GPU types across 40+ datacentres.

Lambda isn’t expensive, it’s selling something else

At $3.99 per GPU-hour, Lambda’s H100 looks 21% more expensive than RunPod’s Secure Cloud SXM. Then you look at what’s attached to it.

Those vCPU, RAM and storage figures are for the whole eight-GPU node. Divide them:

Per GPU, Lambda gives you 225 GiB of RAM against RunPod Secure’s 125 GB, and 2.75 TiB of SSD where RunPod bills storage separately. Add “no egress fees” — Lambda states this explicitly — and the 21% premium starts looking like the price of a different product rather than a worse deal.

That product is multi-GPU training. Lambda’s 1-Click Clusters run 16 to 2,000+ interconnected H100 and B200 GPUs, which is the thing you cannot assemble out of single pods on a marketplace. If your run needs eight GPUs talking to each other at speed, Lambda is not competing with the other two — it’s the only one of the three in that business.

Note the asterisk on its table: prices are “plus applicable sales tax/VAT/GST”, which the other two quote inclusive of nothing in particular. Budget for it.

Where the brief’s assumptions broke

Two predictions worth correcting, because they circulate widely.

“Lambda leads on A100.” It doesn’t. Lambda’s A100 SXM 80GB is $2.79; RunPod’s Secure Cloud A100 SXM 80GB is $1.59 — RunPod is 43% cheaper for the same card. Lambda’s A100 comes with far more CPU and RAM per GPU, so it isn’t a like-for-like loss, but on price alone it’s not close.

“RunPod leads on H100.” Only against Lambda. Vast.ai’s median undercuts RunPod’s Secure Cloud by 17%, and RunPod’s own Community tier undercuts its Secure tier by roughly the same again.

The real pattern isn’t a per-GPU league table. It’s that each vendor wins a different question.

Pick by question, not by card

RunPod’s serverless product deserves its own line, because it’s the one genuinely differentiated offering in this comparison. Neither Vast.ai nor Lambda bills only while a request is processing. For inference endpoints that idle — which is most of them before scale — that difference outweighs every hourly rate above it. We worked the crossover arithmetic in RunPod vs DigitalOcean GPU Droplets: serverless stays cheaper until you’re running a GPU roughly 60% of the month.

Check RunPod pricing → · Read our RunPod review

The multi-provider setup most serious teams end up with

Nobody running GPUs at scale uses one vendor, and the split usually falls the same way:

• Development and experimentation on the cheapest interruptible capacity — Vast.ai interruptible or RunPod Community. Work that can die and restart belongs on the cheapest hardware available.
• Training runs on Lambda, or on reserved capacity, because a multi-day run that gets preempted at hour 40 costs more than the discount saved.
• Production inference on RunPod Serverless or a warm pod, where the duty cycle decides which.

The cost of that arrangement is real and rarely counted: three accounts, three billing relationships, three sets of images and environment quirks, and no single dashboard. For a small team it is usually not worth it — pick one, and revisit when the GPU bill becomes a line item someone asks about. The orchestration overhead is the reason single-vendor setups persist long after the arithmetic says otherwise.

The verdict

RunPod is the sensible default for most people, not because it wins on price — it doesn’t, consistently — but because it covers the widest range of situations without changing platform. Three reliability tiers, a serverless product, per-second billing, thousands of GPUs across 30+ regions. If you want one account rather than three, this is the one to have.

Vast.ai is the cheapest, provided you judge it on the median rather than the floor and can absorb marketplace variance. For batch work, experimentation and anything fault-tolerant, the interruptible tier is the best price in this comparison by a wide margin.

Lambda is not competing on hourly rate and shouldn’t be judged on it. It sells eight-GPU nodes with 225 GiB of RAM and 2.75 TiB of SSD per GPU, no egress fees, and interconnected clusters up to 2,000+ GPUs. For serious training runs that’s the product; for single-GPU inference it’s the wrong shape and the wrong price.

How we picked

Rates are each vendor’s published pricing read in August 2026 and visible in the screenshots above: Vast.ai’s live marketplace page showing floor and median together, and Lambda’s 8×-node table. RunPod’s Secure Cloud figures — H100 SXM $3.29, H100 PCIe $2.89, A100 SXM $1.59, A100 PCIe $1.39 — come from its pricing page, with Community Cloud rates as published alongside.

Vast.ai prices move, and faster than you’d expect. That is the point of the platform, and it means this table has a shorter shelf life than the others. The figures here were read on 16 August 2026 and match the screenshot above; re-checking the same page later that day, the H100 SXM median had moved from $2.72 to $3.07 and the B200 median from $6.34 to $6.88. If you see different medians quoted in our RunPod vs Vast.ai comparison, that is why — both are accurate as captured, hours apart. The floor and median we quote are a snapshot; check the live page before acting on them. It’s also why we’ve used the median rather than the floor throughout — quoting a marketplace’s best-case rate against two vendors’ standing rates isn’t a comparison.

The Lambda per-GPU figures are our arithmetic: its table publishes node-level vCPU, RAM and storage against a per-GPU price, so we divided by eight to make the comparison like-for-like. Its prices also exclude sales tax/VAT/GST, which its own footnote states and which the quoted rates do not include.

What we have not done: we have not rented these GPUs, benchmarked throughput, or measured how often Vast.ai’s floor price is actually available — the last of which is the number that would most change the verdict, and it isn’t published by anyone. We also haven’t tested serverless cold-start times or cluster provisioning latency. Treat the tables as a cost model, not a performance review.

Only RunPod runs an affiliate programme we participate in; Vast.ai and Lambda do not, and we have no product pages for them to link to. That’s worth knowing given Vast.ai wins the price comparison above on our own numbers and Lambda wins the training one.

FAQ

Which is cheapest: RunPod, Vast.ai or Lambda?

Vast.ai, on median marketplace pricing — its H100 SXM median of $2.72/hr is about 17% below RunPod’s Secure Cloud. Its advertised $1.60 floor is not a rate card; the gap between floor and median is roughly 70%. Lambda is the most expensive per GPU-hour but includes substantially more CPU, RAM and storage.

Why is Vast.ai’s “from” price so much lower than what I get quoted?

Because it’s a marketplace. Hosts set their own prices, so the floor is whatever the cheapest available machine costs at that moment, on hardware and in a location you may not want. Vast publishes the median alongside it, which is the more useful planning number.

Is Lambda Labs worth the higher hourly rate?

For multi-GPU training, yes — it sells whole eight-GPU nodes with interconnect, 1-Click Clusters up to 2,000+ GPUs, 225 GiB of RAM and 2.75 TiB of SSD per GPU, and no egress fees. For single-GPU inference you’re paying for resources you won’t use.

Which one scales to zero?

Only RunPod, via Serverless. Neither Vast.ai nor Lambda bills purely per request, so an idle endpoint on either keeps costing money. For inference that isn’t continuously busy, that’s usually the deciding factor.

What’s the catch with interruptible and Community tiers?

They run on cheaper or third-party capacity and can be reclaimed or prove less reliable. Vast.ai advertises interruptible instances at 50%+ below on-demand; RunPod’s Community Cloud runs on third-party hosts. Both are excellent for work that can retry and unsuitable for anything with an uptime commitment or a compliance requirement.

Do any of them charge for egress?

Lambda explicitly states no egress fees. RunPod bills storage separately, including a volume disk rate that is higher when idle than running. Vast.ai’s charges vary by host, which is one more thing the marketplace model makes you check per machine.

Should I use more than one provider?

Most teams running GPUs seriously do — cheap interruptible capacity for experiments, reserved or Lambda for training, serverless for production inference. For a small team the orchestration overhead of three accounts usually outweighs the saving. Pick one until the GPU bill is large enough that someone asks about it.