If Lambda shows "out of capacity" for the instance you want, the table below is the quickest way out. Whether you know it as Lambda Labs or Lambda Cloud, the alternatives are the same. The table gives the cheapest current price for the A100, H100, H200 and B200 at Lambda and a set of comparable providers, and it only shows offers that are in stock right now. Choose by the shape of the job, not by brand. The rule at the end of this page sorts that into five cases.
| Provider | A100 $/GPU-hr | H100 $/GPU-hr | H200 $/GPU-hr | B200 $/GPU-hr |
|---|---|---|---|---|
| Lambda Labs | $1.99 | none in stock | none in stock | none in stock |
| Hyperstack | none in stock | $2.50 | none in stock | none in stock |
| VERDA | $1.27 | $3.37 | $4.41 | none in stock |
| Massed Compute | $1.35 | $2.73 | $3.62 | none in stock |
| Ori | none in stock | $2.90 | $3.50 | none in stock |
| Vultr | $2.80 | none in stock | none in stock | none in stock |
| Crusoe | none in stock | none in stock | none in stock | none in stock |
| CoreWeave | none in stock | none in stock | none in stock | none in stock |
| QuantaCloud | $1.48 | $2.59 | $3.43 | none in stock |
| Scaleway | none in stock | $3.29 | none in stock | none in stock |
How to read the table
We recheck stock about every minute. A price in a cell means that provider had that GPU in stock, on demand, within the last 15 minutes. An empty cell means it had none, or does not sell that GPU at all. That is the point for a reader who arrives because Lambda is sold out: a price list tells you what a GPU would cost, and this table tells you what you can start today.
Lambda has its own row, so you can see whether the capacity you wanted has come back. If it has, and Lambda already fits your work, you may not need this page.
The table leaves out peer-to-peer marketplaces. Vast.ai's documentation describes hosts who "list their machines, configure prices", and Runpod's describes a Community Cloud that connects "individual compute providers to users" (both read on 21 September 2026). Those can be good value, and our Vast.ai alternatives article explains how marketplaces work. A Lambda user is used to hardware the provider runs itself, so this page compares like with like.
Being in the table is not an endorsement. We rank nobody first and let the prices speak. One disclosure: QuantaCloud is run by the owner of this site. It gets the same treatment as every other row.
What Lambda does well
A price list you can read in a minute. Lambda's pricing page sells three lines: Instances, 1-Click Clusters and Superclusters. Instances are priced per GPU per hour in shapes of one, two, four or eight GPUs. On Lambda's price list of 21 September 2026, a single H100 SXM cost about 8% more per GPU-hour than the same GPU in an 8-GPU instance, and a single H100 PCIe cost about a quarter less than a single H100 SXM. Current figures are in the table above. There are no tiers to decode and no bidding.
Simple billing mechanics. Lambda's billing documentation says instances are "billed in one-minute increments" and that "you are not charged for ingress or egress". Many clouds charge to move data out. Our data egress reference lists who does.
Real multi-GPU nodes. The same list has 8-GPU H100 SXM, B200 and A100 SXM 80 GB instances. The 8-GPU shape carries the lowest per-GPU rate on the list.
A long history. Lambda says it was founded in 2012.
Live prices and stock for Lambda alone are on our Lambda page.
Where it falls short
Availability, above all. Lambda describes its instances as "self-serve, first-come access". It sells no reservation below cluster scale, so when an instance type is taken, it is gone until someone terminates. "Out of capacity" is the most common complaint in community reviews. We have to be plain about the evidence: no Lambda document quantifies it, and Lambda's status page reports incidents, not sold-out events. This is also not only a Lambda problem. SemiAnalysis wrote on 2 April 2026 that "On-Demand GPU rental capacity is sold out across all GPU types".
You cannot stop an instance. Lambda's documentation says: "At the moment, Lambda Cloud instances can only be launched, restarted, or terminated." Shutting down from inside the machine puts it into Alert status "and billing will continue". Instances are billed "regardless if they're actively being used". To stop paying you terminate, and then you are back in the queue for capacity.
Persistence costs extra and is tied to a region. Data survives termination only on a Lambda filesystem, priced per GiB per month and billed for as long as the filesystem exists, mounted or not. It "must also reside in the same region" as the instance. If the only H100 in stock is in another region, your data is not there.
Gaps in the range. On 21 September 2026 the instance list had no H200 and no spot or interruptible tier. The H200 carries 141 GB of memory against 80 GB on the H100 SXM, so anyone sizing a model to the H200 has to look elsewhere.
Clusters cost more per GPU, not less. 1-Click Clusters run from 16 to more than 2,000 H100 or B200 GPUs on Quantum-2 InfiniBand. The smallest commitment is 16 GPUs for two weeks, billed weekly. Lambda's own price list on 21 September 2026 showed its 16-GPU cluster rate at about one and a half times its 8-GPU on-demand rate per GPU, for both the H100 and the B200. Terms of a year or more are by quote, and reserved capacity at Lambda's "lowest prices" means contacting sales.
New accounts start with a quota. Lambda's documentation says new accounts have a limit on instances that "increases automatically as you pay your invoices".
What the same GPUs cost across every provider
The table at the top is limited to the providers named in it. This one takes the cheapest in-stock offer across every provider we track and shows how many providers have each GPU available now. Peer-to-peer offers are excluded here too, but Vast.ai's datacenter-hosted listings are included.
| GPU | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|
| H100 | $2.50 | Hyperstack | 11 |
| H200 | $3.43 | QuantaCloud | 8 |
| B200 | $3.75 | Packet.ai | 3 |
If the cheapest row is far below the comparable providers above, check who the provider is before you switch. A marketplace listing, even a datacenter one, is a different product from an instance the provider runs itself.
Alternatives by need
A single H100 now
This is the easiest case. Take the cheapest H100 cell in the top table that is not empty, then check which H100 it is. Providers sell SXM, PCIe and NVL variants, and Lambda itself lists two at different prices. On one GPU there is no GPU-to-GPU traffic, so NVLink, the main thing the SXM version adds, does nothing for you. Memory size can differ between variants, so check it on the offer.
| Variant | VRAM | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|---|
| H100 PCIe | 80 GB | $2.50 | Hyperstack | 6 |
| H100 SXM5 | 80 GB | $2.79 | Lyceum | 7 |
| H100 NVL | 94 GB | $3.11 | Massed Compute | 3 |
Before you launch, check two things on the provider's own site: whether a stopped instance keeps billing, and what persistent storage costs. If stop and resume matters to you, this may be an upgrade on Lambda.
An 8-GPU node
The prices in the table are per GPU. They do not tell you whether a provider sells eight GPUs in one machine, so open the provider's page and confirm the shape. Ask whether it is an HGX board with NVLink between all eight GPUs or a server of PCIe cards. For tensor-parallel work across eight GPUs that difference matters more than a small gap in price.
One example we could verify: on 21 September 2026 Vultr's public plan API listed bare metal servers with eight H100, eight A100 SXM and eight B200 GPUs. We have not verified node shapes for the other providers ourselves, so check before you plan around one. Each provider's page has its current GPUs, prices and stock, in alphabetical order: CoreWeave, Crusoe, Hyperstack, Massed Compute, Ori, QuantaCloud, Scaleway, Verda and Vultr.
A cluster with InfiniBand
This is a different purchase. SemiAnalysis noted on 2 April 2026 that most of the GPU rental market is transacted on contracts of six months or longer. CoreWeave's annual report, filed 2 March 2026, describes its committed contracts as typically "take-or-pay" with an initial period that "generally ranges from one to six years". Expect a sales conversation, not a launch button. Lambda's two-week minimum is short by that standard.
SemiAnalysis rates GPU providers for exactly this kind of cluster work in its ClusterMAX system. The rating says little about renting a single GPU, and our neocloud explainer covers how to read it. Compare quotes on our GPU cluster pricing page.
The cheapest A100
Lambda's list has the A100 in 40 GB form at every instance size, with the 80 GB SXM version only as an 8-GPU instance. If you want one or two 80 GB A100s, Lambda does not sell that shape, and the A100 column in the top table is where to look. Check the memory size on the offer, because the two versions are priced differently everywhere.
An H100 is sometimes close enough in price to be the better buy. It adds FP8 support, which the A100 lacks. The A100 vs H100 comparison has the live gap.
| Spec | A100 | H100 | H200 | B200 |
|---|---|---|---|---|
| VRAM | 40 to 80 GB | 80 to 94 GB | 141 GB | 180 to 192 GB |
| Memory bandwidth | 2,039 GB/s | 3,350 GB/s | 4,800 GB/s | 8,000 GB/s |
| FP16 (dense) | 312 TFLOPS | 989 TFLOPS | 989 TFLOPS | 2,250 TFLOPS |
| FP16 (with sparsity) | 624 TFLOPS | 1,979 TFLOPS | 1,979 TFLOPS | 4,500 TFLOPS |
| FP8 (dense) | Not published | 1,979 TFLOPS | 1,979 TFLOPS | 4,500 TFLOPS |
| FP8 (with sparsity) | Not published | 3,958 TFLOPS | 3,958 TFLOPS | 9,000 TFLOPS |
| Interconnect | NVLink, PCIe 4.0, InfiniBand | NVLink, PCIe 5.0, InfiniBand | NVLink, PCIe 5.0, InfiniBand | NVLink, PCIe 6.0, InfiniBand |
Decision rule
- Lambda's row shows your GPU in stock, and Lambda fits your work. Stay. Nothing on this page is a reason to move a working setup.
- You need one to four GPUs today. Take the cheapest non-empty cell for your GPU in the top table. Confirm the variant and memory size, check how the provider bills a stopped instance, then launch. Do not wait for Lambda to free up. An idle team usually costs more than a small difference in hourly price.
- You need a full 8-GPU node. Do the same, but confirm the node shape and the NVLink topology on the provider's site before you move any data.
- You need 16 or more GPUs with InfiniBand. Get Lambda's 1-Click Cluster price and at least two other quotes. Weigh the minimum term as heavily as the rate.
- You need an H200, a spot tier, or the ability to pause. Lambda does not sell these as instances today. Use the table and move.
Whichever case is yours, keep your data somewhere that does not depend on one provider's region. Then the next sold-out notice costs you minutes, not a day.
Sources
Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.
All accessed 21 September 2026 unless a publication date is given.
- Lambda pricing: product lines, instance and 1-Click Cluster prices, "first-come access" wording
- Lambda 1-Click Clusters: cluster sizes and InfiniBand
- Lambda docs, creating and managing instances: no stop or pause, new account quota
- Lambda docs, billing: one-minute billing, filesystem billing, egress, weekly cluster billing
- Lambda docs, on-demand cloud: filesystem region rule
- Lambda status page: reports incidents, not capacity
- Lambda, agreement with Microsoft (3 November 2025): founding year
- SemiAnalysis, The Great GPU Shortage (2 April 2026)
- SemiAnalysis ClusterMAX overview
- CoreWeave annual report for 2025 (filed 2 March 2026)
- Vultr bare metal plan API
- Vast.ai FAQ and Runpod Pods overview: marketplace models