Lambda GPU pricing: H100, B200 and A100 per hour

Lambda charges $4.29 an hour for a single NVIDIA H100 SXM, $3.99 per GPU when you rent all eight in one machine, and $6.69 to $6.99 per GPU for a B200, as of October 2026. Prices are per GPU per hour, billed by the minute, with sales tax added and no charge for data in or out. Run one H100 SXM around the clock for 30 days and the bill is about $3,089 before tax.

We haven’t rented Lambda’s machines ourselves. The figures come from Lambda’s pricing page and its billing and instance docs, read on October 31, 2026, and from the site’s own price data for Lambda (checked October 25, 2026), which matched what the pricing page showed. Prices for other providers are from our data files, with the dates given below.

How much does a Lambda GPU cost per hour?

Lambda sells self-serve “Instances” in sizes of 1, 2, 4 or 8 GPUs. The price is per GPU, and it drops a little as the instance gets bigger. Here is the full on-demand rate card from Lambda’s pricing page.

GPU (memory)	1 GPU	2 GPUs	4 GPUs	8 GPUs
B200 SXM6 (180 GB)	$6.99	$6.89	$6.79	$6.69
H100 SXM (80 GB)	$4.29	$4.19	$4.09	$3.99
H100 PCIe (80 GB)	$3.29	–	–	–
GH200 (96 GB)	$2.29	–	–	–
A100 SXM (80 GB)	–	–	–	$2.79
A100 SXM (40 GB)	$1.99	–	–	$1.99
A100 PCIe (40 GB)	$1.99	$1.99	$1.99	–
A10 (24 GB)	$1.29	–	–	–
A6000 (48 GB)	$1.09	$1.09	$1.09	–
Tesla V100 (16 GB)	–	–	–	$0.79
Quadro RTX 6000 (24 GB)	$0.69	–	–	–

All prices are per GPU per hour, in US dollars, “plus applicable sales tax/VAT/GST”, as of October 31, 2026. A dash means Lambda doesn’t list that size.

A few things stand out:

• The 80 GB A100 only comes in the 8-GPU size. If you want a single A100, Lambda’s options are the 40 GB SXM or PCIe cards at $1.99.
• The H100 PCIe is $1.00 an hour cheaper than the single H100 SXM ($3.29 against $4.29), with the same 80 GB of memory. The SXM version has faster links between GPUs in the same server, which matters most when you use several GPUs together; Lambda’s docs say SXM “offers improved bandwidth between the NVIDIA GPUs in a single physical server.”
• The GH200 at $2.29 pairs 96 GB of GPU memory with 64 vCPUs and 432 GiB of RAM, according to Lambda’s instance table. It’s the cheapest way on Lambda to get more than 80 GB on one GPU.
• “Starting at $0.79/hr” in Lambda’s marketing refers to the V100, which you can only rent as an 8-GPU machine, so the smallest bill for that card is $6.32 an hour (8 × $0.79). The cheapest single GPU is the Quadro RTX 6000 at $0.69.

For the current figures next to other providers, see our Lambda page and the GPU price index.

What comes with each instance?

The hourly price includes the CPU, RAM and local SSD that come with each instance size. Lambda doesn’t let you pick them separately. From the pricing page:

Instance	vCPUs	RAM	Local SSD	Price per hour (whole instance)
1x H100 SXM	26	225 GiB	2.75 TiB	$4.29
1x H100 PCIe	26	225 GiB	1 TiB	$3.29
8x H100 SXM	208	1,800 GiB	22 TiB	$31.92
1x B200	26	360 GiB	2.75 TiB	$6.99
8x B200	208	2,900 GiB	22 TiB	$53.52
1x GH200	64	432 GiB	4 TiB	$2.29
1x A10	30	226 GiB	1.3 TiB	$1.29

The whole-instance price is the per-GPU price times the number of GPUs: 8 × $3.99 = $31.92 an hour for eight H100s, and 8 × $6.69 = $53.52 for eight B200s.

Each instance runs Ubuntu 22.04 with Lambda Stack by default, which Lambda describes as a set of preinstalled drivers and frameworks including CUDA, cuDNN, PyTorch, TensorFlow and JAX, plus JupyterLab. Lambda lists instances in regions in the US, Europe, the Middle East and Asia, including us-east-1, us-west-1, us-south-1, europe-central-1 and asia-northeast-1. Which GPUs are free in which region at any moment isn’t on the pricing page; Lambda describes instance access as “self-serve, first-come.”

How does Lambda bill you?

Lambda’s billing docs set out the rules:

• Per minute. On-demand instances are “priced by hourly usage” and billed “in one-minute increments.”
• From launch to termination. Billing starts when the instance launches and passes health checks, and stops when you terminate it. Lambda says instances are billed “for as long as they’re running, regardless if they’re actively being used.”
• No stop button. Lambda’s docs say instances “can only be launched, restarted, or terminated.” You can’t pause one to save money and pick up later. To keep your work you need a filesystem (below) or your own backups.
• Don’t shut down from inside. Lambda warns that running sudo shutdown -h now or a similar command inside the instance puts it into “Alert” status, “and billing will continue.” Terminate from the console or the API.
• Weekly invoices. You get an invoice each week for the previous week’s usage.
• Sales tax is added, based on the location in your billing details.
• New accounts have an instance quota, which Lambda says “increases automatically as you pay your invoices.”
• Refunds come as credits. Lambda issues refunds as service credits for future Lambda Cloud use. They can’t be turned into cash and expire 12 months after they’re issued.
• No free tier. Lambda doesn’t offer free GPU hours; our data lists no free plan.

What does storage cost, and are there egress fees?

Lambda calls its persistent storage “filesystems.” Its billing docs say they’re billed “per GiB used per month in one-hour increments,” that billing continues “as long as a filesystem exists, even if it’s not mounted,” and that there’s “no minimum storage period.”

The docs give $0.20 per GiB per month as a worked example, but add that the rate “might not reflect current pricing” and that “the actual price will be displayed when you create your filesystem.” Lambda’s public pricing page doesn’t list a storage rate, so check the price shown in the console before you rely on it.

Two rules matter more than the rate:

• You aren’t charged for data transfer. Lambda’s billing docs say: “you are not charged for ingress or egress.” That applies to filesystems, and our data for Lambda lists no egress fee either. Moving a large dataset or model in and out costs nothing extra.
• A filesystem must be attached when you create the instance. Lambda’s filesystem docs say you “can’t attach filesystems to an instance or 1CC after the instance or 1CC has been created,” and it has to be in the same region. Plan this before you launch, or you’ll be copying data by hand.

Lambda also says you can’t set a usage limit on a filesystem, so a large dataset left behind keeps billing until you delete it.

What does a real month on Lambda cost?

Here are four common uses, priced at Lambda’s on-demand rates (October 2026). Sales tax is extra in each case.

A weekend fine-tuning run on one H100. 48 hours on a single H100 SXM: 48 × $4.29 = $205.92. On the H100 PCIe instead: 48 × $3.29 = $157.92, a saving of $48. If you keep a 500 GiB filesystem for the month and the console shows the docs’ example rate of $0.20, that adds 500 × $0.20 = $100.

A short 8-GPU training job. Ten hours on an 8x H100 SXM instance: 10 × $31.92 = $319.20. The same ten hours on eight separate single-GPU H100s would cost 10 × 8 × $4.29 = $343.20, so the 8x instance saves $24 and gives you the fast GPU-to-GPU links.

An always-on inference server. One H100 SXM for 30 days (720 hours): 720 × $4.29 = $3,088.80 a month. A GH200 for the same month: 720 × $2.29 = $1,648.80. An A10 for a smaller model: 720 × $1.29 = $928.80.

A cheap learning box. A single Quadro RTX 6000 (24 GB) for 40 hours of experiments in a month: 40 × $0.69 = $27.60.

Remember that a running instance bills whether you use it or not. A single H100 left on over a 60-hour weekend by mistake costs 60 × $4.29 = $257.40.

What are 1-Click Clusters, and what do they cost?

1-Click Clusters are Lambda’s reserved multi-node clusters, “from 16 to 2,000+ NVIDIA B200 or H100 GPUs,” for terms from two weeks to a year. You book them through Lambda’s team rather than launching them yourself. Prices from the pricing page, per GPU per hour, plus tax:

Cluster	Term	GPUs	Price per GPU-hour
H100	2 weeks – 1 year	16	$6.16
H100	2 weeks – 1 year	64	$5.85
H100	2 weeks – 1 year	256	$5.54
B200	2 weeks – 1 year	16	$9.86
B200	2 weeks – 1 year	64	$9.36
B200	2 weeks – 1 year	256+	$8.87

Lambda doesn’t list a price for terms over a year; it asks you to talk to its team.

Lambda’s docs say clusters are billed “in weekly increments according to the terms of your reservation,” and that you have ten days to pay the invoice once the reservation is approved.

Worked example: the smallest H100 cluster for two weeks. 16 GPUs × $6.16 = $98.56 an hour. Two weeks is 336 hours, so 336 × $98.56 = $33,116.16, billed as two weekly amounts of $16,558.08. The same 16 GPUs as two on-demand 8x H100 instances would list at 16 × $3.99 = $63.84 an hour, or $21,450.24 for 336 hours. The cluster costs more per GPU-hour. What you pay for is a reserved term on what Lambda calls a “production-ready” cluster, where on-demand instances are first-come. Lambda’s pricing page doesn’t break down what the cluster price includes beyond that.

How does Lambda compare with RunPod, Vast.ai and others?

Lambda is a data-center provider with fixed prices. It isn’t the cheapest per hour, but it has no egress fees and its larger instances come with a lot of CPU, RAM and local disk. Here are the same GPUs elsewhere, per GPU per hour, on demand.

GPU	Lambda	RunPod Secure	RunPod Community	Vast.ai median	Hyperstack	DigitalOcean
H100 80GB SXM	$4.29 (1x), $3.99 (8x)	$3.99	$2.69	$2.40	$3.20 (8-GPU VM only)	$4.41 (HGX H100)
H100 80GB PCIe	$3.29	$2.89	$1.99	$2.47	$2.50	–
A100 80GB	$2.79 (SXM, 8x only)	$1.79	$1.39 (SXM)	$0.84 (SXM4)	$1.35 (PCIe)	–
B200	$6.99 (1x), $6.69 (8x)	$7.99	$5.98	$9.38 (192 GB)	$6.00 (8-GPU VM only)	–

Sources: Lambda’s pricing page (October 31, 2026); our data for RunPod (checked October 18, 2026), Vast.ai (median offers, October 25, 2026), Hyperstack (October 18, 2026) and DigitalOcean GPU Droplets (October 25, 2026). TensorDock’s homepage advertises H100 SXM5 “from $2.25/hr” and A100 SXM4 from $1.80, but it has no public price list, so treat those as indicative.

Worked example: one H100 SXM for a 720-hour month.

• Lambda, 1x: 720 × $4.29 = $3,088.80
• RunPod Secure Cloud: 720 × $3.99 = $2,872.80
• DigitalOcean: 720 × $4.41 = $3,175.20
• Vast.ai median offer: 720 × $2.40 = $1,728.00

So for a single H100, Lambda costs $216 a month more than RunPod’s Secure Cloud and about $1,361 more than Vast.ai’s median host. In an 8-GPU instance its price falls to $3.99, level with RunPod Secure. For B200s, Lambda’s single-GPU $6.99 is a dollar under RunPod Secure’s $7.99, and Lambda rents one B200 at a time where Hyperstack sells them only as 8-GPU machines.

For the full cheapest-by-GPU picture, see the cheapest cloud GPU guide. If you’re weighing Lambda against RunPod specifically, our RunPod alternatives guide and the Lambda vs RunPod comparison put them side by side.

What to watch for

• No pause. Lambda instances can only be launched, restarted or terminated. An idle instance bills at full rate. Terminate it when you’re done.
• Shutting down from the terminal doesn’t stop billing. Lambda says the instance goes into “Alert” status and keeps billing. Use the console or API.
• Local SSD isn’t permanent. Keep anything you need on a filesystem or download it, because terminating an instance ends it.
• Attach storage at launch. A filesystem can’t be added to a running instance.
• Storage keeps billing while the filesystem exists, mounted or not, and Lambda doesn’t let you cap it.
• Tax is extra on every price on the page.
• Capacity isn’t guaranteed. On-demand access is first-come. For guaranteed capacity, Lambda points you to reserved clusters or its sales team.
• Refunds are credits, not cash, and they expire after 12 months.
• Some sizes aren’t sold. No single 80 GB A100, no single V100, and the H100 PCIe comes only as one GPU.
• No crypto mining. Lambda’s docs say it prohibits cryptocurrency mining on its on-demand instances.

Which Lambda option should you choose?

• You’re learning or testing code: a Quadro RTX 6000 ($0.69) or A10 ($1.29) is enough for small models. If you only need a cheap GPU and don’t mind a marketplace, Vast.ai or RunPod Community Cloud cost less for an RTX 4090; see our GPU cloud hub.
• You need 80 GB on one GPU for a few hours: the H100 PCIe at $3.29 is Lambda’s cheapest 80 GB single GPU. Pick the H100 SXM ($4.29) only if your tooling benefits from it.
• You need more than 80 GB on one GPU: the GH200 (96 GB, $2.29) or a single B200 (180 GB, $6.99).
• You’re training across 8 GPUs: the 8x H100 SXM ($31.92 an hour) or 8x B200 ($53.52) instances. Per GPU, these are Lambda’s best on-demand prices.
• You move a lot of data in and out: Lambda’s free ingress and egress can outweigh a slightly higher hourly price elsewhere.
• You need 16 or more GPUs for weeks: 1-Click Clusters, from $6.16 per H100-hour or $9.86 per B200-hour at 16 GPUs, with lower rates at 64 and 256.
• Price per hour is all that matters: Lambda isn’t the cheapest for single GPUs. Compare on our Lambda page and the GPU price index before you launch.

Sources

• Lambda: AI cloud pricing, read October 31, 2026
• Lambda Docs: Billing overview, read October 31, 2026
• Lambda Docs: On-Demand Cloud overview, read October 31, 2026
• Lambda Docs: Creating and managing instances, read October 31, 2026
• Lambda Docs: Filesystems, read October 31, 2026
• AskGPU price data for Lambda, RunPod, Vast.ai, Hyperstack, DigitalOcean GPU Droplets and TensorDock (src/data/products.json), checked October 17-25, 2026