A100 GPU price per hour by provider: September 2026

RunPod lists the lowest on-demand price for an Nvidia A100 80 GB of the five GPU clouds checked: $1.59 per GPU-hour on its pricing page, checked 28 September 2026.¹ Lambda sits at the other end at $2.79, which is 75% above RunPod’s rate.²

The spread is wide for one card, and the cheapest rate isn’t always the cheapest way in.

TL;DR

RunPod ($1.59) and Crusoe ($2.00) list the only on-demand A100 80 GB rates at $2 or less among the five clouds ranked.
CoreWeave and Lambda sell the 80 GB card on demand only as an eight-GPU unit, so their smallest on-demand order costs $21.60 and $22.32 an hour.
Dropping to the 40 GB card saves $0.80 per GPU-hour at Lambda and about $0.40 at Modal. The other three ranked clouds list no on-demand 40 GB A100.
The Vast.ai marketplace median for an A100 SXM4 was $0.99 an hour, 38% below RunPod’s list price. Hosts set those prices and they change hourly, so the median sits beside the ranking, not in it.
Nebius, Together AI and DigitalOcean list no A100 at all.

How this ranking was built

Each provider is ranked by its lowest listed on-demand A100 80 GB price per GPU-hour, checked 28 September 2026. The ranking belongs to our model infrastructure coverage and follows the Glamdring research method.

Source. Each provider’s public pricing page, captured on 28 September 2026.
Field. The on-demand hourly price. Lambda labels it PRICE/GPU/HR;² CoreWeave labels it On-Demand Price (Per Hour).³
Unit. US dollars per GPU-hour, before tax. Lambda states that sales tax, VAT or GST is added to its prices.² The other pages captured show no tax line.
Conversions. Modal prices per second: $0.000694 × 3,600 = $2.4984, shown as $2.50.⁴ CoreWeave prices an eight-GPU node: $21.60 ÷ 8 = $2.70.³
Merges. Where a provider lists both PCIe and SXM versions, the lower price ranks and the other appears in the table.
Exclusions. Spot, reserved, cluster and serverless-inference rates are left out of the rank and noted where they change the reading. The 40 GB card is covered separately.
Row counts. Nine pricing pages captured. One is a marketplace (Vast.ai), leaving eight fixed-price clouds. Three of those list no A100: Nebius,⁵ Together AI⁶ and DigitalOcean.⁷ Five clouds remain and all five are ranked.

This is the first edition of the ranking. The next check of the same pages replaces it.

A100 80 GB on-demand prices, ranked

Rank	Provider	A100 80 GB variant listed	On-demand price (USD per GPU-hour, pre-tax)	Smallest on-demand unit
1	RunPod	SXM and PCIe, same price	$1.59	One GPU (Pod)
2	Crusoe	PCIe; SXM at $2.30	$2.00	Priced per GPU-hour
3	Modal	80 GB, form factor not stated	$2.50	Priced per second
4	CoreWeave	Eight-GPU node	$2.70	Eight GPUs, $21.60 an hour
5	Lambda	SXM, eight-GPU instance	$2.79	Eight GPUs, $22.32 an hour

Prices as listed by RunPod,¹ Crusoe,⁸ Modal,⁴ CoreWeave³ and Lambda,² checked 28 September 2026. Lambda’s eight-GPU hourly figure is its $2.79 rate multiplied by eight.

Which cloud is cheapest for a single A100 80 GB?

RunPod is, at $1.59 per GPU-hour for one A100 SXM Pod with 16 vCPUs and 125 GB of RAM.¹ Its A100 PCIe Pod carries the same $1.59 rate with 8 vCPUs and 117 GB of RAM.¹ Crusoe is next at $2.00 per GPU-hour for the PCIe card.⁸

The order changes once the buying unit enters the sum. CoreWeave lists its on-demand A100 as an eight-GPU node at $21.60 an hour.³ Its single-GPU price of $2.70 applies only to customers of CoreWeave’s inference platform.³ Lambda lists the 80 GB A100 only in its eight-GPU instance, at $2.79 per GPU-hour.² Outside CoreWeave’s inference platform, a team that needs one 80 GB card can’t buy one on demand at either.

At list price, a job that uses 100 GPU-hours costs $159 at RunPod and $279 at Lambda.

Form factor moves the price at one provider only. Crusoe charges $0.30 more for SXM than for PCIe,⁸ while RunPod prices both the same.¹ JarvisLabs’ specification table puts the 80 GB SXM at up to 2,039 GB/s of memory bandwidth against up to 1,935 GB/s for the 80 GB PCIe card.⁹

How much cheaper is the A100 40 GB than the 80 GB?

Between about $0.40 and $0.80 per GPU-hour, at the two ranked clouds that list both. Lambda lists the 40 GB A100 at $1.99 against $2.79 for the 80 GB, so the larger card costs 40% more.² Modal lists the 40 GB at $0.000583 a second, or $2.10 an hour, against $2.50 for the 80 GB: 19% more.⁴

Lambda’s pricing page lists the 40 GB card at $1.99 per GPU-hour in all four of its instance tables, in SXM and PCIe versions, while the 80 GB card appears only in the eight-GPU table.² For a single-card workload at Lambda, the 40 GB A100 is the only A100 on offer.

RunPod,¹ Crusoe⁸ and CoreWeave³ list no on-demand 40 GB A100 on the pages checked.

The extra money buys memory and bandwidth. JarvisLabs lists the 40 GB PCIe card with 40 GB of HBM2 and up to 1,555 GB/s, and the 80 GB PCIe card with 80 GB of HBM2e and up to 1,935 GB/s.⁹ Its own guidance is to start with 40 GB unless you know you need more, and to choose 80 GB for training or fine-tuning models in the 13B to 65B parameter range.⁹ That is one GPU rental company’s view. Test it against your own model size, context length and batch size.

How does the Vast.ai marketplace median compare?

It sits below every ranked list price. Vast.ai showed a median of $0.99 an hour for an A100 SXM4 with 80 GB, with offers from $0.31, and a median of $0.74 for an A100 PCIe with 80 GB, with offers from $0.47.¹⁰ The SXM4 median is $0.60 below RunPod’s list price.

The median stays out of the ranking for three reasons. Vast says its prices are set by the market, not by Vast.¹⁰ It says prices reflect currently available offers and update hourly.¹⁰ And it sells three tiers: on-demand, interruptible capacity that may be reclaimed, and reserved terms.¹⁰ The page doesn’t say which tier the median covers.

The decision changes when the work can survive interruption. Vast describes its interruptible tier as ideal for fault-tolerant workloads that checkpoint and resume.¹⁰ A fixed list price buys predictability; the marketplace median shows what the same card fetches when hosts compete.

How do providers define the hourly price?

Each provider prices a different bundle, so the same “per hour” label covers different purchases. Three things differ.

The unit sold. RunPod and Crusoe price one GPU.^1,8 CoreWeave and Lambda price the 80 GB card inside an eight-GPU unit.^3,2 Modal prices per second.⁴

What comes with the GPU. A RunPod A100 SXM Pod lists 16 vCPUs and 125 GB of RAM,¹ with storage priced separately from $0.05 per GB a month.¹ Modal charges CPU at $0.0000131 per physical core per second and memory at $0.00000222 per GiB per second on top of the GPU.⁴ A CoreWeave A100 node lists 128 vCPUs, 2,048 GB of system RAM and 7.68 TB of local storage.³ Lambda’s eight-GPU A100 80 GB instance lists 240 vCPUs, 1,800 GiB of RAM and 19.5 TiB of SSD.²

The product line. RunPod alone lists three A100 rates: $1.59 for a Pod, $1.79 for an A100 SXM in its multi-GPU Clusters, and $2.72 for an 80 GB worker in its Serverless product.¹ Only the Pod rate is on-demand GPU rental in the sense this ranking uses.

The constraint is the unit, not the headline rate. Compare two quotes only after converting both to the same GPU count, the same bundled CPU and memory, and the same billing increment.

Why do other published A100 prices disagree?

Mostly because they were checked on other dates and priced other configurations. Northflank’s comparison, published on 4 August 2025, listed the 80 GB A100 at $2.17 on RunPod, $3.40 on Modal and $1.79 on Lambda.¹¹ Glamdring Research’s check on 28 September 2026 found $1.59 at RunPod,¹ $2.50 at Modal⁴ and $2.79 at Lambda.²

Two of those moved down and one moved up. The pages don’t say why. Lambda’s current 80 GB rate is for an eight-GPU SXM instance,² and Northflank labelled its Lambda figure as bundled full-node pricing,¹¹ so the two figures may not describe the same configuration.

Bundling is the second cause. Northflank’s own table mixes “GPU only” rates with bundled rates that include CPU, RAM and storage.¹¹ It notes that many platforms list low hourly rates but charge separately for CPU, RAM and storage.¹¹ One figure in that table did hold: it listed Modal’s 40 GB A100 at $2.10,¹¹ which matches Modal’s current per-second rate converted to an hour.⁴

The third cause is speed of change. RunPod dated its pricing page 27 September 2026,¹ and Vast’s marketplace prices update hourly.¹⁰ Any A100 price without a checked date is already a guess.

What the ranking cannot tell you

A list price doesn’t promise capacity. Lambda describes its instances as self-serve with first-come access.² Confirm availability in the console before you plan around a rate.

It doesn’t measure cost per finished job either. JarvisLabs argues that the key metric is cost per completed task, not cost per hour, and reports that the H100 is often 1.5 to 3 times faster than the A100 in many LLM inference setups.⁹ A faster card at a higher hourly rate can finish cheaper. Our explainer on what controls AI inference cost sets out the drivers beyond the GPU-hour.

The ranking covers eight fixed-price clouds and one marketplace. AWS, Google Cloud, Azure, Oracle Cloud, JarvisLabs and Northflank weren’t checked, so their absence says nothing about their prices.

RunPod’s page offers both Community Cloud and Secure Cloud.¹ The captured rate doesn’t state which tier it belongs to, so confirm the tier in the current quote.

The ranked figure is the pre-tax GPU rate. Storage is priced on top where a provider charges for it, as RunPod does,¹ so price the whole workload, not the card.

New price checks go out through the Glamdring research briefing.

Frequently asked questions

How much does an A100 cost per hour?

On 28 September 2026, five GPU clouds listed an on-demand A100 80 GB at between $1.59 per GPU-hour at RunPod¹ and $2.79 at Lambda.² The 40 GB card listed from $1.99 at Lambda² and about $2.10 at Modal.⁴ On the Vast.ai marketplace, the median A100 SXM4 offer was $0.99 an hour.¹⁰ All figures are US dollars before tax.

Is spot or reserved A100 capacity cheaper than on-demand?

Where it’s listed, yes. CoreWeave’s North American spot price for its eight-GPU A100 node was $9.65 an hour against $21.60 on demand, 55% lower.³ Vast says its interruptible tier is 50% or more cheaper and its reserved terms up to 50% off.¹⁰ RunPod and Crusoe route A100 reserved or spot pricing through their sales teams.^1,8 Vast warns that interruptible capacity may be reclaimed, so it suits fault-tolerant work that can checkpoint and resume.¹⁰

Does the A100 hourly price include CPU, RAM and storage?

It depends on the provider. RunPod lists the vCPUs and RAM that come with each A100 Pod and prices storage separately.¹ Modal bills CPU and memory on top of the GPU.⁴ CoreWeave and Lambda price a whole eight-GPU machine with its CPUs, RAM and local storage listed alongside.^3,2

What does an A100 80 GB cost to buy outright?

JarvisLabs lists new street prices of $9,500 to $14,000 for the 80 GB PCIe card and $18,000 to $20,000 for the 80 GB SXM card.⁹ It says these are street prices observed through resellers and vary by region, warranty and seller.⁹ Northflank notes that an eight-GPU A100 node can cost upwards of $150,000 once CPUs, RAM, networking and chassis are added.¹¹