Compare / GPU cloud
On-demand GPU cloud pricing, normalized
Among published like-for-like on-demand rates, RunPod Community Cloud is the cheapest listed non-marketplace H100 SXM ($2.69/GPU-hour), A100 SXM ($1.39), and L40S ($0.79). Vast.ai's official feed on 18 Oct 2026 starts lower (H100 SXM from $1.33, A100 SXM4 from $0.40, L40S from $0.43) but those are marketplace starting offers, not a guaranteed SKU. Modal converts to $3.95 / $2.50 / $1.95 per hour from per-second serverless. Lambda's 1× H100 SXM is $4.29 (8× $3.99). Fly.io GPUs ended after 1 Aug 2026. AWS, GCP, and Azure on-demand dollar rates were not in the official HTML we fetched — em dashes, not guesses.
Prices from each provider's official pricing page or official public feed, checked 18 Oct 2026. USD, excluding tax. Headline cells are on-demand (or serverless converted to an hour). Secure Cloud, interruptible, committed-use, and 1× vs 8× siblings are in the tooltips. Vast.ai is a marketplace: the headline is the official feed's current starting (minimum) offer that day; median is in the tooltip. "Not published" or "unverified" means the official page we fetched did not list a dollar rate. Estimates, not guarantees.
Price per GPU-hour, same accelerator
RunPod Details
| $2.69 Community Cloud H100 SXM. Secure Cloud is higher on the same page. | $1.39 Community Cloud A100 SXM 80GB. Secure Cloud is higher. | $0.79 Community Cloud L40S. RTX 4090 Community is $0.34 (alternate, not L40S). | Per second Community Cloud headline; Secure Cloud in the tooltip |
|---|---|---|---|---|
Lambda Cloud Details
| $4.29 1× H100 SXM 80GB. 8× H100 SXM is $3.99/GPU-hour; 1× H100 PCIe is $3.29. | $2.79 8× A100 SXM 80GB only ($2.79/GPU-hour). No 1× A100 80GB on the page. | — L40S not on lambda.ai/pricing. No RTX 4090 list rate either. | Per minute On-demand instances; 1-minute minimum |
Vast.ai Details
| $1.33 Marketplace starting (official feed min) H100 SXM on 18 Oct 2026 12:30 UTC. Median $2.07. Interruptible is cheaper. | $0.40 Marketplace starting A100 SXM4 80GB. Median $0.71. Feed timestamp 2026-10-18T12:30:14Z. | $0.43 Marketplace starting L40S. Median $0.72. RTX 4090 starting $0.2645 (median not used as headline). | Marketplace Hosts set prices. Headline is the official feed minimum that day, not a guaranteed SKU. |
Modal Details
| $3.95 $0.001097/s × 3,600 = $3.9492. Serverless; you pay only while the container runs. | $2.50 $0.000694/s × 3,600 = $2.4984 for A100 80GB. | $1.95 $0.000542/s × 3,600 = $1.9512 for L40S. | Per second Serverless GPU; converted to an hour for the table |
CoreWeave Details
| $6.16 HGX H100 $49.24/hour for 8 GPUs → $6.16 per GPU. | $2.70 A100 80GB $21.60/hour for 8 GPUs → $2.70 per GPU. | $2.25 L40S $18.00/hour for 8 GPUs → $2.25 per GPU. | Hourly, 8-GPU nodes On-demand list; committed-use is a different quote |
DigitalOcean GPU Droplets Details
| $4.41 H100 GPU Droplet. Paperspace-shaped product on DigitalOcean. | — A100 80GB not on digitalocean.com/pricing/gpu-droplets | $1.57 L40S GPU Droplet. | Hourly GPU Droplets; monthly cap not used as the headline |
Fly.io Details
| — GPU Machines ended after 1 Aug 2026. No current H100 list rate. | — GPU Machines ended after 1 Aug 2026. No current A100 list rate. | — GPU Machines ended after 1 Aug 2026. No current L40S list rate. | Ended Fly.io GPU Machines were withdrawn after 1 Aug 2026 |
Amazon EC2 Details
| — p5 (H100) on-demand $/hour not in the official HTML we fetched (on-demand table is often JavaScript). | — p4 (A100) on-demand $/hour not in the official HTML we fetched. | — g6e (L40S) on-demand $/hour not in the official HTML we fetched. | On-demand (unverified $) Official instance pages list SKUs; dollar rates live in a JS price table |
Google Cloud Details
| — A3 (H100) on-demand $/hour not in the official HTML we fetched (region picker). | — A2 (A100) on-demand $/hour not in the official HTML we fetched. | — G2 (L4, not L40S) on-demand $/hour not in the official HTML we fetched. L40S is not the G2 GPU. | On-demand (unverified $) Accelerator-optimized VMs; dollars sit behind a region picker |
Microsoft Azure Details
| — ND H100 v5 on-demand $/hour not in the official HTML we fetched (signed-in calculator). | — ND A100 v4 on-demand $/hour not in the official HTML we fetched. | — NC / L40S-class on-demand $/hour not in the official HTML we fetched. | On-demand (unverified $) N-series VM pricing HTML often hides dollars behind a calculator |
What you actually get
| RunPod | Community H100 SXM 80GB | Community A100 SXM 80GB | On-demand pod; Secure Cloud is a higher listed SKU | RTX 4090 Community $0.34/hour (tooltip alternate) | RunPod Community and Secure datacenters |
|---|---|---|---|---|---|
| Lambda Cloud | 1× H100 SXM 80GB ($4.29); 8× $3.99/GPU | 8× A100 SXM 80GB ($2.79/GPU) | On-demand; reserved clusters are a different quote | No L40S or 4090 list rate | Lambda Cloud regions on the pricing page |
| Vast.ai | H100 SXM (starting offer) | A100 SXM4 80GB (starting offer) | Marketplace; offer can disappear. Interruptible vs on-demand is host-specific. | RTX 4090 starting $0.26 (tooltip alternate) | Wherever hosts list that GPU |
| Modal | H100 (serverless) | A100 80GB (serverless) | Per-second container; no idle VM unless you keep it warm | $30/month included compute on Starter (not subtracted from the headline hour) | Modal; some regions multiply the base rate |
| CoreWeave | HGX H100 (8-GPU node / 8) | A100 80GB (8-GPU node / 8) | On-demand list on the public pricing page | Committed-use discounts not used as the headline | CoreWeave regions on the pricing page |
| DigitalOcean GPU Droplets | H100 GPU Droplet | A100 80GB not listed | On-demand Droplet | No A100 80GB SKU on the GPU Droplets page | DigitalOcean GPU regions |
| Fly.io | Not offered | Not offered | Discontinued after 1 Aug 2026 | None — product ended | n/a |
| Amazon EC2 | p5 (H100) — rate unverified | p4d/p4de (A100) — rate unverified | On-demand / spot / reserved; only on-demand was in scope | Dollar amounts omitted rather than guessed from blogs | AWS regions; price varies |
| Google Cloud | A3 (H100) — rate unverified | A2 (A100) — rate unverified | On-demand / spot / committed; only on-demand was in scope | G2 is L4, not L40S — not treated as like-for-like even if a dollar appeared | GCP regions; price varies |
| Microsoft Azure | ND H100 v5 — rate unverified | ND A100 v4 — rate unverified | On-demand / spot / reserved; only on-demand was in scope | Dollar amounts omitted rather than guessed from blogs | Azure regions; price varies |
Best for
- RunPod
The cheapest published non-marketplace like-for-like H100 / A100 / L40S on-demand rates.
- Lambda Cloud
An 8× H100 SXM box at $3.99/GPU-hour when you want a single vendor cluster, not a marketplace bid.
- Vast.ai
A floor check — only if you accept a marketplace starting bid instead of a listed Cloud SKU.
- Modal
Short GPU jobs billed by the second in the same place as Modal CPU/sandboxes.
- CoreWeave
A listed 8-GPU HGX node when you already run on CoreWeave and will take the $6.16 H100.
- DigitalOcean GPU Droplets
An H100 or L40S Droplet on a familiar DO account when you do not need A100 80GB.
- Fly.io
Not for new GPU work. Use Fly for CPU Machines; rent GPUs elsewhere.
- Amazon EC2
Reference only until the official on-demand table is readable as HTML.
- Google Cloud
Reference only until the official GPU price table is readable as HTML.
- Microsoft Azure
Reference only until the official VM price table is readable as HTML.
A $2.69 Community H100 vs a marketplace starting bid vs a per-second serverless GPU
Rows are providers, not individual listings. The headline is USD per GPU-hour for that accelerator. RunPod Community is the listed cheap Cloud SKU; Secure Cloud is higher and lives in the tooltip. Lambda publishes 1× and 8× H100 SXM plus PCIe; the H100 headline is the 1× SXM so it matches a single-GPU rental, with the 8× and PCIe rates in the tooltip. CoreWeave list prices are 8-GPU nodes divided by 8.
Modal bills per second. The hour column is the published per-second rate × 3,600. That is the right compare if the GPU is up for a full hour; shorter jobs cost less because they stop at the second. DigitalOcean GPU Droplets are the current Paperspace-shaped product: H100 $4.41, L40S $1.57; no A100 80GB on the page we fetched.
Vast.ai is a marketplace. Hosts set prices. The official public feed (18 Oct 2026 12:30 UTC) gives a current minimum, p10, and median per GPU type. The headline is that starting minimum. It may be interruptible, in a specific region, or gone an hour later — treat it as a floor, not a quote. Fly.io ended GPU machines after 1 Aug 2026. AWS p5/p4/g6e, GCP A3/A2/G2, and Azure ND/NC dollar rates did not appear in the official HTML we fetched (region pickers / signed-in calculators), so those cells are em dashes.
What would a real project cost?
Estimates, not guarantees. Built from the published rates above (checked 18 Oct 2026); check each provider's current pricing before you commit. Tax is not included. Numbers marked unverified or not published were left out rather than guessed.
- Small: fine-tuning experiments. 20 GPU-hours a month of L40S (or the published L40S-class sibling). Evening LoRA runs, then stop. Tax not included.
- Medium: steady inference box. 200 GPU-hours a month of A100 80GB — one card up about 6.5 hours a day, or a smaller always-on fraction. On-demand list, no reserved discount. Tax not included.
- Large: multi-GPU training run. 640 H100-hours a month: 8× H100 80GB × 80 hours (a long weekend cluster, or shorter bursts that add up). On-demand list. Tax not included.
| RunPod | $15.80 Community L40S | $278.00 Community A100 SXM | $1,721.60 Community H100 SXM |
|---|---|---|---|
| Lambda Cloud | — No L40S list rate | $558.00 8× A100 80GB | $2,553.60 8× H100 SXM at $3.99/GPU |
| Vast.ai | $8.56 marketplace starting L40S; median ~$14.40 | $80.46 marketplace starting A100; median ~$142 | $853.95 marketplace starting H100; median ~$1,324.80 |
| Modal | $39.00 L40S serverless × 20 h | $500.00 A100 80GB × 200 h | $2,528.00 H100 × 640 h |
| CoreWeave | $45.00 L40S | $540.00 A100 80GB | $3,942.40 H100 |
| DigitalOcean GPU Droplets | $31.40 L40S | — No A100 80GB list rate | $2,822.40 H100 |
| Fly.io | — GPU Machines ended after 1 Aug 2026 | — GPU Machines ended after 1 Aug 2026 | — GPU Machines ended after 1 Aug 2026 |
| Amazon EC2 | — On-demand $ not in fetched HTML | — On-demand $ not in fetched HTML | — On-demand $ not in fetched HTML |
| Google Cloud | — On-demand $ not in fetched HTML | — On-demand $ not in fetched HTML | — On-demand $ not in fetched HTML |
| Microsoft Azure | — On-demand $ not in fetched HTML | — On-demand $ not in fetched HTML | — On-demand $ not in fetched HTML |
Small: fine-tuning experiments
Try first: RunPod Community L40S at about $15.80 — a listed Cloud SKU, per-second, cheapest published non-marketplace L40S.
Runner-up: Vast.ai starting L40S at about $8.56 if you will take a marketplace bid (median that day was $0.72/hour → about $14.40). DigitalOcean L40S is about $31.40.
Avoid: CoreWeave L40S at about $45 and Modal at about $39 for 20 full hours. Lambda has no L40S. Fly GPUs ended. AWS/GCP/Azure dollars unverified.
Medium: steady inference box
Try first: RunPod Community A100 SXM at about $278.00 — listed Cloud SKU, not a marketplace floor.
Runner-up: Vast.ai starting A100 SXM4 at about $80.46 if you will take a marketplace bid (median $0.71 → about $142). Lambda 8× A100 is about $558.00. Modal is about $500.00.
Avoid: Treating Vast's $80 starting bid as a quote, CoreWeave at about $540 when RunPod is $278, and DigitalOcean / Fly / hyperscalers (no verified A100 80GB $).
Large: multi-GPU training run
Try first: RunPod Community H100 SXM at about $1,721.60 — listed per-GPU Cloud rate for 640 hours.
Runner-up: Vast.ai starting H100 SXM at about $853.95 if you will take a marketplace bid (median $2.07 → about $1,324.80). Lambda 8× H100 SXM at $3.99/GPU is about $2,553.60. Modal is about $2,528.00.
Avoid: CoreWeave at about $3,942.40 and DigitalOcean at about $2,822.40 when RunPod is $1,721.60. Fly GPUs ended. AWS/GCP/Azure dollars unverified.
Show the math
- Project — Small: fine-tuning experiments
- 20 GPU-hours a month of L40S (or the published L40S-class sibling). Evening LoRA runs, then stop. Tax not included. L40S × 20 h: RunPod $0.79 → $15.80. Vast min $0.4281 → $8.56 (median $0.72 → $14.40). Modal $1.95 → $39.00. CoreWeave $2.25 → $45.00. DigitalOcean $1.57 → $31.40. Lambda / Fly / AWS / GCP / Azure omitted.
- Project — Medium: steady inference box
- 200 GPU-hours a month of A100 80GB — one card up about 6.5 hours a day, or a smaller always-on fraction. On-demand list, no reserved discount. Tax not included. A100 80GB × 200 h: RunPod $1.39 → $278.00. Vast min $0.4023 → $80.46 (median $0.71 → $142). Modal $2.50 → $500.00. Lambda $2.79 → $558.00. CoreWeave $2.70 → $540.00. DigitalOcean / Fly / AWS / GCP / Azure omitted.
- Project — Large: multi-GPU training run
- 640 H100-hours a month: 8× H100 80GB × 80 hours (a long weekend cluster, or shorter bursts that add up). On-demand list. Tax not included. H100 × 640 h: RunPod $2.69 → $1,721.60. Vast min $1.3343 → $853.95 (median $2.07 → $1,324.80). Lambda 8× $3.99 → $2,553.60 (1× $4.29 would be $2,745.60). Modal $3.95 → $2,528.00. DigitalOcean $4.41 → $2,822.40. CoreWeave $6.16 → $3,942.40. Fly / AWS / GCP / Azure omitted.
- RunPod
- Official Community Cloud: H100 SXM $2.69, A100 SXM $1.39, L40S $0.79, RTX 4090 $0.34 per GPU-hour. Secure Cloud is higher — tooltip only. Billed per second. Source: https://www.runpod.io/pricing (checked 18 Oct 2026)
- Lambda Cloud
- Official: 1× H100 SXM $4.29/GPU-hour; 8× H100 SXM $3.99; 1× H100 PCIe $3.29; 8× A100 SXM 80GB $2.79. Billed per minute. L40S not listed. Source: https://lambda.ai/pricing (checked 18 Oct 2026)
- Vast.ai
- Official public feed 2026-10-18T12:30:14Z: H100 SXM min $1.3343 (median $2.07), A100 SXM4 min $0.4023 (median $0.71), L40S min $0.4281 (median $0.72), RTX 4090 min $0.2645. Headline = min; median in the tooltip. Source: https://vast.ai/pricing (checked 18 Oct 2026)
- Modal
- Official: H100 $0.001097/s → $3.95/h; A100 80GB $0.000694/s → $2.50/h; L40S $0.000542/s → $1.95/h. Starter includes $30/month compute — not removed from the hourly cell. Source: https://modal.com/pricing (checked 18 Oct 2026)
- CoreWeave
- Official on-demand: HGX H100 $49.24/8 = $6.16 per GPU-hour; A100 80GB $21.60/8 = $2.70; L40S $18.00/8 = $2.25. Source: https://www.coreweave.com/pricing (checked 18 Oct 2026)
- DigitalOcean GPU Droplets
- Official GPU Droplets: H100 $4.41/hour, L40S $1.57/hour. A100 80GB not listed — omitted. Source: https://www.digitalocean.com/pricing/gpu-droplets (checked 18 Oct 2026)
- Fly.io
- Official docs: GPU Machines ended after 1 Aug 2026. No current like-for-like H100/A100/L40S rate — omitted. Source: https://fly.io/docs/about/pricing/ (checked 18 Oct 2026)
- Amazon EC2
- Official pages list p5/p4/g6e SKUs. The fetched on-demand HTML did not include dollar amounts. Cells are em dashes. Source: https://aws.amazon.com/ec2/pricing/on-demand/ (checked 18 Oct 2026)
- Google Cloud
- Official accelerator pages list A3/A2/G2 SKUs. The fetched HTML did not include dollar amounts. Cells are em dashes. Source: https://cloud.google.com/compute/gpus-pricing (checked 18 Oct 2026)
- Microsoft Azure
- Official N-series pages list ND H100 v5 / ND A100 v4 / NC SKUs. The fetched HTML did not include dollar amounts. Cells are em dashes. Source: https://azure.microsoft.com/pricing/details/virtual-machines/ (checked 18 Oct 2026)
FAQ
- What is the cheapest H100 per hour?
- Vast.ai's official feed started at $1.33/GPU-hour on 18 Oct 2026, but that is a marketplace minimum. The cheapest published non-marketplace H100 SXM is RunPod Community at $2.69. Lambda 1× SXM is $4.29; 8× is $3.99. Modal is $3.95. CoreWeave is $6.16. DigitalOcean is $4.41.
- Why is Vast.ai cheaper and marked marketplace?
- Hosts set prices. The official feed publishes a current min, p10, and median. The headline is the minimum that day. The card may be interruptible, remote, or gone later. Median H100 that snapshot was $2.07 — still under RunPod, still not a Cloud SKU.
- Did Fly.io shut down GPUs?
- Yes. Official pricing docs say GPU Machines ended after 1 Aug 2026. Those cells are em dashes.
- Why are AWS, GCP, and Azure em dashes?
- Their official on-demand pages list the SKUs (p5, A3, ND H100 v5, …) but the fetched HTML did not include dollar amounts — JavaScript tables, region pickers, or signed-in calculators. Guessing from blogs is not allowed, so the cells are em dashes.
- How did you convert Modal to per hour?
- Published per-second rate × 3,600. H100 $0.001097 × 3,600 = $3.95. You only pay while the container runs, so a 10-minute job is one-sixth of that hour.
- What would a real project cost?
- Estimates from published rates: 20 L40S hours are about $16 on RunPod, $9 starting on Vast, $31 on DigitalOcean, $39 on Modal. 200 A100 80GB hours are about $278 on RunPod, $80 starting on Vast, $500 on Modal, $558 on Lambda. 640 H100 hours are about $1,722 on RunPod, $854 starting on Vast, $2,528 on Modal, $2,554 on Lambda 8×.
Related pages
Stack of the Week
Get one stack a week that replaces a SaaS bill.