Cloud GPU Pricing Compared: RunPod, Lambda, Vast.ai and Paperspace

Cloud GPU pricing compared: RunPod, Lambda, Vast.ai and Paperspace rates for RTX 4090, A100, H100 and B200, plus the billing details that change the bill.

By Galactic News Editorial · Published 2026-11-08T01:35:13+00:00 · Updated 2026-11-08T01:35:13+00:00

Cloud GPU pricing compared across four providers comes down to one question: what does an hour of the card you need actually cost, and what else does the bill include? RunPod, Lambda, Vast.ai, and Paperspace all rent NVIDIA GPUs by the hour, but they sell them in very different ways. RunPod offers fixed tiers. Lambda sells whole nodes. Vast.ai is a marketplace where hosts set prices. Paperspace bundles GPUs with notebooks and a subscription. This guide compares their published rates, read on November 8, 2026, and explains which provider fits which workload.

Every price below is a published list price in US dollars per GPU-hour, before tax. Prices change often, and marketplace prices change by the hour. Check the provider page before you rent. The rates here are a snapshot for comparison, not a quote.

Cloud GPU pricing compared at a glance

Read the table with care. RunPod's two numbers are its Secure Cloud and Community Cloud tiers. Lambda's H100 and A100 prices are per GPU inside 8-GPU nodes. Vast.ai's numbers are marketplace ranges, not fixed rates. Paperspace's A100 price sits behind a paid subscription. The cheapest number in a row is rarely the whole story.

RunPod: fixed tiers, per-second billing

RunPod splits its inventory into three tiers. Community Cloud runs on third-party host hardware, with no formal service agreement and the lowest rates. Secure Cloud runs in RunPod-operated data centers with an agreement, and costs more. Serverless bills inference workers by usage instead of renting a whole GPU.

On the pricing page read today, Secure Cloud lists the RTX 4090 at $0.89 per hour, the A100 PCIe and A100 SXM at $1.79, the H100 PCIe at $2.89, the H100 SXM at $3.99, the H200 at $5.29, the B200 at $7.99, and the B300 at $8.99. Community Cloud starts lower: the RTX 4090 from $0.34 per hour, the A100 PCIe from $1.19, the A100 SXM from $1.39, the H100 PCIe from $1.99, and the H100 SXM from $2.69. Serverless rates run higher per active hour: $1.34 for an RTX 4090 worker, $2.72 for an A100, and $5.39 for an H100.

Pods bill per second, so a ten-minute test costs about a sixth of the hourly rate. Storage is separate. Container disk costs $0.10 per GB per month, and volume disk runs $0.10 per GB per month while the pod runs, or $0.20 when idle. That idle storage charge is the line that surprises new users. Stop the pod, but keep the volume, and the meter keeps running.

RunPod fits developers who want a fixed rate card and short sessions. The Community tier is the cheapest fixed-rate RTX 4090 and A100 on this list, when it is in stock. The trade-off is availability. Community inventory depends on hosts, and popular cards can sell out.

Lambda: whole nodes, no egress fees

Lambda sells GPU instances in fixed node shapes of 1, 2, 4, and 8 GPUs. The per-GPU rate eases slightly as the node grows. On-demand H100 SXM costs $4.29 per GPU-hour as a single card, $4.19 at 2x, $4.09 at 4x, and $3.99 at 8x. The H100 PCIe single-GPU instance costs $3.29 per hour. The A100 40 GB costs a flat $1.99 per hour at every node size. The A100 80 GB SXM costs $2.79 per GPU-hour in an 8x node. The B200 costs $6.99 per hour as a single card and $6.69 at 8x. These are the rates published in October 2026.

Lambda's pricing page now shows a new rate card that takes effect on November 16, 2026: HGX B200 at $7.69, HGX H100 at $4.59, HGX A100 80 GB at $3.19, A100 40 GB at $1.99, and V100 at $0.79 per GPU-hour. If you are budgeting past that date, use the new card, not the old one. The page notes that current pricing is visible inside a Lambda Cloud account until the change takes effect.

Two billing facts set Lambda apart. Instances bill per minute, and Lambda charges no data egress fees. That matters for training runs that move large datasets in and out. Lambda also sells 1-Click Clusters: committed multi-week reservations of 16 to 2,000 or more GPUs. A 16-GPU H100 cluster runs $6.16 per GPU-hour, which is more than the on-demand rate, not less. Clusters buy guaranteed capacity and fast interconnect, not a discount. Lambda fits teams that need reliable multi-GPU capacity and predictable billing, and who can live without consumer cards like the RTX 4090. Its catalog stops at the A100 on the low end, apart from a legacy V100.

Vast.ai: the cheapest sticker, the most variance

Vast.ai is not a cloud provider in the usual sense. It is a marketplace where independent hosts list their GPUs, and prices are set by supply and demand across more than 20,000 GPUs. There is no fixed rate card. You search the marketplace, filter by model, memory, price, and host reliability, and launch an instance. You can start with as little as $5 of credit.

That structure produces the lowest sticker prices on this list. Recent marketplace readings put the RTX 4090 at about $0.48 per hour typical, with listings from about $0.22. The A100 80 GB runs about $0.50 to $0.80 per hour. The H100 runs about $0.90 to $1.87 per hour on interruptible listings, and about $1.50 to $2.27 on datacenter-verified hosts. Interruptible instances run 30 to 50 percent cheaper than on-demand listings on the same host, because the host can reclaim them.

The catch is reliability, and it is host-specific. A datacenter host with a strong uptime history is dependable. A budget peer host can disconnect without warning. That is why the same GPU can cost half as much on one host as on another. For fault-tolerant work, such as batch inference or checkpointed training, the savings are real. For a long training run with no checkpointing, a disconnection can cost more than the discount saved. Budget above the listed rate, and favor verified hosts for anything you cannot afford to restart.

Paperspace: notebooks first, GPUs second

Paperspace, owned by DigitalOcean since 2023, sells a different product from the other three. Its Gradient platform provides managed Jupyter notebook workspaces, persistent machines, and deployment endpoints. You pick a GPU, open a notebook, and start working. The audience is data scientists and small teams who want notebooks, not infrastructure.

DigitalOcean's official Paperspace pricing documentation, last verified July 13, 2026, lists these hourly machine rates: GPU+ (M4000) at $0.45, A4000 at $0.76, A5000 at $1.38, A6000 at $1.89, V100 at $2.30, A100 at $3.09, A100-80G at $3.18, and H100 (PCIe) at $5.95 per hour. All Paperspace resources bill per hour.

The subscription is the part buyers miss. Gradient plans cost $0 for Free, $8 per month for Pro, and $39 per month for Growth. The A100 tiers, and most workstation cards from the A5000 up, require the Growth plan. The H100 carries no subscription requirement in the pricing table. The Growth plan accrues at $0.058 per hour of use, capped at $39 per month. Add that to the hourly rate when you compare Paperspace with a provider that charges no subscription.

Note that DigitalOcean is folding Paperspace into its GPU Droplets product, which has its own rate card. After an August 1, 2026 price change, on-demand HGX H100 on Droplets costs $4.41 per GPU-hour, and H200 costs $4.47. Check which console you are standing in before you compare prices, because the two product lines do not share a rate card.

What else the hourly rate leaves out

The hourly GPU price is one line of the bill. Storage, egress, and idle time are the other three. RunPod charges for idle volume storage. Lambda charges no egress fees, but persistent storage still bills while instances are stopped. Vast.ai listings vary: some hosts charge for storage, some do not, and the listing shows the terms. Paperspace charges $0.29 per GB per month for storage above the plan allowance.

Billing granularity matters too. RunPod and Vast.ai bill per second on most listings, which rewards short sessions. Lambda bills per minute. Paperspace bills per hour on machines, which makes a 20-minute notebook session cost a full hour. If your sessions are short and frequent, per-second billing can cut the bill by a third or more. If your sessions run for days, granularity stops mattering.

Availability is the last hidden cost. RunPod Community stock varies. Lambda's single-GPU instances are frequently out of stock, and its H100 and A100 SXM cards are sold mainly as 8-GPU nodes, so a small job can mean paying for eight GPUs. Vast.ai availability changes by the hour. Paperspace H100 supply is limited to three regions. Check availability, not just price, before you commit to a provider for a deadline.

Which provider fits which job

Match the provider to the workload, not to the cheapest row in the table. Short experiments and inference on open models fit RunPod Community Cloud or Vast.ai, where the RTX 4090 costs well under a dollar an hour. Multi-GPU training with a deadline fits Lambda, where the node is guaranteed and the interconnect is built in. Notebook-based exploration and teaching fit Paperspace, where the environment is ready in a minute and the GPU is almost incidental.

A practical rule: if the job can restart from a checkpoint, take the cheapest marketplace rate. If the job cannot restart, pay for the fixed tier. If the job needs eight GPUs talking to each other, buy the node, not eight single cards.

Bottom line

Cloud GPU pricing compared honestly shows a wide spread, not a single winner. RunPod has the cheapest fixed-rate RTX 4090 and A100, at $0.34 and $1.19 per hour on Community Cloud. Lambda has the most predictable multi-GPU pricing, with H100 SXM at $3.99 to $4.29 per hour today, rising to $4.59 on November 16. Vast.ai has the lowest sticker prices of all, with RTX 4090 listings from about $0.22 per hour, but reliability varies by host. Paperspace has the highest headline H100 rate at $5.95 per hour, and sells notebooks rather than raw compute. Pick the provider whose billing model matches your workload, add storage and idle time to the math, and re-check the pricing page before you rent. Rates in this market move every month.