RunPod vs Vast.ai: GPU prices, September 2026
Vast.ai's displayed H100 prices were lower in the 28 September 2026 capture. RunPod listed lower A40, B200 and B300 rates; B200 memory labels differ. Compare the chosen rental's storage, transfer and capacity terms before deciding on total cost.
Vast.ai had lower displayed H100 GPU-hour prices in the providers’ pricing pages captured 28 September 2026.1,2 RunPod listed lower B200 and B300 rates.1,2 It also beat Vast.ai’s A40 price.1,2 The cheaper rental depends on the GPU and the storage, transfer and capacity terms attached to it.3
TL;DR
- Check the actual Vast.ai offer before budgeting: hosts set prices, and the displayed offers update hourly.3,2
- Both services publish per-second compute billing; compare their storage and transfer charges separately.4,5
- Stopping compute can leave storage charges running on either service.1,5
- RunPod describes formal SLAs for enterprise workloads and SLA-backed Reserved Clusters; confirm the agreement for the capacity you rent.4,1
For the wider category, browse our compute infrastructure research.
How were these prices compared?
We compared RunPod’s displayed Pods rates with Vast.ai’s displayed starting offers and medians, using prices captured 28 September 2026.1,2 The table selects ten overlapping GPU models. It compares published prices, with no performance weighting or provider ranking.
RunPod’s Pods are dedicated GPU instances, while Vast.ai describes a marketplace where hosts set prices.1,3 A marketplace median is a reference point for comparing offers, not a fixed quote for the machine you will rent.3
The RunPod column reproduces the displayed Pods rates. It does not reconstruct separate Secure Cloud and Community Cloud quotes. We also keep Vast.ai’s starting offers and medians separate rather than assigning either figure to a particular host tier or rental type.
Prices retain the providers’ dollar ($) notation.1,2 Confirm currency and tax at checkout before using these displayed prices as a final quote. The comparison follows our research methodology.
Official Vast instance-pricing and payment-billing documentation was captured separately on 1 October 2026.3,5 RunPod’s own comparison statements about billing and capacity were also checked on 1 October 2026.4 These later terms are separate from the September GPU-price snapshot.
Which GPUs cost less on each service?
Vast.ai’s displayed medians were lower for H100 PCIe and H100 SXM, while RunPod was lower for B200 and B300.1,2,1 Vast.ai also displayed a lower RTX 4090 median.1,2 RunPod was lower for A40.1,2 The comparison changes with the GPU model.
All figures below are displayed dollars per GPU-hour, checked 28 September 2026.1,2
| GPU | RunPod listed Pods rate ($/GPU-hour) | Vast.ai from ($/GPU-hour) | Vast.ai median ($/GPU-hour) |
|---|---|---|---|
| H100 PCIe, 80 GB | 2.891 | 1.872 | 2.532 |
| H100 SXM, 80 GB | 3.491 | 1.732 | 2.162 |
| A100 PCIe, 80 GB | 1.591 | 0.472 | 0.742 |
| RTX 5090 | 0.991 | 0.332 | 0.632 |
| RTX 4090 | 0.741 | 0.202 | 0.512 |
| RTX 3090 | 0.501 | 0.112 | 0.192 |
| L40S | 1.091 | 0.472 | 0.792 |
| A40 | 0.491 | 0.792 | 0.792 |
| B200 | 6.791 | 9.382 | 10.632 |
| B300 | 7.891 | 8.752 | 8.752 |
The B200 memory labels differ: RunPod lists 180 GB and Vast.ai lists 192 GB.1,2 Treat that row as a comparison of B200-labelled rates and confirm the available memory in the actual configurations.
The H100 comparison keeps PCIe and SXM separate: RunPod lists them at $2.89 and $3.49 respectively, against Vast.ai medians of $2.53 and $2.16.1,2 Use the matching variant when comparing other suppliers in our H100 rental-price comparison.
A40 provides the clearest counterexample to a universal Vast.ai price advantage.1,2 RunPod’s $0.49 listed rate was below both Vast.ai’s $0.79 starting offer and its $0.79 median.1,2
How do billing and data transfer differ?
RunPod and Vast.ai both publish per-second compute billing, but their stated transfer policies differ.4,5 RunPod’s own comparison says it charges no ingress or egress fees.4 Vast.ai’s billing documentation charges for each byte transferred at rates that vary by machine.5
Vast.ai’s instance-pricing documentation includes both upload and download traffic in bandwidth charges.3 Before renting, record the chosen host’s compute rate, allocated storage price and transfer rates together. A GPU-price card alone leaves those storage and bandwidth charges outside the comparison.3
Vast.ai also requires prepaid credits.5 Its payment documentation says instances stop when the balance reaches zero or below.5 Keeping an account funded is therefore part of keeping the workload running, even when the chosen capacity is on-demand.5
What storage costs remain when compute stops?
RunPod lists an idle volume-disk charge, and Vast.ai bills allocated storage while an instance is stopped but remains online.1,5 On either service, a stopped GPU can leave a storage bill.1,5
RunPod’s storage rates were listed as follows on 28 September 2026.1
| RunPod storage type or state | Listed price ($/GB/month) |
|---|---|
| Container disk | 0.101 |
| Volume disk, running | 0.101 |
| Volume disk, idle | 0.201 |
| Standard network storage, under 1 TB | 0.071 |
| Standard network storage, over 1 TB | 0.051 |
| High-performance network storage | 0.141 |
Keep the running and idle volume-disk prices separate when budgeting gaps between jobs. Our RunPod pricing breakdown covers the service’s pricing structure in more detail.
Vast.ai lists storage in $/GB/hour and bandwidth in $/TB, with rates that vary by machine.5 Match the billing periods before comparing a Vast.ai storage quote with RunPod’s monthly rates.
Vast.ai’s 1 October instance-pricing documentation says storage costs vary by host and are typically higher for stopped instances than running ones.3 It instructs users to delete instances completely to cease instance-storage billing.3 Copy off anything you need before deleting a rental.
Vast.ai’s detailed payment documentation makes an additional distinction: offline instances accrue neither active-rental nor storage charges.5 “Stopped” and “offline” therefore have different billing treatment.5 Use the instance’s actual state and the host’s storage rate when calculating retained-storage cost.
What changes between secure and community capacity?
Secure Cloud and Community Cloud labels require different capacity checks on RunPod; Vast.ai requires an offer-specific check of host terms and rental priority.4,3 Compare the capacity and agreement attached to the GPU before treating two displayed prices as equivalent rentals.
RunPod places consumer GPUs on Community Cloud and says inventory is visible in the console at deployment.4 Its own comparison describes supply through verified data centre partners and formal SLAs available for enterprise workloads.4 It also says Secure Cloud has no active workload monitoring.4 These are RunPod’s published descriptions, not an independently measured uptime comparison.
RunPod’s pricing page specifically describes Reserved Clusters as having guaranteed availability and SLA-backed uptime.1 That product-level description is insufficient to establish the same agreement for every Community Cloud or on-demand Pod. Confirm the SLA’s coverage for the actual rental.
Vast.ai’s official instance-pricing documentation lists on-demand and reserved rentals as high-priority access; interruptible instances are low-priority and may be paused.3 Its price page markets on-demand as “Guaranteed uptime” and “No interruptions”.2 Those statements do not supply a measured uptime comparison or the compensation terms for a particular host.
Vast.ai also identifies host reliability as a pricing factor and says higher reliability typically correlates with higher prices.3 Keep host reliability and the right to interrupt an instance as separate selection criteria. A lower marketplace price does not establish that either matches the RunPod capacity you are considering.
Which service should you choose for your workload?
Glamdring Research’s verdict is to compare Vast.ai offers first for the H100 variants, while keeping RunPod in the shortlist for B200 and B300.1,2,1 For A40, the captured RunPod rate was lower even than Vast.ai’s starting offer.1,2 This recommendation concerns displayed compute prices, not completed-job cost.
For a restartable batch job, consider interruptible access only after confirming that the job can resume and that its outputs survive the interruption. Vast.ai publishes interruptible access as low-priority capacity that may be paused.3 The discount is a different rental condition from high-priority on-demand access.3
For a workload that must keep serving users, request the capacity commitment, outage remedy and support terms alongside the price. RunPod’s published enterprise-SLA and Reserved Clusters descriptions provide a reason to examine those products; they do not prove that any RunPod rental will finish a job more cheaply.4,1
For a transfer-heavy workload, compare the chosen Vast.ai machine’s bandwidth charges with RunPod’s stated no-ingress/no-egress policy.5,4 For intermittent work, compare the cost of retaining storage during idle periods.1,5 Use those terms to test whether the lower compute rate remains the lower bill.
What can these prices not tell you?
RunPod’s listed rates and Vast.ai’s marketplace prices do not establish the cost of a completed workload: storage and deployment choices affect the bill.1,3 The table compares displayed GPU-hour prices without a workload benchmark.
The comparison does not match CPU allocation, system RAM, storage performance, networking or region across actual instances. It also does not assign Vast.ai’s price cards to a verified-host tier or a specific rental priority. Confirm those details in a current matched offer.
B200’s differing memory labels remain a configuration limit.1,2 Other GPUs with unresolved variant matching are outside the selected table. No percentage premium for restarts or universal reliability winner follows from these prices.
The decision changes when a matched offer, its storage and transfer charges, or its capacity agreement changes. Recheck those terms before committing to a rental.
Frequently asked questions
Can you rent without a long-term commitment?
Yes. RunPod describes on-demand access with no commitment, and Vast.ai’s price page says no long-term contract is required.4,2 Vast.ai’s reserved option instead uses discounted rates with prepayment.3 Compare that commitment separately from a flexible rental.
What happens if a Vast.ai account runs out of credits?
Vast.ai stops instances when the credit balance reaches zero or below and schedules instances, storage volumes and data for deletion unless the balance becomes positive.5 Stopped instances detach from their GPUs but continue to incur disk-storage charges.5
A saved credit card is periodically charged to cover a negative balance and protect instances from deletion.5 Without that backup payment method, instances and stored data can be destroyed.5 The documented grace buffer depends on historical daily spend; it has no universal duration.5
Can you compare RunPod Serverless with these Vast.ai rates?
Only after changing the pricing basis. RunPod’s Serverless charges inference workers by usage, while the table uses its dedicated Pods rates.1 Vast Serverless charges the underlying instance costs for compute, storage and bandwidth without a separate pricing tier.3 Compare the actual worker or instance usage for the same serving workload before choosing on price.
Sources checked
Company-owned pages establish what a company says. They do not prove a market conclusion. Each source is dated so readers can judge each claim.
