Vast AI pricing: Vast.ai September 2026
Use Vast.ai's reported GPU median as a budget reference, then price the chosen host's compute, storage and transfers. The 28 September 2026 price snapshot is not a current instance quote; billing documentation was checked separately on 1 October 2026.
Vast.ai’s reported GPU-hour medians were $0.19 for RTX 3090, $0.51 for RTX 4090 and $2.16 for H100 SXM on its pricing page captured 28 September 2026.1 The rental bill also includes storage and bandwidth under official documentation captured 1 October 2026.2 Use the median as a budget reference, then check the chosen host’s complete price.
TL;DR
- A starting price and a median describe different points in Vast.ai’s reported offers; the captured cards don’t provide a full price range.1
- Vast.ai distinguishes on-demand, reserved and interruptible rentals, with different commitment and interruption terms.2 Choose the rental form before comparing quotes.
- Stopping an instance leaves storage charges running; Vast.ai’s billing documentation treats offline instances separately.3
- Hosts set bandwidth rates, and charges cover both upload and download traffic.2 Budget transfers alongside compute.
How were the prices checked?
The GPU figures come from Vast.ai’s pricing page captured 28 September 2026.1 Official Vast instance-pricing and payment-billing documentation was captured separately on 1 October 2026.2,3 The pricing page reports available offers and says prices update hourly.1 Dollar signs below preserve the source’s notation; the captures do not specify a currency code or tax treatment.1,3
The table retains the source fields from and median, without combining them into a range.1 It is a selected model-by-model comparison, with consumer cards and enterprise GPU variants kept separate. Our research methodology explains the source-bound approach.
What does Vast.ai cost by GPU?
Vast.ai reported an RTX 4090 starting price of $0.20 per GPU-hour and a median of $0.51 in the 28 September capture.1 Those are the card’s from and median values, not the lowest and highest prices.1 Compare the same GPU variant and rental form when requesting a current quote.
| GPU model | From ($/GPU-hour, as displayed) | Median ($/GPU-hour, as displayed) |
|---|---|---|
| RTX 3060 | $0.03 | $0.071 |
| RTX 3090 | $0.11 | $0.191 |
| RTX 4090 | $0.20 | $0.511 |
| RTX 5090 | $0.33 | $0.631 |
| RTX 5060 Ti | $0.09 | $0.171 |
| RTX 5080 | $0.20 | $0.331 |
| L4 | $0.16 | $0.321 |
| L40 | $0.33 | $0.571 |
| L40S | $0.47 | $0.791 |
| RTX A6000 | $0.29 | $0.521 |
| RTX PRO 6000 WS | $1.00 | $1.391 |
| A100 PCIE | $0.47 | $0.741 |
| A100 SXM4 | $0.31 | $0.991 |
| H100 SXM | $1.73 | $2.161 |
| H100 PCIE | $1.87 | $2.531 |
| H100 NVL | $2.27 | $2.761 |
| H200 | $2.63 | $4.301 |
| H200 NVL | $2.67 | $4.001 |
| B200 | $9.38 | $10.631 |
| B300 | $8.75 | $8.751 |
B300’s starting price and median were both $8.75, while B200’s reported median was $10.63.1 That ordering alone gives no performance or availability comparison.1 For quotes beyond Vast.ai, use the H100 rental-price comparison or B200 rental-price comparison.
How does Vast.ai pricing work?
Vast.ai is a marketplace where hosts set their own prices, and the total rental cost combines compute, storage and bandwidth.2 Its instance-pricing documentation says rates vary with supply and demand, machine specifications, location and host reliability.2 Check the selected offer’s terms before treating a reported median as a rental quote.
The documentation captured on 1 October distinguishes these rental forms:2
| Rental form | Pricing and access described by Vast.ai | Budget decision |
|---|---|---|
| On-demand | Fixed pricing with high-priority, guaranteed resources.2 | Check the chosen instance’s rate and configuration. |
| Reserved | Discounted rates with prepayment and high-priority access.2 | Confirm the commitment before paying upfront. |
| Interruptible | Low-priority access at the lowest cost; instances may be paused.2 | Accept it only if a pause is tolerable. |
“Fixed pricing” for on-demand describes that rental form within a host-priced marketplace.2 It doesn’t establish a single platform-wide rate for every machine with the same GPU.2
Vast.ai’s active rental charge accrues for every second in the active/connected state.3 Check whether a displayed offer is priced for the whole instance or per GPU before applying it to a multi-GPU budget.
What does Vast.ai charge for storage?
Vast.ai charges for allocated storage while an instance is stopped, and storage rates vary by host.2,3 Its payment-billing documentation lists storage in $/GB/hr and ties the charge to the size of the storage allocation.3 Budget the allocation and how long you retain it, including stopped time.
The payment-billing documentation gives the more specific state distinction:3
| Instance state | Active rental charge | Instance storage charge |
|---|---|---|
| Active/connected | Charged per second.3 | Charged per second while online.3 |
| Stopped but online | No active rental charge while stopped.3 | Continues despite the stopped state.3 |
| Offline | No active rental charge.3 | No instance storage charge.3 |
Stopped and offline therefore have different billing treatment.3 The instance-pricing guide also says storage is typically more expensive for stopped instances than running ones.2 Confirm both rates if you plan to pause a rental between jobs.
Vast.ai instructs users to delete instances completely to cease instance storage billing.2 Copy off any data you need before deleting an instance.
What does Vast.ai charge for bandwidth?
Vast.ai’s bandwidth rates vary by host, and charges apply to both upload and download traffic per byte transferred.2 Its payment-billing documentation displays bandwidth prices in $/TB and says transfers are chargeable regardless of instance state.3 Check both directions when selecting a host.
Treat dataset uploads, model downloads and exported results as separate transfer requirements in your budget. A blanket assumption of free inbound or outbound traffic conflicts with Vast.ai’s documented host-specific charging model.2,3 Use the actual offer’s transfer rates before deciding whether a lower compute price saves money overall.
Vast.ai says hovering over the price on the console’s Search or Instance page reveals the pricing breakdown.3 Inspect that breakdown before renting and when reviewing an existing instance.
How should you budget a Vast.ai rental?
Glamdring Research recommends budgeting the chosen Vast.ai offer’s compute, storage and bandwidth together.3 Vast.ai documents separate charges for those components, with storage continuing when an online instance is stopped.3 Compare the full bill for your intended rental lifecycle, including retained data and transfers.
Use the offer’s actual rates in this cost model: active instance rate multiplied by active time, plus storage allocation multiplied by its rate and billable time, plus uploaded and downloaded data multiplied by their respective transfer rates.3 The components and billing periods follow Vast.ai’s payment-billing documentation.3
For a rental you run, stop and later delete, compute accrues during active/connected time and storage continues through stopped online time.3 Transfers remain chargeable per byte, even outside the active state.3 Before accepting the quote:
- Match the GPU variant and GPU count to the workload.
- Record the active rental rate and rental form.
- Check allocated storage and the running and stopped storage rates.
- Estimate upload and download volumes separately.
- Confirm currency, tax treatment and the available configuration.
- Plan the data export and instance deletion.
Vast.ai requires prepaid credits, and saved-card autobilling can replenish the balance.3 Include the credit balance in your operating plan as well as the hourly rate.
Subscribe to our research for new compute and infrastructure analysis.
What the price snapshot cannot tell you
Vast.ai’s 28 September 2026 price capture records the company’s reported offers at that time; its pricing page says those offers update hourly.1 Use a current instance quote for a purchase decision. The table supplies price references, without a performance test or an uptime assessment.
The captured pricing cards do not explain how Vast.ai calculates the displayed median or provide the complete spread of offers for each model.1 They also do not identify a rental form for each price card.1 Keep those limits attached to any comparison with another provider’s quote.
Further compute infrastructure research covers the wider buying decision.
Frequently asked questions
What happens if my Vast.ai credits run out?
Vast.ai says instances stop automatically when the balance reaches zero, while stopped storage continues to accrue charges.3 Its billing documentation describes a grace period based on account spending history, without a fixed duration.3 A saved card can be charged to cover a negative balance; without a saved card, instances and stored data face deletion unless you restore the balance.3 Keep a positive balance if you need to retain the instance and its data.
Can I get a refund for unused credits?
Vast.ai says spent credits are non-refundable, while unused credits can usually be refunded on request through its website chat.3 The documentation captured 1 October 2026 excludes refunds for credits bought through legacy Coinbase Commerce.3 Confirm the payment route before adding credits you may not use.
Does Vast.ai serverless have an extra platform fee?
Vast.ai’s instance-pricing documentation captured 1 October 2026 says serverless has no separate pricing tier or additional fee: users pay the underlying instance compute, storage and bandwidth costs.2 Budget those components even when serverless manages instance scaling.
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.
