Vast.ai Pricing: GPU Rates, Storage and Bandwidth
Vast.ai pricing verdict: Vast.ai pricing is a live marketplace, not a rate card: hosts set GPU, storage and bandwidth prices per offer, and Vast publishes no fixed H100 or RTX 4090 price. In Vast’s official search API at 09:50 Europe/Paris on November 10, 2026, single verified on-demand GPUs had median asks of $0.229 per hour (RTX 3090), $0.535 (RTX 4090), $1.069 (A100 SXM4) and $3.261 (H100 SXM). Your bill is that hourly rate plus storage billed while the instance exists and bandwidth billed per byte.
Ideal for: buyers who can re-check offers before each rental, jobs that tolerate picking a host by price and reliability, and checkpointed batch work that can use interruptible bids.
Avoid if: you need a quote that stays fixed for a budget cycle, you will leave stopped instances parked for weeks (storage keeps billing), or you move many terabytes per month without checking each host’s bandwidth rate.
- Spread beats list price: at 09:50 the cheapest verified H100 SXM ask was $2.202 per hour against a $3.261 median — a 48% gap inside one marketplace.
- Interruptible gap is smaller on consumer cards than the “50%+” headline: median minimum bids sat 26–30% below median on-demand asks for RTX 3090, 4090 and 5090 in the same snapshot.
- Reserved discounts are host-dependent: docs say “up to 50%”, the API prepay reference says “typically 40%”, and the highest host discount field we saw was 0.4.
By the Unfair Advantage editorial team · Updated November 10, 2026. Method: unauthenticated queries to Vast.ai’s documented search offers endpoint (console.vast.ai/api/v0/bundles) at several timestamps on November 10, 2026 (Europe/Paris), filtered to verified, rentable, unrented, single-GPU offers, for on-demand and interruptible (bid) pricing; plus Vast.ai’s pricing page and documentation (pricing, pricing overview, instance types, reserved instances, find and rent, billing, billing FAQ, serverless pricing, search offers reference, API index) fetched the same morning. We did not create an account, rent an instance, place a bid or run a benchmark. Medians are our arithmetic on the offers the API returned (at most 64 per query); they are observations with a timestamp, not prices Vast.ai promises. For whether Vast.ai is the right platform at all, read our live Vast.ai review on cheap GPUs and host trade-offs; this page only covers what you pay.
How does Vast.ai pricing work?
Vast.ai pricing has three meters: GPU compute billed per second while the instance is active, storage billed continuously while the instance exists, and bandwidth billed per byte sent or received. The pricing overview states that Vast.ai “is a marketplace, not a fixed-price provider” and that hosts set their own rates.
- Compute: the host’s hourly ask for the GPU slot, charged by the second. Docs list GPU model, GPU count, host reliability, location and market conditions as drivers.
- Storage: a host-set rate in dollars per GB per month on the disk you allocate at creation. The default disk is 10 GB and it cannot be resized after creation (Find & Rent docs).
- Bandwidth: host-set upload and download rates, quoted in the API in dollars per GB. The offer card shows them only on hover.
The price on an offer card is “the GPU rental price plus the hourly cost of the storage allocated,” with bandwidth shown separately on hover (Find & Rent docs). When you rent, a contract is created on the offer’s current terms, so the rate you accept is the one you pay for that rental; the next renter may see a different price. Vast.ai Serverless adds no surcharge: you pay the underlying worker compute, storage and bandwidth (serverless pricing docs).
Which Vast.ai pricing model fits your job?
Pick the rental type from how much interruption your job can absorb, then pick the host from price, reliability and bandwidth rate.
On-demand or reserved when…
- You serve inference or run a training job that cannot pause: on-demand and reserved are “high priority, never interrupted” (instance types docs).
- You will keep the same host for weeks: reserved prepays credits on that instance for a discount.
- You need a known end date: each offer card shows a maximum rental duration.
Interruptible when…
- Your job checkpoints to disk or cloud storage and resumes cleanly.
- You accept that a higher bid or any on-demand renter pauses you.
- You price the bid, not the floor: the API’s
min_bidis the minimum to bid, not a guarantee of running.
What do Vast.ai GPUs cost per hour today?
Vast.ai GPU rates on November 10, 2026 ranged from about $0.12 per hour for the cheapest verified RTX 3090 to $12.51 for a single B200 offer, in the official search API. The table shows medians per model across our snapshots; a single median per model is not a quote, and offer counts for datacenter cards were small.
| GPU (1×, verified) | On-demand median $/h | Lowest on-demand ask $/h | Interruptible min-bid median $/h | On-demand offers |
|---|---|---|---|---|
| H100 SXM | $2.564–$3.261 | $2.138 | $1.000–$1.353 | 6–7 |
| H100 NVL | $2.562 | $2.562 | no offer | 0–1 |
| H200 | $4.344–$4.804 | $4.344 | $1.316–$2.105 | 1–2 |
| B200 | $10.006–$12.510 | $7.502 | $7.969–$10.000 | 1–2 |
| A100 SXM4 | $1.069–$1.086 | $0.376 | $0.467–$0.700 | 9–10 |
| A100 PCIE | $0.610–$0.736 | $0.403 | $0.301–$0.484 | 7 |
| L40S | $0.801 | $0.429 | $0.533 | 5 |
| RTX PRO 6000 WS | $1.474–$1.607 | $1.309 | $1.267–$1.333 | 13–14 |
| RTX 5090 | $0.716–$0.746 | $0.410 | $0.533 | 47–51 |
| RTX 4090 | $0.535–$0.539 | $0.367 | $0.357–$0.387 | 36–40 |
| RTX 3090 | $0.229–$0.263 | $0.116 | $0.160 | 28–30 |
Snapshots: 09:50, 09:52, 09:55 Europe/Paris, November 10, 2026, Vast.ai search offers API. Ranges show the lowest and highest median across snapshots. RTX PRO 6000 WS is the workstation edition as named in the API.
How to read it: “On-demand median” is the median dph_total (the GPU rate plus the storage cost the API bundles into the offer price) of verified, rentable, single-GPU offers. “Interruptible min-bid median” is the median min_bid on bid offers for the same filters. “Lowest ask” is the cheapest on-demand offer seen in any snapshot. Offer counts are the on-demand offers returned per snapshot. Storage and bandwidth rates are covered in their own sections below.
Two minutes can move a median. Between 09:50 and 09:52 Europe/Paris, the H100 SXM on-demand median dropped from $3.261 to $2.564 while the lowest ask stayed at $2.202, because the set of unrented offers changed; by 09:55 a $2.138 ask had appeared. That is why the Vast.ai docs tell buyers to “always query current rates rather than relying on a quote.”
How do on-demand, interruptible and reserved prices compare?
Interruptible is the cheapest Vast.ai rental type, reserved sits between, and on-demand is the reference price. Vast.ai’s headline figures are “often 50%+ cheaper” for interruptible and “up to 50%” off for reserved; the observed gaps on November 10, 2026 depended heavily on the GPU model.
| Type | Official rule | Observed Nov 10, 09:50 | Use it for |
|---|---|---|---|
| On-demand | Host-set price, guaranteed resources, high priority | RTX 4090 median $0.535/h; H100 SXM median $3.261/h | Production, time-critical jobs |
| Interruptible | Bidding; paused if outbid or if on-demand is requested; “often 50%+ cheaper” | Median min-bid 30% below on-demand (RTX 4090), 26% (RTX 5090), 30% (RTX 3090), 69% (H100 SXM, 5 offers) | Checkpointed batch, hyperparameter search |
| Reserved | Prepay on an existing on-demand instance; “up to 50%”; API reference “typically 40%” | Highest host discount field 0.4; 31% of RTX 4090 offers listed any discount at 09:52 | Multi-week steady workloads on one host |
The interruptible gap above compares two different offer sets (bid offers versus on-demand offers) and uses the minimum bid, so treat it as a floor on savings, not a forecast. The CLI reference describes the default reserved discount function as 0.2 for one month and 0.3 for three months of prepaid usage.
Reserved refund rule, quoted from the docs: early cancellation claws back the discount already received plus a 5% fee. Vast.ai’s own example: on-demand $1 per hour is $720 per month; a one-month reservation costs $576; cancelling after 15 days refunds about $212. Prepaid credits are locked to that instance and cannot migrate between hosts; if the host is deverified for failing its service level, the full balance can be withdrawn without penalty. Adding credits mid-term recalculates the discount and can lower it.
How much does Vast.ai storage cost?
Vast.ai storage costs a host-set rate per GB per month, billed every second the instance exists, including while it is stopped. At 09:50 Europe/Paris on November 10, 2026, sampled medians ranged from $0.149 per GB per month (H200 offers) to $0.937 (the single B200 offer), with most consumer and A100 offers at $0.20–$0.27.
- A stopped instance is not free: 200 GB at the $0.30 H100 SXM median from that snapshot is $60 per month before any GPU time.
- Stopped can cost more: the instance pricing docs say storage costs “are typically higher for stopped instances than running instances.”
- Size once: disk size is fixed at creation; over-allocating is paid every hour, under-allocating means a new instance.
- Delete to stop the meter: only deleting the instance ends storage billing.
How much does Vast.ai bandwidth cost?
Vast.ai bandwidth is billed per byte, both directions, at a rate each host sets. The search offers reference defines inet_up_cost and inet_down_cost in dollars per GB. At 09:50 Europe/Paris on November 10, 2026, median upload rates by model ranged from about $3.43 per TB (H200 offers) to $37.50 (B200), with RTX 4090 at $12.29 and A100 PCIe at $4.00. The highest single rate we saw was $102.40 per TB on an H100 NVL offer.
Bandwidth rarely matters for a short fine-tune; it decides the bill for dataset pulls, model downloads repeated on fresh instances, and inference that streams media. Filter by bandwidth cost in the search interface, or hover the price on the offer card, before you choose a host for a data-heavy job. The billing FAQ adds that bandwidth is charged “regardless of what state” the instance is in.
How do you estimate a monthly Vast.ai bill?
Estimate a Vast.ai bill as GPU hours × hourly ask + allocated GB × storage rate × months the instance exists + TB transferred × bandwidth rate. The worked examples below use the 09:50 Europe/Paris snapshot of November 10, 2026; swap in the offer you pick.
| Scenario | Inputs (median offer) | Formula | Estimate |
|---|---|---|---|
| RTX 4090, always on for 30 days | $0.535/h; 100 GB at $0.267; 0.5 TB egress at $12.29 | 720 × 0.535 + 100 × 0.267 + 0.5 × 12.29 | ≈ $418 / month |
| H100 SXM, 200 GPU-hours | median $3.261/h vs lowest $2.202/h | 200 × rate | $652 at median, $440 at the lowest ask |
| Parked H100 SXM instance | 200 GB at $0.30, stopped all month | 200 × 0.30 | $60 / month, no GPU time |
| RTX 3090 always on | $0.229/h median vs $0.116/h lowest | 720 × rate | $165 at median, $84 at the lowest ask |
Assumptions: 720 hours per month, storage rate as listed on the offer while running (docs warn the stopped rate may be higher), egress at the upload rate, taxes not included. Vast.ai’s pricing page also links a pricing calculator; we did not use it, because it loads in the browser and our figures come from the API.
How do Vast.ai credits and payments work?
Vast.ai is prepaid: you add credits before starting any instance, and the balance updates about every few seconds. The billing guide lists card payments through Stripe and crypto payments through BitPay and Crypto.com; autobilling works with a saved card only.
- At $0.00: instances stop automatically and storage keeps billing, even on a negative balance.
- With a saved card: the card is charged automatically to cover the negative balance, which protects instances from deletion.
- Without a saved card: instances and stored data are scheduled for deletion; accounts with usage history get a grace buffer sized on average daily spend.
- Not charged: instance compute while a standard instance is loading (Find & Rent docs; Serverless workers in the Loading state are billed), and active rental or storage on instances that are offline (billing FAQ).
The billing guide recommends setting the autobilling threshold near your daily or weekly spend and a low-balance email at about 75% of that threshold.
Where do Vast.ai’s official pricing pages disagree?
Four official statements checked on November 10, 2026 do not line up. This page follows the more specific document in each case and tells you to confirm in the console.
- Maximum reserved discount: the pricing page, pricing docs and reserved-instances guide say “up to 50%”; the API index entry for prepay instance says “Maximum discount is typically 40%.” The highest host discount field in our API sample was 0.4.
- Lock-in: the pricing page says “No long-term contracts required… switch GPU types anytime without penalties,” while the reserved-instances guide charges a discount clawback plus a 5% fee on early cancellation. No penalty applies to on-demand; one applies to reserved.
- How to reserve: the pricing page’s Reserved card shows “Contact Sales” and “1, 3, or 6 month terms”; the docs describe self-serve conversion through the instance discount badge or the CLI prepay command with any amount.
- Zero balance: the billing guide says instances, volumes and data “will be scheduled for deletion” unless you add credits; the billing FAQ says instances are “stopped automatically, but not destroyed.” The instance pricing page reconciles them by payment method on file — plan for the stricter reading.
What pricing pitfalls should Vast.ai buyers watch?
- Quoting a median as a price: medians moved by $0.70 per hour on H100 SXM in two minutes; screenshot the offer you rent.
- Parked instances: stopped instances keep billing storage; delete what you no longer need.
- Bandwidth outliers: one offer listed $102.40 per TB upload; filter or hover before data-heavy jobs.
- Bid floor optimism:
min_bidis the entry price, not the winning bid when demand rises. - Topping up a reservation in small amounts: each credit addition recalculates the discount and can lower it.
- Unverified hosts: they are hidden by default for a reason; our medians cover verified offers only.
- Expired rentals: expired on-demand instances may be deleted 48 hours after expiration and cannot restart while expired.
Is Vast.ai pricing a good deal for your workload?
| Buyer | Verdict | Why |
|---|---|---|
| Hobby or research user on consumer GPUs | Yes — on-demand or interruptible | RTX 3090/4090 medians $0.229–$0.535 per hour at 09:50 |
| Checkpointed batch job | Yes — interruptible | Median min-bids 26–30% below on-demand on consumer cards |
| Steady multi-week training on one host | Yes — reserved, after testing the host | Discount up to 0.4 seen; credits locked to the host; 5% fee on early exit |
| Team needing a fixed annual GPU budget | Weak fit | No fixed rate card; prices move by the minute |
| Data-heavy pipeline | Only after filtering bandwidth | Upload rates from about $3.43 to $102.40 per TB in one snapshot |
Key takeaways
- Vast.ai pricing is set per offer by hosts, so Vast.ai publishes no fixed GPU price list; the docs tell buyers to query current rates instead of relying on a quote (pricing overview, November 10, 2026).
- A single verified on-demand RTX 4090 on Vast.ai had a median ask of $0.535 per hour at 09:50 Europe/Paris on November 10, 2026, in Vast.ai’s official search API.
- The cheapest verified single H100 SXM on Vast.ai was $2.202 per hour at 09:50 Europe/Paris on November 10, 2026, 48% below that snapshot’s $3.261 median.
- Vast.ai storage bills while an instance exists, including when stopped; only deleting the instance stops it (Vast.ai billing FAQ, checked November 10, 2026).
- Vast.ai reserved discounts are described as “up to 50%” in guides but “typically 40%” in the API prepay reference, and early cancellation claws back the discount plus a 5% fee (docs checked November 10, 2026).
- Vast.ai accepts card payments through Stripe and crypto through BitPay and Crypto.com, with autobilling limited to saved cards (billing guide, November 10, 2026).
Our verdict: Vast.ai pricing rewards buyers who shop each rental: compare the lowest verified ask with the median, size disk once, check bandwidth, and delete instances you stop using. If you need a fixed rate card, this pricing model is the wrong fit regardless of how low today’s floor looks. For platform trade-offs beyond price, see our Vast.ai review.
FAQ
How much does Vast.ai cost per hour?
There is no fixed Vast.ai rate, because hosts set their own prices. In the official search API on November 10, 2026 at 09:50 Europe/Paris, single verified on-demand GPUs had median asks of $0.229 per hour for an RTX 3090, $0.535 for an RTX 4090, $1.069 for an A100 SXM4 and $3.261 for an H100 SXM. Re-query before you rent.
Why do Vast.ai prices vary so much?
Prices vary because each host sets its own rate. Vast.ai docs name supply and demand in the host’s region, operating costs, machine specs, reliability and competition as the drivers. The spread is wide: at 09:50 Europe/Paris on November 10, 2026, the cheapest verified single H100 SXM ask was $2.202 per hour while the median was $3.261.
How much does Vast.ai storage cost?
Storage is host-priced in dollars per GB per month and runs while the instance exists, even when it is stopped. Across the verified on-demand offers sampled at 09:50 Europe/Paris on November 10, 2026, medians sat between $0.149 and $0.937 per GB per month depending on the GPU model. Only deleting the instance stops storage billing, per Vast.ai docs.
Does Vast.ai charge for bandwidth?
Yes. Vast.ai bills bandwidth per byte for both upload and download, at rates each host sets. In the official API at 09:50 Europe/Paris on November 10, 2026, median upload rates on sampled single-GPU offers ranged from about $3.43 to $37.50 per TB by model, and one H100 NVL offer listed $102.40 per TB. Check the hover breakdown before data-heavy jobs.
How do Vast.ai reserved discounts work?
You rent on-demand first, then prepay credits on that instance for a lower rate. Docs say up to 50% off, while the API prepay reference says the maximum discount is typically 40%. Early cancellation claws back the discount received plus a 5% fee, and credits stay locked to that host (Vast.ai docs, checked November 10, 2026).
What happens when my Vast.ai balance hits zero?
Your instances stop and storage keeps billing. The billing guide says instances, volumes and data are then scheduled for deletion unless you top up, with a grace period sized on average daily spend; a saved card is charged automatically. The billing FAQ says instances are stopped but not destroyed. Keep a positive balance (docs checked November 10, 2026).
Did Unfair Advantage rent GPUs for this page?
No. We did not open an account, rent an instance or run a benchmark. Rates come from unauthenticated queries to Vast.ai’s documented search offers endpoint on November 10, 2026 (Europe/Paris), and rules come from Vast.ai’s pricing page and documentation fetched the same day. Prices move by the minute, so treat every figure as a dated observation.
Sources
Sources are listed as plain URLs, without links, so this page sends no clicks to vendors other than through the disclosed affiliate button. Times are Europe/Paris on November 10, 2026.
- Vast.ai search offers endpoint, https://console.vast.ai/api/v0/bundles/ (unauthenticated queries, verified + rentable + unrented + 1 GPU, on-demand and bid; snapshots 09:50, 09:52, 09:55)
- Vast.ai pricing page, https://vast.ai/pricing (instance type cards, “No Lock-In” wording; checked 09:49)
- Vast.ai pricing overview, https://docs.vast.ai/guides/pricing.md (checked 09:51)
- Vast.ai instance pricing, https://docs.vast.ai/guides/instances/pricing.md (checked 09:50)
- Vast.ai instance types, https://docs.vast.ai/guides/instances/choosing/instance-types.md (checked 09:50)
- Vast.ai reserved instances, https://docs.vast.ai/guides/instances/choosing/reserved-instances.md (refund example, 5% fee; checked 09:50)
- Vast.ai finding and renting instances, https://docs.vast.ai/guides/instances/choosing/find-and-rent.md (offer card price, disk size, tiers; checked 09:50)
- Vast.ai billing guide, https://docs.vast.ai/guides/reference/billing.md and billing FAQ, https://docs.vast.ai/guides/reference/faq/billing.md (checked 09:51)
- Vast.ai serverless pricing, https://docs.vast.ai/guides/serverless/pricing.md (checked 09:51)
- Vast.ai search offers CLI reference, https://docs.vast.ai/cli/reference/search-offers.md (field units; checked 09:51)
- Vast.ai documentation index, https://docs.vast.ai/llms.txt (prepay instance entry, “typically 40%”; checked 09:49)