Vast.ai reviews: prices and limits, September 2026
Assess the specific Vast.ai offer and the whole-job cost. Published rates and review aggregates do not establish the selected host's performance; interruption, idle storage and credit-balance rules change the rental decision.
Vast.ai warrants an offer-level comparison: its published prices vary by host, and the GPU rate leaves storage and bandwidth costs to account for separately.1 The review record, checked 28 September 2026, showed Trustpilot at 3.9/5 from 256 reviews and G2 with no reviews.2,3 Neither establishes performance for the machine you intend to rent.
No first-hand use. GPU prices and ratings retain the September checked date; the billing terms below were captured on 1 October 2026.
TL;DR
- Compare the published
fromprice with themedian: the RTX 4090 card showed $0.20/hr and $0.51/hr respectively.4 A budget based only on the minimum needs a matching offer. - G2’s displayed
0/5(0)accompanies an empty review sample. Treat it as missing customer evidence.3 Trustpilot’s aggregate includes both five-star and one-star reviews.2 Keep the sample and checked date beside the rating. - Vast.ai describes interruptible rentals as preemptible and says they may be reclaimed.4 Choose that tier only when the job can tolerate interruption.
- The later billing documentation says stopped instances continue to incur storage charges; offline instances are exempt from active rental and storage charges.5 Budget for the instance’s state as well as its runtime.
How does Vast.ai pricing work?
Vast.ai uses a GPU marketplace where hosts set prices; its documentation says the total includes compute, storage and bandwidth.1 Compare the exact GPU variant and the selected offer’s charges when assessing compute infrastructure options.
These are selected cards from Vast.ai’s published pricing, checked 28 September 2026.4 Figures retain the source’s $ notation; the captured text does not specify a currency code or tax treatment.4
| GPU, as listed | Listed VRAM | Published from ($/hr) | Published median ($/hr) |
|---|---|---|---|
| RTX 3090 | 24GB | 0.11 | 0.194 |
| RTX 4090 | 24GB | 0.20 | 0.514 |
| RTX 5090 | 32GB | 0.33 | 0.634 |
| A100 SXM4 | 80GB | 0.31 | 0.994 |
| H100 SXM | 80GB | 1.73 | 2.164 |
| H100 PCIE | 80GB | 1.87 | 2.534 |
| H200 | 141GB | 2.63 | 4.304 |
| B200 | 192GB | 9.38 | 10.634 |
The cards show why the cheapest displayed offer and the median deserve separate columns: Vast.ai listed an A100 SXM4 from $0.31/hr alongside a $0.99/hr median.4 Preserve both values when comparing offers.
GPU names also need their suffixes. Vast.ai listed separate H100 SXM, H100 NVL and H100 PCIE cards, each with its own rates.4 Compare the named variant in the H100 rental-price comparison; use the B200 rental-price comparison for a separate hardware decision.
Vast.ai says its displayed prices reflect currently available offers and update hourly.4 Treat the September figures as a dated comparison, then check the current offer. They do not reserve a machine or establish its performance.
What do the Trustpilot and G2 ratings establish?
Vast.ai’s Trustpilot record provides a dated customer-opinion aggregate, while its G2 profile had no reviews on 28 September 2026.2,3 Keep those evidence types separate when deciding how much confidence to place in public reviews.
| Source | Displayed record, checked 28 September 2026 | Interpretation |
|---|---|---|
| Trustpilot | TrustScore 3.9/5; 256 reviews2 | A published opinion aggregate, with the sample size attached |
| G2 | 0/5(0); “See all 0 Vast AI reviews”3 | No customer-review sample to interpret |
Trustpilot displayed 78% five-star, 4% four-star, 1% three-star, 2% two-star and 15% one-star reviews.2 Reporting only the five-star share would conceal the negative responses; reporting only the one-star share would conceal the majority’s positive ratings.
The profile also said Vast.ai invites customers to review, whether positive or negative.2 Trustpilot states that it screens reviews but does not fact-check their specific claims.2 Read the aggregate as opinion evidence. A rating cannot substitute for a test of your workload on your chosen host.
G2’s empty sample warrants particular care. Its displayed zero supplies no basis for describing customer satisfaction as zero.3 The distinction follows our research methodology: retain the source’s figure and explain what it can establish.
Which rental terms limit the lowest prices?
Vast.ai’s interruptible tier permits the host capacity to be reclaimed, so the advertised price advantage comes with a continuity constraint.4 A job that cannot resume after interruption needs a different rental decision.
The pricing page, checked 28 September 2026, distinguished the following terms.4 These are Vast.ai’s published descriptions, not independently measured uptime results.
| Rental type | Published terms | Decision consequence |
|---|---|---|
| On-demand | Per-second billing; “No interruptions”; “Guaranteed uptime”4 | Confirm the applicable contract before relying on a continuity claim |
| Interruptible | “50%+ cheaper”; preemptible and may be reclaimed; described for fault-tolerant workloads4 | Plan how to preserve work and resume it |
| Reserved | “Up to 50% Off”; long-term commitment; 1, 3 or 6 month terms; guaranteed capacity4 | Assess the commitment as well as the hourly price |
The discount language is conditional. “50%+ cheaper” and “Up to 50% Off” are company statements on the pricing page, not a discount calculated against matched offers in this review.4 Request the relevant offer and its terms before assigning either saving to a budget.
What can keep the bill running after compute stops?
Vast.ai’s billing documentation, captured on 1 October 2026, says storage charges continue for stopped online instances, while offline instances incur neither active rental nor storage charges.5 Include the instance state when estimating idle costs.
The reference documentation separates active rental in $/hr, storage in $/GB/hr and bandwidth in $/TB.5 It charges active rental by the second in the active/connected state and bandwidth for bytes sent or received regardless of instance state.5 Rates vary from machine to machine.5 Inspect the selected offer’s breakdown before uploading data or leaving storage allocated.
The instance-pricing guide says deleting an instance ends its storage billing.1 An offline instance falls under the separate exemption in the detailed billing reference.5 Keep the stopped, offline and deleted states distinct when estimating costs.
Credits add another limit. Vast.ai requires prepayment and says instances stop automatically when the balance reaches zero or below.5 Its documentation says data and storage volumes are scheduled for deletion unless credits are added; an account-history-based grace period can allow a negative balance before deletion.5 It specifies no universal fixed duration for that buffer.5 A saved credit card may be charged automatically to cover the negative balance, while accounts without one face destruction of instances and stored data.5 Treat balance management and data retention as part of the rental decision.
What should you verify before renting?
Glamdring Research’s verdict is to assess a specific Vast.ai offer against the whole job: published GPU prices and review aggregates leave the selected machine’s performance unresolved. The host-set pricing model and separate compute, storage and bandwidth charges make offer-level checking necessary.1 Ratings add context within the limits of their samples.2,3
Before committing, use the published terms to define the checks:
- Match the GPU model and suffix to the required memory. The listed H100 variants and VRAM values differ from the H200 and B200 cards.4
- Read all cost components. Vast.ai says compute pricing varies with factors including GPU quantity, host reliability and location, while transfer charges cover uploads and downloads.1
- Choose rental continuity deliberately. An interruptible offer may be reclaimed; reserved pricing carries a commitment.4
- Check what happens after stopping and when credits run out. Storage billing and scheduled deletion remain relevant after compute stops.5
Use the cloud GPU provider comparison to frame the alternative offers. A representative workload test should establish whether the selected machine’s transfer behaviour and completed-job cost meet your requirements. Let those results determine the rental decision.
Subscribe to our research for new evidence reviews.
Frequently asked questions
Why is Vast.ai so cheap?
Vast.ai attributes its pricing to competition between hosts in a marketplace, with rates determined by supply and demand.1 The September pricing page also advertised lower interruptible rates in exchange for reclaimable capacity.4 Those explanations support comparing offers; they do not establish that Vast.ai is the cheapest provider for every workload.
Can you get a refund for unused credits?
Vast.ai’s refund documentation, captured on 1 October 2026, says customers can request a refund of unspent credits through website chat and that it can refund them in most cases.5 It states there are no refunds after credits are spent and excludes legacy Coinbase Commerce purchases from refunds.5 Refundability therefore depends on the credits and payment method, not simply whether a rental disappointed you.
Does a listed GPU always have a current offer?
No. Vast.ai’s GB10 card, checked 28 September 2026, displayed an indicative recent-market price of approximately $0.31/hr and explicitly said “No current offers” and “indicative price, not a live offer”.4 A listed model and a rentable offer require separate checks.
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.
