Vast.ai Review: Cheap GPUs and Host Trade-offs
Vast.ai review verdict: Vast.ai is worth using when cost per GPU-hour is your bottleneck and your team can pick a host from an offer card, because hosts set their own prices: one on-demand H100 SXM had a median ask of $2.27/hr on November 9, 2026 (third-party tracker of Vast’s public offers API), against $3.99/hr on RunPod Secure Cloud.
Ideal for: ML teams running Docker fine-tunes, hyperparameter sweeps or batch jobs that checkpoint, and that will filter hosts by reliability, disk and bandwidth.
Avoid if: procurement needs one fixed rate card and one operator SLA on every node, or you run always-on customer inference that nobody will monitor.
- Prices move within the day: the same tracker logged H100 SXM medians between $2.27 and $3.40 on November 9, 2026.
- Two invoice lines people miss: storage bills while an instance exists, even stopped, and bandwidth is host-priced per byte (Vast.ai pricing docs, checked November 9, 2026).
- Reserved is a host commitment: up to 50% off by prepaying, but credits are locked to one host and early cancellation claws back the discount plus a 5% fee.
By the Unfair Advantage editorial team · Updated November 9, 2026. Method: analysis of Vast.ai’s official pricing page and documentation (pricing, concepts, instance types, find and rent, reserved instances, billing), re-checked November 9, 2026 (Europe/Paris), plus a dated third-party snapshot of Vast’s public offers API. We did not rent a GPU or run a benchmark. Our own query to Vast’s public offers API from our server returned HTTP 403 on November 9, 2026, so we cite the tracker instead of computing medians ourselves.
Who is Vast.ai for, and who should skip it?
Vast.ai is for teams that already know they need GPU hours and will trade a uniform product for a lower hourly rate. It is the wrong buy when what you need is a contractual SLA rather than cheaper hours.
Worth it when…
- Your job runs in Docker and checkpoints, and you care more about $/GPU-hour than about one vendor SKU.
- You want H100, A100 or RTX 4090-class cards without hyperscaler list prices, and someone will compare offer cards (reliability, disk speed, bandwidth price, max duration).
- Interruptible or short on-demand bursts fit the work: sweeps, checkpointed fine-tunes, overnight batch jobs.
- You accept that the market is the price list: rates move during the day and the cheapest card is not always the right card.
Wrong buy when…
- Procurement needs a fixed published rate card, compliance paperwork from one operator, or a contractual SLA on every node.
- The workload cannot tolerate a pause, host churn or variable egress, such as customer-facing inference with hard latency targets.
- Nobody on the team can read an offer card (PCIe lanes, disk MB/s, upload and download price, a reliability score that starts at 60% on new machines).
- You needed web data for an LLM pipeline, not compute: start with our comparison of scraping APIs for LLMs.
If the GPU job sits inside a wider pipeline, orchestration is a separate purchase; our n8n review for workflow automation covers that layer. Vast.ai rents the compute; it does not run your workflow.
What do you rent on Vast.ai?
You rent an offer from a host, not a product from Vast. Vast.ai’s concepts docs (checked November 9, 2026) define hosts as anyone listing GPU hardware, from a single gaming PC to Tier 4 datacenters, each setting its own price and verification level.
| Term | What Vast.ai’s docs say (Nov 9, 2026) |
|---|---|
| Machine | One physical host system; it can publish several offers (GPU slices, storage, bandwidth). |
| Offer | A rentable configuration: GPU model, count and RAM; CPU, RAM, disk, bandwidth; price, max rental duration, location; DLPerf and reliability scores. |
| DLPerf | A Vast-defined score approximating deep learning throughput for a GPU and host pair, for comparing offers with different GPUs. |
| Instance | What you get when you accept an offer: almost always a Docker container (a few are virtual machines), billed per second while running plus storage while it exists. |
| Rental contract | The agreement for one instance, with a max duration shown on the offer card and a type: on-demand, reserved or interruptible. |
Vast.ai’s find-and-rent docs are explicit about the limit of that model: Vast is “only a marketplace” and does not manage or provide the hardware, so when no offer matches your filters there is no fallback capacity to call on. During demand surges, an offer can disappear between search and the Rent click.
Vast.ai’s pricing page (checked November 9, 2026) says prices are set by supply and demand across 40+ data centers, with per-second billing and 68+ GPU types from the RTX 3060 to the B200.
How much does Vast.ai cost per GPU-hour?
There is no fixed Vast.ai price: the docs say hosts set their own rates, and your bill is GPU compute plus storage plus bandwidth. The figures below are a dated third-party snapshot, not a quote.
Snapshot of Vast single-GPU on-demand asks (November 9, 2026)
The openpromt GPU price index queries Vast’s public API for one-GPU on-demand offers, excludes hosts with reliability below 0.9, and reads up to 64 offers per model. Values from its update of November 9, 2026, 18:25 UTC (19:25 Paris), next to RunPod’s published prices (page updated November 4, 2026, checked November 9):
| GPU | Vast median / low (offers) | RunPod Community / Secure | Read |
|---|---|---|---|
| H100 SXM | $2.27 / $2.20 (14) | $2.69 / $3.99 | Vast median under both RunPod tiers |
| H100 PCIe / NVL | $2.55 / $2.34 (16) | PCIe $1.99 / $2.89; NVL $2.59 / $3.19 | Mixed bucket: compare variant by variant |
| H200 | $4.34 / $3.57 (33) | $3.59 / $5.29 | Above RunPod Community, under Secure |
| A100 80GB | $1.07 / $0.84 (19) | PCIe $1.19, SXM $1.39 / $1.79 | Vast median under both RunPod tiers |
| RTX 4090 | $0.47 / $0.31 (64) | $0.34 / $0.89 | Above RunPod Community, under Secure |
| RTX 5090 | $0.54 / $0.37 (64) | $0.69 / $1.19 | Vast median under both RunPod tiers |
Two limits on that table. First, the tracker caps reads at 64 offers per model, so the RTX 4090 and 5090 counts are a ceiling, not the whole market. Second, the median moves within hours: the tracker’s change log for November 9, 2026 shows the H100 SXM median at $2.27, $2.34, $2.50, $2.87 and $3.40 at different points of the day.
How published fixed prices compare
RunPod’s pricing page, updated November 4, 2026 and checked November 9, lists one H100 SXM at $2.69/hr on Community Cloud and $3.99/hr on Secure Cloud, billed per second with no commitment.
Lambda’s pricing page, checked November 9, 2026, lists instance prices that take effect November 16, 2026 at 00:00 UTC, plus tax: $4.89/GPU-hr for a 1× H100 SXM and $3.79 for a 1× H100 PCIe. It says current prices until then are visible only after login, so we quote none.
Hour math for a 48-hour H100 SXM fine-tune
| Option | Rate used | 48 hours | Read |
|---|---|---|---|
| Vast, median at 18:25 UTC | $2.27 | ≈ $109 | Lowest line; tracker median, add host extras |
| Vast, day’s highest median | $3.40 | ≈ $163 | Same GPU, different hour |
| RunPod Community | $2.69 | ≈ $129 | Published price |
| RunPod Secure | $3.99 | ≈ $192 | Published price |
| Lambda 1× (from Nov 16) | $4.89 | ≈ $235 + tax | Published future price |
The Vast lines are tracker medians, not invoices. On an offer card, the headline price is the GPU rate plus the hourly cost of the allocated storage, and bandwidth is priced separately (Vast.ai find-and-rent docs, checked November 9, 2026). A job that pulls a large dataset and pushes checkpoints can erase the gap with a fixed-price provider.
Storage, bandwidth and credits
- Storage: billed continuously while the instance exists, running or stopped, and typically higher when stopped. Default disk is 10 GB and the size cannot be changed after creation. Delete the instance to stop the charge.
- Bandwidth: host-set rates for upload and download, charged per byte.
- Credits: prepaid. At a zero balance, instances stop; with a saved card the card is charged automatically; without one, instances and data are scheduled for deletion unless you add credits promptly. Payment options: cards through Stripe, crypto through BitPay and Crypto.com.
Should you pick on-demand, interruptible or reserved?
Pick on-demand for anything time-sensitive, interruptible only for checkpointed work, and reserved only once you trust a specific host for weeks. The type is fixed at creation, except that on-demand can be converted to reserved.
| Type | Priority and price | Catch | Best for |
|---|---|---|---|
| On-demand | High; fixed host price | Contract ends at the offer’s max duration; expired instances may be deleted 48 hours after expiry | Time-critical jobs, inference |
| Interruptible | Low (bid); often 50%+ cheaper | Paused when outbid or when on-demand claims the GPU; cannot be switched to on-demand | Sweeps, preprocessing, checkpointed training |
| Reserved | High; up to 50% off by prepaying | Credits locked to one instance and host; early exit costs the discount plus 5% | Steady multi-week usage on a host you trust |
What a reserved discount costs to undo
Vast.ai’s reserved-instances docs (checked November 9, 2026) give the worked example: an on-demand $1/hr instance is $720 a month; a one-month reservation costs $576; cancelling after 15 days refunds about $212 once the discount is clawed back and the 5% fee applied. The CLI documents a default discount of 0.2 for one month and 0.3 for three months, and the prepay API reference says the maximum discount is typically 40%, below the “up to 50%” headline.
Two more rules from the same page: adding small amounts of credit mid-term recalculates the discount and can lower it, and if the host is deverified you can withdraw the full balance without penalty.
How do verified hosts and Secure Cloud differ?
Secure Cloud is the only tier with a facility claim: Vast confirms those machines sit in a datacenter with a Tier 2/3 rating or ISO 27001. Vast.ai’s find-and-rent docs (checked November 9, 2026) list three tiers:
- Unverified: typically new machines not yet tested by Vast; hidden from search results by default.
- Verified: machines that passed Vast’s internal tests.
- Secure Cloud (Datacenter): verified and confirmed in a certified datacenter; blue label; recommended by Vast for production.
The reliability score measures a machine’s historical uptime and health, and every machine starts at 60%. Vast’s own pricing docs note that lower reliability scores may offer better rates, which is exactly the trade you are making when you sort by price alone.
Two filters matter more than the headline rate. Max duration: a 72-hour fine-tune on an offer with a 48-hour contract is an eviction risk, since hosts may extend or stop the instance when the contract ends. Multi-GPU topology: offer cards show PCIe version and lanes; eight cheap GPUs with slow interconnect can take longer, in wall-clock terms, than fewer GPUs on a better host.
How does a Vast.ai rental run, step by step?
The documented flow is credits, template, disk size, filters, offer card, rent, then delete when done. In order:
- Add credits. No instance starts without a balance.
- Choose a template (Docker image, launch mode, on-start script).
- Set disk size before you rent. Default 10 GB; it cannot be resized later and bills even when stopped.
- Filter by GPU model and count, verification tier, reliability, location, network speed and max duration; sort by price last.
- Hover the offer card price for the breakdown: the headline is GPU plus allocated storage; bandwidth is listed separately.
- Rent on-demand or interruptible. The type cannot be changed afterwards, except on-demand to reserved.
- Watch startup. The docs say it is usually under 5 minutes; past 20 minutes, try another machine. Loading time is not charged.
- Download artifacts, then delete the instance. Stopping is not deleting.
When should you choose RunPod, Lambda or a hyperscaler instead?
Choose a fixed-price provider when a predictable product matters more than the lowest hourly rate. Alternatives are named in text only, with no outbound vendor links.
- RunPod: published Community and Secure prices, plus Serverless and Clusters. Its Community tier was cheaper than the tracked Vast median for the RTX 4090 and H200 on November 9, 2026.
- Lambda: a classic instance menu with published prices; note the November 16, 2026 price change on its page.
- Hyperscalers (AWS, GCP, Azure): when IAM, VPC and procurement paperwork matter more than $/hr.
- Managed inference APIs: when you do not want to operate CUDA, Docker or model servers at all.
What are the common Vast.ai buying mistakes?
- Stopping instead of deleting. Storage keeps billing until the instance is deleted.
- Sorting by price only. Slow disk, expensive egress or a low reliability score cost wall-clock time and retries.
- Interruptible without checkpoints. A pause without a saved state loses the work.
- Ignoring max duration on long jobs; expired instances may be deleted 48 hours after expiry.
- Reserving before trusting the host. Reserved credits cannot move to another host.
- Topping up a reservation in small amounts, which recalculates and can lower the discount.
- Treating a tracker median as a quote. The H100 SXM median moved between $2.27 and $3.40 on November 9, 2026.
- Running out of credits without a saved card, which puts instances and data on a deletion schedule.
Decision: should you rent GPUs on Vast.ai?
Rent on Vast.ai when the job is elastic, checkpointed and cost-driven; pay a fixed price elsewhere when the job needs a guarantee. The table maps each situation to one choice.
| If your situation is… | Decision |
|---|---|
| Checkpointed fine-tunes or batch jobs; you will filter for verified or Secure Cloud hosts | Use Vast.ai on-demand. Move to interruptible once checkpoints and external storage are in place. |
| Lowest possible $/hr and the job can be paused | Vast.ai interruptible, budgeting engineering time as well as GPU time. |
| Steady multi-week usage on a host that has already held up | Convert on-demand to reserved, knowing early exit costs the discount plus 5%. |
| Customer-facing inference with hard latency targets or single-vendor compliance | Prefer a fixed-price fleet (RunPod Secure, Lambda, hyperscaler), or Vast Secure Cloud only with your own redundancy. |
| You do not want to operate Docker GPU instances | Do not use Vast.ai. Use a managed inference API. |
Key takeaways (checked November 9, 2026)
- Vast.ai is a GPU marketplace rather than a fixed-price cloud, because hosts set their own rates and Vast publishes no static H100 price (Vast.ai pricing docs, November 9, 2026).
- A single on-demand H100 SXM on Vast.ai had a median ask of $2.27 per GPU-hour at 18:25 UTC on November 9, 2026, below RunPod’s $3.99 Secure Cloud price, according to the openpromt tracker of Vast’s public API and RunPod’s pricing page.
- A stopped Vast.ai instance is not free, because storage is billed continuously while the instance exists and stops only when it is deleted (Vast.ai pricing docs, November 9, 2026).
- Vast.ai interruptible instances suit only checkpointed jobs, because they are often 50%+ cheaper than on-demand but may be paused when outbid or when on-demand renters claim the GPU (Vast.ai instance-types docs, November 9, 2026).
- A Vast.ai reserved instance is a commitment to one host, because prepaid credits are locked to that instance, cannot migrate between hosts, and early cancellation claws back the discount plus a 5% fee (Vast.ai reserved-instances docs, November 9, 2026).
- Vast.ai Secure Cloud is the production-leaning tier, because Vast confirms those machines are in datacenters with a Tier 2/3 rating or ISO 27001, while unverified machines are hidden from search by default (Vast.ai find-and-rent docs, November 9, 2026).
Our verdict: Vast.ai is a coherent, well-documented marketplace for teams that will read an offer card: host-set prices, per-second compute, continuous storage, per-byte bandwidth, three verification tiers and three rental types. It is not a discounted copy of a fixed-price GPU cloud. If cheap H100, A100 or RTX 4090 hours for checkpointed work are the goal, compare the live offers; if one SKU and one SLA are the goal, pay a published price elsewhere.
FAQ
Is Vast.ai worth it?
Yes for teams that need elastic GPU hours for checkpointed fine-tunes or batch jobs and can choose hosts from offer cards. No if you need one fixed rate card, one operator SLA, or always-on customer inference. Vast.ai’s pricing docs, checked November 9, 2026, describe host-set marketplace rates, per-second compute, continuous storage and per-byte bandwidth.
How much does an H100 cost on Vast.ai?
Vast.ai publishes no fixed H100 price. On November 9, 2026 at 18:25 UTC, the openpromt tracker of Vast’s public API showed a median of $2.27 per GPU-hour for one on-demand H100 SXM (minimum $2.20, 14 offers). The same tracker logged medians between $2.27 and $3.40 that day, so check live offers before you budget.
Is Vast.ai cheaper than RunPod?
Often, but not on every GPU. On November 9, 2026, the tracked Vast median for one H100 SXM ($2.27) and one A100 80GB ($1.07) sat below both RunPod tiers, while the RTX 4090 median ($0.47) was above RunPod Community’s $0.34 and below Secure Cloud’s $0.89 (RunPod page updated November 4, 2026).
Is interruptible safe for fine-tuning?
Only with checkpoints and outputs saved outside the instance. Vast.ai’s instance-types docs say interruptible instances are often 50%+ cheaper than on-demand but may be paused when outbid or when an on-demand renter claims the GPU; data stays on disk, but the instance does not run until it regains priority.
Do stopped Vast.ai instances still cost money?
Yes. Vast.ai’s pricing docs say storage is billed continuously while an instance exists, running or stopped, and is typically higher when stopped. Only deleting the instance stops storage billing. Disk size is fixed at creation, with a 10 GB default.
Can you get a refund on a Vast.ai reserved instance?
Partly. Vast.ai’s reserved-instances docs say early cancellation claws back the discount already received plus a 5% fee. In their example, a $576 one-month reservation cancelled after 15 days refunds about $212. If the host is deverified, the full balance can be withdrawn without penalty.
How do you pay for Vast.ai?
With prepaid credits. Vast.ai’s billing docs list credit cards through Stripe and crypto through BitPay and Crypto.com. Instances stop when the balance reaches zero; without a saved card, instances and data are scheduled for deletion unless you add credits promptly. Auto-billing works with a saved card only.
Did Unfair Advantage rent GPUs for this review?
No. We did not open a Vast.ai account, rent an instance or run a benchmark. This review relies on Vast.ai’s official pricing page and docs, a dated third-party snapshot of Vast’s public offers API, and RunPod and Lambda pricing pages, all checked November 9, 2026.
Sources
Sources are listed as plain URLs, without links, so this page sends no clicks to vendors other than through the disclosed affiliate button. Dates are Europe/Paris verification dates unless a source stamps another zone.
- Vast.ai pricing page, vast.ai/pricing (checked November 9, 2026)
- Vast.ai docs, Pricing and Pricing overview, docs.vast.ai/guides/instances/pricing and docs.vast.ai/guides/pricing (checked November 9, 2026)
- Vast.ai docs, Concepts, docs.vast.ai/guides/concepts (checked November 9, 2026)
- Vast.ai docs, Instance Types, docs.vast.ai/guides/instances/choosing/instance-types (checked November 9, 2026)
- Vast.ai docs, Finding & Renting Instances, docs.vast.ai/guides/instances/choosing/find-and-rent (checked November 9, 2026)
- Vast.ai docs, Reserved Instances, docs.vast.ai/guides/instances/choosing/reserved-instances, and prepay instance API reference, docs.vast.ai/api-reference/instances/prepay-instance (checked November 9, 2026)
- Vast.ai docs, Billing, docs.vast.ai/guides/reference/billing (checked November 9, 2026)
- openpromt GPU rental price index, openpromt.com/en/tracker/gpu (update of November 9, 2026, 18:25 UTC; checked November 9, 2026)
- RunPod pricing page, runpod.io/pricing (“Updated November 4, 2026”; checked November 9, 2026), text comparison only
- Lambda pricing page, lambda.ai/pricing (checked November 9, 2026; prices take effect November 16, 2026, 00:00 UTC), text comparison only