Vast.ai

What Vast.ai actually is

Vast.ai is a marketplace, not a cloud provider, and everything that follows flows from that distinction. Independent hosts — everyone from a gamer with a spare RTX 4090 to colocation operators with idle datacenter racks — list GPU machines on the platform and set their own prices. Vast.ai provides the matching, billing, container runtime, verification, and reputation system; the hardware belongs to someone else. The company was incorporated in 2016 in Delaware, launched publicly in 2018, and is headquartered in Los Angeles with a lean team of roughly 40 people.

The pitch is price. Because host-set rates compete with each other in real time, GPU-hours on Vast.ai routinely undercut every fixed-price alternative: on the day we checked, H100 SXM offers started at $1.73 per GPU-hour with a market median of $2.13, RTX 4090 offers appeared from $0.14, and the cheapest listings on the entire platform (older consumer cards like the GTX 1060) cleared around $0.03 per hour. Interruptible bidding goes lower still — a $0.69 minimum bid on H100 SXM was visible in the public API. For researchers and teams paying hyperscaler rates, the arithmetic is blunt: this is the floor of the GPU rental market in /benchmarks terms of dollars per accelerator-hour.

The inventory is broad — the pricing page listed 81 GPU card types across hosts in 52 countries, from consumer GeForce cards through A100, H100, H200, and B200 datacenter parts, and Vast was the first rental marketplace to list AMD Instinct accelerators alongside NVIDIA. You rent by the second, run your own container, and stop paying when you stop the instance. There are no subscriptions and no minimum commitments; accounts are prepaid with credits.

Pricing: what you really pay

Three rental types, all billed per second. On-demand is a fixed host-set price with guaranteed access — the host cannot preempt you while the contract runs. Interruptible is a bidding market: you name a maximum bid, higher bidders take priority, and lower-priority instances are paused (processes killed, disk preserved) when outbid or when an on-demand renter claims the machine. The pricing page markets interruptible as "50%+ cheaper" and positions it for batch training with checkpoint/resume. Reserved converts an on-demand rental to pre-paid terms of 1, 3, or 6 months for "up to 50% off," with funds locked to that instance (withdrawable if the machine is deverified).

The headline GPU rate is not the whole bill. Vast.ai's own docs are explicit: total cost combines GPU compute, storage, and bandwidth. Storage is billed continuously while the instance exists — including while it is stopped — at host-set rates (third-party price surveys have seen anything from $0.0013 to $1.00 per GB-month depending on host). Bandwidth is metered and host-priced, and comparison shopping has surfaced egress rates as high as $68.27/TB on expensive hosts; many hosts are far cheaper, but you must check the offer card. GPUs are released when you stop an instance, but disk charges continue until you delete it. Run out of credits and instances stop automatically — which, mid-training, is exactly as painful as it sounds.

Our practical advice: treat the search filters as part of the price. Sort by total cost including storage and bandwidth, not by GPU-hour, and set minimums for reliability score (we'd use 0.95+ for anything unattended) and network speed. The $1 sign-up credit is enough to validate the workflow before real money goes in.

Performance & infrastructure

There is no Vast.ai datacenter map, because Vast.ai owns no datacenters. Machines run in hosts' locations — the platform reports hosts in 52 countries — so EU, US, and APAC coverage exists in the sense that you can find hosts there, not in the sense that a single operator guarantees regional tiers. On the plus side, that means genuinely global supply and real data-residency choice (pick a host in Germany if you need German hardware). On the minus side, consistency is your problem.

Vast.ai's quality layer is a three-tier verification system: unverified machines (filtered out of search by default), verified machines (automated checks: CUDA 12+, reliability score ≥90%, ≥500 Mbps symmetric network, ≥7 GB VRAM per GPU, port and storage minimums), and datacenter-tier labels. Every machine carries a reliability score (historical uptime; new machines start at 60%) and a DLPerf score for cross-hardware deep-learning performance comparison. Verification criteria are only partially disclosed and partly demand-driven, which we consider a transparency weakness.

Host variance is the headline operational risk, and third-party testers keep confirming it. GAX's 2026 sampling of eight H100 SXM hosts running an identical Llama 70B inference workload measured a median 1,752 tok/s against a best-host 1,879 and a worst-host 1,683 — an 11.7% spread on identical SKU specs, with one host showing visible packet loss and another capped at 100 Mbps egress. Their 12 interruptible jobs saw two preemptions (both on hosts scoring below 9.7 reliability), and one host vanished mid-job with no warning. Platform-side, the booking and bidding surface has been solidly available in the same tracking (~99.9%), but median host uptime for high-scoring hosts sat near 98.3%. The gap between those numbers is the product: the market is reliable, the machines are somebody's side business.

Per-second billing with no minimum charge is genuinely consumer-friendly, and provisioning speed is good (GAX measured SSH-ready in 94–176 seconds depending on host). Multi-GPU nodes, host-side max contract durations (visible per offer), and VM/container support vary by host — again, read the card.

Where it falls short

No SLA. None. There is no provider-level uptime guarantee, no technical account manager, and no recourse when a host disappears beyond re-provisioning elsewhere. Vast.ai's docs handle deverification (a full reserved-balance withdrawal if a machine fails its implicit SLA), which is better than nothing, but the support model is forum and email, not live enterprise support. For compliance-heavy or customer-facing production work, that disqualifies the platform regardless of price — see /category/gpu-cloud for managed alternatives with contractual guarantees.

Interruptible economics require engineering discipline: checkpoint aggressively or lose hours of work when outbid. Storage billing on stopped instances trips up newcomers, and the cheap end of the market (residential hosts, old consumer cards) delivers correspondingly cheap reliability. Vast.ai has certifications a marketplace of this age rarely has (SOC 2 Type II and HIPAA attestation, per third-party comparisons), but a certification on the platform is not a guarantee about the host's garage.

Finally, pricing is a live market: expect week-to-week swings of roughly ±10% and no price locks outside reserved terms. Budget forecasting is median-based, not quote-based.

Who should pick this

We'd pick Vast.ai when the workload is fault-tolerant or checkpointable and the budget is real: ML training runs, fine-tuning, rendering, batch inference, research experiments, and burst capacity when every managed cloud says "capacity coming soon." The marketplace always has something to rent, and the interruptible tier's dollar-per-compute is unmatched. Hobbyists get a legitimate entry point — a $0.19–0.29/hour RTX 4090-class machine is cheaper than the electricity of running one at home in many countries.

We'd steer away anyone running production inference SLAs, unattended multi-week jobs without checkpointing, or workloads with regulatory requirements that demand a single accountable operator. In those cases, paying 30–100% more at Lambda, CoreWeave, or a hyperscaler buys you exactly the guarantee you're giving up here.

The discipline that makes Vast.ai work: filter hard (reliability ≥0.95, datacenter tier, ≥1 Gbps both directions), benchmark a candidate host for an hour before committing a long job, checkpoint every N steps on interruptible, and delete stopped instances so storage stops billing. Buyers who do this capture most of the savings; buyers who don't subsidize the market with surprise downtime.

Key takeaways

• Peer-to-peer GPU marketplace: hosts in 52 countries set their own prices; Vast.ai runs the matching, billing, and verification layer.
• Cheapest GPU-hours on the public market — H100 SXM from $1.73/hour (median $2.13), RTX 4090 from $0.14, old consumer cards near $0.03/hour at the extreme low end.
• Three rental types: on-demand (no preemption), interruptible (bidding, 50%+ cheaper), reserved (up to 50% off, 1–6 month terms); all per-second.
• Total cost includes host-set storage (billed even on stopped instances) and metered bandwidth — check the offer card, not just the GPU rate.
• No SLA, no owned datacenters, host variance measured at ±11.7% on identical SKUs — fine for checkpointable batch work, wrong for production SLAs.
• Our view: the best dollars-per-GPU-hour anywhere if you filter hosts like your job depends on it, because it does.

Sources

• https://vast.ai/pricing
• https://docs.vast.ai/documentation/instances/rental-types
• https://aiwiki.ai/wiki/vast_ai
• https://gpubank.io/compare/gpubank-vs-vast-ai
• https://gaxonline.com/rankings/gpu-cloud/vast-ai
• https://deepwiki.com/vast-ai/docs/2.1-two-sided-marketplace