What Vast.ai sells — and what it buys
GPU-hours through a two-sided marketplace — buyers filter 17,000+ host-owned GPUs by spec, price, location and reliability score, from consumer RTX 4090s (~$0.34/hr in early 2026) to datacenter H100s and H200s, and run workloads in Docker containers with per-second billing. Three rental models: fixed-price on-demand, interruptible spot auctions on idle capacity, and reserved commitments — plus a serverless orchestration layer that autoscales inference-style workloads across the marketplace. And a second product, sold to the other side: a business-in-a-box for hosts — hardware sourcing through vetted suppliers, financing qualified on platform earnings, presold enterprise demand, and a verified-datacenter certification tier.
Nothing — the limit case of this ladder's buy side. Supply is listed, not bought: 1,400+ independent providers, from single-workstation hobbyists to data-center operators, own the hardware, set their own prices, and compete on rate. What Vast.ai builds instead of buying is trust infrastructure — reliability scoring, datacenter verification, SOC 2 Type I compliance, fraud protection, and the 24/7 support layer that hosts cite as the platform's main value.
Scale proof
The bet
That liquidity beats ownership. Every owned fleet on this ladder carries capex, depreciation and utilization risk; Vast.ai's wager is that aggregating idle supply from providers who already sank the capex drives the marginal price of a GPU-hour toward electricity plus host margin — structurally below what any fleet owner can sustain. The second act is the Amazon-marketplace play: sell the sellers the operating system — hardware sourcing, financing underwritten by platform earnings history the way Amazon Lending underwrites merchants, capacity presold before it's racked — so supply compounds without Vast buying a single GPU. The utilization numbers argue the flywheel is turning: at 98.6% B200 fleet utilization, hosts' on-demand earnings often beat average reserved rates — spot liquidity out-yielding commitment, which is the thesis in one statistic. The known cost of the model is variance: no contractual SLAs on community hosts, and effective prices 20–40% above the listed floor on unverified hosts once downtime is counted — the gap the verified-datacenter tier and the serverless orchestration layer (which does provide autoscaling) exist to close.
Natural counterparty
On the demand side: cost-sensitive researchers, batch and experiment workloads, and long-tail developers who can tolerate variance or restrict themselves to verified hosts — the same population Runpod courts, minus the managed layer. On the supply side — where the marketplace's real counterparties live: GPU owners of every scale seeking demand for idle hardware, data-center operators preselling capacity before it comes online, hardware suppliers and financing partners plugged into Vast's sourcing and lending funnel. Vast.ai transacts with almost everyone on this ladder's buy side without ever holding inventory itself.
Buyer fit
You're running short experiments, batch jobs or price-sensitive training where the lowest listed rates anywhere plus per-second exit beat consistency — or you can restrict to verified datacenter hosts for jobs that need predictable uptime at a still-competitive $1.50–1.87/hr for H100s.
You need contractual SLAs and guaranteed uptime for production inference or multi-day training — community hosts carry no uptime guarantee, and third-party guidance is to budget 30–50% above the lowest listed rate for predictable costs. For autoscaled serving, Vast's serverless orchestration covers the basics; teams wanting a fully managed platform layer still sit better on other rungs of the ladder.
Five fields we track but don’t publish
Everything above is public or provider-verified — third-party figures are sourced as noted, and Vast.ai reviewed this card on 10 July 2026 (autoscaling and host-earnings figures updated per their feedback). The fields below change weekly and are verified directly with each provider — they live in a private supply-demand ledger, not on this page.