Vast.ai
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,887 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,887 words
Add missing citations, update stale details, or suggest a clearer explanation.
Vast.ai is a cloud-based GPU rental marketplace that connects independent hardware operators (called hosts) with developers, researchers and companies looking to rent compute by the second.[^1][^2] The platform was incorporated in 2016 by Jake Cannell and Christian Horne and launched publicly in 2018, positioning itself as a peer-to-peer alternative to the fixed-price GPU offerings of hyperscale clouds such as AWS, Google Cloud and Microsoft Azure.[^3][^4] Each rental runs inside a Linux Docker container with exclusive access to one or more GPUs, supports both fixed (on-demand) and interruptible (bid-priced) modes, and is reachable over SSH, Jupyter or a user-supplied entrypoint.[^5][^6] The marketplace inventory mixes consumer-grade cards (the GeForce RTX 3090 and RTX 4090 are heavily represented) with data-center GPUs such as the NVIDIA A100, NVIDIA H100 and NVIDIA H200, and in 2024 the company became the first GPU rental marketplace to list AMD Radeon and Instinct accelerators alongside its NVIDIA fleet.[^7][^8]
Jake Cannell (CEO) and Christian Horne co-founded Vast.ai Inc., a Delaware C corporation, on 28 June 2016 after years of writing on LessWrong about compute scaling and the brain-as-universal-learning-machine hypothesis.[^3][^9] Cannell had begun publishing essays in 2010 arguing that intelligence is primarily a function of compute rather than algorithmic novelty, and the founding thesis of Vast.ai was that GPU compute should remain distributed across many independent operators rather than concentrated inside a few hyperscalers.[^3][^9] Travis Cannell later joined as Chief Operating Officer to run the business side as the company scaled.[^3][^7]
The public marketplace went live in 2018, initially seeded with friends, family and Reddit users running GPUs at home.[^4] Vast.ai has remained a relatively small organisation: as of 2024-2025 it employed roughly 40 people, was headquartered in Los Angeles with a secondary office in San Francisco, and operated without large venture rounds, instead reinvesting marketplace take-rate revenue.[^3][^10] In May 2024 the company stated that it was averaging 265% year-over-year growth since 2019 and had grown about 310% in 2024 alone, with over 350 independent hosts and more than 17,000 GPUs listed at that time.[^7][^11] Vast.ai's own homepage in 2026 reports more than 20,000 GPUs across 40+ data centers, 68+ distinct GPU models, and over 700,000 transactions per month.[^1]
The company is legally and organisationally distinct from the New York-based storage vendor VAST Data, which sells enterprise file and object storage to AI factories and is not affiliated with the GPU marketplace.[^2][^12]
| Vast.ai infobox | Value |
|---|---|
| Company type | Private (Delaware C corporation) |
| Founded | 28 June 2016 (incorporated); public launch in 2018 |
| Founders | Jake Cannell, Christian Horne |
| CEO | Jake Cannell |
| COO | Travis Cannell |
| Headquarters | Los Angeles, California; second office in San Francisco |
| Business model | Two-sided GPU rental marketplace (hosts and renters) |
| Approximate GPU inventory | 20,000+ GPUs across 40+ data centers (2026) |
| Container runtime | Docker (Linux containers, exclusive GPU passthrough) |
| Rental modes | On-demand, Interruptible (bid), Reserved |
| Notable firsts | First GPU rental marketplace to support AMD Radeon/Instinct (2024) |
Vast.ai sits between two populations: hosts, who install the Vast hosting daemon on their own machines (ranging from a single workstation with a few consumer cards to professional data-center racks of H100s), and clients/renters, who use the Vast web console, REST API, Python SDK or vastai command-line client to search the live inventory and launch instances.[^1][^2] Hosts decide what hardware to expose, set their own prices for each rental mode, and configure constraints such as minimum rental duration; the platform handles user authentication, billing, search ranking and dispute mediation, and takes a cut of each transaction.[^11][^13]
Pricing is dynamic rather than centrally set. Vast.ai exposes three rental types, all priced per second:
Because prices are set by each host and float with regional supply and demand, the same GPU model can be listed at very different prices across the marketplace; Vast.ai claims that this competitive dynamic makes its consumer-GPU rentals roughly 3-5x cheaper than mainstream alternatives.[^7][^11] Storage continues to accrue charges when an instance is stopped (not just running), so users are advised to destroy instances they no longer need rather than merely pausing them.[^6][^14]
Each Vast.ai instance is a Linux Docker container that the renter configures by supplying a Docker image (either a public one, e.g. from Docker Hub, or a private one with credentials).[^5][^13] The platform supports three launch modes:
ENTRYPOINT/CMD. This is appropriate for long-running training jobs, model servers and batch workloads.[^5]GPUs are passed through to the container with NVIDIA's GPU driver stack already installed on the host, so renters do not need to install GPU drivers themselves; they typically use images such as the official PyTorch or CUDA containers as their base.[^5][^13] CPU and RAM are allocated proportionally to the GPU fraction the renter has rented, and storage is set at instance creation and cannot be resized afterward.[^13] Because instances are themselves containers, nested Docker (Docker-in-Docker) is not supported.[^5]
Each Vast.ai host typically has one public IPv4 address that is shared by all of the host's instances. Vast.ai partitions a subset of the host's external ports to each container, and each internal port the renter opens (for example, 22 for SSH or 8080 for an inference server) is mapped to a pseudo-random external port on the host's IP.[^14] An instance can request up to roughly 64 mapped ports. For users who require predictable IP and port assignments, some hosts offer dedicated/static-IP machines as a premium option.[^14]
The marketplace runs an explicit reputation and verification system. Hosts that pass Vast.ai's vetting are tagged as certified data centres and receive a visible blue verification badge as well as algorithmic preference in search ranking, so that enterprise renters and high-priced jobs naturally migrate to them.[^15] Vast.ai vets certified hosts against documentation of ownership, source of funds and third-party compliance certifications (ISO 27001 is a baseline; some carry healthcare, finance or government certifications).[^15] Vast.ai itself completed SOC 2 Type 2 attestation, which an independent auditor checks against the AICPA Trust Services Criteria over a sustained observation window.[^15][^16] For workloads handling sensitive data, the platform offers a Secure Cloud tier that restricts the user to inventory hosted exclusively at vetted data centres.[^15] In Vast.ai's own security documentation, the company acknowledges that, in principle, hosts have administrative control over the physical machines they operate and can therefore reach files on the local disk; renters with sensitive data are explicitly advised to use verified Secure Cloud hosts and to apply their own encryption.[^17]
Vast.ai's catalogue spans both consumer and data-center GPUs. According to the company, the marketplace lists 68+ distinct GPU models ranging from older consumer cards such as the GeForce RTX 3060 class up through current data-center hardware such as the NVIDIA B200.[^1] Coverage from Tom's Hardware in 2024 noted that the marketplace was heavily weighted toward consumer cards (the RTX 4090 in particular), with relatively few H100s and A100s and that the largest single H100 instance available at that time was only eight GPUs, which Nat Friedman of the rival service gpulist.ai pointed to as the reason his service explicitly targeted multi-hundred GPU clusters instead.[^18]
Beyond NVIDIA, on 2 May 2024 Vast.ai announced via press release that it had become the first GPU rental marketplace to support AMD hardware, adding Radeon (consumer) and Instinct (data-center, including the MI series) product lines to its inventory alongside NVIDIA.[^7] The company has subsequently expanded its enterprise-focused product line (including reserved instances, cluster offerings, virtual machines and a serverless orchestration layer), while keeping the core marketplace consumer-friendly.[^1][^7]
The economics of Vast.ai make it most attractive in workloads where the cheap unit cost of a consumer or commodity GPU outweighs the convenience and uptime guarantees of a managed cloud:
Vast.ai also runs vertical programmes aimed at startups (the Vast.ai Startup Program) and educational institutions (GPUs for EDUs) that bundle credits and prioritised support for those audiences.[^1]
The GPU cloud market spans pure marketplaces, hybrid models that combine first-party data centres with third-party inventory, and fully owned clouds focused on enterprise reliability. The most common direct comparisons for Vast.ai are RunPod, Lambda Labs and TensorDock.
| Provider | Founded | Model | Hardware focus | Key positioning |
|---|---|---|---|---|
| Vast.ai | 2016 (public 2018) | Pure peer-to-peer marketplace, dynamic pricing, bidding for interruptible[^2][^3] | Wide mix; heavy consumer-GPU representation alongside data-centre cards; added AMD support in 2024[^7][^18] | Cheapest consumer-GPU rentals in industry comparisons; variable uptime because hosts can reclaim machines[^19] |
| RunPod | 2022 | Hybrid: first-party "Secure Cloud" plus third-party "Community Cloud" marketplace, also offers serverless[^20] | Strong on data-centre GPUs (H100, L40S), with community-cloud commodity tier | Founded by Zhen Lu and Pardeep Singh; reported $120M ARR by 2026; backed by Intel Capital and Dell Technologies Capital seed[^20] |
| Lambda Labs | Cloud platform launched in the late 2010s | Owned-and-operated data centres, ML-focused managed cloud | Premium NVIDIA GPUs (H100, H200, B200, GH200, A100) with InfiniBand multi-node clusters[^21] | Best-known for ML-focused support; reported $425M revenue in 2024 (70% YoY)[^21] |
| TensorDock | 2021 | Peer-to-peer marketplace similar to Vast.ai, also offers reserved bare-metal options[^22] | Mix of consumer and data-center; H100 from ~$2.25/hr in 2025 | Founded by Jonathan Lei; acquired by Voltage Park in March 2025, with Lei becoming GM of On-Demand at Voltage Park[^22] |
Within this group, Vast.ai is the purest marketplace play: it owns essentially no compute itself and instead orchestrates supply from independent hosts. That choice yields the lowest headline prices, particularly for consumer GPUs, but also produces wider variance in network bandwidth, disk performance and uptime than first-party clouds such as Lambda Labs.[^19] RunPod sits between Vast.ai and Lambda by offering both a tightly controlled first-party tier and a community marketplace tier, while TensorDock historically resembled Vast.ai's model and has since been folded under Voltage Park.[^20][^22] Larger first-party AI clouds such as CoreWeave and energy-co-located providers such as Crusoe are not typically direct substitutes for Vast.ai, because they target multi-thousand-GPU H100 reservations for enterprises rather than consumer-priced single-GPU rentals.
For very large clusters (256 GPUs and above), Vast.ai is not the natural fit. As Nat Friedman noted when launching the cluster-focused gpulist.ai service in February 2024, the largest single H100 server then listed on Vast.ai held only eight GPUs, which made Vast.ai unsuitable for the cluster-scale frontier training jobs that gpulist.ai was designed to serve.[^18]
Vast.ai's marketplace structure is responsible for both its low prices and most of the practical criticisms levied against it:
Vast.ai is one of the earliest and most-cited examples of a peer-to-peer marketplace for AI compute, and it pioneered several patterns that later providers (TensorDock, RunPod's Community Cloud and others) imitated: a dynamic dual-tier pricing structure that combines fixed on-demand prices with a bid-based interruptible tier; per-second billing on commodity GPUs; and a host-reputation system that elevates verified data-centre operators within an otherwise open marketplace.[^2][^11][^15] In an industry dominated by hyperscalers and large dedicated AI clouds, it has demonstrated that there is durable demand for a long-tail GPU marketplace optimised for price-sensitive AI/ML practitioners working on consumer-grade hardware, and that this niche can be served profitably without large venture rounds.[^3][^10]