Vast.ai Review

20,000 GPUs, 68+ types, 40+ data centers, 700K transactions a month. That's a liquid market, not a shell of one. $5 to start with no sales call is the right on-ramp for engineers who hate procurement calls.

Two things stand out: the SOC 2 certification, and the interruptible tier at 50%+ off for batch work. That's a mature pricing model for a marketplace — most don't segment preemptible vs guaranteed uptime this cleanly.

The tradeoff: you're renting from independent hosts, not a single accountable data center. Reserved instances cut up to 50% for steady workloads and are self-serve — you prepay an on-demand instance from the console or CLI — but the discount stays on that host, so it doesn't remove the host risk. Fine for experiments and batch training, riskier for workloads that can't tolerate a reclaimed instance.

The architecture is a two-sided marketplace: independent hosts and data centers supply GPUs, Vast arbitrates search, pricing, and SSH access. That means 68+ GPU types and 20,000+ GPUs at spot-market rates, per-second billing, $5 minimum entry — real elasticity, no procurement cycle. But heterogeneous hosts mean heterogeneous reliability; the interruptible tier is explicitly preemptible, and on-demand guarantees uptime only within Vast's own SLA, not a hyperscaler's.

Domain fit is strong for teams doing batch training, fine-tuning, or bursty inference who already checkpoint and can tolerate host-level variance. The serverless layer (vLLM, ComfyUI, autoscale-to-zero) and REST API with OpenAPI spec make this scriptable into CI/CD rather than console-driven, which matters for anyone building agent-based provisioning — the SKILL.md and MCP server card for Claude Code/Cursor are a real signal here.

Three-year risk: your workload-placement logic and cost model get tuned to a marketplace, not a fixed catalog, so migrating off means re-architecting scheduling assumptions, not just swapping an SDK. SOC 2 certification helps the compliance conversation but doesn't replace hyperscaler-grade support contracts.

No SSO tax here — there's no seat model at all. Three modes: On-Demand, Interruptible (50%+ cheaper, preemptible), Reserved (up to 50% off, 1-6 month terms). Start with $5 credit. No sales call required for any of them.

TCO math is different from SaaS. Rent an H100 at market rate, run it 24/7 for a training job, cost scales with hours not headcount. Interruptible workloads cut cost in half but risk reclamation — checkpoint or lose the run. Reserved is a prepay on one instance: you can preview the discounted rate before renting, but cancel early and the discount is clawed back plus a 5% fee.

700K+ transactions monthly, 20,000+ GPUs, 40+ data centers — real marketplace depth, per the vendor's own numbers. SOC 2 certified, which helps procurement. But pricing floats with 68+ GPU types across independent hosts — budget forecasting is harder than a fixed-tier competitor. No published overage rate because there's no overage concept — you just spend more when you rent more.

Search, rent, SSH in. That's the core loop, and it works the way you'd expect from a marketplace rather than a hyperscaler — you're picking a host, not a black box. 20,000+ GPUs across 40+ data centers and 68+ GPU types (RTX 4090s up to B300s) means real choice, but also real variance: interruptible instances get reclaimed, and reliability depends on whose rig you landed on that day.

The CLI and Python SDK (VastAI, SyncClient, AsyncClient) cover scripted provisioning well, and the OpenAPI spec plus REST API mean you're not stuck clicking a dashboard for automation. SKILL.md files for Claude Code and Cursor are a nice touch if you're wiring agents into provisioning — most GPU clouds haven't bothered.

Vast posts regular product updates listing new features and bug fixes, which matters on infra you depend on daily. Reserved instances and per-second billing on-demand cover both steady and bursty workloads reasonably, but SOC 2 aside, the trust model here is still 'read the host's reviews.'

Vast.ai is a marketplace, not a polished cloud console, and it feels like one. You're browsing 20,000+ GPUs across 40+ data centers and independent hosts, picking by price and location, which means quality and latency vary listing to listing. That's the tradeoff of skipping the AWS-style markup: you get per-second billing and interruptible instances at 50%+ off, but you're doing the vetting yourself.

First ten minutes is genuinely low friction — $5 credit, an API key, SSH in, go. That's rare in this category. The CLI, Python SDK, and REST API all exist, and they've even shipped SKILL.md files for Claude Code and Cursor, which tells me someone there is thinking about how agents will actually use this thing, not just humans.

Serverless vLLM and ComfyUI deployment with autoscale-to-zero is the feature that ages well past week one. SOC 2 certification is a nice trust signal for a marketplace model that could otherwise feel sketchy.

No 'best in class' language on the landing page. Just numbers: 20,000+ GPUs, 68+ types, 40+ data centers, 700K+ transactions a month, $5 minimum. That specificity is rare and it's the kind of claim that's easy to check and hard to fake.

The catch is structural, not marketing. Hosts include independent hardware suppliers alongside data centers — the docs cover host verification and payment for a reason. On-demand pricing floats with the market, interruptible instances can get reclaimed. Fine for batch training, riskier if you assumed cloud-GPU stability like a hyperscaler.

Exit is decent — SSH access, standard containers, REST API with an OpenAPI spec, nothing proprietary trapping your workload. SOC 2 certification is a real trust signal for a marketplace model. Regular product-update posts list new features and bug fixes, so the update cadence is on the record. Agent tooling (SKILL.md, MCP server card) is a genuinely different bet than most GPU clouds are making right now.