Vast.ai

Peer-to-peer GPU rental marketplace; usually the cheapest hourly GPU price on the market, at the cost of host heterogeneity.

1. Core Product / Service

Decentralized GPU cloud where independent hosts (gamers with spare RTX 3090/4090s, ex-crypto-mining ops, small colocation facilities) list hardware and set their own prices. Vast handles user accounts, billing, container orchestration (Docker images via SSH/Jupyter), and 24/7 support; it does not own data centers.

Three pricing tiers as of 2026: On-Demand ($0.29/GPU/hr floor), Reserved ($0.20/GPU/hr, longer-term), and Interruptible (~$0.10/GPU/hr, spot-style bidding that can be preempted by higher bidders) [1]. Headline rates: RTX 3090 from ~$0.12/hr, RTX 4090 from ~$0.29/hr, A100 PCIe from ~$0.52/hr, H100 PCIE from ~$1.47/hr [1][2]. Platform takes ~25–30% commission from hosts [3].

2. Target Users & Pain Points

• ML researchers / indie devs running short experiments, fine-tunes, or batch inference on a budget.
• Teams doing large-scale offline inference where interruption tolerance is fine.
• Anyone priced out of hyperscalers (Vast often runs 50–70% cheaper than AWS/GCP) [2].

Pain points it solves: (a) no GPU quotas / waiting lists, (b) spin-up in seconds, (c) deep inventory of consumer-grade GPUs (3090/4090) that hyperscalers don't offer at all.

Pain points it does not solve: workload reliability depends on the specific host you land on — bandwidth, uptime, and disk I/O vary widely. Unverified hosts carry real risk of downtime and inconsistent performance [2]. Not suitable for production-grade SLAs.

3. Competitive Landscape

• runpod — closest competitor. RunPod splits into Community Cloud (P2P, similar to Vast) and Secure Cloud (vetted DCs). RunPod typically more expensive per hour but stronger DX (serverless, sub-200ms cold starts, autoscaling). Vast wins on raw price; RunPod wins on "just works."
• Salad — consumer-GPU distributed network; even more aggressive on price for inference/prototyping, narrower hardware mix.
• Akash — decentralized container marketplace on a crypto rails; broader use case (any container) but smaller GPU inventory.
• lambda-labs / coreweave / nebius — at the other end of the spectrum: owned data centers, H100/H200 clusters, enterprise SLAs, much higher prices.

Vast's differentiation: largest peer-to-peer GPU inventory (17,000+ GPUs across 1,400+ providers in 500+ locations as of 2026 [4]), and price floor set by spot/interruptible bidding.

4. Unique Observations

• Long-term inference cost analysis (raw notes from 2026-04-12): for steady-state inference workloads, the "always-on" cost of pinning a Vast instance can rival a reserved RunPod box once you factor in eviction risk and migration overhead. The price advantage is real for bursty / preemptible jobs but narrows for production serving — RunPod or runpod Secure Cloud often wins for sustained 24/7 inference.
• The marketplace dynamic means H100 supply on Vast tends to be thin and expensive relative to RTX 4090 supply, because H100 hosts have better commercial alternatives (CoreWeave, hyperscalers). Consumer-card pricing is where Vast is structurally hard to beat.
• See ai-inference-engines and runpod-gpu-inference for related modules.

5. Financials / Funding

• Founded 2018 in Los Angeles, CA. Largely bootstrapped; raised ~$4M in early funding from DRW Holdings and Nazare [5].
• Reported ~$2.2M ARR with ~38 employees circa 2025 (per Crunchbase / public profiles) [5] — small revenue figure suggests these are pre-commission take-rate dollars on platform GMV; underlying transaction volume is much higher. Numbers should be treated as approximate.
• No publicly disclosed late-stage round. Capital efficiency is part of the story: marketplace model means hosts supply the capex.

6. People & Relationships

• Founder / CEO: Jake Cannell.
• COO: Tavis Cannell.
• Investors: DRW Holdings, Nazare (early funding only).
• Competitors: runpod, lambda-labs, coreweave, nebius.

Sources

[1] https://vast.ai/pricing (2026-05-09) — current tier pricing [2] https://getdeploying.com/runpod-vs-vast-ai (2026-05-09) — head-to-head with RunPod, reliability tradeoffs [3] https://vast.ai/hosting (2026-05-09) — host commission terms [4] https://www.fluence.network/blog/best-gpu-rental-marketplaces/ (2026-05-09) — marketplace scale [5] https://www.crunchbase.com/organization/vastai (2026-05-09) — founding, funding, headcount

• local: 2026-04-12.md — long-term inference cost notes