Vast.ai

Vast.ai may be the most capital-efficient company in AI infrastructure: roughly $4M raised since 2018, yet it operates the largest open GPU rental marketplace, with prices that run several times below major clouds because it monetizes hardware others already paid for. The model has quietly compounded into real institutional shape: 2025 brought SOC 2 Type I and Type II certification, a Serverless offering that auto-schedules inference across the fleet, and supply depth no single operator can match — including over 1,000 RTX 5090 machines within months of that GPU's release, plus H200s and B200s from independent data centers. It was also first to bring AMD GPUs to the marketplace model in 2024, widening supply beyond Nvidia. Structurally, Vast wins in both market regimes: in scarcity, demand overflows to it from allocation-constrained clouds; in glut, hosts flood it with cheap capacity and its price advantage widens against owners of depreciating fleets. As inference and fine-tuning workloads — cost-sensitive, interruption-tolerant, embarrassingly parallel — become the volume segment of AI compute, the exchange sitting on the cheapest supply curve takes rake on an enormous flow. A take-rate business with near-zero capex riding the largest hardware buildout in history is a venture outcome hiding in plain sight.

Vast.ai's cheapness is a symptom of what it sells: heterogeneous, unaudited, interruption-prone capacity from anonymous hosts, structurally unsuited to the enterprise workloads where AI budgets actually concentrate. SOC 2 certifies Vast's own software, not the thousands of independent operators holding customer data on consumer GPUs — many running hardware whose datacenter use sits in tension with Nvidia's driver licensing, a supply base with crypto-mining DNA and churn to match. The competitive squeeze is arriving from every direction: Runpod raised orders of magnitude more to professionalize the same community-cloud model, hyperscaler spot pricing gets aggressive whenever utilization dips, and a genuine GPU glut — the scenario bulls celebrate — would collapse marketplace clearing prices and take-rate revenue simultaneously, since Vast earns a percentage of a falling number. Meanwhile the $4M war chest that signals efficiency also means no enterprise sales motion, no compliance organization, no capital to subsidize the serverless product against Modal and Fireworks, and key-person concentration in a small team. The marketplace's defensibility is thinner than it looks: hosts multi-home across Vast, Runpod, and newer aggregators without friction, and renters follow price. A fine, profitable niche business — but the ceiling is the arbitrage, and the arbitrage is closing.