Inferact raises a $150M seed at $800M from a16z and Lightspeed

The company was formed by the creators and core maintainers of vLLM, an open-source inference engine incubated in 2023 at UC Berkeley's Sky Computing Lab under Ion Stoica, a Databricks co-founder. Simon Mo is CEO. Beyond the two lead firms, Sequoia Capital, Altimeter Capital, Redpoint Ventures and ZhenFund backed the round, and Databricks' venture arm and the UC Berkeley Chancellor's Fund also took part. Proceeds are earmarked to commercialize vLLM, scale its platform for enterprise AI inference, and expand adoption, plus hiring and finishing a commercial version. The usage figures come from a16z and the maintainers, not an audit: vLLM is said to run on more than 400,000 GPUs, with over 50 core developers, and production use at Meta, Google and Character.ai. More than 2,000 developers have contributed. Roblox's Assistant also runs on it. The round reflects investor attention shifting from training to serving, as deployment and operating efficiency become central to scaling AI applications. For founders, inference cost is now a funded product category. A default open-source serving layer is getting a commercial steward with a large balance sheet, which affects build-versus-buy choices and the pitch of any rival inference vendor. Teams already on vLLM should watch how the company separates the free project from paid features.