Inferact, the San Francisco startup founded by the core maintainers of vLLM, raised $150 million in a seed round announced on January 22, 2026, valuing the company at $800 million post‑money. Investors include Andreessen Horowitz and Lightspeed Venture Partners, among others.
What Inferact is selling
Inferact is building a commercial business around vLLM while keeping the project open source. vLLM addresses inference efficiency using a technique called PagedAttention, which borrows from virtual memory: rather than reserving large static blocks of RAM per request, the system allocates smaller memory pages on demand so less memory sits idle and the same hardware can serve more requests.
The project supports more than 500 model architectures and over 200 hardware accelerators, and more than 2,000 contributors have worked on it. Amazon, Meta, and Google all use vLLM in their production systems.
On top of the open‑source engine, Inferact plans to offer paid services including serverless offerings that let customers run models without managing servers, and enhanced observability tools to monitor deployments. The company says its commercial focus is tuning inference across three levers: cost, latency, and throughput.
From Berkeley lab to $800 million startup
vLLM began in 2023 at UC Berkeley’s Sky Computing Lab, directed by Ion Stoica, a co‑founder of Databricks. Stoica is a founding member of Inferact. The startup’s founding team also includes Simon Mo, who serves as CEO, alongside Woosuk Kwon, Kaichao You, and Roger Wang.
vLLM became a PyTorch Foundation project in May 2025. Inferact was established later in 2025 with the stated goal of commercializing and scaling vLLM while keeping the open‑source project intact. As of October 2026, Inferact continues contributing upstream to vLLM, including early optimization work for NVIDIA’s Vera Rubin hardware that has delivered up to 7.8x throughput improvements on specified models.
Implications for the AI infrastructure market
The company follows an open‑source commercialization playbook similar to Databricks’ business around Apache Spark, leveraging maintainers’ expertise to build paid products on top of a widely used project. With support for more than 200 accelerators, vLLM positions itself as a cross‑chip layer; at the same time, large cloud providers already run vLLM and offer their own managed AI services. Inferact’s stated advantage is that the people who know the codebase best now work under one roof.
The funding and valuation give Inferact resources to scale commercial offerings while continuing upstream development of vLLM, according to the facts provided.
