Inferact, a San Francisco-based AI infrastructure startup founded by the people who maintain vLLM, has raised $150 million in seed funding. The round, announced on January 22, 2026, values the company at $800 million post-money, with Andreessen Horowitz and Lightspeed Venture Partners among the investors.

The pitch is straightforward: take one of the most widely used open-source tools for running large language models and build a commercial business around it, all without closing the source.

What Inferact Is Actually Selling

vLLM tackles inference efficiency with a technique called PagedAttention. The approach borrows from virtual memory, the mechanism operating systems use to juggle limited RAM across many programs at once. Rather than reserving large, rigid blocks of memory for each request, the system hands out memory in smaller pages as needed. Less memory sits idle, so the same hardware can serve more requests.

The stakes at this layer are structural. Inference is the recurring cost of AI—every request a deployed model serves consumes compute and memory—so efficiency gains compound directly into lower cost per request rather than one-off savings.

The project's reach is broad. vLLM 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.

Inferact plans to layer paid services on top of the open-source project. Those include serverless offerings, which let customers run models without managing servers themselves, and stronger observability tools for monitoring how deployments behave. The company says its 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, which is directed by Ion Stoica, a co-founder of Databricks. Stoica is also a founding member of Inferact. The rest of the founding team includes Simon Mo, who serves as CEO, joined by Woosuk Kwon, Kaichao You, and Roger Wang.

In May 2025, vLLM became a PyTorch Foundation project. Foundation governance keeps the roadmap vendor-neutral—a detail that matters commercially, since adopters can standardize on vLLM without depending on any single vendor. Inferact was established later in 2025 with a stated goal of commercializing and scaling vLLM while keeping the open-source model intact.

As of October 2026, Inferact continues contributing upstream to vLLM, including early optimization work for NVIDIA's Vera Rubin hardware. That effort has delivered throughput improvements of up to 7.8x on specified models.

What This Means for the AI Infrastructure Market

The open-source commercialization playbook is familiar. Databricks built a business around Apache Spark, another project with Berkeley roots, and Inferact is running a similar play—with a founding member who has done it before.

With support for more than 200 accelerators, vLLM is positioned as a layer that works across chip vendors. Large cloud providers already run vLLM themselves, and they have deep pockets and their own managed AI services. Inferact's stated advantage is expertise: the people who know the codebase best now work under one roof.

What comes next turns on practical questions: whether the new capital accelerates upstream work like the Vera Rubin optimizations, how the company prices services on software customers can otherwise self-host, and whether the cloud giants that already run vLLM buy from its maintainers or keep building managed alternatives in-house. Those dynamics, more than the headline valuation, will indicate whether the Databricks playbook transfers to inference.