A growing trend in the AI infrastructure space is seeing open source projects evolve into well-funded startups, and RadixArk is the latest example. The company, which emerged from the open source project SGLang, has reportedly been valued at $400 million in a recent funding round led by Accel, according to two people familiar with the matter. TechCrunch could not independently confirm the exact size of the investment. RadixArk was launched in August of last year as a commercial spinout from a research lab at UC Berkeley, where SGLang was originally developed under the guidance of Ion Stoica, co-founder of Databricks. The tool is designed to optimize inference—the process of running AI models after training—making them faster and more efficient on existing hardware. This optimization directly reduces server costs, a major concern for companies deploying AI at scale. The transition from open source to startup has been driven by increasing demand. SGLang is already used by major AI companies like xAI and Cursor to accelerate model inference. Ying Sheng, a key developer of SGLang and former engineer at xAI, has joined RadixArk as co-founder and CEO, according to a LinkedIn post she shared last month. Sheng previously worked as a research scientist at Databricks. She, Accel, and Intel CEO Lip-Bu Tan—another early investor—did not respond to requests for comment. SGLang and its commercial counterpart, RadixArk, focus on improving inference efficiency, a critical component of AI infrastructure. With model training and inference accounting for the bulk of AI computing costs, tools that enhance performance can deliver immediate savings. RadixArk continues to maintain SGLang as an open source project while building proprietary offerings. One such product is Miles, a new framework tailored for reinforcement learning—enabling AI systems to improve over time through experience. While core tools remain free, RadixArk has begun introducing paid hosting services, a person familiar with the company said. RadixArk is not alone in this shift. vLLM, another inference-optimization project from the same UC Berkeley lab, has also transitioned into a commercial venture. According to reports, vLLM is in talks to raise up to $160 million at a $1 billion valuation, with Andreessen Horowitz reportedly leading the round. Though vLLM co-founder Simon Mo called some details of the report “factually inaccurate,” he declined to clarify which parts were incorrect. Andreessen Horowitz did not comment. Both SGLang and vLLM were incubated under Ion Stoica’s lab, which has become a hub for AI infrastructure innovation. Brittany Walker, a general partner at CRV, noted that several large companies already rely on vLLM for inference workloads, and SGLang has gained rapid traction in recent months, despite being relatively new. The broader inference infrastructure market is heating up. Baseten recently raised $300 million at a $5 billion valuation, while Fireworks AI secured $250 million at a $4 billion valuation in October. These developments highlight the growing importance of efficient, scalable inference solutions as AI adoption expands across industries.