For the past three years, SGLang, an open-source project, has processed trillions of tokens daily for companies such as Google, Microsoft, xAI, and Nvidia. Until recently, most people outside the inference community did not know who created it.
RadixArk, a Palo Alto startup bringing SGLang to market, just raised $100 million in seed funding at a $400 million valuation. The round was led by Accel and Spark Capital, with NVentures, Salience Capital, A&E Investment, HOF Capital, Walden Catalyst, AMD, LDVP, WTT Fubon Family, MediaTek, and Databricks joining
Other investors include John Schulman, co-founder of OpenAI; Soumith Chintala, creator of PyTorch; and Thomas Wolf, co-founder of Hugging Face. The CEOs of Intel and Broadcom also joined the round.
RadixArk was founded by Ying Sheng and Banghua Zhu in 2025. Sheng built inference systems for Elon Musk’s Grok models at xAI, and Zhu worked on systems at Nvidia. In 2023, Sheng and he team created SGLang as part of LMSYS research group, a non-profit created by researchers from Stanford, Berkeley, CMU, UCSD, among others.
SGLang became popular in the inference community because of its technical strengths, without any marketing or sales team. Today, it runs on hundreds of thousands of GPUs. Its main competitor is vLLM, another open-source engine from Berkeley that also turned into a funded startup.
SGLang solves a major memory problem in AI inference. Usually, AI models recompute the context for each query, even when most of the prompt is the same. SGLang uses a Radix tree data structure to store previously processed parts, reducing redundant work for new queries. This reduces the per-token computational cost and helps organisations save money when running their own inference.
“Our mission is simple yet ambitious: make frontier-level AI infrastructure open and accessible to everyone. We believe the next generation of AI won’t be defined by who owns the biggest private infrastructure, but by who builds the most meaningful applications on top of shared, world-class systems. We aim to make these systems orders of magnitude cheaper and more accessible, so everyone can build on them,” says Sheng.
The efficiency is at the heart of RadixArk’s mission. It keeps SGLang open and free, but makes money by offering managed hosting, similar to what Databricks and Elastic do.
“RadixArk is building the open foundation for the next era of AI — where companies don’t just consume models, they train and manage them as a core part of product development. By democratising training and inference infrastructure, RadixArk enables any engineer to experiment and innovate at the frontier, fully owning how AI powers their products,” notes Ivan Zhou, partner at Accel.
The new funding will help RadixArk expand to more model types and hardware and grow its managed platform.
