Nvidia and Accel pour $100M into RadixArk, the open-source engine powering half the AI internet — TFN

For the previous three years, SGLang, an open-source challenge, has processed trillions of tokens day by day for corporations similar to Google, Microsoft, xAI, and Nvidia. Till just lately, most individuals exterior the inference group didn’t know who created it.

RadixArk, a Palo Alto startup bringing SGLang to market, simply raised $100 million in seed funding at a $400 million valuation. The spherical was led by Accel and Spark Capital, with NVentures, Salience Capital, A&E Funding, HOF Capital, Walden Catalyst, AMD, LDVP, WTT Fubon Household, MediaTek, and Databricks becoming a member of

Different buyers embrace 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 additionally joined the spherical.

RadixArk was based by Ying Sheng and Banghua Zhu in 2025. Sheng constructed inference methods for Elon Musk’s Grok fashions at xAI, and Zhu labored on methods at Nvidia. In 2023, Sheng and he workforce created SGLang as a part of LMSYS analysis group, a non-profit created by researchers from Stanford, Berkeley, CMU, UCSD, amongst others.

SGLang grew to become well-liked within the inference group due to its technical strengths, with none advertising or gross sales workforce. In the present day, it runs on a whole lot of hundreds of GPUs. Its fundamental competitor is vLLM, one other open-source engine from Berkeley that additionally was a funded startup.

SGLang solves a significant reminiscence downside in AI inference. Normally, AI fashions recompute the context for every question, even when many of the immediate is identical. SGLang makes use of a Radix tree knowledge construction to retailer beforehand processed elements, decreasing redundant work for brand new queries. This reduces the per-token computational value and helps organisations get monetary savings when working their very own inference.

“Our mission is straightforward but formidable: make frontier-level AI infrastructure open and accessible to everybody. We imagine the subsequent era of AI gained’t be outlined by who owns the most important personal infrastructure, however by who builds probably the most significant purposes on prime of shared, world-class methods. We goal to make these methods orders of magnitude cheaper and extra accessible, so everybody can construct on them,” says Sheng.

The effectivity is on the coronary heart of RadixArk’s mission. It retains SGLang open and free, however makes cash by providing managed internet hosting, much like what Databricks and Elastic do.

“RadixArk is constructing the open basis for the subsequent period of AI — the place corporations don’t simply devour fashions, they prepare and handle them as a core a part of product improvement. By democratising coaching and inference infrastructure, RadixArk allows any engineer to experiment and innovate on the frontier, totally proudly owning how AI powers their merchandise,” notes Ivan Zhou, accomplice at Accel.

The brand new funding will assist RadixArk increase to extra mannequin varieties and {hardware} and develop its managed platform.