How SGLang’s $100M Seed Funding Powers the Next‑Gen Open AI Infrastructure

On May 5, AI‑infra startup RadixArk announced a $100 million seed round, valuing the company at $400 million post‑money. The round was led by Accel and Spark Capital and included investors such as NVIDIA’s NVentures, AMD, MediaTek, Broadcom, Intel, as well as industry figures like Intel CEO Chen Liwu and OpenAI co‑founder John Schulman.

SGLang, the open‑source LLM inference project launched in 2023, quickly became a de‑facto standard: it has amassed over 27 k GitHub stars, runs on more than 400 k GPUs, and processes tens of trillions of tokens daily for users including Google, Microsoft, NVIDIA, Oracle, AMD, LinkedIn, xAI and Thinking Machines Lab. The project maintains day‑0 compatibility across major model evolutions (MoE, long‑context, reasoning, multimodal) through innovations such as the RadixAttention prefix‑cache that organizes shared prefixes into a tree and deduplicates KV storage.

The founding team combines deep system and algorithm expertise: CEO Ying Sheng (Stanford PhD, former Databricks and xAI engineer) and CTO Banghua Zhu (UC Berkeley PhD, former Nvidia Principal Research Scientist). A hardware‑vendor technical lead described the team as “the most valuable founder combination in AI‑infra history.”

Beyond inference, RadixArk released the Miles reinforcement‑learning framework in November 2025, targeting large‑scale RL stability and efficiency; it is already used by over 20 teams for MoE model training. In April 2025 the team demonstrated simultaneous support for DeepSeek‑V4 inference and RL training, leveraging system‑level optimizations such as the ShadowRadix prefix cache, a single‑chip Flash Compressor, and a Lightning TopK that reduces latency to 15 µs, while bridging FP8 inference to BF16 training.

The investor mix signals a strategic bet: hardware giants recognize that a hardware‑agnostic, open‑source inference system can unlock higher performance across heterogeneous platforms, while AI leaders anticipate a unified training‑to‑inference infrastructure. RadixArk’s vision is to make AI infrastructure a public utility—reliable, widely accessible, and not locked behind private systems.

To reward the community, RadixArk offers free usage credits on its hosted platform to anyone who stars the SGLang GitHub repository. The company is also hiring across system, model, compiler, kernel, scheduler and evaluation domains.