A new wave of AI infrastructure startups is emerging from university labs and open source communities, and RadixArk is the latest to ride that momentum into a major venture-backed valuation.

According to people familiar with the company, RadixArk, the commercial entity built around the open source project SGLang, has been valued at roughly $400 million in a recent funding round led by Accel. The size of the round has not been confirmed, but the valuation alone is striking for a company that only formally emerged from stealth in the past year.

RadixArk’s core technology, SGLang, is an open source engine designed to make large language models run faster and more cheaply, particularly during inference, the phase when a trained model is actually used to generate outputs. Inference is where AI systems meet real-world demand, and it is also where costs can spiral as companies scale up usage.

SGLang has quickly gained traction among AI developers and infrastructure teams. It is used by companies including xAI and Cursor to accelerate model serving and experimentation. The project began in 2023 inside the UC Berkeley lab of Ion Stoica, the computer science professor best known as a co-founder of Databricks and a serial incubator of infrastructure startups.

As SGLang’s adoption grew, part of the core team maintaining the project transitioned into RadixArk, formalizing a commercial effort around the technology. The company has already raised angel funding from a group of prominent backers, including Intel veteran and investor Lip-Bu Tan, according to people familiar with the matter.

Leading the new company is Ying Sheng, a key contributor to SGLang. Sheng previously worked as an engineer at Elon Musk’s AI startup xAI and earlier as a research scientist at Databricks. She left xAI to become co-founder and CEO of RadixArk, a move she disclosed in a public LinkedIn post. Sheng and the company’s investors have not publicly commented on the latest funding or valuation.

RadixArk’s focus is squarely on inference optimization. While training large models attracts much of the public attention, inference is where the economics of AI are being fought. Every query to a chatbot, every code completion, every AI-generated image or document is an inference call that consumes GPU time and energy. For companies operating at scale, even small efficiency gains can translate into millions of dollars in savings.

SGLang and RadixArk aim to deliver those gains by improving how models are scheduled, batched, and executed on existing hardware. By squeezing more useful work out of the same GPUs, they promise to reduce latency and cost per token, making it more feasible for startups and enterprises to deploy sophisticated models in production.

RadixArk is not alone in turning an academic open source project into a high-value commercial platform. Another inference engine, vLLM, has followed a similar path. Also incubated in Stoica’s UC Berkeley lab, vLLM has become one of the most widely used systems for serving large language models efficiently. It underpins inference workloads at several major technology companies and cloud providers.

vLLM has been in talks to raise a large venture round that could value the company at around $1 billion, according to prior reporting. Multiple people familiar with those discussions have said that Andreessen Horowitz is leading the investment, though the final terms have not been publicly disclosed. vLLM co-founder Simon Mo has pushed back on some of the reported details, calling them factually inaccurate without specifying which elements he disputes.

The parallel trajectories of SGLang and vLLM underscore how Stoica’s lab has become a pipeline for commercial AI infrastructure ventures. Databricks itself emerged from his work on Apache Spark, and the pattern is repeating in the era of generative AI: build a powerful open source engine, attract a developer community, then spin out a company to offer managed services and enterprise features.

For RadixArk, the open source SGLang project remains central. The company continues to develop SGLang as a freely available AI model engine, while layering commercial offerings on top. One of those is Miles, a specialized framework for reinforcement learning. Miles is designed to help organizations train models that can improve over time based on feedback and interaction, a capability that is increasingly important for applications like autonomous agents, recommendation systems, and adaptive copilots.

Most of RadixArk’s tools are still free to use, but the company has begun charging for hosting and managed services, according to a person familiar with its plans. That model mirrors the playbook of many successful open source infrastructure companies: keep the core project open to maximize adoption, then monetize through cloud offerings, support, and advanced features.

The timing for such a strategy appears favorable. Inference infrastructure has become one of the hottest segments in AI investing as the industry shifts from proof-of-concept experiments to large-scale deployment. While model training remains capital intensive, the recurring cost of serving models to millions of users is what determines long-term margins for AI-native businesses.

Investors are betting heavily that the companies that can make inference cheaper, faster, and more reliable will capture a significant share of that value. Baseten, which provides tools and infrastructure for deploying AI models, recently raised a large round at a multibillion-dollar valuation. Fireworks AI, another inference-focused startup, has also secured hundreds of millions of dollars in funding at a multibillion-dollar valuation.

These companies, along with RadixArk and vLLM, are competing to become the default layer that developers reach for when they need to serve models in production. Some focus on providing a full-stack platform, while others concentrate on the low-level performance optimizations that can be integrated into existing workflows. All are chasing the same underlying opportunity: the explosion of demand for inference capacity as generative AI moves into mainstream products.

RadixArk’s $400 million valuation suggests that investors see SGLang as a credible contender in that race. Its roots in a respected academic lab, its early adoption by high-profile AI teams, and its dual focus on open source and commercial offerings give it a familiar but powerful narrative in the infrastructure world.