NextFin News - In a significant move for the artificial intelligence infrastructure sector, the team behind the open-source project SGLang has officially spun out to form RadixArk, a commercial entity that has reportedly secured a $400 million valuation. According to TechCrunch, the funding round was led by Accel, with participation from high-profile angel investors including Intel CEO Lip-Bu Tan. The transition, which took place on January 21, 2026, marks the formal commercialization of one of the industry’s most critical tools for accelerating AI model inference.
RadixArk is led by Ying Sheng, a former engineer at Elon Musk’s xAI and a research scientist at Databricks, who now serves as co-founder and CEO. The startup originated within the UC Berkeley laboratory of Ion Stoica, a co-founder of Databricks and a prolific figure in the distributed systems space. While SGLang remains a celebrated open-source engine used by major players like xAI and Cursor, RadixArk is expanding its portfolio with "Miles," a specialized framework designed for reinforcement learning. The company has already begun implementing a hybrid business model, offering hosting services for a fee while maintaining its core open-source contributions.
The emergence of RadixArk at such a high valuation reflects a fundamental shift in the AI investment thesis. For the past two years, the market was dominated by "training-heavy" narratives, where the primary goal was building larger and more capable foundational models. However, as these models move into production, the focus has shifted toward the "inference wall"—the massive computational and financial cost of running these models at scale. Inference often accounts for up to 90% of the total lifecycle cost of an AI model, making optimization tools like SGLang and its rival vLLM indispensable for enterprise profitability.
The competitive landscape is intensifying rapidly. vLLM, another UC Berkeley-born project, is reportedly in talks to raise over $160 million at a $1 billion valuation, with Andreessen Horowitz rumored to be leading the round. This surge in capital is not limited to Berkeley spin-outs; other inference infrastructure startups like Baseten and Fireworks AI have recently secured hundreds of millions in funding at multi-billion dollar valuations. The common thread among these companies is the promise of hardware-agnostic efficiency, allowing developers to squeeze more performance out of existing GPU clusters—a critical capability given the ongoing global semiconductor supply constraints.
From a macroeconomic perspective, the success of RadixArk is also tied to the broader industrial policy of the current administration. U.S. President Trump has consistently emphasized the need for American dominance in the AI and semiconductor sectors. The involvement of Tan, who recently reached an agreement with U.S. President Trump to strengthen domestic semiconductor manufacturing under the CHIPS Act, suggests that software-level optimization is now viewed as a strategic pillar of national technological sovereignty. By making AI models run more efficiently, companies like RadixArk effectively increase the "virtual capacity" of the nation’s compute resources without requiring immediate physical expansion of data centers.
Looking ahead, the trajectory for RadixArk and the broader inference market suggests a move toward "vertical integration of the stack." As Sheng and her team develop Miles, they are positioning RadixArk not just as a performance layer, but as a continuous improvement engine. The integration of reinforcement learning into the inference pipeline allows models to become smarter through usage, creating a feedback loop that could eventually replace traditional, static fine-tuning methods. For the venture capital community, the $400 million bet on RadixArk is a bet that the future of AI belongs not to those who build the biggest models, but to those who can run them most efficiently in the real world.
Explore more exclusive insights at nextfin.ai.
