RadixArk spins off SGLang at $400M valuation amid inference boom - AI News Today Recency

# RadixArk Spins Off SGLang at $400M Valuation Amid Inference Boom

The AI infrastructure landscape is experiencing a major shift as RadixArk, a newly commercialized startup built on the popular open-source SGLang project, has secured a $400 million valuation in a funding round led by Accel[1]. This spinoff represents a significant moment in the race to optimize AI model inference, as companies increasingly recognize that inference optimization can deliver enormous cost savings almost immediately.

RadixArk emerged from SGLang, an open-source tool originally developed in 2023 inside UC Berkeley's lab by Databricks co-founder Ion Stoica[1]. The company was officially announced last August and has now attracted major backing from prominent investors, including Intel CEO Lip-Bu Tan, who provided angel capital in earlier rounds[1]. The transition of SGLang's core team to RadixArk signals a maturation of the inference optimization market, where demand for faster and more efficient AI model serving has become critical for enterprises managing massive server costs.

The Rise of Inference Optimization as a Market Opportunity

Inference processing — the computational work required to run trained AI models and generate predictions — has become one of the largest cost drivers in AI operations[1]. Unlike model training, which happens once, inference occurs continuously as users interact with AI systems. This means optimization tools that accelerate inference can create immediate and substantial savings for companies deploying AI at scale.

The market opportunity is so compelling that competing inference optimization projects are also transitioning to commercial ventures. vLLM, a more mature inference optimization project, has similarly shifted from open-source to a startup model and is in discussions about raising upwards of $160 million at a $1 billion valuation[1]. This trend underscores how critical inference optimization has become in the broader AI infrastructure ecosystem.

RadixArk's Technology and Business Model

RadixArk is building a comprehensive infrastructure platform centered on two core technologies: SGLang for inference and Miles for reinforcement learning training[1]. The company describes itself as "infrastructure-first," aiming to democratize access to the advanced systems that currently exist only within the largest AI companies[2].

SGLang is positioned as "the fastest, most flexible open engine for serving modern models," and RadixArk has committed to continuing its development as an open-source project[2]. Miles, the company's reinforcement learning framework, brings similar rigor to post-training that modern serving engines brought to inference optimization[2]. While most of RadixArk's tools remain free, the company has begun charging fees for hosting services, creating a sustainable business model while maintaining the open-source foundation[1].

The company's long-term vision is ambitious: making frontier AI infrastructure 10x cheaper and 10x more accessible than it currently is[2]. RadixArk envisions a future where world-class infrastructure becomes a shared foundation rather than a competitive advantage locked within individual companies[2].

Key Leadership and Strategic Backing

Ying Sheng, a key contributor to SGLang and former engineer at xAI (Elon Musk's AI startup), has transitioned to become RadixArk's co-founder and CEO[1]. Sheng previously worked as a research scientist at Databricks, bringing deep expertise in both infrastructure engineering and AI systems. Her move from xAI to lead RadixArk demonstrates the talent flow occurring as inference optimization becomes a standalone business priority.

The investor backing reflects confidence in the market opportunity. Beyond Accel's lead investment, Intel CEO Lip-Bu Tan's participation signals hardware manufacturers' interest in software solutions that optimize their chips' utilization[1]. This alignment between hardware and software optimization creates a natural partnership dynamic in the emerging inference infrastructure market.

Frequently Asked Questions

What is SGLang and why does it matter?

SGLang is an open-source serving framework that optimizes how AI models run on hardware[4]. It makes models execute faster and more efficiently without requiring more powerful computers, directly reducing the operational costs of running AI systems[1]. Companies like xAI and Cursor use SGLang to accelerate their AI operations[1].

How does RadixArk make money if SGLang remains open-source?

While SGLang continues as a free, open-source project, RadixArk charges for managed hosting services and infrastructure tooling built on top of the core engine[1]. This model allows the company to monetize while maintaining community trust and contribution to the open-source foundation[1].

What is the difference between SGLang and vLLM?

Both SGLang and vLLM are inference optimization tools, but vLLM is described as the more mature project[1]. SGLang is positioned as faster and more flexible, while vLLM has established broader adoption. The fact that both projects have spun into commercial ventures demonstrates the market's recognition of inference optimization's critical importance[1].

What does Miles do, and how does it fit into RadixArk's strategy?

Miles is RadixArk's open-source framework for large-scale reinforcement learning training[1]. While SGLang handles the inference side, Miles addresses the training side, allowing RadixArk to provide a more complete infrastructure solution for companies building and deploying frontier AI models[2].

Why is inference optimization becoming a standalone business?

Inference represents a massive portion of server costs for AI services, often exceeding training costs in production systems[1]. Tools that optimize inference can deliver enormous savings almost immediately, making them highly valuable to enterprises. This economic reality has made inference optimization attractive enough to support dedicated companies[1].

What is RadixArk's ultimate goal?

RadixArk aims to make frontier AI infrastructure 10x cheaper and 10x more accessible, shifting the competitive advantage from proprietary internal systems to world-class shared infrastructure[2]. The company believes the next generation of AI breakthroughs will be defined by who builds the best applications on top of shared systems, not by who owns the best private infrastructure[2].