RadixArk Raises $100M to Open Frontier AI Infrastructure

RadixArk officially launched on May 5, 2026 with $100 million in seed funding at a $400 million post-money valuation, a round led by Accel and co-led by Spark Capital with participation from a broad syndicate of corporate and angel backers.

The company says its mission is to build open, production-ready infrastructure for frontier AI so core systems are engineered once and shared widely rather than rebuilt at every lab and company.

RadixArk is rooted in two open-source projects: SGLang, an inference engine, and Miles, a reinforcement learning framework for post-training and large-scale MoE workloads. The startup is positioning those projects as the foundations for a managed platform.

SGLang, the inference engine at the center of RadixArk’s work, is already described by the company as powering trillions of tokens daily and running on hundreds of thousands of GPUs for organizations including Google, Microsoft, NVIDIA, Oracle, AMD, LinkedIn, xAI, Thinking Machines Lab, and humans&.

Miles is explicitly aimed at enterprise reinforcement learning and large-scale mixture-of-experts training, with early roadmaps emphasizing true on-policy RL, memory improvements, and integrations with training backends such as Megatron and FSDP. The project is available on GitHub and was introduced last November by the LMSys-aligned team.

RadixArk was founded by Ying Sheng and Banghua Zhu, veterans of xAI and NVIDIA, who previously helped develop the SGLang project in open communities before incorporating the work into a commercial company.

The startup says it will use the seed capital to expand SGLang and Miles, add support for new model architectures and hardware platforms, and build managed infrastructure and tooling for teams developing AI at production scale. That includes both training and inference capabilities.

Investors span traditional venture firms and strategic corporate backers, reflecting the infrastructure and silicon angle: Accel, Spark Capital, NVentures (NVIDIA’s VC arm), AMD, MediaTek, Salience Capital, HOF Capital, Walden Catalyst and others are listed in the round. Several prominent AI researchers and executives also participated as angels.

RadixArk’s pitch taps into a broader industry trend: open-source inference and training runtimes are being commercialized as companies race to reduce the cost and engineering duplication of running large models in production. The team argues that shared, battle-tested systems will accelerate product development across the AI landscape.

Technically, RadixArk highlights SGLang’s broad model and hardware compatibility and Miles’ focus on production stability as differentiators. SGLang advertises day‑zero support for many model families and GPUs, while Miles promotes tooling for on-policy RL and MoE examples on newer hardware.

Still, execution will be heavy lifting: building and operating a managed stack that supports training, post-training RL, and low-latency inference across multiple clouds and chips requires deep engineering, close hardware partnerships, and enterprise sales. The company’s mix of strategic investors may help bridge those gaps but does not eliminate competitive or operational risk.

For enterprises and labs, the immediate promise is practical: a common, open stack could cut redundant engineering work, reduce vendor lock-in by supporting multiple accelerators, and speed the path from research to production. RadixArk says that is exactly the goal as it scales SGLang and Miles into managed offerings.