RadixArk raises $100M seed to scale inference engine

RadixArk announced a $100 million seed financing at a $400 million valuation, per reporting in The Wall Street Journal and a May 5 investor note from HOF Capital. The company is built around an open-source inference engine called SGLang, which HOF Capital's writeup says already serves "trillions of tokens a day" for large users. The Wall Street Journal reports that SGLang functions as a middle layer between models and hardware and that RadixArk has attracted investment tied to Nvidia.

Reporting by The Wall Street Journal describes SGLang as a software layer that aims to reduce AI inference memory overhead by better utilizing short-term memory, which in turn lowers overall compute requirements during inference and training. HOF Capital's investor note frames the RadixArk stack as an open inference engine plus an accompanying framework for large-scale reinforcement learning, and highlights the engine's existing high-throughput token serving.

Editorial analysis: Investors and practitioners have increasingly focused on the infrastructure layer between accelerators and model runtimes as a lever to reduce costs and latency. Open-source inference engines have become a common mechanism for organizations that prefer self-hosting over per-token APIs, and substantial seed capital for such projects reflects belief among some backers that software-level efficiency can materially change cloud and on-prem compute economics.

Editorial analysis: A $100 million seed at a $400 million valuation is unusually large for a seed round, indicating strong investor conviction in the market opportunity for inference optimization. Nvidia-linked funding, as reported by the WSJ, places this effort among a cohort of startups seeking to complement or extend accelerator hardware through software and middleware. For enterprises running high-volume or latency-sensitive workloads, lower memory usage and improved token throughput could reduce operating costs or enable new real-time applications.

Editorial analysis: Teams building high-throughput production systems should watch open inference-engine developments because they affect choices around model placement, batching strategies, and hardware provisioning. Software that reduces memory pressure can change tradeoffs between model size, context window, and inference batch size, which in turn affects latency and per-request cost.

Editorial analysis: Observers should look for:

• •SGLang release cadence and adoption signals in public repositories and release notes
• •performance benchmarks from neutral testers comparing memory use and throughput against mainstream runtimes
• •partnerships or OEM agreements tying middleware to specific accelerators or cloud offerings. HOF Capital's investor note and the WSJ report are the primary published sources describing the round and the product at launch; RadixArk has not been quoted directly in the scraped coverage available to LDS

Editorial analysis: The round underscores persistent investor interest in software-first approaches to inference efficiency. If SGLang delivers on the performance claims reported by investors and journalists, it could accelerate adoption of self-hosted inference stacks among users with sustained token volume.