RadixArk raises $100M seed to make AI inference faster and cheaper

What's the deal? RadixArk, a Palo Alto-based AI infrastructure startup, has launched with $100M in seed funding at a $400M valuation. The round was led by Accel and co-led by Spark Capital, with participation from Nvidia's NVenturesDealroom has a profile for this one. Try Dealroom →, Broadcom chief executive Hock TanDealroom has a profile for this one. Try Dealroom →, AMDDealroom has a profile for this one. Try Dealroom →, and others.

RadixArk was founded by the core developers of SGLangDealroom has a profile for this one. Try Dealroom →, an open-source inference engine they built within the LMSYS OrgDealroom has a profile for this one. Try Dealroom → ecosystem — a UC BerkeleyDealroom has a profile for this one. Try Dealroom →-affiliated nonprofit that also incubated Chatbot ArenaDealroom has a profile for this one. Try Dealroom →.

CEO Ying ShengDealroom has a profile for this one. Try Dealroom → left xAIDealroom has a profile for this one. Try Dealroom → in August 2025 to co-found the company. SGLang sits as a software layer between AI models and the hardware they run on, making inference faster and cheaper through smarter management of KV cache — the short-term memory AI models use to avoid reprocessing identical information.

By organising cached data into a structure called a Radix tree, SGLang recognises when a new query overlaps with something already processed and skips redundant computation, cutting the cost of running AI models substantially.

Why now? AI inference has become the dominant and fastest-growing AI workload as deployment scales globally. The constraint is no longer model capability but the cost and efficiency of serving those models at scale. Hardware alone cannot solve the problem fast enough — software optimisation has become equally critical. SGLang already has a large developer community built over two years of open-source development, giving RadixArk a distribution advantage most seed-stage companies take years to build.

Nvidia's participation through NVentures is strategically notable. The chip giant has a direct interest in software that makes its GPUs more efficient, extending their effective capacity without requiring additional hardware investment.

What could go wrong? RadixArk operates in a layer that large cloud providers and AI labs are increasingly building in-house. If Amazon, Google, or the frontier AI labs optimise their own inference stacks at the same layer SGLang targets, RadixArk's competitive position could erode.

Open-source projects also face a structural challenge: widespread adoption does not automatically translate into revenue, and converting a popular open-source tool into a sustainable commercial business requires a go-to-market motion that is distinct from building good software.

The signal: RadixArk's $100M seed reflects a broader recognition that the AI infrastructure stack is not finished — and that software efficiency gains are as valuable as hardware advances in the race to scale AI at lower cost.

The angel investor list is unusually credible for a seed-stage company. Among others, it includes:

The list signals that some of the AI industry's most respected builders believe RadixArk is solving a real and urgent problem.

Sources:
RadixArk
Business Wire
Wall Street Journal
Pulse2
Let's Data Science
San Jose Business Journal
Ivan Zhou, X 
Ying Sheng, LinkedIn
Ying Sheng, LinkedIn

Image credit:
RadixArk

J.V.