The numbers arrived with the kind of precision that makes venture capitalists wince: $100 million raised, $400 million valuation. RadixArk, barely six months old as a corporate entity, announced its seed round on May 5, 2026, with the sort of metrics that either signal the next essential piece of AI infrastructure or the frothiest excesses of a market that's forgotten how to count.
The Palo Alto startup emerged last fall from UC Berkeley's LMSys lab—the same research ecosystem that's become something of a factory for AI infrastructure companies. What RadixArk brought with it was SGLang, an open-source inference engine that the company says now powers hundreds of thousands of GPUs for a client list including Google, Microsoft, NVIDIA, Oracle, and xAI. Whether those deployments translate into a durable business remains the question investors are betting $100 million to answer.
Accel led the round. Spark Capital co-led. The roster beyond that reads less like a cap table and more like a technical validation stamp: NVentures (NVIDIA's venture arm), AMD, MediaTek, Salience Capital, HOF Capital, and a collection of others who've made fortunes—or at least reputations—in chips and infrastructure. The angels might be even more telling. Intel CEO Lip-Bu Tan, Broadcom's Hock Tan, xAI co-founder Igor Babuschkin, OpenAI's John Schulman, PyTorch creator Soumith Chintala. When the people who build the picks and shovels start investing in the mine, you pay attention.
Speed-Running from Lab to Unicorn Territory
RadixArk incorporated in Delaware on October 27, 2025. By January, TechCrunch had already caught wind of the valuation—around $400 million—and reported that the team had begun charging for hosting services before the ink was dry on the corporate paperwork. There's a certain audacity to that timeline. Most startups spend their first year figuring out what they're selling; RadixArk arrived knowing exactly what it had.
The company is led by CEO Ying Sheng, who cut his teeth as an engineer at xAI and a research scientist at Databricks, and CTO Banghua Zhu, a UC Berkeley PhD who co-founded Nexusflow AI in 2023 and is somehow also joining the University of Washington as an assistant professor. That dual-track academic-commercial path isn't unusual in AI, but it does raise eyebrows about where priorities will land when research deadlines collide with product roadmaps.
By early May, LinkedIn showed 25 employees (the platform's 11-50 band leaves room for interpretation), with active postings for infrastructure engineers, full-stack developers, and cluster operators. Building out fast, in other words—perhaps faster than the founders initially expected.
The Technical Pitch: Making Inference Cheaper and Faster
SGLang first appeared in a December 2023 paper as a solution to one of AI's most expensive problems: inference. The engine uses techniques like RadixAttention for KV-cache reuse and compressed finite state machines to speed up structured output decoding. If that sounds like jargon, here's the translation: it helps companies run large language models more efficiently, which means lower costs and faster responses. In an industry where compute bills can run into the millions monthly, shaving seconds and dollars off each query adds up.
The GitHub repository has accumulated roughly 27,000 stars as of early May 2026—a meaningful signal in open source, though stars don't pay salaries. RadixArk says SGLang offers day-zero compatibility with most major model families: Llama, Qwen, DeepSeek, GLM, Gemma, Mistral. It runs on NVIDIA and AMD GPUs, Intel CPUs, even Google TPUs. That cross-platform flexibility matters when cloud providers are locked in their own hardware wars.
The company also stewards Miles, an open-source reinforcement learning framework introduced last November. In March, RadixArk added ROCm support for Miles, enabling deployment on AMD's Instinct MI300 and MI350-class clusters. AMD didn't just invest; they integrated. That's the kind of partnership that suggests deeper strategic alignment, not just a check written at a partner meeting.
RadixArk claims SGLang processes "trillions of tokens daily" across "hundreds of thousands of GPUs worldwide." Those figures come from the company, not independent auditors, and in the startup world, self-reported scale has a way of bending under scrutiny. Named deployments include LinkedIn, Nebius, Thinking Machines Lab, and humans&, alongside the hyperscaler clients. Real enough, but whether those installations represent deep integration or experimental deployments remains unclear.
The $100 Million Question
The funding will go toward expanding SGLang, adding support for what the company calls "frontier hardware" and emerging model architectures, and building an end-to-end managed platform spanning training, fine-tuning, reinforcement learning, and large-scale inference. RadixArk pitches customer ownership and control—no lock-in, no proprietary traps. It's a message tailored for enterprises exhausted by vendor dependencies and cloud bills that spiral without warning.
The competitive context is brutal. Fireworks AI raised $250 million at a $4 billion valuation late last year. Baseten reportedly closed $300 million at $5 billion in January. The vLLM project—another Berkeley spinout and direct SGLang competitor—was said to be negotiating roughly $160 million at around $1 billion, though vLLM's co-founder later disputed some of those details. RadixArk's $400 million valuation lands it squarely in the middle tier of a market that's consolidating faster than anyone anticipated six months ago.
Whether $100 million at that price is shrewd or reckless depends entirely on execution. SGLang's open-source momentum is legitimate. The investor roster suggests serious technical due diligence. The Berkeley pedigree aligns with previous infrastructure winners—companies that became essential because they solved real problems elegantly. But there's a canyon between GitHub stars and sustainable revenue, between cloud deployments and margin-positive managed services.
RadixArk is betting it can cross that divide. So are Accel, Spark, and a long list of investors who've seen this movie before. Some of them made fortunes on similar bets. Others are still waiting for the credits to roll.
