The speed was striking, even by Silicon Valley's occasionally breathless standards.
Behnam Neyshabur and Harsh Mehta left Anthropic in December 2025. Six months later, they'd assembled a 20-person research team plucked from the industry's most prestigious labs and closed a significant seed round of $200 million—one of the larger seed rounds in recent memory—led by Andreessen Horowitz and Kleiner Perkins, with NVIDIA joining as a strategic investor.
Mirendil—the startup the pair founded to automate frontier AI research itself—announced the funding on June 25, putting the finishing touches on what industry watchers had been anticipating since whispers began circulating in March about a deal in the works. At the time, those early reports pegged the valuation discussions around $1 billion, though the final terms were not disclosed. The company hasn't confirmed where the round ultimately priced, but the investor lineup and check size speak to the kind of conviction typically reserved for teams with proven track records at the bleeding edge of machine learning.
Which this team has. Neyshabur, now CEO, was a principal research scientist at Anthropic, where he co-led the Discovery team. Before that, he spent years at Google and DeepMind working on the thorny problems of implicit regularization and generalization in deep learning—the kind of foundational work that doesn't always make headlines but shapes how models actually learn. Mehta, Mirendil's CTO, came from the same Anthropic cohort.
They didn't leave alone. The founding roster reads like a who's-who of frontier AI development: researchers and engineers from Anthropic, xAI, Google DeepMind, and OpenAI. Shayan Salehian joined from xAI, where he'd worked on post-training, reasoning systems, and the agent infrastructure powering Grok models. Tara Rezaei, just 23 and a recent MIT graduate, arrived with stints as an early student researcher at OpenAI and an Olympiad medal to her name. It's the sort of talent concentration that makes investors sit up straight.
The Pitch: AI Labs That Don't Need to Become AI Labs
Mirendil's thesis hinges on a specific friction point in scientific research today. Drug discovery labs, chemistry departments, biotech startups, robotics companies—they all need cutting-edge AI capabilities. The problem? Building and maintaining those capabilities means essentially becoming a frontier AI lab yourself. Hiring the talent, securing the compute, training the models, iterating on architecture. It's expensive, time-consuming, and orthogonal to the actual scientific work.
Mirendil wants to be the infrastructure layer. The company is training what it describes as frontier models specialized for AI research and development, then wrapping them in tooling designed to make R&D faster and more autonomous. "Democratizing frontier AI R&D to accelerate science and technology," as the mission statement goes—though in practice, that means selling access to the kind of systems that have historically been the exclusive domain of a handful of well-funded labs in San Francisco and London.
It's an audacious goal, perhaps more so than the founders initially let on. Industry reporting from June suggests Neyshabur had already initiated and led automated pre-training R&D efforts at Anthropic before his departure. That experience likely shaped Mirendil's roadmap, though the company has been circumspect about technical specifics.
Kleiner Perkins general partner Mamoon Hamid, writing on June 24 to announce his firm's participation, emphasized the team's experience "building large-scale AI systems" and noted that Kleiner backed Mirendil "from day one." The timeline checks out—Mirendil registered as a Delaware corporation on December 10, 2025, immediately following the founders' departure from Anthropic. Mirendil's LinkedIn states an employee range of 11-50, consistent with active hiring, and the careers page suggests aggressive hiring is underway across research and engineering.
Seed Rounds That Make Series Bs Look Modest
Nine-figure seed rounds remain rare enough to warrant a double-take. Mistral AI's €105 million seed in 2023 set records across Europe and signaled a new willingness among investors to write huge checks at the earliest stages—provided the team had the right pedigree and the market had the right appetite. Mirendil's $200 million puts it in similar territory, a reflection of both the capital intensity required to train frontier models and the frenzy around anything adjacent to AI research automation.
The full roster of participating investors hasn't been disclosed. Beyond Andreessen Horowitz, Kleiner Perkins, and NVIDIA, the company notes only that "others" joined the round. That vagueness is typical for deals of this size, where strategic investors or family offices often participate under the radar.
What happens next remains largely opaque. Mirendil hasn't laid out a specific deployment plan for the $200 million—team scaling and infrastructure buildout are the safe bets—but the broader trajectory seems clear. The company is positioning itself to sell AI research capabilities to organizations that can't or won't build them in-house. Whether that thesis holds, and whether six months is enough runway from exit to product-market fit, will become clearer as the company moves from fundraising mode to actually shipping.
For now, the round itself is the story. A testament to how quickly experienced teams can move when the capital markets are open and the hype cycle is running hot.
