Mirendil is in talks to raise up to $1 billion at a $5 billion valuation -- the AI-research startup still has no product, three months after its $200 million seed

Mirendil is negotiating to raise as much as $1 billion at a $5 billion valuation -- a fivefold jump from the $1 billion valuation it closed just three months ago, before the company has shipped anything a customer can use. Kleiner Perkins is in talks to lead the new round, with Andreessen Horowitz, which co-led the original seed, discussing whether to join again.

Mirendil was founded in December 2025 by Behnam Neyshabur and Harsh Mehta, along with co-founders Shayan Salehian and Tara Rezaei, a 23-year-old MIT graduate. Neyshabur, now CEO, spent more than five years at Google DeepMind co-leading reasoning research for Gemini before a stint at Anthropic; Mehta built internal automation tools for Anthropic's own research workflows. The roughly 20-person team, split across a downtown San Francisco office, is trying to build AI models that are themselves expert at AI research and development -- automating the loop of proposing experiments, running code, interpreting results, managing compute and comparing checkpoints with progressively less human input.

The 20-person founding team is drawn from across the frontier-lab landscape, not just Anthropic: reporting on the round has named staff who came from xAI, Google DeepMind and OpenAI as well, alongside Neyshabur and Mehta's own Anthropic and DeepMind backgrounds. That breadth is itself part of the pitch to investors -- a team that has seen how research actually gets done inside four different frontier labs, now building a fifth option that isn't a lab at all, but infrastructure any of them (or their competitors) could eventually rent.

Run the arithmetic on the number itself: a $5 billion valuation on a 20-person company works out to roughly $250 million of paper value per employee, before a single dollar of revenue. That is an extreme even by the standards of a year that has already produced several eye-popping AI valuations -- and it is priced almost entirely on the credibility of two named researchers and a thesis, since there is no product, usage, or revenue figure anywhere in the reporting to point to instead.

Three months is not long for a private valuation to grow fivefold with nothing new to show a customer. “Three months ago, Mirendil didn't have a product,” Tech Funding News wrote of the company. “It still doesn't.” The pitch instead rests on the founders' pedigree and a thesis: that the next real edge in AI won't come from a bigger chatbot, but from a system that can run its own science.

The mission statement points at a real gap: building and running a frontier-scale model is now expensive enough that a drug-discovery startup or a materials-science lab can't easily justify doing it themselves, even when AI-driven research would help. Mirendil's pitch is to be the shared frontier-AI capability those labs rent instead of build -- the same logic that made cloud computing a bigger business than most of the companies that first tried to run their own data centers. Nvidia's participation in the June seed fits that framing: a chipmaker backing a company whose entire premise is that demand for frontier-scale compute is about to broaden well past the handful of labs that currently buy it in bulk.

The company's own stated mission, per its website: “Any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab.” That's the bet Mirendil is selling investors -- that it can rent out the missing capability rather than make every biotech and materials-science startup build a frontier lab of its own.

“Any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab.” — Mirendil, company mission statement

The timing lines up with a claim Anthropic made about its own research five days before Mirendil's talks became public: the company's R&D Automation Index, published Sept. 17, put Claude at AL4 -- “leads” -- for 26% of Anthropic's own measured AI research and development work. Mirendil is effectively betting that capability is worth building as a standalone company rather than an internal lab tool -- and that a team of twenty, operating outside any frontier lab's existing compute and data advantages, can get there faster.

Nothing about the new round is finished. Kleiner Perkins has not confirmed terms publicly, Mirendil has not commented, and both Bloomberg and PYMNTS describe the talks as ongoing rather than closed. A round this size, moving this fast, on a company with nothing shipped, is either an early, correct bet on where the industry's next edge comes from -- or a valuation set almost entirely by two résumés and a moment when investors are chasing anything adjacent to frontier AI.