Two researchers left Anthropic in December 2025. Six months later, they’re sitting on $200 million and a $1 billion valuation for a company that doesn’t have a public product yet. Mirendil is one of the most closely watched bets in AI right now, and the reason comes down to one idea: using AI to build better AI, faster.
What Mirendil Is Actually Building
Founders Behnam Neyshabur and Harsh Mehta aren’t newcomers chasing a trend. Neyshabur spent years working on AI for science at both Google and Anthropic, while Mehta built the internal automation platform Anthropic’s own research teams relied on.
Automating the Research Loop
Their pitch centers on recursive self-improvement, systems that can propose experiments, write and run code, debug failures, and decide what to try next with minimal human steering. Every major AI lab has built some version of this internally. None of them sell it.
That gap is the business. Mirendil wants to hand smaller labs, universities, and research institutions the same kind of tooling that OpenAI, Google DeepMind, and Anthropic keep locked inside their own walls.
The Funding and the Firepower Behind It
The seed round itself is unusual. Andreessen Horowitz and Kleiner Perkins co-led the $200 million raise, with NVIDIA also chipping in, an eyebrow-raising amount of capital for a company that hasn’t shipped a product.
Where the Money Goes
Mirendil says the funding will go toward securing high-performance GPU compute clusters, building out scientific data-ingestion pipelines, and hiring researchers pulled from Anthropic, OpenAI, Google DeepMind, and xAI. Compute is clearly the bottleneck the founders are racing to solve first, since training systems that can iterate on their own research is enormously resource-hungry.
That compute-first priority is worth watching. As Mirendil scales, its choice of cloud infrastructure partner, whether that’s Google Cloud’s TPU fleet, NVIDIA-backed GPU clusters, or a multi-cloud mix, will shape how fast it can move from research demo to real product.
Why the Industry Is Paying Attention
A16z’s Matt Bornstein framed the logic bluntly: major labs are “rational economic actors” for keeping this technology to themselves. Mirendil is betting there’s room for an independent player willing to sell what the labs won’t.
It’s not uncontested territory. Anthropic itself has pointed to recursive self-improvement as a genuine risk, the concern being a model that rewrites its own code with too little oversight. Mirendil’s founders see it differently, calling it a supervisable shortcut to faster science rather than a danger to avoid.
The Takeaway
Mirendil is pre-product, pre-revenue, and already valued at $1 billion, a bet that says more about investor conviction than proven results. Whether the self-improving AI space delivers on that promise, or runs into the same safety questions the big labs have been wrestling with internally, is the story worth following from here.
