The archive · AI & Models · Strategic decision · 2026
Mirendil's $200M bet that self-improving AI can automate scientific discovery
Ex-Anthropic researchers raise one of AI's biggest seed rounds — $200M at $1B from a16z and Kleiner Perkins — to sell recursive self-improving AI for science.
Mirendil
What the business is
A San Francisco AI lab building recursively self-improving models and domain-specific AI tools for scientific discovery, targeting fields like medicine, biology and materials science.
Starting capital：$200M seed at ~$1B valuation (June 2026); a16z and Kleiner Perkins led, Nvidia participated
How it started
Co-founders Behnam Neyshabur and Harsh Mehta met at Google in 2019, joined Anthropic together at the end of 2024 and resigned after the release of Claude Opus 4.5 in December 2025. They founded Mirendil — from the Elvish for 'friend of discovery' — to build AI that iteratively improves itself, betting that is the fastest path to accelerating scientific research. On 24 June 2026 the company announced a $200 million seed round at a $1 billion valuation, led by Andreessen Horowitz and Kleiner Perkins with Nvidia participating; the founding team also includes ex-xAI early member Shayan Salehian and MIT graduate Tara Rezaei.
What happened
In August 2026 Mirendil signed a multiyear Google Cloud partnership worth upward of $100 million — roughly half its seed — for access to Google TPUs and Nvidia GPUs to train self-improving AI. Neyshabur told TechCrunch the ambition is an AI that can eventually take on the work of an entire frontier AI lab: scientists point it at a problem such as Alzheimer's disease and it keeps improving its own knowledge and performance.
No ending yet — it is still running.
Background
Mirendil's bet is that recursive self-improvement — AI systems that iteratively improve their own knowledge and performance — can be turned into a product for science. Founded in 2026 by Behnam Neyshabur and Harsh Mehta, two researchers who left Anthropic after the release of Claude Opus 4.5, the startup aims to give scientists domain-specific AI that automates research in medicine, biology and materials science, pointing it at problems like Alzheimer's and letting it keep getting better with time.
The company announced a $200 million seed round on 24 June 2026 at a $1 billion valuation, led by Andreessen Horowitz and Kleiner Perkins with Nvidia participating — among the largest AI seed rounds on record. The founding team, including ex-xAI early member Shayan Salehian and MIT graduate Tara Rezaei, operates from a downtown San Francisco office with roughly 20 technical staff.
In August 2026 Mirendil signed a multiyear Google Cloud partnership worth upward of $100 million — about half its seed — for Google TPUs and Nvidia GPUs, with co-founder Harsh Mehta arguing flexible multi-chip training lowers cost for Mirendil and its customers. The company named itself from the Elvish for 'friend of discovery' and planned to release models and products within months to gather user feedback.
What has to be true
- Neyshabur and Mehta spent a year at Anthropic building the very capability — self-improving AI — that frontier labs restrict, then bet they could sell it with oversight.
- The $200M seed at a $1B valuation from a16z, Kleiner Perkins and Nvidia shows investors treating recursive self-improvement as the next frontier-AI wedge.
- Scientific discovery is a measurable market: the pitch is automating research rather than replacing researchers, which broadens the addressable buyers.
- The Google Cloud deal ties compute access to the product, letting Mirendil claim both infrastructure and software differentiation before shipping.
What can be applied
When the biggest labs treat a capability as too dangerous to sell, a startup can raise enormous capital on the opposite bet — that the same capability, productized with guardrails, is a business.
Aftermath
As of September 2026 Mirendil remains in building phase: a ~20-person lab with a $200 million seed, a $1 billion valuation and a $100M+ Google Cloud compute deal, still pre-revenue and pre-launch. The company has said it plans to release models and products to gather user feedback; its success will hinge on whether self-improving systems can deliver results in real scientific workflows without the safeguards frontier labs say justify keeping the capability internal.
Sources
- Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI
- Former Anthropic Employee Raises $200 Million to Develop AI capable of Building AI
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card