Mirendil signs $100M-plus Google Cloud deal to automate AI research
The multiyear contract is worth at least half Mirendil's seed financing, although the frontier AI lab has yet to release a public product, benchmark or customer deployment.
By Ryan Merket · Published
Why it matters
Mirendil is making a cloud commitment worth at least half its $200 million seed financing before releasing a model, benchmark or commercial product. The agreement shows how much infrastructure automated AI research can require and how early cloud providers are competing for newly funded labs.
Mirendil co-founders Behnam Neyshabur and Harsh Mehta have signed a multiyear Google Cloud agreement worth more than $100 million, giving the young frontier AI lab access to TPUs, NVIDIA GPUs and managed training clusters for its effort to automate AI research.
The infrastructure deal arrives before Mirendil has released a public model, benchmark or commercial product. The San Francisco-based company has also disclosed no customers, usage figures, revenue or pricing, leaving investors and prospective scientific users without public evidence that its automated research loop works outside the lab.
TechCrunch reported the agreement on August 6. Neyshabur disclosed its value to the publication. Neither side has published the exact contract length, payment schedule, usage minimums or amount of any cloud credits.
The deal turns Mirendil's founding thesis into a large infrastructure commitment less than two months after its public launch. Mirendil announced a $200 million seed round on June 24, led by Andreessen Horowitz and Kleiner Perkins, followed by an investment from NVIDIA and other backers. TechCrunch reported that the financing valued Mirendil at $1 billion.
At more than $100 million, the Google Cloud agreement is worth at least half as much as the capital Mirendil raised. That comparison does not establish how much cash Mirendil will pay upfront because the companies have not disclosed the contract's financial structure. It does illustrate the cost of pursuing automated frontier research before Mirendil has supplied public product or performance data.
A research loop built from experience inside major labs
Neyshabur and Mehta developed their approach after working inside Google and Anthropic. Neyshabur served as a senior staff research scientist and Gemini team lead and co-led Google's Blueshift group before joining Anthropic in late 2024, according to his biography. His earlier research included postdoctoral work with Yann LeCun at New York University and Sanjeev Arora at the Institute for Advanced Study.
At Anthropic, Neyshabur co-led a Discovery group pursuing an AI scientist or engineer capable of long-horizon technical work, according to his biography. Mehta built the first version of Anthropic's internal automated AI research platform as a one-person project before the effort expanded, Andreessen Horowitz said.
Their co-founders extend that experience across other model-development teams. Shayan Salehian worked as a core machine-learning engineer at xAI, according to Andreessen Horowitz materials. Tara Rezaei is also identified as a co-founder in the investor's account. Mirendil says its founding group includes 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind and OpenAI.
Mirendil is designing models to generate research ideas, implement experiments, evaluate results and feed those findings into the next cycle of work. The intended system goes beyond a coding assistant. Mirendil wants to rebuild the lab around agents that can run parts of the research process continuously, with less human coordination between experiments.
"You can have a self-improving AI where you can point a problem at it and it keeps getting better with time," Neyshabur told TechCrunch. He used Alzheimer's disease as an example of a field where a system could accumulate relevant knowledge and improve its performance over repeated research cycles.
Mirendil has described systems that improve a research process and their performance on assigned problems. It has not publicly demonstrated unrestricted recursive self-improvement, in which an AI independently redesigns its core capabilities through successive generations. The Google Cloud contract supplies infrastructure for Mirendil to pursue its research plan; it does not validate the technical thesis.
Google gives Mirendil two chip paths
Mirendil's infrastructure choice reflects the variety of work inside that proposed loop. Pre-training, post-training, reinforcement learning, inference and parallel experiment execution can place different demands on accelerators and supporting systems.
In Mirendil's announcement, Mehta and Neyshabur said its software would match workloads with suitable available hardware. The agreement covers access to Google TPUs, NVIDIA GPUs and managed training clusters, according to TechCrunch. Using both accelerator families could reduce the pressure to design Mirendil's research stack around one chip supplier.
The arrangement gives Google Cloud a newly financed AI lab as cloud providers compete for startups whose compute needs could expand. Mirendil gains access to cloud infrastructure without first building data centers. The commercial outcome remains untested: Mirendil has announced no customers, public product, pricing or general availability.
Neyshabur told TechCrunch that Mirendil's software could eventually help Google Cloud customers use its hardware more effectively. That remains a prospective distribution path rather than an announced product or sales agreement.
Rivals are funding different versions of automated research
Mirendil is entering a well-funded field. Recursive Superintelligence, which is also pursuing systems that automate ideation, implementation and validation, emerged with $650 million and recruited researchers including Richard Socher and Peter Norvig. Its financing gives it substantially more initial capital than Mirendil, although neither company's public fundraising establishes that recursive or self-accelerating research systems work reliably.
Other companies are packaging narrower versions of the idea. Autoscience raised $14 million in March 2026 and presents its automated AI research lab as a managed service. rekursiv.ai, an AI-research automation startup in its Summer 2026 batch, is focused on AI scientists and machine-learning research. Mirendil's stated target extends from automating AI development to eventually serving laboratories in drug discovery, chemistry, biology and robotics.
The next milestone is evidence
Mirendil's founders have assembled experience from leading AI labs, a $200 million seed round and a cloud contract worth more than $100 million. Investors are funding the team before product evidence has arrived, while Google Cloud is securing a large agreement with a laboratory whose future workloads could grow if its approach succeeds.
The technical burden extends across the research system. Mirendil would need agents that can design and execute experiments, evaluations that distinguish genuine advances from plausible-looking failures, and infrastructure capable of coordinating repeated runs. Scientific users would also need reproducible results before relying on the system for consequential research.
The Google Cloud deal gives Neyshabur, Mehta, Salehian and Rezaei access to the chip and cluster options required to pursue that work. Mirendil's next meaningful proof will come from a model, benchmark, reproducible research result or outside deployment showing that its automated loop can deliver measurable progress.