Mirendil Secures $100M+ Google Cloud Deal to Build Self-Improving AI Systems

An emerging AI lab co-founded by researchers from Anthropic has inked a multi-year partnership with Google Cloud valued at more than $100 million to fuel its research into self-improving AI systems. The deal reflects a competitive arms race among cloud infrastructure providers vying to partner with advanced AI startups while those startups race to lock in computing resources as they expand.

The Partnership & Computing Resources#

Mirendil will gain access to Google's tensor processing units (TPUs), Nvidia graphics processing units (GPUs), and managed training clusters—infrastructure essential for developing and refining recursive self-improving AI. Co-founder Benham Neyshabur revealed the partnership value represents roughly half the company's seed funding haul from June, when Mirendil raised capital at a $1 billion valuation.

The infrastructure deal exemplifies a broader dynamic: cloud giants are making substantial commitments to emerging AI companies, while those startups are aggressively securing compute capacity to support their scaling ambitions.

What Self-Improving AI Aims to Do#

Self-improving, or recursive, AI refers to systems capable of iteratively refining their own performance without constant human intervention. Mirendil's vision centers on automating research workflows across scientific domains—from medicine and biology to materials science. Rather than requiring scientists to manually guide every step, the company imagines pointing an AI system at a research problem and letting it progressively deepen its expertise and capabilities over time.

Neyshabur described the concept as: "You can have a self-improving AI where you can point a problem at it and it keeps getting better with time." He offered a concrete example: designing an AI system that continuously pursues Alzheimer's disease research, setting ambitious goals and allowing the system to advance incrementally toward them.

Hardware Flexibility as a Competitive Edge#

Co-founder Harsh Mehta emphasized that effective training of these systems hinges on intelligent workload orchestration—matching computational tasks to the hardware best suited to handle them. Google's diverse accelerator portfolio allows Mirendil to optimize performance while controlling costs, benefiting not only the startup but also customers who will eventually use its software layer atop Google's infrastructure.

Google's AI chief technologist Amin Vahdat positioned the partnership within a broader shift: AI progress now depends less on raw chip performance and more on "orchestrating entire systems of intelligence and breaking through the physical constraints of scaling."