The deal underscores a dual trend in the artificial intelligence sector: cloud providers are aggressively courting emerging AI labs with hefty infrastructure commitments, while startups are racing to lock down AI compute capacity. For Mirendil, the $100 million-plus commitment represents roughly half of the seed funding it raised at a $1 billion valuation in late June. The agreement highlights how quickly capital is moving from funding model development to securing the physical infrastructure needed to run it.
Scaling Recursive Self-Improvement
Mirendil focuses on recursive self-improvement, an AI concept where systems continuously learn and enhance their own performance. Co-founded by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, the startup aims to build AI capable of eventually taking over the workload of an entire frontier AI lab. A handful of other startups, including Recursive Superintelligence and Ricursive Intelligence, have also recently launched to tackle this specific challenge.
Neyshabur, Mirendil’s CEO, envisions a system that mimics human scientists by accumulating knowledge and gradually improving its output. “You can have a self-improving AI where you can point a problem at it and it keeps getting better with time,” he said. The technology could eventually automate significant portions of scientific research across fields like medicine, biology, and materials science.
“How can we have an AI system that keeps doing research, keeps improving its own knowledge and performance when it comes to Alzheimer’s disease?” Neyshabur asked. “This technology allows us to set goals that are ambitious for AI, and the AI would keep making progress.”
Hardware Flexibility and Google Cloud Strategy
Training self-improving AI demands enormous computing power, making infrastructure costs a primary barrier to entry. The Google Cloud deal provides Mirendil with access to both Google’s custom TPUs and Nvidia GPUs, alongside managed training clusters. Mehta noted that modern training requires matching specific workloads to the right hardware to optimize efficiency and control spiraling costs.
“These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips,” Mehta said. “[Google] provides multiple kinds of chips … This flexibility allows us to ultimately mix and match workloads with the right kind of accelerators, and then lower the cost not just for us, but also for our customers using our systems.”
This hardware flexibility forms the core of Google’s AI infrastructure pitch. Amin Vahdat, SVP and chief technologist of AI and infrastructure at Google, stated that AI advancement now extends beyond chip-level performance. The focus has shifted to how providers “orchestrate entire systems of intelligence and break through the physical constraints of scaling.”
What Happens Next
Mirendil’s software layer acts as an optimizer for Google’s hardware, potentially giving the cloud provider a competitive edge against rivals. In exchange for the compute capacity, Google secures a strategic partner building frontier recursive self-improving AI. As Mirendil scales its models, Google will likely position this technology to sell to enterprise customers seeking advanced automation.
Watch for Mirendil to leverage its new compute resources to hit specific research milestones in medicine and materials science. If the startup successfully demonstrates an AI system that continuously learns without human intervention, it could trigger a wave of similar infrastructure deals across the industry as competitors scramble to match its capabilities.
— David Kim, technology desk, AXO News