Mirendil Secures $100M-Plus Google Cloud Deal for Self-Improving AI
Mirendil signs a $100M-plus Google Cloud partnership for TPUs, Nvidia GPUs and managed clusters to advance recursive self-improving AI research at scale.
Summary
Verified facts: AI research startup Mirendil has entered a multiyear cloud-computing partnership with Google Cloud valued at more than $100 million, co-founder and CEO Behnam Neyshabur told TechCrunch. The agreement, reported on August 6, 2026, will provide Mirendil with Google Tensor Processing Units, Nvidia graphics processors and managed training clusters. The companies did not disclose the deal’s precise value, duration, spending schedule or capacity commitments.
The infrastructure commitment is substantial relative to Mirendil’s financing. TechCrunch reported that its value is approximately half the seed capital the startup raised in late June 2026, when investors valued the company at $1 billion. Mirendil was founded by former Anthropic researchers, including Neyshabur and co-founder Harsh Mehta. The agreement illustrates how young AI companies are securing multiple forms of scarce computing capacity while cloud providers compete to attract promising model developers with large, long-term infrastructure packages.
Company-stated goal: Mirendil is researching recursive self-improvement, an approach in which an AI system repeatedly expands its knowledge or enhances its performance. The startup’s ultimate ambition is to build software capable of performing work associated with an entire frontier AI laboratory. Neyshabur said such systems could pursue difficult scientific objectives over time, citing Alzheimer’s disease as an example, while the company also sees possible applications in medicine, biology and materials science. These are prospective use cases rather than demonstrated results; the article supplies no benchmarks, scientific discoveries or deployed customer examples validating those claims.
The hardware mix is central to the partnership. Mehta argued that advanced AI training increasingly depends on assigning different workloads to the accelerators best suited to them. Access to both TPUs and Nvidia GPUs could let Mirendil optimize tasks across chip types, potentially improving utilization and reducing costs for itself and future customers. Google AI infrastructure executive Amin Vahdat similarly framed progress as a systems-orchestration challenge rather than a contest based only on individual chip performance. The arrangement could benefit both sides strategically. Mirendil gains access to costly infrastructure needed for compute-intensive experimentation without depending on one accelerator architecture. Google Cloud gains a heavily funded AI startup that could increase demand for its TPUs while also using Nvidia hardware available through its platform. TechCrunch further suggests that Mirendil’s software layer could help customers use Google’s infrastructure more efficiently and that Google may eventually offer resulting capabilities to enterprise clients, although no commercial product or resale plan was formally detailed.
What happens next remains uncertain. Mirendil must show that recursive improvement can produce reliable, measurable gains rather than merely consuming increasing amounts of compute. It will also need to translate research into safe and economically useful systems. The report does not specify when training will begin at scale, when products might reach customers, how intellectual property will be divided, or whether Google receives preferential access. The deal therefore secures important resources, but it does not establish that Mirendil can achieve its broad scientific or frontier-lab ambitions.
Positives
- The multiyear agreement gives Mirendil access to Google TPUs, Nvidia GPUs and managed training clusters rather than limiting the startup to a single accelerator architecture.
- Google Cloud is committing infrastructure valued at more than $100 million, giving Mirendil substantial capacity for compute-intensive AI research.
- Mirendil says its workload-orchestration software can match tasks with appropriate chips, which could improve hardware utilization and lower operating costs.
- The partnership gives Google Cloud a strategic relationship with a $1 billion AI startup founded by former Anthropic researchers.
- Mirendil is targeting potential applications in medicine, biology and materials science, including sustained AI research related to Alzheimer’s disease.
Risks & concerns
- The article provides no benchmarks or independent evidence showing that Mirendil’s systems can improve themselves reliably or perform the work of a frontier AI laboratory.
- Neither company disclosed the agreement’s exact value, duration, capacity allocation, pricing structure or spending timetable.
- Mirendil’s scientific ambitions, including possible work on Alzheimer’s disease, remain proposed use cases rather than reported discoveries or validated deployments.
- A commitment exceeding $100 million represents roughly half the seed funding Mirendil raised in late June 2026, highlighting the high capital demands of its research strategy.
- The report does not establish when Mirendil will launch a commercial product or whether Google Cloud will ultimately offer its technology to enterprise customers.