Mirendil has signed a multiyear partnership with Google Cloud worth more than $100 million, according to co-founder and CEO Behnam Neyshabur, giving the young AI lab access to Google TPUs, Nvidia GPUs, and managed training clusters for its research into self-improving AI.

The agreement is significant because it converts a large share of Mirendil’s recently raised capital into computing capacity at a time when frontier AI companies are competing for scarce accelerators. The company says its long-term goal is to build systems that can repeatedly improve their knowledge and performance while conducting research, potentially automating parts of scientific discovery and AI development.

A major infrastructure commitment for a young lab

Neyshabur told TechCrunch that the Google Cloud deal is worth upward of $100 million. That would represent roughly half of the seed funding Mirendil raised in late June at a $1 billion valuation, based on the same report.

The arrangement gives Mirendil access to more than one type of accelerator rather than tying its workloads to a single chip platform. Google’s TPUs and Nvidia GPUs will be paired with managed training clusters, allowing the startup to run different stages of its work on hardware suited to each workload.

The deal reflects a broader shift in the economics of frontier AI. For companies developing large models or complex agentic systems, access to computing can be as strategically important as access to talent or venture capital. Cloud providers, meanwhile, are using infrastructure commitments to attract startups whose future systems could become important sources of enterprise demand.

There is no indication in the available reporting that Mirendil has launched a commercial product or disclosed customers tied to the agreement. The immediate purpose appears to be research and systems development rather than a near-term software rollout.

What Mirendil means by self-improvement

Mirendil is working on what the industry commonly calls recursive self-improvement: AI systems that iteratively improve their own capabilities, knowledge, or methods. The concept has drawn attention from established labs and newer startups, but it remains a research objective rather than a demonstrated product category with agreed performance standards.

Neyshabur described a system that could receive a difficult problem and continue improving as it investigates the subject. He gave Alzheimer’s disease as an example, suggesting that such a system could accumulate knowledge, learn about a domain, and make progress toward a defined research goal over time.

Mirendil’s stated targets extend beyond medicine. The company believes its approach could support work in biology, materials science, and AI research. Those ambitions would require more than a capable model: they would require reliable evaluation, long-running task management, tools for gathering and testing evidence, and mechanisms that prevent an automated research loop from reinforcing its own errors.

The Google Cloud arrangement therefore matters not simply because of its dollar value. It gives Mirendil the capacity to experiment with the orchestration layer around self-improving AI, including how models, tools, evaluators, and compute resources interact across extended research runs.

Evidence and claims remain limited

The central facts in this report come from TechCrunch’s interview with Mirendil’s CEO and from comments attributed to Google and Mirendil co-founder Harsh Mehta. The two other items in the source cluster provide headlines or summaries but no independently usable article text, so they do not add confirmation or technical detail.

The reported deal value, hardware access, and intended research direction should therefore be understood as company-disclosed or media-reported facts, not as independently audited figures. The available evidence does not establish how much compute Mirendil will receive, over what exact schedule, or whether the commitment represents reserved capacity, cloud spending, credits, or a combination of arrangements.

Google senior vice president and chief technologist of AI and infrastructure Amin Vahdat said AI progress increasingly depends on orchestrating complete systems rather than improving chip-level performance alone. That statement is Google’s characterization of its infrastructure strategy, not evidence that Mirendil’s research has achieved a particular capability.

Similarly, Mirendil’s claim that its software can help customers extract more value from Google hardware is a company position. No benchmark, model evaluation, production workload, or customer deployment was provided in the supplied reporting to verify that claim. The same caution applies to the company’s ambition to perform work comparable to an entire frontier AI lab.

Implications for builders and enterprise buyers

For AI builders, the agreement highlights the importance of workload flexibility. Training, inference, synthetic data generation, evaluation, and tool-use experiments may not have the same hardware requirements. A platform that can move workloads between Google TPUs and Nvidia GPUs could reduce dependence on one accelerator family, but it also introduces engineering complexity around software compatibility, scheduling, observability, and cost control.

The arrangement may also offer a useful signal about how early-stage AI companies are financing infrastructure. A large cloud commitment can provide capacity and technical support, yet it can consume capital before a product has established revenue. Founders considering similar deals will need to weigh guaranteed access against contractual lock-in, utilization risk, and the possibility that hardware efficiency improvements could change the economics during a multiyear term.

For enterprises, the more important question is whether self-improving AI can be made governable in real workflows. A system that continues changing its strategies while pursuing a research goal would need strong audit trails, bounded permissions, reproducible results, human review, and safeguards against fabricated or weakly supported findings. Higher compute availability does not solve those reliability and safety problems.

The deal also gives Google a potential strategic relationship with a lab focused on a high-impact research direction. If Mirendil’s systems mature, Google could gain a partner whose software is designed around its infrastructure while also helping demonstrate how cloud platforms support long-running AI research. That is a possibility, not a disclosed commercial outcome.

What to watch next

The first signal will be whether Mirendil publishes technical details about its system architecture, training methods, evaluation framework, or compute usage. Those disclosures would help distinguish a general research vision from a reproducible approach to recursive self-improvement.

Builders should also watch for evidence that the startup is using both Google TPUs and Nvidia GPUs in production-scale experiments, rather than merely retaining theoretical access to both. Useful indicators would include workload-level cost comparisons, throughput data, and results from independent evaluations.

Further funding, named enterprise customers, research publications, or a public demonstration of a system that improves over extended tasks would clarify Mirendil’s commercial and technical trajectory. Conversely, a lack of measurable progress would reinforce the view that the deal is primarily a bet on a demanding research program.

Creati.ai perspective

Mirendil’s Google Cloud commitment shows how the next generation of AI labs may compete: not only by training larger models, but by securing the infrastructure needed to run persistent research loops. The size of the deal is notable, but the harder test will be whether additional compute produces dependable improvements rather than simply longer experiments.

For the market, the story is best read as an infrastructure and execution bet. Mirendil has secured access to important hardware and cloud systems; it has not yet demonstrated, in the available evidence, that self-improving AI can deliver the scientific or commercial outcomes it envisions.