Mirendil Signs More Than $100M Google Cloud Deal to Power Self-Improving AI

Updated August 6, 2026 6:23 pm

AI lab Mirendil has struck a multi-year Google Cloud partnership worth more than $100 million to secure the computing power it needs for self-improving AI research. The deal matters because it highlights two of the most important forces in today’s AI race: startups are locking in massive infrastructure agreements, and cloud providers are using those commitments to deepen their position in frontier AI.

Mirendil co-founder and chief executive Benham Neyshabur confirmed the scale of the agreement in an interview with TechCrunch, saying the contract is roughly half the size of the startup’s seed financing, which closed in late June at a $1 billion valuation. The new arrangement gives Mirendil access to Google’s tensor processing units, Nvidia graphics processors and managed training clusters as it works toward AI systems that can improve their own capabilities over time.

The company’s ambition is unusually sweeping: Mirendil wants to build technology that can eventually do the work of an entire frontier AI lab. That puts it in one of the most ambitious corners of the industry, where researchers are pursuing recursive self-improvement, a long-discussed concept in which AI systems iteratively refine their own performance and knowledge.

For Google, the partnership is also strategic. It gives the company a marquee startup customer focused on one of the field’s most attention-grabbing research directions, while showcasing a cloud pitch that increasingly emphasizes not only raw chip performance but also the orchestration of complex training systems across multiple kinds of hardware.

What Mirendil is trying to build

Mirendil is pursuing what it calls self-improving AI, a research direction that aims to create systems capable of making themselves better over time with limited human intervention. The idea is not simply to build a more powerful chatbot or assistant, but to develop a model that can plan, test, learn and iterate in ways that resemble a scientific research loop.

The lab says that approach could eventually accelerate work in medicine, biology and materials science, where progress often depends on sustained experimentation and expertise accumulation. In Mirendil’s view, a system that can keep refining itself may be able to help researchers work through difficult open problems, including diseases such as Alzheimer’s.

“You can have a self-improving AI where you can point a problem at it and it keeps getting better with time,” Neyshabur said.

He added that the vision is to set ambitious scientific goals and then let the system continue improving its own knowledge and output over time, rather than treating model development as a one-off training event.

Why the Google Cloud deal matters

The agreement underscores how central compute has become to AI competitiveness. Building and training large models now requires access to scarce, expensive and highly specialized infrastructure, and the biggest cloud vendors are increasingly using capacity deals to win startup loyalty before rivals do.

For Mirendil, the deal solves a practical problem: self-improving AI research is compute-hungry, and the startup needs reliable access to hardware that can support large-scale experimentation. For Google, the contract adds another sophisticated AI lab to its customer base at a time when cloud competition is intensifying across both enterprise and frontier AI workloads.

The arrangement also reflects how AI companies are changing the way they buy infrastructure. Instead of relying on ad hoc capacity purchases, many startups are now signing longer-term deals to guarantee access to large amounts of hardware, even if that means committing significant portions of their funding to compute before they have a mature product.

How the hardware mix helps

The answer is flexibility: Mirendil says it will use both Google TPUs and Nvidia GPUs, along with managed training clusters, to match different workloads to different chips. That matters because frontier AI systems often require different kinds of acceleration depending on whether the task involves training, inference, experimentation or systems orchestration.

Harsh Mehta, Mirendil’s co-founder, said the company’s software and systems layer is designed to assign the right jobs to the right accelerators, which should improve efficiency and lower costs. In his telling, the value is not simply access to more hardware, but access to the best mix of hardware for a broad range of research tasks.

Mehta said the ability to mix and match workloads across different chips can reduce costs for Mirendil and, eventually, for the customers that use its systems.

That is also the pitch Google wants to make to the broader market: that AI progress increasingly depends on well-orchestrated systems rather than isolated chip benchmarks.

How Google is positioning itself in the AI infrastructure race

Google has been pushing a broader message that cloud leadership in AI is no longer only about having the fastest individual accelerator. Instead, the company argues that the winning formula will come from system-level design: integrating compute, storage, networking, orchestration and software in a way that makes large-scale AI training more efficient.

Amin Vahdat, Google’s senior vice president and chief technologist of AI and infrastructure, framed the challenge in those terms, saying progress now depends on how intelligence systems are orchestrated and how the industry overcomes the physical limits of scaling.

Vahdat said AI advancement is no longer just about chip-level performance, but about how entire systems of intelligence are coordinated under real-world constraints.

That message helps explain why deals like Mirendil’s matter so much to cloud providers. These agreements are not only revenue commitments; they are also proof points that a cloud platform can support the most demanding research workloads in the market.

What is recursive self-improvement?

Recursive self-improvement is a concept in which an AI system improves its own capabilities, then uses those gains to improve further. In theory, that creates a feedback loop in which the system becomes progressively more capable without requiring the same amount of outside human innovation at each step.

The idea has long been associated with frontier AI speculation, but it is now becoming a concrete research agenda for a small number of startups and established labs. Anthropic, where Mirendil’s founders previously worked, is among the major organizations that have explored related questions. A handful of startups, including Recursive Superintelligence and Ricursive Intelligence, have also recently emerged around the same goal.

The concept is scientifically and operationally difficult. Even if a system can improve at one task, it does not automatically follow that it can safely or reliably redesign itself. That is one reason the field remains early and highly experimental. But it is also one reason the compute demands are so high: the work involves repeated training cycles, evaluation runs and systems-level experimentation.

Why the field is drawing startup attention

Self-improving AI has become attractive to founders because it points to a potentially transformative outcome: automation of part of the scientific method itself. If a model can test hypotheses, update its own understanding and keep refining performance, it could speed up discovery across a range of disciplines.

That promise has made recursive self-improvement a magnet for talent from leading labs and a natural target for cloud providers looking to lock in the next generation of AI infrastructure demand.

How much compute does this really require?

The short answer is: a great deal. Mirendil’s new partnership is worth more than $100 million because the company needs sustained access to large-scale training capacity, not just occasional bursts of GPU time.

Frontier AI research typically depends on three broad ingredients:

• large clusters of accelerators for model training and experimentation;
• networking and orchestration tools to manage distributed workloads; and
• software layers that help maximize the efficiency of every training run.

That combination is expensive to assemble and difficult to obtain on short notice. As a result, startups increasingly treat compute contracts as a core part of their go-to-market and research strategy, especially if they are trying to move faster than better-funded rivals.

Mirendil’s funding round helps explain why it can make such a commitment so early. The company raised seed capital in late June at a $1 billion valuation, giving it the financial flexibility to pursue a major infrastructure bet before product revenue becomes meaningful.

Why cloud giants are chasing startup AI labs

Cloud companies see major AI startups as strategic customers because they can generate large, long-duration spending and offer public validation of the platform itself. A startup building cutting-edge models also becomes a reference point for other customers deciding where to train their own systems.

For Google, signing Mirendil helps reinforce the idea that its infrastructure is good enough for the most demanding AI research programs. It also places Google in a stronger position to sell similar capabilities to enterprise customers that want to run custom models, fine-tune systems or deploy advanced AI workloads.

These kinds of deals are increasingly common because they benefit both sides in the short term, even if the long-term commercial payoff is less certain. Startups get guaranteed access to scarce hardware. Cloud providers get predictable revenue and a relationship with a potentially important future AI platform.

Who stands to benefit most?

In the near term, Mirendil gains the most operational leverage because the agreement gives it enough compute to keep pushing its research agenda forward. Over time, Google may benefit from the software and systems tools Mirendil builds on top of its cloud, especially if those tools help unlock more efficient use of Google’s hardware.

There is also a broader ecosystem effect. If Mirendil succeeds in turning self-improving AI into a repeatable research method, it could create demand for an entirely new category of infrastructure services tuned for iterative model improvement rather than one-time training.

What makes Mirendil different from other AI startups?

Mirendil is unusual because it is not positioning itself primarily as a chatbot company, a consumer application or a general-purpose foundation model provider. Instead, it is staking its future on a research thesis: that AI systems can be designed to improve themselves in ways that make them more useful for scientific discovery.

That emphasis changes the economics of the business. Rather than focusing first on user growth or enterprise deployments, Mirendil is investing heavily in infrastructure and experimentation. The payoff could be enormous if the approach works, but the technical risk is equally high.

The company’s founders’ background at Anthropic also matters, since it signals familiarity with frontier model development and alignment-focused thinking. That pedigree can help attract investors, recruits and partners who want exposure to top-tier AI talent.

What happens next?

The next phase for Mirendil will likely be less about publicity and more about execution. The company now has funding, a major cloud partner and the hardware access needed to push its research agenda. The question is whether it can turn the idea of self-improving AI into something practical, reliable and eventually commercially meaningful.

For the wider industry, the deal is another reminder that the AI race is becoming a contest over infrastructure as much as intelligence. The companies best positioned to win may be those that can combine research ambition, systems engineering and long-term access to compute at scale.

In that sense, Mirendil’s agreement with Google Cloud is more than a procurement story. It is a snapshot of how the next generation of AI labs may be built: capital-intensive, cloud-dependent and organized around the belief that machines can eventually help design smarter versions of themselves.

Whether that vision proves realistic remains an open question. But for now, Mirendil has secured one of the clearest signs yet that major cloud providers are willing to bet big on the startups trying to get there first.