Google Cloud eyes enterprise distribution in $100M deal with self-improving AI builder Mirendil
Mirendil, an AI lab founded by former Anthropic researchers, has secured a multi-year Google Cloud partnership valued at over $100 million to power its self-improving AI systems. For institutional investors and enterprise customers, the deal signals both a validation of recursive self-improvement as a near-term commercial technology and Google’s strategic pivot toward infrastructure plays that lock in hardware adoption through software partnerships.
- Mirendil raised $200 million at $1 billion valuation in June; Google deal represents roughly 50% of seed capital committed to single infrastructure agreement
- Partnership grants access to Google TPUs and Nvidia GPUs with managed training clusters, emphasizing hardware flexibility and workload optimization across chips
- Self-improving AI already showing measurable commercial traction: Anthropic reports 80% of code merged in May 2026 written by Claude models, up from single digits in early 2025
- $100M+ Google Cloud commitment as percentage of Mirendil’s entire seed funding round
- 80% Anthropic code merged by AI models as of May 2026, versus prior single-digit baseline
- $1B Mirendil valuation at seed stage, one of largest AI startup seed rounds on record
Mirendil, the AI research lab founded by former Anthropic researchers, has committed to a multi-year partnership with Google Cloud worth more than $100 million, securing computational infrastructure for its self-improving AI models.
CEO Behnam Neyshabur confirmed the deal, which provides access to Google’s Tensor Processing Units (TPUs) and Nvidia graphics processors alongside managed training clusters.
The agreement represents a substantial portion of the company’s recently closed seed funding and underscores growing institutional confidence in recursive self-improvement as a near-term technology capable of automating research workflows across medicine, biology, chemistry, and materials science.
Mirendil commits half its seed capital to single Google Cloud infrastructure deal
The $100 million Google commitment amounts to roughly 50 percent of Mirendil’s $200 million seed round, which closed in late June at a $1 billion valuation. That round, led by Andreessen Horowitz with participation from Kleiner Perkins and Nvidia, represented one of the largest seed financings for an AI startup to date.
By allocating such a substantial portion of early capital to a single infrastructure partner, Mirendil’s founders are signaling both the computational intensity of their research agenda and confidence in Google’s ability to deliver the horsepower required for recursive self-improvement at scale.
The concentration of spend reflects the economics of frontier AI development, where training and inference costs dominate operational budgets. For Mirendil, the Google arrangement locks in multi-year access to heterogeneous compute, both TPUs and Nvidia GPUs, at a time when capacity remains constrained and pricing volatile.
Cofounder Harsh Mehta emphasized that the company’s models are architected to distribute workloads efficiently across different chip types, lowering costs not only for Mirendil’s internal research but for enterprise customers licensing its systems.
Self-improving AI moves from research concept to measurable productivity metric
Recursive self-improvement refers to AI systems that iteratively refine and redeploy themselves, becoming progressively more capable without constant human retraining or intervention. Mirendil’s stated ambition is to build models capable of performing the full range of work conducted by a frontier research laboratory, from hypothesis generation through experimentation and iteration.
CEO Neyshabur, a former Anthropic research scientist and co-inventor of the Sharpness-Aware Minimization optimization algorithm, articulated the vision plainly: “You can have a self-improving AI where you can point a problem at it, and it keeps getting better with time.” His example was Alzheimer’s research, a domain where self-improving systems could autonomously pursue open-ended scientific goals without constant human guidance.
Anthropic has already demonstrated measurable traction with this model. As of May 2026, more than 80 percent of code merged into Anthropic’s systems was written by its Claude models, compared to single-digit percentages in early 2025.
The company acknowledged that this metric likely overstates true productivity gains, yet confirmed that the acceleration is genuine: engineers are now merging approximately eight times as much code per day as they were a year prior.
This data point carries weight for institutional investors evaluating whether self-improvement remains theoretical or has entered the realm of practical enterprise deployment.
Mirendil’s cofounder Harsh Mehta contributed to Anthropic’s internal automation efforts, giving the company direct operational experience with self-improving workflows. That institutional knowledge, combined with access to Anthropic’s research legacy, positions Mirendil to compress the timeline from prototype to production deployment.
The Google partnership accelerates that path by removing infrastructure barriers.
Google positions self-improving AI as hardware lock-in and enterprise upsell opportunity
Google’s $100 million commitment is not purely philanthropic. For the cloud giant, Mirendil represents a strategic customer whose software layer can drive adoption of Google’s own silicon. Neyshabur noted that Mirendil’s workload optimization capabilities help customers extract more performance from Google’s chips, providing Google with a competitive pitch against rival cloud providers.
In a market where raw chip performance is increasingly commoditized, software that can orchestrate heterogeneous compute more efficiently becomes a differentiator worth funding directly.
Amin Vahdat, Google’s senior vice president and chief technologist for AI and infrastructure, framed the partnership in systemic terms. Progress in AI infrastructure, he indicated, is no longer driven solely by chip-level performance but by the ability to orchestrate entire systems, balancing compute, memory, networking, and scheduling across complex workloads.
Mirendil’s software architecture, tested against both TPUs and Nvidia GPUs, offers Google a vehicle to demonstrate that thesis to enterprise customers. When those customers adopt Mirendil’s models, they simultaneously adopt Google Cloud as the optimal deployment platform.
The deal also grants Google early exposure to recursive self-improvement at scale. Should Mirendil’s research succeed, Google gains both a strategic partner and a template for understanding how self-improving systems operate within its cloud infrastructure.
That knowledge compounds in value as self-improving AI moves from research labs into commercial deployments across pharmaceuticals, materials science, and research institutions.
Institutional investors face new metrics for evaluating self-improving AI runway and unit economics
The Mirendil-Google deal establishes a practical pricing baseline for self-improving AI infrastructure partnerships. At $100 million over multiple years, the deal implies annualized commitment in the tens of millions of dollars for a company still in early-stage product development.
Investors evaluating AI infrastructure plays must now assess whether early-stage AI firms can sustain multi-year infrastructure commitments while simultaneously building enterprise sales organizations and navigating regulatory scrutiny.
Anthropic’s disclosed metrics on code generation and engineer productivity suggest that self-improving systems do accelerate output measurably, even as the company cautioned against overstating gains.
For institutional investors, the critical question is whether Mirendil’s models will achieve similar or superior productivity metrics in domains like drug discovery or materials science, where economic value per unit of acceleration is orders of magnitude higher than software development.
A self-improving system that accelerates therapeutic discovery by even 10 percent could justify substantially larger infrastructure spending.
The heterogeneous compute architecture, spanning both Google TPUs and Nvidia GPUs, also signals a broader industry shift. Rather than betting on a single chip manufacturer, Mirendil’s design assumes a multi-vendor landscape and optimizes for portability.
This approach reduces lock-in risk for enterprise customers and suggests that the AI infrastructure market is moving toward abstraction layers and scheduler-level optimization, not raw silicon dominance.
Watch for Mirendil’s first disclosed customer deployments and performance benchmarks in pharmaceuticals or materials science over the next 12-18 months. Those metrics will determine whether Mirendil’s self-improving architecture translates competitive advantage in research workflows into enterprise revenue sufficient to justify its $1 billion valuation and the implied $50+ million annual spend on Google infrastructure. Google’s willingness to expand or renew the commitment at higher levels would signal internal validation of recursive self-improvement’s commercial viability; conversely, any public shift to alternative compute partners would suggest the technology faces unforeseen deployment challenges.
