A startup founded by former Anthropic researchers just signed a major infrastructure deal that signals where frontier AI research is heading next. Mirendil, which raised $200 million in seed funding at a $1 billion valuation less than two months ago, has inked a $100 million-plus multiyear partnership with Google Cloud to access the compute it needs to scale what it calls “self-improving AI.”

The deal, reported exclusively by TechCrunch on August 6, 2026, gives Mirendil access to both Google’s TPUs and Nvidia GPUs through Google Cloud’s managed training clusters. For a company with a vision this ambitious, the compute infrastructure is everything.

Who Built This and Why It Matters

Mirendil was founded by Behnam Neyshabur and Harsh Mehta, both former Anthropic researchers. Neyshabur previously worked on AI for science at Google before joining Anthropic. Mehta built the first version of Anthropic’s internal autoresearch platform, the tooling Anthropic itself used to accelerate its own AI research.

The team has also drawn researchers from xAI, Google DeepMind, and OpenAI.

Their investors took notice early. The $200 million seed round was led by Andreessen Horowitz and Kleiner Perkins, with Nvidia also participating. Nvidia investing in a company building self-improving AI while also supplying the compute via Google Cloud partnership is not a coincidence.

What Self-Improving AI Actually Is

Recursive self-improvement, sometimes called self-improving AI, refers to AI systems that iteratively improve themselves. The system identifies its own weaknesses, designs improvements, and integrates them without requiring a full human-led retraining cycle each time.

Mirendil’s specific framing is “AI for AI for Science.” The vision is not to build a general-purpose model that competes with Claude or GPT. Instead, the company wants to build a platform that lets research labs in fields like medicine, materials science, and biology develop and iteratively improve their own specialized AI models.

The argument is that the current model, where every research team has to prompt and fine-tune a general-purpose foundation model, is a bottleneck. Mirendil wants to give those teams the infrastructure to build AI that gets better at their specific domain problems over time.

What the Google Cloud Deal Means

The $100 million-plus compute commitment from Google Cloud is both practical and strategic.

On the practical side, self-improving AI is extremely compute-intensive. Systems that evaluate themselves, propose improvements, and run training cycles repeatedly need far more infrastructure than systems that run a single training job. The managed training clusters give Mirendil the flexibility to scale those workloads without managing infrastructure themselves.

On the strategic side, Google gets a seat at the table with one of the most-watched frontier AI research efforts of 2026. The startup’s long-term vision, as described by its founders, is to eventually automate the work of an entire frontier AI research lab. If that materialises, Google wants the relationship established early.

Why Data and AI Professionals Should Pay Attention

The Mirendil story matters beyond the funding numbers for two reasons.

First, it signals that the next phase of AI development is moving toward specialization at scale. General-purpose models are powerful, but the teams doing domain-specific AI research are increasingly building purpose-built systems rather than prompting foundation models and hoping for the best. For data professionals, this points toward a future where domain expertise is more valuable, not less, because it becomes the input that shapes what the AI learns to do.

Second, recursive self-improvement, if it works as described, changes the pace of AI capability growth. Systems that improve themselves can compound in ways that externally-trained systems cannot. The researchers working on this are not casual observers. Neyshabur and Mehta built the internal research infrastructure at one of the most safety-focused AI labs in the world before leaving to do this. They know what they are chasing.

What This Means for Business

For most businesses, Mirendil is not a vendor you will be calling next quarter. The company has no product and the technology it is building is frontier research.

But the deal matters as a signal about where serious AI infrastructure investment is flowing. The combination of former safety-focused AI researchers, tier-one investors, and a major cloud compute commitment pointed at self-improving AI suggests this is a space the industry’s most informed players believe will matter significantly over the next two to five years.

For organisations building AI strategy today, the lesson from Mirendil is to watch what the researchers, not just the vendors, are doing. The research direction of today tends to become the enterprise product of tomorrow.