Investing in Nunu.Ai

Training tomorrow's robots in today's games - the virtual-to-physical AI revolution

Somewhere in a game development studio, a tester is repeatedly jumping against a wall, clicking endlessly searching for that one pixel where a character might clip through.

This tedious, repetitive and invisible labour powers a $300+ billion dollar industry.

Humans performing robot-like tasks because… ironically, robots weren't smart enough to do it.

But what if AI agents could do this job?

And what if as the AI that learns to grab a virtual sword from a virtual table solves the exact same spatial reasoning problem as a warehouse robot trying to grab a real package from a real shelf?

This is Nunu.ai - our latest investment – whose agents have learned to navigate the complex physics of Hogwarts Legacy and Assassin's Creed. And, then, in a demonstration that dropped jaws, took the same AI agent trained only in gaming and transferred it to a physical robot that navigated a real space, identified a real Coca-Cola bottle, and retrieved it.

This is why we’re part of their $6 million seed round co-led by a16z speedrun, Y Combinator and TIRTA Ventures. Because, this convergence is the next era of computing.

Co-investors include:
Y Combinator, a16z, Hartmann Capital, Earthling, TRITA Ventures, nrw & Factorial Funds

Gaming as a Gateway to “Physical AI”

Think about it. When you play a game, you're constantly making intuitive decisions. You know instinctively that you can probably climb that ledge, that water will slow you down, and that glass breaks when hit.

Nunu’s agents are also figuring this out.

Nunu automated the Full 30-Minute Tutorial in Hogwarts Legacy - you can see the AI Agent reasoning in the top right.

From the simplicity of Roblox to the realistic environments of Hogwart’s Legacy. Each time, they adapt immediately to the new physics, new rules, and new objectives, even completing specific testing tasks and finding bugs.

Nunu’s agent broke the world record by completing three gym challenges in Pokémon Emerald, surpassing all previous AI attempts. It repeatedly showed smart long-term planning by buying Pokéballs and healing items when anticipating big fights. Check out this screenshot of May stocking up (or find the moment in the VOD here):

For studios, this is significant. Roboto Games, a 13-person studio developing Stormforge (an open-world survival game), implemented Nunu's agents which automatically run over 400 tests monthly, examining everything from crafting systems to building mechanics, character creation to combat. The agents have identified critical bugs like broken character naming, item splitting issues, and unexpected building damage – all without human intervention.

For the real world, games provide the perfect training ground because modern game engines are essentially sophisticated physics simulators. When an AI learns to drive a car in a game, it's learning fundamental principles about timing and spatial relationships that transfer surprisingly well to real robotics.

That's the virtuous cycle that made this investment a no-brainer for us - games make better AI agents, better agents make more capable robots, and more capable robots change everything from warehouses to elder care.

Investment Rationale

"Before we have a million robots in the physical world, we will first see a billion embodied agents in virtual worlds." This observation from NVIDIA's Dr Jim Fan captures exactly why Nunu.ai represents a billion dollar opportunity.

The math is straightforward:

First, there's the immediate market – game testing.

This enables Nunu to capture value across two growing markets. Game studios currently allocate roughly 10% of their development budgets to QA testing - for AAA titles like GTA 6 (with its reported $2 billion budget), that's hundreds of millions in testing costs. Nunu’s agents can run hundreds of tests continuously, 24/7, finding bugs that exhausted human testers might miss. Several game studios are already using their platform, proving the market need is real and urgent. The global game testing market alone is $1+ billion and growing rapidly as games become increasingly complex, requiring more thorough testing across multiple platforms and devices.

At the other end of the spectrum, indie game developers face mounting pressure to deliver polished experiences with limited resources. For them, Nunu offers a solution that dramatically reduces both time and cost. Studios like Roboto Games have seen 50% reduction in QA costs and reclaimed 160 hours of development time monthly after implementing Nunu's AI testing.

Additionally, the recent trend of 'vibe coding' games (cc: Pieter Levels, $1M ARR AI-generated game), may grow the demand for solo devs or lean teams to 'in-source' game testing to 24/7 agents.

But game testing is just the beginning

As Dr Fan points out, games provide "the next trillion high-quality tokens to train our foundation models." Unlike the static text used to train language models, these are active learning experiences where agents explore complex physics environments. In Fan's words, "An agent's capabilities are upper-bounded by the complexity of the world it lives in." And games are becoming increasingly complex simulations of reality.

This is the aforementioned a virtuous cycle - agents trained in games become smarter, these smarter agents can transfer skills to physical robots, and the global robotics market - projected to hit $170 billion by 2030 - becomes accessible.

By starting with gaming, Nunu has found a brilliant go-to-market strategy. While competitors struggle with the limitations of physical robots - expensive hardware, dangerous testing environments, slow iteration cycles - Nunu can rapidly evolve their AI in virtual worlds where failures have zero consequences and you can simulate thousands of years of robot experience in days.

This is a massive head start.

Their demonstration of transferring game-trained AI to a physical robot that successfully retrieved objects provides early proof that the pathway is already working.

What makes this particularly compelling is the breadth of tasks their 'Rob the Robot' project achieved. In just three months, the same agent playing Hogwarts Legacy could control navigate IRL spaces, locate objects, respond to verbal commands, interact with home automation systems, and perform manipulation tasks like retrieving a Coca-Cola can. Despite limits in areas like latency and complex manipulation, it validated a core thesis - the skills learned in virtual environments can transfer to physical embodiments.

It's all too common for Europe's top technical talent to be absorbed by tech giants, so it's refreshing to back European founders of such exceptional talent with the boldness to build something transformative on their own terms.

All three founders- Jan Schnyder, Kyrill Hux, and Nicolas Muntwyler - met during their studies at ETH one of Europe's premier technical universities, where they had access to world-class robotics and AI departments that helped shape their vision for Nunu.ai.

The team has already demonstrated their ability to gain traction in the US game-studios market, has successfully been through Y Combinator and secured investment from a16z's speedrun fund.

We're excited to see how they continue to evolve from game testing to physical AI applications, and we look forward to supporting them on this journey.

If you're building or investing in anything in the physical AI or spatial computing space, we'd love to hear from you.