• Viewpoints
  • Posts
  • AI's Leap from Information to Atoms

AI's Leap from Information to Atoms

The Convergence of Spatial Computing, Physical Understanding, and Robotic Systems

Watch a child play with a ball. Without conscious thought, they know how hard to throw it, how it will bounce off the wall, how it will roll to a stop. This intuitive physics - mastered in our earliest years - remains completely foreign to our most sophisticated AI systems.

While ChatGPT can write essays or explain Newton's laws, it has never truly understood what it means for something to be heavy, bouncy, or smooth.

But could this change?

AI is now pushing for its most consequential transformation - the leap from the realm of pure information into the messy, atomic reality of the physical world.

The physical world - the actual stuff of life - manufacturing, logistics, healthcare, construction - represents a $104 trillion market, that dwarfs the digital economy which is worth roughly $11.5 trillion annually.

The venture ecosystem has picked up the scent of something massive.

After all, Google announced a dedicated physical world modelling team, NVIDIA's Jensen Huang declared their Cosmos model to be "for the physical world what ChatGPT is for words," and a16z has written a check into World Labs $230M round.

Something fundamental is shifting.

As investors at FOV Ventures focused on spatial computing, we see this as a pivotal moment. Companies that can bridge the gap between AI's digital capabilities and physical world applications are poised to create enormous value.

The Convergence Making Physical AI Possible

Just as language models hit their breakthrough moment when computing power, data availability, and algorithmic advances aligned, Physical AI is experiencing its own perfect storm of enabling factors.

The NVIDIA Effect

Over the last decade, NVIDIA's GPUs have improved AI computing power by a staggering 1 million times.

But beyond just raw horsepower, the breakthrough has come from the complete stack: hardware acceleration, software optimisations, networking, and system architecture.

Jevons' Paradox suggests as computing becomes more efficient and affordable, the demand for it only increases. Physical AI, with its need to process complex physics simulations and real-world data, is the next massive compute challenge.

Previous AI models were limited by their training data - text for language models, images for vision models. But Physical AI is benefiting from true multi-modal training. Nvidia's Cosmos, trained on 20 million hours of video, represents a new approach to giving AI systems real-world understanding.

By combining video, audio, physics simulations, and sensor data, these systems are developing something akin to common sense about how the physical world works.

For startups this a real opportunity to attack industries vertically, as NVIDIA wants companies to build solutions on Cosmos as they push into world models & AI-powered robotics.

World foundation models help robots build spatial intelligence capabilities by simulating virtual environments

The Tech Stack - Value in Physical AI

World Models & Physics Engines

The foundation of Physical AI centres on systems that can truly understand and predict physical interactions. NVIDIA's Cosmos (and others liked it), trained on 20 million hours of video, is just the beginning. The real value will emerge from platforms that can:

  • Create generalisable physics understanding that transfers across domains.

  • Enable efficient compute that doesn't require datacenter-scale resources.

  • Bridge the gap between simulation and reality.

  • Handle complex multi-object interactions and material properties.

The computational challenge here is staggering. Current approaches demand massive resources for even simple physical predictions. The winners will likely combine novel differentiable physics engines with optimised computational architectures.

Training Infrastructure:

While the industry focuses on end applications, there is investment potential in training infrastructure. The current approach of repurposing game engines like Unreal and Unity, while clever, isn't sustainable. We need:

  • Specialised physics simulation environments e.g. Sandbox for self driving cars.

  • Robust synthetic data generation platforms.

  • Hardware-in-the-loop testing capabilities.

  • Efficient data pipelines for physical world data.

The "sim2real" gap remains one of the hardest challenges in robotics and Physical AI. Companies that can reduce this gap through domain randomisation, adaptive simulation, and real-world calibration will become essential platform players.

Hardware Abstraction & Control - The Integration Layer

The interface between AI systems and physical hardware is crucial yet often overlooked. The challenge is about creating abstractions that are both powerful and practical. Key developments include:

  • Universal control interfaces that work across hardware platforms

  • Real-time sensor fusion and processing systems

  • Safety-critical control architectures

  • Multi-robot coordination frameworks

The most promising approaches combine classical control theory with modern AI. Rather than learning everything from scratch, these systems use physical models for bounds while using AI for optimisation. This hybrid approach is especially valuable in safety-critical applications.

Archetype AI's Newton foundation model can summarise real-time radar, audio sensor data, and language to provide a rich contextual understanding of human behaviour - all without cameras. By fusing basic sensor data with context, Newton demonstrates that AI can achieve human-like understanding of fluid physical situations.

Development Platforms: Making Physical AI Accessible

For Physical AI to reach widespread adoption, we need sophisticated development tools that can:

  • Abstract away complexity while maintaining control.

  • Enable rapid prototyping and iteration.

  • Support both simulation and real-world deployment.

  • Provide robust testing and validation frameworks.

The winning platforms will emerge from teams that understand both the technical challenges of physical systems and practical developer needs. Think "Vercel for Physical AI" – making it dramatically easier to develop and deploy physical world applications while maintaining the precision these systems require.

Investing In The Convergence

At FOV Ventures, we've placed our chips on this convergence.

As investors focused on spatial computing, our journey began in the digital entertainment stack - gaming engines, studios, and VR products. What’s uncovering is that these are precisely the proving grounds for AI's physical understanding.

A by-product of development of real time technologies in the gaming industry has created built the perfect toolkit for physical AI. Unreal Engine and Unity - created to entertain - have evolved into sophisticated physics simulation platforms with years of optimisation for consumer hardware, vast libraries of pre-built physics interactions, and real-time performance requirements that force efficient solutions.

This accidental gift from gaming creates a virtuous cycle: better simulations enable smarter physical systems, which in turn inform more accurate simulations. Game companies solving NPC navigation are tackling the same fundamental problems as warehouse robots. VR developers crafting hand interactions are solving challenges identical to robotic grasping.

This convergence is precisely why we invested in Nunu.ai.

They're building multimodal agents that test and play games - infusing AI into game development from QA through to player simulation. Their debut trailer showed these agents navigating games like Roblox, Genshin Impact and Assassin's Creed with an uncanny understanding of how to interact with these virtual worlds.

With those successes the team transferred an AI agent—trained solely in virtual environments—to a real-world robot body. It navigated successfully. This validates what Jim Fan at NVIDIA observed: "Before we have a million robots in the physical world, we will first see a billion embodied agents in virtual worlds."

Starting in gaming gives Nunu a massive head start.

Game environments allow them to build agents and perform actions we're years away from achieving in the real world. They can 'tweak' physics in ways that help agents overcome limitations they'd face in reality. Through this virtual-to-real training process, they're accelerating progress in areas like complex object manipulation - something that's just a button tap in games but remarkably difficult for physical robots.

We also an upcoming investment announcement in the space - so stay tuned.

Ultimately, real question for investors is where this technology will first create billion-dollar companies. Manufacturers want systems that adapt to existing environments, not complete factory redesigns.

The winners will bridge the gap between simulation and reality. Like how AWS changed how companies think about computing resources, the most valuable physical AI companies will create the software layer that makes physical operations programmable.

Beyond technical excellence, solutions need to address clear, funded problems. The best companies understand both the physics of their problem and the economics of their solution.

If you're building or investing at the intersection of AI and the physical world, we'd love to hear from you - reach out at https://www.fov.ventures/