When Funding Outpaces Understanding

Every technology wave tends to attract capital faster than it attracts clarity.

The current cycle around physical AI is no exception. The narrative is clear: robots capable of operating in real-world environments, world models that can simulate physical reality and AI systems that can reason about space.

The ambition is justified. But ambition alone doesn’t determine where companies should focus their effort.

In most AI discussions today, attention gravitates toward the visible layers of the stack: robots, foundation models and large-scale compute infrastructure. These are easy to demonstrate and benchmark and will naturally generate headlines.

The invisible layers, the systems that feed and structure data about the physical world, rarely receive the same attention. Yet, those are the layers that will determine whether physical AI actually scales.

This is something I’ve seen repeatedly while building Nfinite. Our work focused on turning visual inputs into structured 3D representations of real-world objects and environments, enabling simulations and synthetic datasets that models can actually learn from.

Over time it became clear that the hardest problems weren’t model architectures or GPUs. They were data realism, spatial consistency and long-horizon validation.

The Founder’s Dilemma During Technology Booms

One of the hardest decisions for founders during technology booms is deciding what not to build.

Capital tends to reward what looks impressive today. But the companies that endure often focus on what will be indispensable tomorrow.

Instead of asking "What will investors fund right now?" founders need to ask "What will the ecosystem eventually depend on?"

History offers plenty of examples. Take the history of internet for example, content delivery networks and database architecture became foundational long before most users understood they existed. Physical AI is likely to follow the same pattern.

As models improve and robots become more capable, their performance will depend less on raw intelligence and more on the fidelity of the environments they learn from.

Training systems to operate in the real world requires structured representations of geometry, physics and spatial relationships. These are not things that can simply be scraped from the internet. They have to be constructed.

The Signal Founders Should Watch

When a technology sector begins to mature, the signals of progress start shifting.

Early in a cycle, progress is measured through demos and benchmarks. Later on, the focus shifts toward reliability, economics and deployment.

Systems must operate across thousands of environments rather than one curated lab scenario. They must tolerate noise, variation and unpredictability. They must improve without requiring exponentially larger budgets. This transition is where many well-funded projects struggle.

The bottleneck moves from intelligence to representation, from how powerful the models are to how accurately they understand the environments they operate within.

In physical AI, that bottleneck is already increasingly visible in the availability of structured, spatially coherent data.

Why Some Companies Choose The Harder Path

This reality shapes an important decision for founders. Do you build where the spotlight currently sits, or where the system will eventually depend?

The first option usually attracts faster funding and clearer narratives. The second often involves longer timelines, and a harder story to explain. But historically, it’s the second category that ends up defining the platform layer of new technology ecosystems.

As robotics and world models advance, the industry is recognizing that training systems to operate in physical environments requires more than improved neural networks. It requires a consistent representation of the world those systems inhabit.

That representation does not yet exist at the scale the industry will require. And building it may turn out to be one of the defining infrastructure challenges of the next decade.

When A Founder Decides To Start Again

For me, this realization eventually led to a difficult but familiar decision. After building Nfinite for years—working at the intersection of visual data and AI training pipelines—it became clear that the industry was heading toward a structural gap. The systems being funded to reason about the physical world were advancing rapidly, but the infrastructure required to feed them reliable spatial data was still fragmented.

That observation ultimately pushed our team to start building again. Not another model or robotics platform, but infrastructure designed to support them. The goal wasn’t to chase the excitement around physical AI, but to focus on the part of the stack that tends to be overlooked when capital moves quickly.​

Building In The Noise

Technology booms create enormous opportunities. They also create enormous distraction. For founders, the challenge is learning to distinguish between the two. The companies that endure through multiple technology cycles tend to share one trait: They build for the constraints of the system rather than the narratives of the moment.

In physical AI, those constraints are becoming clearer every year. Systems must understand space, interact with objects and operate in environments that are messy and unpredictable. It requires discipline in how we build the infrastructure those systems depend on. And during moments like that, discipline becomes even more important. Because the companies that survive the gold rush are rarely the ones chasing it. They’re the ones building what everyone else will eventually need.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?