When five of the most consequential people in AI infrastructure sit down together, they usually talk about what's working. At TechCrunch in early May 2026, they talked about what's not.

The conversation — featuring ASML CEO Christophe Fouquet, Google Cloud's chief architect, Applied Intuition's CEO, Logical Intelligence's founder, and Perplexity's infrastructure lead — covered five problems that, taken together, suggest the AI boom is running into physical limits that no amount of capital can immediately solve.

The chip shortage isn't ending anytime soon

ASML's Fouquet put a number on it: the market will remain "supply limited" for the next 2 to 5 years. ASML makes the machines that make the chips that make AI possible, and it's not building them fast enough to meet current demand curves. The EUV lithography machines that ASML alone can produce take years to manufacture. Even with massive capital investment in new fabs, the constraint is not money — it's time and specialized manufacturing capacity.

This has consequences beyond GPU availability. Every AI company planning infrastructure is effectively placing a bet on a supply chain they can't control.

Energy is the new real estate

Google Cloud is exploring orbital data centers — literally putting compute in space — to solve an energy problem that can't be solved on Earth at the required scale. The heat dissipation challenge in space is formidable, but the underlying point is straightforward: the power requirements for frontier AI training and inference are growing faster than grid capacity in any single location.

This is a physical constraint that differentiates AI infrastructure planning from traditional cloud computing in a fundamental way. You can colocate servers anywhere there's power; you can't colocate them where power doesn't exist.

Physical AI has a data problem that synthetic data can't solve

Applied Intuition's CEO — whose company works on autonomous vehicle AI — pointed out that physical AI (robots, autonomous vehicles, drones) requires real-world data that no amount of simulation can fully replicate. Synthetic data works for text and code. It's insufficient for a robot that needs to navigate novel physical environments.

This creates a data moat for companies that have been collecting physical-world sensor data for years, and a significant barrier for new entrants.

Geopolitics is infrastructure

The geopolitical dimension of physical AI — countries resisting foreign-controlled autonomous systems operating within their borders — is emerging as a serious constraint on global AI deployment. This isn't hypothetical: it's already affecting where data centers can be built, which cloud providers can operate in which markets, and what AI systems governments will allow in critical infrastructure.

The most uncomfortable question: is the LLM architecture wrong for physical AI?

Logical Intelligence's founder raised what may be the most important — and most ignored — question: whether the transformer architecture that powers today's LLMs is fundamentally suited to the physical world. LLMs excel at reasoning about language and code. Whether that architecture extends naturally to robot navigation, autonomous vehicles, and physical manipulation is genuinely unclear.

This doesn't mean LLMs are failing — it's a much more specific critique: the dominant AI architecture may need augmentation or replacement for physical AI applications. That's a different problem than making the next GPT version better at writing emails.

What this means for AI infrastructure planning

The convergence of these five challenges — chip shortages, energy constraints, data limitations, geopolitical friction, and architectural uncertainty — is a signal that the easy part of building AI infrastructure is behind us. The next phase requires solving physical problems, not just software ones.

Companies building on top of AI infrastructure should be paying attention to which providers are addressing these constraints directly, not just adding more compute. The AI companies that win the next decade will be the ones solving the physical limits, not just scaling the software.

Sources: TechCrunch