Five executives spanning the AI stack met at the Milken Global Conference to discuss where the industry’s architectural assumptions break down. The panel included ASML’s CEO Christophe Fouquet, Google Cloud COO Francis deSouza, Applied Intuition CEO Qasar Younis, Perplexity CBO Dimitry Shevelenko, and Logical Intelligence founder Eve Bodnia. Here’s what they said about bottlenecks, system design, and trade-offs.
**Chip supply is the first hard constraint.** Fouquet stated plainly that even with massive fab acceleration, the market will remain supply-constrained for three to five years. Hyperscalers won’t receive all the silicon they’ve ordered. deSouza backed this with numbers: Google Cloud’s backlog nearly doubled in one quarter, from $250B to $460B. Demand is real; delivery is not.
**Data is a separate, harder bottleneck for physical AI.** Younis explained that Applied Intuition’s autonomy systems for vehicles and defense hardware require real-world data. Synthetic simulation cannot close the gap. “You have to find it from the real world,” he said, noting that fully synthetic training for physical-world models remains years away.
**Energy constraints are driving exotic architecture choices.** deSouza confirmed Google is seriously evaluating orbital data centers. Space offers abundant solar energy but complicates cooling—radiation is the only heat dissipation method in a vacuum, far less efficient than terrestrial liquid or air systems. Google’s strategy of co-engineering TPUs, models, and agents yields better flops per watt, but as Fouquet noted, “Nothing can be priceless.”
**A different inference paradigm is emerging.** Bodnia’s Logical Intelligence builds energy-based models (EBMs) that learn underlying data rules rather than predicting next tokens. Her largest model runs 200 million parameters—orders of magnitude smaller than leading LLMs—and updates knowledge incrementally without full retraining. She argues EBMs are more natural for robotics and chip design, where physical rules matter more than linguistic patterns.
**Agent permissions demand granularity.** Shevelenko described Perplexity’s evolution into a “digital worker” platform. Enterprise administrators can set read-only or read-write access per connector. The computer-use agent, Comet, presents a plan for approval before acting. Some users find this friction annoying, but Shevelenko considers it essential for security hygiene, especially after joining Lazard’s board.
**Geopolitical sovereignty now shapes physical AI deployment.** Younis observed that fewer nations can field a robotaxi than possess nuclear weapons. Countries resist foreign-controlled physical AI systems operating within their borders. Fouquet added that China’s model-layer progress is real but constrained by the lack of EUV lithography—older hardware compounds the disadvantage regardless of software optimization.
**The generation question drew pragmatic answers.** deSouza pointed to unsolved problems in neurology and climate that more capable tools could address. Shevelenko noted that entry-level jobs may shrink, but individual agency expands. Younis highlighted labor shortages in agriculture and mining—the average U.S. farmer is 58—where physical AI fills existing voids rather than displacing willing workers.
Source: TechCrunch