Five AI infrastructure leaders discuss mounting constraints in chip supply, energy, and model architecture

Five executives representing different tiers of the AI supply chain convened at the Milken Global Conference to discuss technical and infrastructure challenges constraining the industry's expansion. ASML's CEO emphasized that despite manufacturing acceleration, chip supply will lag demand for several years. Google Cloud's leadership revealed the scale of unfulfilled demand: quarterly backlog had nearly doubled from $250 billion to $460 billion in a single period, driven partly by 63% year-over-year revenue growth.

Energy consumption emerged as the second-order constraint. Google is actively investigating orbital data centers to address thermal and power limitations, though the absence of convection in space introduces engineering complexities around radiative cooling that differ substantially from current liquid and air-based systems. The implication, articulated by Google Cloud's COO, is that efficiency gains come from vertical integration—designing chips, models, and inference pipelines as a single system rather than assembling off-the-shelf components.

Physical AI autonomy work faces a distinct bottleneck: the scarcity of real-world training data. Applied Intuition's CEO noted that synthetic simulation, despite improvements, cannot fully replace on-site data collection from vehicles, drones, and industrial equipment. This suggests a near-term constraint on scaling autonomous systems independent of chip or energy limits.

A contrasting technical direction was represented by Logical Intelligence, which is building systems on energy-based models rather than large language model architectures. The approach uses substantially fewer parameters (200 million versus hundreds of billions) and claims orders-of-magnitude speed improvements while maintaining the ability to update knowledge incrementally. The company's framing suggests this model class may be better suited to domains requiring physical reasoning rather than linguistic pattern matching.