Supply chain constraints and architectural shifts define AI future at Milken Global Conference

Five executives representing critical layers of the artificial intelligence supply chain convened at the Milken Global Conference in Beverly Hills to address emerging bottlenecks and architectural shifts. The panel highlighted that advanced semiconductor manufacturing, specifically extreme ultraviolet lithography, faces a multi-year shortage that will restrict hardware availability for hyperscalers. Simultaneously, energy limitations are driving innovation in cooling solutions, with Google exploring orbital data centres to meet the demands of its expanding infrastructure.

Christophe Fouquet, chief executive of ASML, stated there will be a supply-limited market for advanced chips for the next two to five years. This constraint means hyperscalers will not receive all the hardware they are paying for, despite a huge acceleration in manufacturing efforts. Fouquet noted that while the United States holds the data, computing access, chips, and talent, other nations lack access to the necessary lithography machines to manufacture the most advanced semiconductors required for top-tier models.

Francis deSouza, chief operating officer of Google Cloud, confirmed that the company is seriously exploring data centres in space to access abundant energy and solve cooling issues. He noted that while radiation cooling in a vacuum is slower than air or liquid systems, it is a viable path. DeSouza highlighted that Google Cloud's revenue crossed $20 billion last quarter with growth of 63%, while its backlog nearly doubled in a single quarter, underscoring the reality of the demand.

The discussion also pivoted to the unique challenges of physical artificial intelligence, where real-world data collection and national sovereignty barriers pose significant hurdles compared to digital models. Qasar Younis, co-founder and chief executive of Applied Intuition, identified real-world data collection as the primary constraint for physical AI. He argued that synthetic simulation cannot fully close the gap for training models that operate in the physical world, noting that almost every country refuses foreign-controlled intelligence systems within their borders.

Eve Bodnia, a quantum physicist at Logical Intelligence, presented an alternative to large language models, arguing that energy-based models better understand physical rules and can update knowledge without full retraining. She described her company's largest model as running thousands of times faster than current large language models, designed to grasp the rules of the world rather than just linguistic patterns.

Dimitry Shevelenko, chief business officer of Perplexity, described the evolution of AI from search tools to autonomous digital workers requiring granular security controls. He emphasised that enterprise security requires strict protocols, including human approval for actions taken by agents, to protect corporate systems. Younis further noted that in domains like agriculture and mining, physical AI is filling a void of chronic labour shortages rather than displacing willing workers.