AI Bottlenecks Highlighted at Milken: Chips, Energy and New Architectures Shape the Debate
Milken panel warns of AI bottlenecks – chip shortages, energy and cooling limits, and architectural shifts – reshaping cloud strategies and national policy.
A high-profile panel at the Milken Global Conference in Beverly Hills this week put AI bottlenecks at the center of the industry’s immediate challenges. Executives and founders from chip makers, cloud platforms, robotics and alternative AI startups laid out converging constraints—from silicon supply to power and data—that they say will limit growth for years. Their comments sketched a picture of an AI ecosystem that is as much constrained by physical systems as it is driven by algorithms and capital.
Chip supply limits to persist for years
Christophe Fouquet, CEO of ASML, warned that the industry faces a prolonged period of supply constraints for advanced semiconductors. He argued that even with rapid investments in manufacturing, the market will remain supply-limited for the next two to five years. That shortage, he said, directly affects hyperscalers and others who are racing to deploy more compute for AI workloads.
Panelists noted that chip scarcity cascades through the stack: without access to the most advanced lithography tools and nodes, cloud providers, startups and national programs will face practical limits on scaling. The imbalance between demand and fabrication capacity, they said, is already reshaping procurement and long-term planning at large tech firms.
Energy and cooling force new data center strategies
Google Cloud’s chief operating officer explained that energy and heat removal are emerging as the next major bottlenecks after silicon. The company is exploring unconventional options, including orbital data centers, to access more abundant energy and rethink cooling strategies. Moving compute off-planet, however, introduces new engineering hurdles because space lacks the convection mechanisms that terrestrial data centers rely on.
Speakers emphasized that co-design—aligning chips, models and infrastructure—improves efficiency in flops per watt and buys more runway before energy limits bite. Integrated stacks that pair custom accelerators with optimized models can reduce energy per computation in ways off-the-shelf configurations cannot, they said.
Real-world data remains the slowest ingredient for autonomy
For companies building physical AI systems, raw data from the real world is the hardest constraint to overcome. The founder of an autonomy-focused firm explained that no amount of synthetic simulation fully substitutes for observations gathered by machines operating in live environments. Training systems that control vehicles, drones and industrial equipment require diverse, high-quality real-world events that take time and resources to collect.
Panelists predicted that physical AI will therefore evolve at a different pace than purely digital models, with progress tied to field deployments, sensor networks and tightly controlled testing programs. That timeline has implications for safety certification, regulatory compliance and the business models of companies that must amortize costly data-collection campaigns.
Energy-based models challenge prevailing architectures
A founder and physicist on the stage argued that alternative model families could mitigate some of the pressures on scale and energy. Her startup is pursuing energy-based models (EBMs), which aim to learn underlying rules rather than predict token sequences, and she claimed these models can run far smaller and faster than contemporary large language models. EBMs are designed to update incrementally as data changes, potentially avoiding expensive full-model retraining cycles.
Speakers framed this as a growing architectural debate: whether continued scale-up of current models is the best path, or whether different abstractions could deliver better performance per watt for domains like robotics, chip design and control systems. The conversation signaled increasing interest in plural approaches to AI that trade raw parameter counts for structural efficiency.
Enterprise agents prompt new questions about control and trust
Perplexity’s executive explained how the company is shifting from search to “digital worker” agents that act on behalf of knowledge workers. These systems can access enterprise tools and data, but their adoption raises operational security and governance questions. The firm emphasized fine-grained permissioning, saying administrators can limit connectors and designate read-only versus read-write access to preserve control.
Speakers highlighted a tension between productivity gains and the need for safeguards: agents that act autonomously must present plans, request approvals when required, and operate within auditable policies. Trust, they argued, will be built through granular controls and conservative defaults, particularly for institutions guarding long-standing client relationships.
Physical AI raises sovereignty and geopolitical concerns
Panelists also drew a geopolitical line between digital and physical AI, arguing that machines that operate in the physical world create sovereignty issues that software alone never did. Nations are more likely to restrict hardware, data collection and fielded systems that operate within their borders, and companies deploying autonomous equipment must navigate divergent regulatory regimes. The discussion noted that leading-edge manufacturing technologies remain geographically concentrated, which limits some countries’ ability to field the most advanced systems.
Speakers said this dynamic will shape international competitiveness: top-of-stack software innovation matters, but without equivalent access to chips, data and infrastructure the advantages are incomplete. That interplay is driving policy debates and firms’ decisions about where to locate manufacturing, data centers and testing facilities.
Across the session, a recurring theme emerged: the AI era is not only a software story. It is a physical systems challenge that requires coordination across manufacturing, energy, data collection and governance. Executives at the table agreed that capital and technical talent alone will not automatically clear these bottlenecks; targeted investments, new architectures and regulatory clarity will be needed to translate ambition into deployable systems.
Longer term, panelists suggested, the stresses revealed by today’s bottlenecks could catalyze innovation in chip design, model architectures and energy use that ultimately makes AI more sustainable and broadly accessible. In the near term, however, firms and governments alike are preparing for a constrained landscape where procurement, integration and policy choices will determine who can scale AI most effectively.
