The AI supply chain is hitting hard walls. Here’s what five top executives told us.

Five leaders spanning the entire AI supply chain gathered at the Milken Global Conference in Beverly Hills this week to discuss where the industry is breaking. From chip shortages to orbital data centers, the conversation revealed a sector grappling with real physical limits.

Christophe Fouquet, CEO of ASML, started with a blunt warning: chip manufacturing is accelerating fast, but supply will lag demand for the next two to five years. Hyperscalers like Google, Microsoft, and Amazon won’t get all the chips they’ve ordered. Francis deSouza, COO of Google Cloud, backed this up with numbers: Google Cloud’s revenue hit $20 billion last quarter, up 63%, and its backlog nearly doubled to $460 billion. “The demand is real,” he said.

Qasar Younis, CEO of Applied Intuition, pointed to a different bottleneck: real-world data for physical AI. His company builds autonomy for cars, drones, and defense vehicles. Synthetic data helps, but you can’t fully train models on the physical world without sending machines into the field. “You have to find it from the real world,” he explained.

Energy is the next constraint. DeSouza confirmed Google is exploring space-based data centers for better access to solar power, though heat dissipation in a vacuum is a challenge. He stressed that Google’s custom TPU chip-to-model integration delivers far better energy efficiency than off-the-shelf setups.

Eve Bodnia, a quantum physicist turned founder, argued the industry’s core architecture may be wrong. Her startup, Logical Intelligence, builds energy-based models that learn rules, not just token patterns. Her largest model runs 200 million parameters—compared to hundreds of billions in LLMs—and claims it’s thousands of times faster and adapts without retraining. “Language is a user interface,” she said. “Reasoning itself isn’t attached to language.”

Dimitry Shevelenko of Perplexity described the shift from search to “digital workers.” Perplexity Computer lets users direct a hundred AI agents, but control is granular: administrators set read-only or read-write permissions, and agents must present plans for approval. “Granularity is the bedrock of good security hygiene,” he noted.

Younis raised a geopolitical point: physical AI ties directly to national sovereignty. Fewer countries can operate a robotaxi than possess nuclear weapons, he said. Fouquet added that China’s AI progress is real but constrained by a lack of EUV lithography for advanced chips.

When asked about impacts on the next generation’s critical thinking, the panel was optimistic. DeSouza sees AI tackling problems like neurological diseases and grid infrastructure. Shevelenko said the entry-level job may fade, but the ability to launch something independently has never been more accessible. Younis noted that physical AI fills labor shortages in farming, mining, and trucking—jobs people don’t want anyway.

Source: TechCrunch