Earlier this week, a remarkable gathering took place at the Milken Global Conference in Beverly Hills. Five influential figures, each touching a critical layer of the artificial intelligence supply chain, convened to discuss the pressing challenges and future trajectories of the AI economy. Their candid conversation, moderated by TechCrunch, delved into everything from persistent chip shortages to radical ideas like orbital data centers and even questioned the fundamental architecture underpinning much of today’s AI advancements.
The esteemed panel featured Christophe Fouquet, CEO of ASML, the Dutch company with a monopoly on crucial extreme ultraviolet lithography machines, essential for modern chip manufacturing. He was joined by Francis deSouza, COO of Google Cloud, who is spearheading one of the tech giant’s most significant infrastructure investments. Also contributing were Qasar Younis, co-founder and CEO of Applied Intuition, a prominent physical AI company, and Dimitry Shevelenko, Chief Business Officer of Perplexity, known for its AI-native search-to-agents platform.
Rounding out the expert lineup was Eve Bodnia, a quantum physicist and founder of Logical Intelligence. Bodnia has courageously ventured from academia to challenge the very foundational architecture that much of the AI industry currently takes for granted, promising a fresh perspective on intelligence.
The AI Boom’s Physical Bottlenecks
The current AI boom is encountering tangible physical limits, starting much further down the technology stack than many might realize. Christophe Fouquet from ASML was unequivocal, describing a “huge acceleration of chips manufacturing” but expressing his firm belief that for the next two, three, maybe five years, the market will remain supply-limited. This means even hyperscalers like Google, Microsoft, Amazon, and Meta won’t receive all the advanced chips they’re demanding.
Francis deSouza underscored the immense scale and rapid growth of this demand, revealing staggering figures from Google Cloud. Its revenue topped $20 billion last quarter, growing 63%, while its backlog of committed yet undelivered revenue nearly doubled from $250 billion to an astounding $460 billion in a single quarter. DeSouza calmly affirmed, “The demand is real.”
For Qasar Younis, the primary constraint for Applied Intuition, which builds autonomy systems for a range of physical vehicles from cars to defense drones, lies elsewhere. Their bottleneck isn’t silicon, but rather the real-world data that can only be gathered by deploying machines and observing their interactions. He emphasized that no amount of synthetic simulation can fully bridge this gap, predicting a long road ahead before models running on the physical world can be entirely trained synthetically.
Beyond chips and data, energy emerges as another colossal hurdle. DeSouza confirmed that Google is seriously exploring data centers in space as a radical solution to energy constraints, noting the potential for more abundant power access in orbit. While acknowledging the complex engineering challenges of heat dissipation in a vacuum, Google considers this a legitimate future path.
Google’s strategy of vertically integrating its entire AI stack, from custom TPU chips to models and agents, offers significant energy efficiency advantages. DeSouza highlighted that “Running Gemini on TPUs is much more energy efficient than any other configuration.” This integrated approach allows chip designers to anticipate model requirements, providing a major competitive edge in a world where energy availability increasingly limits AI’s potential.
Reimagining AI: New Architectures and Applications
While much of the industry focuses on scaling and optimizing large language models, Eve Bodnia of Logical Intelligence is forging a different path. Her startup is built on energy-based models (EBMs), an AI class that aims to understand the underlying rules of data rather than just predicting the next token. She posits this approach is closer to how the human brain truly functions, viewing language as merely an interface for deeper reasoning.
Bodnia’s largest EBM boasts a mere 200 million parameters, a stark contrast to the hundreds of billions found in leading LLMs, yet she claims it runs thousands of times faster. Crucially, these models are designed to update their knowledge as data evolves, eliminating the need for costly and time-consuming retraining from scratch. For applications like chip design, robotics, or any domain requiring a grasp of physical rules over linguistic patterns, EBMs present a compelling, more natural fit.
Dimitry Shevelenko detailed Perplexity’s evolution from a search product into what they now call a “digital worker.” Their newest offering, Perplexity Computer, is envisioned not as a tool for a knowledge worker, but as a team of AI staff that a knowledge worker directs. Shevelenko articulated this opportunity as “every day you wake up and you have a hundred staff on your team,” prompting users to consider how to maximize this newfound capacity.
This powerful concept naturally raises questions about control and security, which Shevelenko addressed through the principle of granularity. Enterprise administrators can meticulously specify not only which connectors and tools an agent can access but also whether those permissions are read-only or read-write. This distinction is vital for agents operating within corporate systems, ensuring robust security hygiene.
Geopolitical Dynamics and AI’s Future
Qasar Younis offered a geopolitically significant insight: physical AI is intrinsically linked with national sovereignty in a way purely digital AI never was. He explained that while the internet initially spread as American technology, pushback only arose at the application layer when offline consequences became apparent. Physical AI, however, manifests directly in the real world through autonomous vehicles, defense drones, and agricultural machinery, making it an immediate concern for governments.
Younis observed a consistent global sentiment: “Almost consistently, every country is saying: we don’t want this intelligence in a physical form in our borders, controlled by another country.” He startlingly noted that fewer nations can currently field a robotaxi than possess nuclear weapons, highlighting the strategic significance of this technology.
Christophe Fouquet contributed a different perspective on global AI competition, particularly regarding China. While acknowledging China’s real progress in AI—exemplified by DeepSeek’s recent release—he stressed that this advancement is ultimately constrained at the hardware level. Without access to advanced EUV lithography, Chinese chipmakers cannot produce the most cutting-edge semiconductors. Models built on older hardware, he argued, operate at a compounding disadvantage regardless of software sophistication.
Fouquet summarized the current landscape by stating, “Today, in the United States, you have the data, you have the computing access, you have the chips, you have the talent. China does a very good job on the top of the stack, but is lacking some elements below.” This highlights a critical imbalance in the global AI supply chain.
As the panel neared its conclusion, an audience member posed a crucial question: would all this AI innovation impact the next generation’s capacity for critical thinking? The panelists offered generally optimistic, though not naive, responses.
DeSouza envisioned AI as a tool to tackle previously insurmountable challenges, like neurological diseases, greenhouse gas removal, and neglected infrastructure. He expressed hope that “This should unleash us to the next level of creativity,” suggesting AI will elevate human potential rather than diminish it. Shevelenko echoed this sentiment, arguing that while some entry-level jobs might fade, the ability to launch independent ventures has never been more accessible. He noted, “the constraint is your own curiosity and agency.”
Younis drew a sharp distinction between knowledge work and physical labor, pointing to chronic and growing labor shortages in sectors like farming, mining, and long-haul trucking. In these domains, physical AI isn’t displacing willing workers; it’s filling a crucial void that people are increasingly unwilling to occupy. This perspective suggests AI will address profound societal needs rather than simply replacing human jobs.
Source: TechCrunch – AI
