AI at the Crossroads: Supply Chains, Energy Limits, and the Future of Intelligence

Beverly Hills, California – The artificial intelligence revolution is colliding with the hard realities of physics, geopolitics, and human cognition, according to a panel of industry leaders at this week’s Milken Global Conference. From semiconductor shortages to orbital data centers and radical new AI architectures, the discussion laid bare the challenges—and existential questions—facing the sector as it reshapes economies and societies worldwide.

The panel, moderated by TechCrunch, brought together five executives whose companies span every layer of the AI supply chain:

• Christophe Fouquet, CEO of ASML, the Dutch firm that holds a monopoly on the extreme ultraviolet (EUV) lithography machines essential for manufacturing cutting-edge chips.
• Francis deSouza, COO of Google Cloud, overseeing one of the largest AI infrastructure expansions in corporate history.
• Qasar Younis, CEO of Applied Intuition, a $15 billion AI company specializing in autonomous vehicles and defense systems.
• Dimitry Shevelenko, Chief Business Officer of Perplexity, the AI-native search firm now pivoting toward autonomous digital agents.
• Eve Bodnia, a quantum physicist and founder of Logical Intelligence, a startup challenging the foundational assumptions of modern AI.

Their insights painted a picture of an industry racing against its own limitations—and grappling with the societal implications of its success.

The Physical Limits of AI’s Expansion

1. The Chip Bottleneck

The AI boom is straining global semiconductor supply chains, with demand far outstripping manufacturing capacity. Fouquet warned that despite massive investments in chip production, the market will remain “supply-limited” for the next three to five years.

“Hyperscalers like Google, Microsoft, and Amazon won’t get all the chips they’ve ordered—it’s simply impossible,” he said.

DeSouza underscored the scale of the issue, revealing that Google Cloud’s revenue surged to $20 billion last quarter, growing 63% year-over-year, while its backlog—committed but undelivered revenue—nearly doubled from $250 billion to $460 billion in just three months. “The demand is real,” he said, “and it’s accelerating faster than supply can keep up.”

For Younis, the bottleneck isn’t just chips—it’s real-world data. Applied Intuition builds AI for autonomous vehicles and defense systems, where synthetic simulations can’t fully replace physical testing. “You have to gather data from the real world,” he said. “No amount of simulation closes that gap entirely.”

2. The Energy Crunch

If chips are the first constraint, energy is the looming second. DeSouza confirmed that Google is seriously exploring space-based data centers to circumvent terrestrial power limitations.

“In orbit, you get access to more abundant energy,” he said. But the engineering hurdles are immense: space’s vacuum eliminates convection, forcing engineers to rely solely on radiation for cooling—a far slower and more complex process than traditional air or liquid cooling.

Google’s strategy of vertically integrating its AI stack—from custom TPU chips to models like Gemini—gives it an efficiency edge, DeSouza argued. “Running Gemini on TPUs is far more energy-efficient than off-the-shelf hardware because the chips are optimized for the models.”

Fouquet echoed the point, warning that AI’s exponential growth comes at a cost. “Nothing can be priceless,” he said. “More compute means more energy, and energy has a price.”

Rethinking AI’s Foundations

While most of the industry focuses on scaling large language models (LLMs), Bodnia’s startup, Logical Intelligence, is pursuing a fundamentally different approach: energy-based models (EBMs).

Unlike LLMs, which predict the next word in a sequence, EBMs attempt to infer the underlying rules of data—a method Bodnia claims is closer to human cognition. “Language is just an interface between brains,” she said. “Reasoning itself isn’t tied to words.”

Her largest model contains just 200 million parameters—a fraction of the hundreds of billions in leading LLMs—yet she says it runs thousands of times faster and updates dynamically without full retraining.

For applications like chip design or robotics, where understanding physical laws matters more than linguistic patterns, EBMs could be a better fit. “When you drive a car, you’re not searching for linguistic patterns—you’re interpreting real-world rules,” she said.

Her work signals a growing debate in AI: Is sheer scale enough, or do we need smarter architectures?

Agents, Trust, and the Future of Work

Shevelenko outlined Perplexity’s evolution from a search engine into an AI “digital worker.” Its latest product, Perplexity Computer, is designed not as a tool but as an autonomous assistant—akin to having “a hundred staff on your team.”

But autonomy raises security concerns. Shevelenko emphasized granular control: enterprises can restrict agents to read-only access or require approval before taking actions. “Some users find the friction annoying, but it’s essential for trust,” he said, citing his experience on Lazard’s board, where security is paramount.

Geopolitics and Physical AI

Younis delivered one of the panel’s most provocative insights: physical AI—autonomous cars, drones, mining equipment—is inherently geopolitical in ways digital AI never was.

While the internet spread globally with minimal resistance, physical AI forces governments to confront safety, data sovereignty, and control. “Almost every country is saying: we don’t want foreign-controlled AI operating inside our borders,” Younis said. He noted that fewer nations can deploy robotaxis than possess nuclear weapons.

Fouquet added that China’s AI progress, while impressive, is constrained by semiconductor restrictions. Without EUV lithography, Chinese firms can’t produce the most advanced chips, leaving them at a compounding disadvantage.

Will AI Stifle Human Thinking?

An audience member posed the thorniest question: Could AI erode critical thinking in future generations?

The panelists struck an optimistic tone. DeSouza pointed to AI’s potential to solve intractable problems—neurological diseases, climate change, aging infrastructure. “This should unleash the next level of human creativity,” he said.

Shevelenko argued that AI lowers barriers to entrepreneurship: “With tools like Perplexity Computer, the only limit is your curiosity.”

Younis highlighted labor shortages in farming, trucking, and mining, where AI isn’t replacing workers but filling gaps left by an aging workforce. “People don’t want these jobs,” he said. “AI is stepping in where humans won’t.”

Conclusion: A Future Shaped by Constraints

The discussion revealed an industry at an inflection point—one where technological ambition is increasingly checked by physical, economic, and ethical realities. Whether through orbital data centers, radical new architectures, or geopolitical maneuvering, AI’s next phase will be defined not just by breakthroughs, but by how it navigates its own limits.

As Fouquet put it: “Nothing can be priceless.” The question now is what price the world is willing to pay.