The AI boom is running into a wall, and it’s made of silicon and electricity.
At the Milken Global Conference in Beverly Hills this week, five people who sit at different layers of the AI supply chain gathered to talk about the infrastructure crisis nobody wants to admit is already here. The conversation, involving leaders from ASML, Google Cloud, Applied Intuition, Perplexity, and a quantum physicist building AI from scratch, revealed something uncomfortable: we’re scaling compute faster than we can physically support it.
The problem starts with chips. Christophe Fouquet, CEO of ASML, controls one of the world’s most important choke points. His company manufactures the extreme ultraviolet lithography machines without which modern semiconductors simply don’t exist. He didn’t mince words about what’s coming: despite massive acceleration in chip manufacturing globally, “for the next two, three, maybe five years, the market will be supply limited.” Translation: Google, Microsoft, Amazon, and Meta won’t get all the hardware they’re paying for. Full stop.
The scale of demand is staggering. Francis deSouza, COO of Google Cloud, shared numbers that crystallized just how broken the supply chain has become. Google Cloud’s revenue hit $20 billion last quarter, growing 63 percent year-over-year. But here’s the real problem: their backlog nearly doubled in a single quarter, from $250 billion to $460 billion. “The demand is real,” deSouza said, with the kind of calm that only comes from confronting an existential constraint you can’t immediately fix.
When Chips Aren’t the Bottleneck
For some companies, the silicon shortage is almost beside the point.
Qasar Younis, CEO of Applied Intuition, builds autonomy systems for vehicles, drones, mining equipment, and defense applications. His constraint isn’t chips. It’s data. Real-world data. You can simulate only so much before the models fail to generalize into the physical world. “You have to find it from the real world,” he said at the conference. “There will be a long time before you can fully train models that run on the physical world synthetically.” No amount of synthetic training closes that gap entirely. It’s a problem that scales with ambition, and it can’t be solved with more compute alone.
This matters more than it might seem. Physical AI and national sovereignty are becoming entangled in ways purely digital technology never was. Autonomous vehicles, defense drones, agricultural machines, mining equipment—these systems operate in the real world in ways governments can’t ignore. “Almost consistently, every country is saying: we don’t want this intelligence in a physical form in our borders, controlled by another country,” Younis observed. Fewer nations can currently deploy a robotaxi than possess nuclear weapons. That’s not hyperbole. That’s geopolitics.
The Energy Crisis Nobody Wants to Admit
Behind the chip shortage lurks an even scarier constraint: energy.
DeSouza confirmed that Google is seriously exploring orbital data centers as a response to energy constraints. The logic is simple: space has abundant solar energy, no atmospheric interference, and none of the cooling challenges that plague terrestrial facilities. Of course, even in a vacuum, physics doesn’t cooperate. Without air or liquid to carry heat away, radiation becomes the only option, and it’s exponentially slower and harder to engineer than the cooling systems data centers rely on today. Google is treating it as a legitimate path anyway.
The real insight deSouza offered was about efficiency through integration. Google’s strategy of co-engineering its entire AI stack—from custom TPU chips through models and agents—delivers efficiency gains in flops per watt that companies buying off-the-shelf components simply can’t match. “Running Gemini on TPUs is much more energy efficient than any other configuration,” he explained, because the hardware designers know what the models will demand before they ship. It’s a vertical integration advantage, and it suggests that the future belongs to companies that can engineer at scale across the full stack.
Fouquet echoed the sentiment from a different angle: “Nothing can be priceless.” The industry is investing capital at extraordinary levels, driven by strategic necessity. But more compute demands more energy, and energy always carries a price. At some point, that math stops working.
What If the Architecture Is Wrong?
While the rest of the industry debates scale within the large language model paradigm, Eve Bodnia is building something radically different.
Her startup, Logical Intelligence, is built on energy-based models—a class of AI that doesn’t predict the next token in a sequence but instead attempts to understand the rules underlying data. She argues this approach aligns more closely with how human brains actually work. “Language is a user interface between my brain and yours,” she said. “The reasoning itself is not attached to any language.”
Her largest model runs to 200 million parameters compared to the hundreds of billions in leading LLMs, and she claims it runs thousands of times faster. More importantly, it’s designed to update knowledge as data changes rather than requiring complete retraining. For domains like chip design and robotics where systems need to grasp physical rules rather than linguistic patterns, she argues EBMs are the natural fit. “When you drive a car, you’re not searching for patterns in any language. You look around you, understand the rules about the world around you, and make a decision.”
It’s a compelling argument, especially now that the AI field is beginning to ask whether technology built on scale alone is sustainable. The consensus is cracking.
The New Business Model: Agents, Not Tools
Dimitry Shevelenko spent much of the conversation explaining how Perplexity has evolved from search into what the company now calls a “digital worker.”
Perplexity Computer isn’t designed as a tool you use. It’s designed as a staff you direct. “Every day you wake up and you have a hundred staff on your team,” Shevelenko said, framing the opportunity with genuine ambition. “What are you going to do to make the most of it?”
The control problem, though, isn’t trivial. When agents act inside corporate systems, the stakes get real fast. His answer was granularity. Enterprise administrators can specify not just which connectors agents can access, but whether permissions are read-only or read-write—a distinction that matters enormously. When Perplexity’s computer-use agent takes action, it presents a plan and requests approval first. Some users find the friction annoying. Shevelenko sees it as essential, particularly after joining the board of Lazard, where he’s developed sympathy for the conservative instincts of security leaders protecting 180-year-old brands built entirely on client trust. “Granularity is the bedrock of good security hygiene,” he said.
It’s a revealing moment. Even the most bullish advocates for AI agents understand that the friction matters.
The Geopolitical Layer
Fouquet framed the competitive landscape differently than most silicon valley optimists do.
China’s AI progress is real. DeepSeek’s release earlier this year sent something close to panic through parts of the industry. But that progress is constrained below the model layer. Without access to EUV lithography, Chinese chipmakers cannot manufacture the most advanced semiconductors. Models built on older hardware operate at a compounding disadvantage no matter how sophisticated the software becomes. “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,” Fouquet said.
It’s a revealing asymmetry. The U.S. doesn’t just lead in business and technology. It controls the physical layer that everyone else depends on.
The Question Everyone Avoids
Near the end of the panel, someone asked the uncomfortable question everyone was thinking: Is all of this going to destroy the next generation’s capacity for critical thinking?
The answers were optimistic, as you’d expect. DeSouza immediately pointed to scale: neurological diseases we don’t understand, greenhouse gas removal, deferred grid infrastructure. “This should unleash us to the next level of creativity,” he said. Shevelenko offered something more pragmatic. The entry-level job might be disappearing, but the ability to launch something independently has never been more accessible. “The constraint is your own curiosity and agency.”
Younis drew the sharpest distinction. The average American farmer is 58 years old. Labor shortages in mining, trucking, and agriculture aren’t happening because wages are too low. They’re happening because people don’t want those jobs anymore. Physical AI isn’t displacing willing workers. It’s filling a void that already exists and keeps deepening.
Maybe that’s the real story here. We’re not replacing capability. We’re automating what nobody wants to do anyway, while the infrastructure constraints get worse every quarter, and the companies that can engineer vertically across the entire stack consolidate dominance.
The question isn’t whether AI will change everything. It will. The question is whether we’ll have solved the physics before we’ve solved the policy.
