Five AI economics engineers explain where things start

Earlier this week, five people with expertise at every layer of the AI ​​supply chain sat down at the Milken Global Conference in Beverly Hills, where they spoke with this editor about everything from chip shortages to orbital data centers to the possibility that the entire architecture underpinning the technology is wrong.

On stage with TechCrunch: Christophe Fouquet, CEO of ASML, the Dutch company that has a monopoly on extreme ultraviolet lithography machines without which modern chips would not exist; Francis D’Souza, Google Cloud’s chief operating officer, who is overseeing one of the largest infrastructure bets in the company’s history; Qasr Yunus, co-founder and CEO of Applied Intuition, a $15 billion physical AI company that started in simulation and has since moved into defense; Dmitry Shevelenko, chief business officer at Perplexity, an AI-powered agent search company; And Eve Bodnia, a quantum physicist who left academia to challenge the infrastructure that most of the AI ​​industry takes for granted at her startup Logical Intelligence. (Meta’s former chief AI scientist, Yann LeCun, signed on as founding president of the Technical Research Council earlier this year.)

Here’s what the five said:

The bottlenecks are real

The AI ​​boom is facing difficult physical limits, and the limitations start further down the pack than many may realize. Fouquet was the first to say so, describing the “tremendous acceleration in chip manufacturing,” and expressing his “strong belief” that despite all these efforts, “over the next two, three, maybe five years, the market will be in limited supply,” meaning that the supersized players — Google, Microsoft, Amazon, Meta — won’t get all the chips they are paying for, exactly.

DeSouza highlighted the scale of this problem — and how quickly it’s growing — reminding the audience that Google Cloud’s revenue surpassed $20 billion last quarter, up 63%, while its backlog — revenue committed but not yet delivered — nearly doubled in one quarter, from $250 billion to $460 billion. “The demand is real,” he said with impressive calm.

For Yunus, the constraint comes primarily from elsewhere. Applied Intuition builds autonomous systems for cars, trucks, drones, mining equipment and defense vehicles, and the bottleneck is not the silicon, but the data one can only collect by sending machines out into the real world and watching what happens. “You have to find it from the real world,” he said, and no amount of artificial simulation can fully fill that gap. “It will be a long time before you can fully train models that operate in the physical world artificially.”

The energy problem is also real

If chips are the first bottleneck, power is looming behind them. D’Souza confirmed that Google is exploring data centers in space as a serious response to energy constraints. “You have access to more abundant energy,” he noted. Of course, even in orbit, it’s not simple. D’Souza notes that space is a vacuum, so it eliminates convection, leaving radiation as the only way to shed heat into the surrounding environment (a process that is much slower and more difficult to engineer than the air-and-liquid cooling systems that data centers rely on today). But the company is still treating it as a legitimate route.

Somewhat unsurprisingly, de Souza’s deeper argument was for efficiency through integration. He suggested that Google’s strategy of co-engineering its entire AI stack — from custom TPU chips to models and agents — delivers profits in watts per flip, which a company that buys off-the-shelf components can’t replicate. “Running Gemini on TPUs is more power efficient than any other configuration,” he said, because chip designers know what’s coming in the model before they ship it. In a world where the availability of energy has become a major constraint on how far this technology can go, this type of vertical integration is a major competitive advantage.

Fouquet echoed this point later in the discussion. “There is nothing priceless,” he added. The industry is going through a strange moment right now, investing extraordinary amounts of capital, driven by strategic necessity. But more computing means more power, and more power comes at a price.

A different kind of intelligence

While the rest of the industry debates the size, architecture, and efficiency of inference within the large language model paradigm, Bodnia is building something entirely different.

Her company, Logical Intelligence, is built on so-called energy-based models (EBMs), a class of artificial intelligence that does not predict the next code in a sequence but instead attempts to understand the underlying rules of the data, in a way that she argues is closer to how the human brain actually works. “Language is the user interface between my brain and my brain,” she said. “The inference itself is not tied to any language.”

Its largest model reaches 200 million parameters – compared to hundreds of billion in leading MBAs – and claims to work thousands of times faster. Most importantly, it is designed to update its knowledge as data changes, rather than requiring retraining from scratch.

For chip design, robotics, and other fields where a system needs to understand physical rules rather than linguistic patterns, EBMs are a natural fit. “When you drive a car, you’re not looking for patterns in any language. You’re looking around, understanding the rules about the world around you, and making a decision.” It’s an interesting argument that is likely to attract more attention in the coming months, as the AI ​​field begins to question whether scale alone is enough.

Agents, barriers and trust

Shevelenko spent most of the conversation explaining how Perplexity’s technology has evolved from a research product into something he now calls a “digital agent.” Perplexity Computer, its latest offering, is designed not as a tool for knowledge workers to use, but as a team directed by the knowledge worker. “Every day you wake up and you have a hundred employees on your team,” he said of the opportunity. “What are you going to do to make the most of it?”

It’s a convincing move. It also raises obvious questions about control, so I asked them. His answer was detailed. Enterprise administrators can specify not only which connectors and tools an agent can access, but whether those permissions are read-only or read-write – a very important distinction when agents are working within corporate systems. When Comet, Perplexity’s computer user agent, takes action on a user’s behalf, it presents a plan and asks for approval first. Shevelenko said some users find the friction annoying, but he said he considers it necessary, especially after joining Lazard’s board, where he said he found himself unexpectedly sympathetic to the conservative instincts of a chief information security officer protecting a 180-year-old brand built entirely on customer trust. He added: “Details are the cornerstone of good security hygiene.”

Sovereignty, not just security

Yunus made what may be the panel’s most geopolitically charged observation, which is that physical AI and national sovereignty are intertwined in ways that purely digital AI has not been before.

The Internet initially spread as an American technology and only encountered opposition at the application layer — Ubers and DoorDashes — when the offline consequences became visible. Physical AI is different. Self-driving vehicles, defense drones, mining equipment, agricultural machinery—these things are manifesting in the real world in ways that governments can’t ignore, raising questions about safety, data collection, and who ultimately controls the systems that operate within state borders. “Almost every country is saying: ‘We don’t want this intelligence to be in physical form on our border, controlled by another country,'” he added. He told the crowd that there were fewer countries currently able to use a robotaxi than countries that possess nuclear weapons.

Fouquet framed it a little differently. China’s progress in AI is real, with the release of DeepSeek earlier this year sending something of a panic through parts of the industry, but that progress is restricted below the model layer. Without access to UV lithography, Chinese chipmakers cannot manufacture the most advanced semiconductors, and models built on older hardware operate at a disadvantage no matter how good the software is. “Today, in the United States, you have the data, you have access to computing, you have the chips, you have the talent,” Fouquet said. “China does very well at the top of the pile, but it lacks some of the elements below.”

Generation question

Toward the end of our discussion, an audience member asked the uncomfortable obvious question: Will all of this affect the next generation’s ability to think critically?

The answers were perhaps unsurprisingly optimistic, if not naive. De Souza pointed out the scale of problems that more powerful tools may eventually allow humanity to address. Consider neurological diseases whose biological mechanisms we do not yet understand, the removal of greenhouse gases, and network infrastructure that has been postponed for decades. “This should unleash us to the next level of creativity,” he added.

Shevlenko made a more sobering point: The entry-level job may be on the way out, but the ability to launch something independently has never been easier. “[For] Anyone who has computer confusion. . . The limitation is your curiosity and ability to act.

Yunus made a clear distinction between cognitive work and physical work. He pointed to the fact that the average age of an American farmer is 58, and that labor shortages in mining, long-distance trucking, and agriculture are chronic and growing — not because wages are too low, but because people don’t want these jobs. In these fields, physical AI does not replace willing workers. It fills a void that already exists and which looks to only deepen from here.