Earlier this week, five people who touch every layer of the AI supply chain sat down at the Milken Global Conference in Beverly Hills, where they talked to this editor about everything from chip shortages to orbital data centers.
On stage with TechCrunch: Christophe Fuquet, CEO of ASML, the Dutch company with a monopoly on extreme ultraviolet lithography machines without which modern chips wouldn’t exist; Francis deSouza, COO of Google Cloud, overseeing one of the biggest infrastructure bets in corporate history; Ghasar Younis, co-founder and CEO of Applied Intuition, a $15 billion physical AI company; Dmitry Shevelenko, COO of Perplexity, an AI-agent search company; and Eve Bodnia, a quantum physicist who left academia to challenge the basic architecture of the AI industry, is a natural at the logical intelligence startup. (Ian LeCun, Meta’s former chief artificial intelligence scientist, signed on as founding chairman of the technical research board earlier this year.)
Here’s what five people said:
Bottlenecks are real
The AI boom faces tough physical limits, and the limits start at a lower level than many realize. Fouquet was the first to say this, describing the “tremendous acceleration of chip production” and expressing his “strong belief” that despite all these efforts, “the market will be in limited supply in the next two, three, maybe five years,” meaning that the hyperscalers—Google, Microsoft, Amazon, Meta—will not fully pay for all the chips.
DeSouza underscored just how big and how fast-growing this is, telling the audience that Google Cloud’s revenue surpassed $20 billion last quarter, up 63%, while trailing revenue nearly doubled in one quarter, from $250 billion to $460 billion. “The demand is real,” he said with impressive calm.
For Eunice, the limitation comes primarily from elsewhere. Applied Intuition is building autonomous systems for cars, trucks, drones, mining equipment and defense vehicles, and its bottleneck isn’t silicon—it’s the data that can only be collected by sending machines into the real world and watching what happens. “You have to find it in the real world,” he said, and no synthetic simulation completely closes that gap. “It will be a long time before you can fully train models that work synthetically in the physical world.”
Techcrunch event
San Francisco, CA
|
October 13-15, 2026
The energy problem is also real
If chips are the first bottleneck, it’s energy behind it. DeSouza confirmed that Google is exploring data centers in space as a serious response to energy constraints. “You get access to more abundant energy,” he said. Of course, even in orbit, this is not simple. The space DeSouza observed is a vacuum, thus eliminating convection, leaving radiation as the only way to transfer heat to the environment (a much slower and more difficult-to-engineer process than the air and liquid cooling systems that data centers rely on today). But the company still treats it as a legitimate way.
The deeper argument de Souza made was, somewhat unsurprisingly, about efficiency through integration. Google’s strategy of co-engineering the entire AI stack — from individual TPU chips to models and agents — is paying dividends in flops (more computations per unit of energy) per watt, he suggested. “Running Gemini on TPUs is more energy efficient than any other configuration,” he said, because chip designers know what’s going on in the model before it ships.
Fouquet made a similar point later in the discussion. “Nothing is priceless,” he said. The industry is at a strange point right now, investing extraordinary amounts of capital out of 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 discusses scale, architecture, and inference efficiency within the broad language model paradigm, Bodnia is building something very different.
His company, Logical Intelligence, is built on energy-based models (EBM), a class of artificial intelligence that doesn’t predict the next cue in a sequence, but rather tries to understand the rules underlying the data. “Language is the user interface between my brain and yours,” he said. “Reason itself is not tied to any language.”
Its largest model has 200 million parameters—compared to hundreds of billions in leading LLMs—and it claims to run thousands of times faster. More importantly, it is designed to update its knowledge as information changes, rather than requiring retraining from scratch.
For chip design, robotics, and other fields where the system needs to grasp physical rules rather than linguistic patterns, he argues that EBMs are a more natural fit. “When you’re driving 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 decisions.” It’s an interesting debate, and one that will gain more attention in the coming months, given that the field of AI is beginning to question whether scale alone is enough.
Agents, custodians and trusts
Shevelenko spent much of the conversation explaining how Confusion evolved from a search product to what he now calls a “digital worker.” Perplexity Computer, its newest offering, is designed not as a tool used by a knowledge worker, but as a staff led by a knowledge worker. “You wake up every day and you have a hundred employees on your team,” he said. “What are you going to do to make the most of it?”
It has an attractive pitch; which raises obvious questions about control, so I asked them. His answer was: granularity. Enterprise administrators can specify not only which connectors and tools an agent can access, but also whether those permissions are read-only or read-write—a distinction that is critical when agents operate on enterprise systems. When Comet, Perplexity’s computer usage agent, performs actions on behalf of the user, it presents a plan and first requests approval. Some users find that friction annoying, Shevelenko said, but he said he finds it important, especially after joining Lazard’s board, where he finds himself unexpectedly sympathetic to the conservative instincts of a CISO who has built a 180-year-old brand built entirely on customer trust. “Granularity is the foundation of good safety hygiene,” he said.
Sovereignty, not just security
Younis offered the panel’s most geopolitically charged observation, namely that physical AI and national sovereignty are intertwined in a way that purely digital AI never is.
The Internet first took off as an American technology, and when offline results showed up, it only faced pushback at the application level—Ubers and DoorDashes. Physical AI is different. Autonomous vehicles, defense drones, mining equipment, agricultural machinery — these are manifesting in real-world ways that governments can’t ignore, raising questions about security, data collection, and ultimately who controls systems operating within a country’s borders. “Almost consistently, every country is saying: we don’t want this intelligence to be physically controlled on our borders by another country.” He told the crowd that fewer countries than those possessing nuclear weapons could currently use robotic missiles.
Fouquet framed it a little differently. China’s AI progress is real — the release of DeepSeek earlier this year sent some parts of the industry into something close to panic — but that progress is limited below the model layer. Without access to EUV lithography, Chinese chipmakers can’t produce cutting-edge semiconductors, and models built on older hardware are at a compound disadvantage, no matter how good the software. “In the United States today, you have the data, you have the computing power, you have the chips, you have the talent. China is doing very well at the top of the stack, but some elements are missing at the bottom,” Fuquet said.
A generational question
Near the end of our panel, an audience member asked an obviously concerned question: will all of this affect the next generation’s ability to think critically?
As you might expect from people who have built their careers on this technology, the responses were upbeat. DeSouza immediately pointed to the scale of the problems that more powerful tools might eventually allow humanity to solve. Think neurological diseases whose biological mechanisms we do not yet understand, greenhouse gas emissions, and network infrastructure that has been delayed for decades. “It should take us to the next level of creativity,” he said.
Shevelenko made a more pragmatic point: entry-level work may be disappearing, but the opportunity to launch something independently has never been more accessible. “(For) anyone with a Confusion Computer… the limit is your own interest and agency.”
Eunice drew the sharpest distinction between knowledge work and physical labor. He noted that the average American farmer is 58 years old, and that labor shortages in mining, long-haul trucking and agriculture are chronic and growing — not because wages are too low, but because people don’t want those jobs. In these domains, physical AI does not replace willing workers. It fills a gap that already exists and is only going to get deeper from here.
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
