Earlier this week, five people from every layer of the AI supply chain gathered at the Milken Global Conference in Beverly Hills to discuss everything from chip shortages to orbital data centers to the possibility that the entire architecture behind the technology is wrong.
On stage at TechCrunch: Christophe Fouquet, CEO of ASML. ASML is a Dutch company that holds exclusive rights to extreme ultraviolet lithography equipment that would not exist without modern chips. Google Cloud COO Francis D’Souza is overseeing one of the biggest infrastructure bets in the company’s history. Qasar Younis, co-founder and CEO of Applied Intuition, a $15 billion physics AI company that started in simulation and then moved into defense. Dimitry Shevelenko, chief business officer of Perplexity, an AI-native search agent company. And quantum physicist Yves Bodnia left academia to challenge the basic architecture that most of the AI industry takes for granted with her startup, Logical Intelligence. (Yan LeCun, former chief AI scientist at Meta, signed on as founding chair of the technology research committee earlier this year.)
What the five people said is as follows.
The bottleneck is real
The AI boom has reached its physical limits, and its limitations begin further down the line than many realize. Fouquet was the first to say this, describing a “huge acceleration in chip manufacturing” and expressing his “strong belief” that despite their efforts, “the market will be limited in supply for the next two to three, maybe five years,” meaning hyperscalers like Google, Microsoft, Amazon, and Meta won’t be able to get what they pay for full stop.
DeSouza emphasized how big the problem is and how fast it is growing, reminding the audience that Google Cloud’s revenue increased 63% last quarter to more than $20 billion, while its backlog (revenue that has been promised 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 Yunis, the constraints come primarily from elsewhere. Applied Intuition builds autonomous systems for cars, trucks, drones, mining equipment, and defense vehicles. His bottleneck isn’t silicon. It’s data that can only be collected by sending machines into the real world and seeing what happens. “You have to find it from the real world,” he says, and no amount of synthetic simulation can completely close that gap. “It will take a long time before we can fully train a model that operates comprehensively on the physical world.”
tech crunch event
San Francisco, California
|
October 13-15, 2026
Energy problems are also serious
If chips are the first bottleneck, energy is looming close behind. DeSouza acknowledged that Google is considering data centers in space as a serious response to energy constraints. “We will have access to more abundant energy,” he noted. Of course, even in orbit, it’s not that simple. Because the universe DeSouza observed is a vacuum, there is no convection, and the only way to release heat to the surrounding environment is through radiation (a process much slower and more difficult to design than the air and liquid cooling systems that today’s data centers rely on). But the company still treats it as a legitimate pass.
The deeper discussion de Sousa made, somewhat unsurprisingly, was about efficiency through integration. Google’s strategy to co-engineer complete AI stacks, from custom TPU chips to models and agents, offers benefits in flops per watt (more computation per unit of energy) that companies that buy off-the-shelf components can’t match, he suggested. “Running Gemini on a TPU is much more energy efficient than other configurations,” he said, because chip designers know what the models will contain before they ship.
Fouquet made a similar point later in his discussion. “Nothing can be precious,” he said. The industry is currently in a strange time, with extraordinary amounts of capital being invested driven by strategic necessity. But more computing means more energy, and more energy comes at a price.
different types of intelligence
While other industries are debating scale, architecture, and inference efficiency within the large-scale language model paradigm, Bodnia is building something completely different.
Her company, Logical Intelligence, is built on so-called energy-based models (EBMs), a type of AI that tries to understand the underlying rules of data rather than predicting the next token in a series, which she claims is in some ways closer to how the human brain actually works. “Language is the user interface between my brain and your brain,” she said. “Inference itself does not belong to any language.”
Her largest models run with 200 million parameters, compared to the hundreds of billions of parameters in leading LLMs, and she claims they run thousands of times faster. More importantly, it is designed to update its knowledge as data changes, rather than requiring retraining from the beginning.
She argues that EBM is a more natural fit in chip design, robotics, and other fields where systems need to grasp physical rules rather than linguistic patterns. “When you drive a car, you’re not looking for patterns in language; you’re looking around, understanding the rules about the world around you, and making decisions.” This is an interesting discussion, and one that’s likely to receive more attention in the coming months, as the AI field begins to question whether scale alone is enough.
Agents, guardrails, and trust
Schevelenko spent much of the conversation explaining how Perplexity evolved from a search product to what is now called the “digital worker.” Its newest product, the Perplexity Computer, is designed not as a tool for knowledge workers to use, but as a staff member for knowledge workers to direct. “Every day I wake up, I have 100 people on my team,” he said of the opportunity. “What are you going to do to make the most of it?”
That’s a convincing sales pitch. It also raised obvious questions about controls, so I asked them. His answer was “granularity”. Enterprise administrators can specify which connectors and tools agents can access, as well as whether those permissions are read-only or read/write. This distinction is very important when agents operate within enterprise systems. When Comet, Perplexity’s computing agent, performs an action on your behalf, it first presents a plan and asks for your approval. Shevelenko said that while some find this friction annoying, he believes it is essential, especially since joining Lazard’s board, and said he has unexpectedly come to empathize with the CISO’s conservative instincts to protect a 180-year-old brand built entirely on customer trust. “Granularity is the foundation of good security hygiene,” he said.
Sovereignty, not just security
Younis offered what was likely the most geopolitical view of the panel: that physical AI and national sovereignty are intertwined in a way that never existed with purely digital AI.
The Internet initially became popular as an American technology, but only faced backlash at the application layer (Uber, DoorDashes) when offline effects became visible. Physical AI is different. Self-driving cars, defense drones, mining equipment, agricultural machinery, and more are coming into the real world in ways that governments can’t ignore, raising questions about safety, data collection, and who ultimately controls the systems operating within their borders. “Almost consistently, every country has said this: We don’t want this information to be physically on our borders and controlled by other countries,” he told the audience, adding that there are currently fewer countries that can deploy robotaxis than there are nuclear weapons.
Fouquet painted it a little differently. China’s AI advances are real, and the release of DeepSeek earlier this year caused near panic in some parts of the industry. However, its progress is limited below the model layer. Without access to EUV lithography, Chinese chipmakers cannot produce cutting-edge semiconductors, and models built on older hardware operate at even worse conditions, no matter how good the software. “Today, the United States has the data, it has access to computing, it has the chips, it has the talent. China is doing a very good job at the top of that, but it’s missing some pieces below that,” Fouquet said.
generational issues
Near the end of our panel, someone in the audience asked a clearly uncomfortable question. Will all this impact the critical thinking skills of the next generation?
As you might expect from people who have staked their careers on this technology, the answer was optimistic. DeSouza was quick to point out the scale of the problem that humanity could eventually address with more powerful tools. Consider neurological diseases whose biological mechanisms are not yet understood, greenhouse gas removal, and power grid infrastructure that has been delayed for decades. “This should take us to the next level of creativity,” he said.
Mr. Shebelenko made a more pragmatic point. Entry-level jobs may be disappearing, but the ability to start something independent is more accessible than ever. “For those who own a Perplexity Computer, the only constraint is your own curiosity and independence.”
Younis made the clearest distinction between knowledge work and manual labor. He pointed to the fact 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 the wages are too low, but because people don’t want these jobs. Physical AI will not replace motivated workers in these fields. It’s filling a void that already exists, and it looks like it’s only going to get deeper.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
