Five architects of the AI economy explain where the wheels are coming off

Earlier this week, five influential figures in the AI supply chain convened at the Milken Global Conference in Beverly Hills. They discussed a range of topics with this editor, including chip shortages, orbital data centers, and the potential flaws in the foundational architecture of AI technology.

Joining JS on stage were: Christophe Fouquet, CEO of ASML, the Dutch company with a monopoly on extreme ultraviolet lithography machines essential for modern chips; Francis deSouza, COO of Google Cloud, who is managing one of the largest infrastructure investments in corporate history; Qasar Younis, co-founder and CEO of Applied Intuition, a $15 billion physical AI company transitioning from simulation to defense; Dimitry Shevelenko, chief business officer of Perplexity, an AI-native search-to-agents company; and Eve Bodnia, a quantum physicist who left academia to challenge conventional AI architecture at her startup, Logical Intelligence. (Yan LeCun, Meta’s former chief AI scientist, joined as founding chair of its technical research board earlier this year.)

Here’s what the five had to say:

The bottlenecks are real

The AI industry is confronting significant physical constraints, which become apparent further down the stack than many might expect. Fouquet highlighted a “huge acceleration of chip manufacturing,” expressing a “strong belief” that the market will be supply-limited for the next two to five years, meaning major tech companies like Google, Microsoft, Amazon, and Meta will not receive all the chips they have paid for.

DeSouza emphasized the rapid growth and seriousness of this issue, noting that Google Cloud’s revenue surpassed $20 billion last quarter, with a growth rate of 63%. The company’s backlog, representing committed but undelivered revenue, nearly doubled from $250 billion to $460 billion in just one quarter. “The demand is real,” he stated calmly.

Younis identified a different constraint for Applied Intuition, which develops autonomy systems for various vehicles. His main bottleneck is not silicon, but the data that can only be gathered by observing real-world machine interactions. “You have to find it from the real world,” he said, indicating that synthetic simulation cannot entirely bridge this gap. “It will take a long time to fully train models for the physical world synthetically.”

The energy problem is also real

Following chips, energy emerges as the next pressing issue. DeSouza confirmed that Google is seriously considering space-based data centers to address energy constraints. “You get access to more abundant energy,” he noted. However, space presents its own challenges, as the vacuum environment negates convection, leaving radiation as the sole method for heat dissipation, which is slower and more complex than current air and liquid cooling systems. Nevertheless, the company sees this as a viable path.

De Souza further argued for efficiency through integration. Google’s approach of co-engineering its entire AI stack, from custom TPU chips to models and agents, maximizes computational efficiency per unit of energy, which off-the-shelf components cannot match. “Running Gemini on TPUs is much more energy efficient than any other configuration,” he stated, as chip designers are aware of model requirements before deployment.

Fouquet echoed this sentiment later, stating, “Nothing can be priceless.” The industry is in a unique phase, heavily investing capital due to strategic necessity. However, increased computation demands more energy, which comes at a cost.

A different kind of intelligence

While the industry focuses on scale, architecture, and efficiency within the large language model framework, Bodnia is pursuing a distinct approach. Her company, Logical Intelligence, employs energy-based models (EBMs), which aim to understand the rules underlying data rather than predicting the next token in a sequence. She claims this approach more closely mirrors human cognitive processes. “Language is a user interface between my brain and yours,” she explained. “The reasoning itself is not attached to any language.”

Her largest model comprises 200 million parameters, significantly fewer than leading LLMs’ billions, yet she claims it operates thousands of times faster. Crucially, it updates its knowledge as data changes, eliminating the need for retraining from scratch.

In chip design, robotics, and other fields requiring comprehension of physical laws rather than linguistic patterns, EBMs may provide a more natural fit. “When driving a car, you’re not searching for patterns in any language. You look around, understand the environment, and make decisions.” This perspective is likely to gain attention as the AI sector questions whether scale alone suffices.

Agents, guardrails, and trust

Shevelenko detailed how Perplexity has transitioned from a search tool to a “digital worker.” The latest product, Perplexity Computer, is designed as a team directed by a knowledge worker rather than a tool they use. “Every day you wake up and you have a hundred staff on your team. What are you going to do to make the most of it?” he asked.

While the concept is compelling, it raises control concerns, which I addressed. Shevelenko’s response was granularity. Enterprise administrators can control not only which connectors and tools an agent can access but also set read-only or read-write permissions, crucial when agents operate within corporate systems. When Comet, Perplexity’s computer-use agent, acts on a user’s behalf, it presents a plan for approval first. Some users find this friction annoying, but Shevelenko considers it essential, especially after joining Lazard’s board, where he empathized with a CISO’s conservative instincts protecting a 180-year-old brand built on client trust. “Granularity is the bedrock of good security hygiene,” he said.

Sovereignty, not just safety

Younis highlighted a geopolitically charged observation: physical AI and national sovereignty are intertwined in ways digital AI never was.

The internet spread as American technology and faced resistance only at the application layer, like Uber and DoorDash, when offline consequences became apparent. Physical AI differs. Autonomous vehicles, defense drones, mining equipment, and agricultural machines manifest in the real world in ways governments cannot ignore, raising safety, data collection, and control questions within national borders. “Almost consistently, every country is saying: we don’t want this intelligence in a physical form in our borders, controlled by another country.” Fewer nations can field a robotaxi than possess nuclear weapons, he noted.

Fouquet offered another perspective. China’s AI advances are undeniable—DeepSeek’s release earlier this year caused industry panic. However, progress is limited below the model layer. Without EUV lithography access, Chinese chipmakers cannot produce the most advanced semiconductors, and models on older hardware face compounded disadvantages, regardless of software quality. “Today, in the United States, you have the data, computing access, chips, and talent. China excels at the top of the stack but lacks some elements below,” Fouquet explained.

The generation question

Near the panel’s conclusion, an audience member posed the uncomfortable question: will this impact the next generation’s critical thinking ability?

The panelists were optimistic, as expected from those invested in this technology. DeSouza pointed to the scale of problems that more powerful tools might help solve, such as neurological diseases with unknown biological mechanisms, greenhouse gas removal, and long-deferred grid infrastructure improvements. “This should unleash us to the next level of creativity,” he said.

Shevelenko offered a pragmatic view: entry-level jobs may be vanishing, but launching something independently has never been easier. “[For] anybody who has Perplexity Computer . . . the constraint is your own curiosity and agency.”

Younis distinguished between knowledge work and physical labor, noting the average American farmer is 58 years old and labor shortages in mining, long-haul trucking, and agriculture persist—not due to low wages, but because people avoid these jobs. In these areas, physical AI fills a pre-existing void that is expected to grow.