AI Architects Reveal the AI Economy's Cracks

A recent gathering at the Milken Global Conference in Beverly Hills brought together five pivotal figures spanning the entire artificial intelligence supply chain. These industry leaders engaged in a comprehensive discussion, exploring diverse topics from the critical scarcity of advanced chips to the ambitious prospect of orbital data centers, and even questioning the fundamental architectural principles underpinning current AI technology.

The distinguished panel, hosted by TechCrunch, featured Christophe Fouquet, CEO of ASML, the Dutch firm holding a global monopoly on the extreme ultraviolet lithography machines essential for modern chip manufacturing; Francis deSouza, COO of Google Cloud, who is spearheading one of the largest infrastructure investments in corporate history; Qasar Younis, co-founder and CEO of Applied Intuition, a $15 billion physical AI company that originated in simulation before expanding into defense applications; Dimitry Shevelenko, Chief Business Officer of Perplexity, an AI-native company evolving from search to agent-based services; and Eve Bodnia, a quantum physicist who departed academia to establish Logical Intelligence, a startup challenging the foundational AI architecture widely adopted by the industry. Notably, Meta’s former chief AI scientist, Yann LeCun, joined Logical Intelligence as founding chair of its technical research board earlier this year.

Here are the key insights shared by the panelists:

The AI sector is confronting tangible physical limitations, with constraints emerging deeper within the technology stack than many might perceive. Christophe Fouquet initiated this point, acknowledging a "huge acceleration of chips manufacturing," yet expressing his "strong belief" that despite these efforts, "for the next two, three, maybe five years, the market will be supply limited." This implies that hyperscale cloud providers—such as Google, Microsoft, Amazon, and Meta—will not fully receive the volume of chips they are currently procuring.

Francis deSouza underscored the magnitude and rapid growth of this challenge, informing the audience that Google Cloud’s revenue surpassed $20 billion last quarter, achieving 63% growth. Furthermore, its backlog—representing committed but not yet delivered revenue—nearly doubled in a single quarter, surging from $250 billion to $460 billion. DeSouza stated with remarkable composure, "The demand is real."

For Qasar Younis, the primary constraint originates elsewhere. Applied Intuition develops autonomy systems for a range of vehicles including cars, trucks, drones, mining equipment, and defense vehicles. Younis identified his bottleneck not as silicon, but as the data that can only be acquired by deploying machines into the real world and observing their interactions. He emphasized, "You have to find it from the real world," asserting that synthetic simulation cannot fully bridge this gap. "There will be a long time before you can fully train models that run on the physical world synthetically."

Beyond chip scarcity, energy emerges as another significant and looming bottleneck. Francis deSouza confirmed that Google is seriously exploring the concept of data centers in space as a potential solution to energy constraints. He noted, "You get access to more abundant energy." However, even in orbit, the challenges are substantial. DeSouza pointed out that space is a vacuum, which eliminates convection, leaving radiation as the sole method for dissipating heat into the surrounding environment—a process far slower and more complex to engineer than the air and liquid cooling systems currently used in terrestrial data centers. Despite these hurdles, Google continues to consider it a legitimate path forward.

DeSouza’s more profound argument, though perhaps unsurprising, centered on achieving efficiency through vertical integration. He suggested that Google’s strategy of co-engineering its entire AI stack—from custom TPU chips to advanced models and agents—yields significant dividends in "watts per flop" that companies relying on off-the-shelf components simply cannot replicate. He explained, "Running Gemini on TPUs is much more energy efficient than any other configuration," because chip designers have prior knowledge of the model’s requirements before its release. In an environment where energy availability is increasingly restricting the potential of AI technology, this form of vertical integration offers a substantial competitive advantage.

Christophe Fouquet later reinforced this point, stating, "Nothing can be priceless." He described the industry as being in a peculiar phase, characterized by extraordinary capital investment driven by strategic imperative. Yet, increased computational power necessitates greater energy consumption, and that energy comes with an escalating cost.

While much of the industry focuses on scaling, architectural design, and inference efficiency within the large language model (LLM) paradigm, Eve Bodnia is pioneering a fundamentally different approach. Her company, Logical Intelligence, is developing what are known as energy-based models (EBMs)—a class of AI that, instead of predicting the next token in a sequence, aims to comprehend the underlying rules governing data. Bodnia argues this method more closely resembles how the human brain functions. She articulated, "Language is a user interface between my brain and yours," adding, "The reasoning itself is not attached to any language."

Bodnia’s largest model operates with 200 million parameters—a stark contrast to the hundreds of billions found in leading LLMs—and she asserts it runs thousands of times faster. Crucially, it is engineered to dynamically update its knowledge as data evolves, eliminating the need for complete retraining from scratch. For applications such as chip design, robotics, and other domains where systems must grasp physical rules rather than linguistic patterns, Bodnia contends that EBMs are a more natural fit. She offered an analogy: "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." This is a compelling argument that is poised to gain more traction as the AI field increasingly questions whether scale alone is sufficient.

Dimitry Shevelenko dedicated a significant portion of the discussion to detailing Perplexity's evolution from a traditional search product into what it now terms a "digital worker." Perplexity Computer, the company’s latest offering, is designed not as a tool for a knowledge worker to use, but rather as a virtual staff that a knowledge worker directs. Reflecting on the potential, he remarked, "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?"

While this pitch is compelling, it naturally raises questions about control, which Shevelenko addressed by emphasizing "granularity." He explained that enterprise administrators can precisely specify not only which connectors and tools an agent can access, but also whether those permissions are read-only or read-write—a critical distinction when agents operate within corporate systems. When Comet, Perplexity’s computer-use agent, prepares to take actions on a user’s behalf, it first presents a plan and requests approval. Shevelenko acknowledged that some users find this friction inconvenient, but he considers it essential, particularly after joining the board of Lazard. There, he said he found himself unexpectedly sympathetic to the conservative instincts of a CISO tasked with protecting a 180-year-old brand built entirely on client trust. He concluded, "Granularity is the bedrock of good security hygiene."

Qasar Younis offered what might have been the panel’s most geopolitically charged insight, positing that physical AI and national sovereignty are inextricably linked in ways that purely digital AI never was. The internet, initially a spread of American technology, encountered significant pushback primarily at the application layer—with services like Uber and DoorDash—once its offline consequences became apparent. Physical AI presents a different scenario. Autonomous vehicles, defense drones, mining equipment, and agricultural machinery manifest in the real world in ways that governments cannot overlook, prompting serious questions about safety, data collection, and ultimately, who controls systems operating within a nation’s borders. Younis observed, "Almost consistently, every country is saying: we don’t want this intelligence in a physical form in our borders, controlled by another country." He starkly noted that fewer nations currently possess the capability to field a robotaxi than possess nuclear weapons.

Christophe Fouquet framed the issue somewhat differently. While acknowledging that China’s AI progress is real—with DeepSeek’s release earlier this year causing considerable concern within parts of the industry—he asserted that this progress is constrained below the model layer. Without access to EUV lithography, Chinese chipmakers cannot manufacture the most advanced semiconductors. Consequently, models built on older hardware operate at a compounding disadvantage, regardless of how sophisticated the software becomes. Fouquet summarized the current landscape: "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."

Towards the conclusion of the panel, an audience member posed the inherently uncomfortable question: will the widespread adoption of AI negatively impact the next generation’s capacity for critical thinking? The responses, perhaps predictably, were optimistic, though not without nuance. Francis deSouza highlighted the immense scale of problems that more powerful AI tools might finally enable humanity to address. He cited examples such as neurological diseases whose biological mechanisms remain poorly understood, the complex challenge of greenhouse gas removal, and the modernization of grid infrastructure that has been deferred for decades. DeSouza expressed confidence, stating, "This should unleash us to the next level of creativity."

Dimitry Shevelenko offered a more pragmatic perspective: while entry-level jobs may diminish, the ability to launch independent ventures has never been more accessible. He suggested that for "anybody who has Perplexity Computer . . . the constraint is your own curiosity and agency."

Qasar Younis drew a sharp distinction between knowledge work and physical labor. He pointed out that the average American farmer is 58 years old, and that labor shortages in critical sectors such as mining, long-haul trucking, and agriculture are chronic and worsening—not because wages are too low, but because people are increasingly unwilling to undertake these jobs. In these domains, Younis argued, physical AI is not displacing willing workers; rather, it is filling an existing void that appears only set to deepen in the future.