The Physical Wall: Why AI’s Future Depends on Hardware, Energy, and Sovereignty
The narrative surrounding artificial intelligence has long been dominated by software breakthroughs, token counts, and benchmark races. However, a recent high-level discussion at the Milken Global Conference in Beverly Hills revealed a different reality: the AI boom is colliding with hard physical limits. From semiconductor shortages and energy deficits to the geopolitics of “physical AI,” the industry is shifting its focus from the virtual layer to the infrastructure that sustains it.
- Supply Constraints: Chip manufacturing is expected to remain supply-limited for the next two to five years.
- Energy Innovation: Energy constraints are pushing hyperscalers to explore radical solutions, including orbital data centers.
- Architectural Shifts: Energy-based models (EBMs) are emerging as a leaner, faster alternative to massive Large Language Models (LLMs).
- AI Sovereignty: Governments are increasingly viewing physical AI (robotics, drones) as a matter of national security and sovereignty.
The Hardware Bottleneck: Silicon and Real-World Data
While the demand for compute is skyrocketing, the ability to produce the necessary hardware is struggling to keep pace. Christophe Fouquet, CEO of ASML—the sole provider of the extreme ultraviolet (EUV) lithography machines essential for advanced chips—warned that the market will likely remain supply-limited for the next two to five years. This means that even the largest hyperscalers, including Google, Microsoft, Amazon, and Meta, may not receive all the chips they have already paid for.
The scale of this demand is evident in the financial data. Francis deSouza, COO of Google Cloud, noted that the division’s revenue crossed $20 billion last quarter with 63% growth. More telling is the backlog of committed but undelivered revenue, which nearly doubled in a single quarter, jumping from $250 billion to $460 billion.
However, for those building AI that interacts with the physical world, the bottleneck isn’t just silicon—it’s data. Qasar Younis, CEO of Applied Intuition, argues that synthetic simulations cannot fully replace real-world observation. For autonomy systems in defense, mining, and transport, the only way to truly train models is to send machines into the environment and gather empirical data.
The Energy Crisis and the Move to Orbit
Energy is the looming shadow behind the chip shortage. The power requirements for modern AI are so immense that Google is exploring the possibility of data centers in space. According to deSouza, orbital centers would provide access to more abundant energy, though they introduce significant engineering hurdles. Because space is a vacuum, convection is impossible, meaning heat must be shed via radiation—a much slower and more complex process than the liquid and air cooling used on Earth.
To combat these inefficiencies, Google is pursuing a strategy of vertical integration. By co-engineering the entire stack—from custom TPU (Tensor Processing Unit) chips to the models themselves—the company can optimize “flops per watt.” DeSouza asserts that running Gemini on TPUs is significantly more energy-efficient than using off-the-shelf components because the chip designers can anticipate the model’s needs before it is even shipped.
Beyond the LLM: The Rise of Energy-Based Models
While the industry continues to scale LLMs into the trillions of parameters, some are questioning if the foundational architecture is wrong. Eve Bodnia, founder of Logical Intelligence, is developing Energy-Based Models (EBMs). Unlike LLMs, which predict the next token in a sequence, EBMs attempt to understand the underlying rules of data, mimicking human reasoning more closely.
The efficiency gains are stark. Bodnia’s largest model runs on 200 million parameters—a fraction of the hundreds of billions found in leading LLMs—and she claims it operates thousands of times faster. Because EBMs are designed to update knowledge as data changes without requiring a full retrain from scratch, they are particularly suited for robotics and chip design, where grasping physical rules is more important than linguistic patterns.
Digital Workers and the Security of Agency
The evolution of AI is moving from “tools” to “agents.” Dimitry Shevelenko, chief business officer of Perplexity, describes the shift toward “digital workers.” With the introduction of Perplexity Computer, the goal is to provide users with a virtual staff that can be directed rather than a tool that must be operated.
This shift toward agency introduces critical security risks. Shevelenko emphasizes that “granularity is the bedrock of good security hygiene.” To maintain trust, enterprise administrators must have precise control over agent permissions, distinguishing between read-only and read-write access. For instance, Perplexity’s computer-use agent, Comet, is designed to present a plan and seek user approval before taking action, ensuring a human remains in the loop for critical corporate tasks.
AI Sovereignty and the Geopolitical Divide
The intersection of AI and physical hardware has turned technology into a matter of national sovereignty. Qasar Younis points out that while digital AI faced pushback primarily at the application layer, physical AI—such as autonomous defense drones and agricultural machinery—operates within a nation’s borders. This has led many countries to insist that the intelligence controlling physical systems within their territory must not be controlled by a foreign power.
This geopolitical tension is further exacerbated by the semiconductor supply chain. Christophe Fouquet notes that while China has made significant progress at the top of the AI stack (software and models), it remains constrained by the lack of EUV lithography. Without access to the most advanced chip-making equipment, Chinese chipmakers cannot produce the semiconductors required to compete on a level playing field, giving the U.S. A compounding advantage in computing access and talent.
The Human Impact: Creativity and Labor
Despite concerns that AI might erode critical thinking, industry leaders remain optimistic. Francis deSouza suggests that more powerful AI tools will unleash a new level of human creativity, allowing us to solve previously intractable problems in neurological disease, greenhouse gas removal, and aging grid infrastructure.
While entry-level knowledge work may be disrupted, the impact on physical labor is different. Qasar Younis highlights chronic labor shortages in long-haul trucking, mining, and agriculture—sectors where the average American farmer is 58 years old. In these domains, physical AI is not displacing willing workers but is instead filling a void that has existed for years.
Frequently Asked Questions
Why is EUV lithography so important for AI?
EUV lithography is the only method capable of printing the incredibly compact features required for the most advanced semiconductors. Without it, chipmakers cannot produce the high-performance GPUs and TPUs that power modern AI models.
What are Energy-Based Models (EBMs)?
EBMs are a class of AI that focuses on understanding the rules and constraints of data rather than predicting the next token in a sequence. They are generally much smaller and faster than LLMs and are better suited for physical world applications.
What is “AI Sovereignty”?
AI sovereignty refers to a nation’s desire to control the AI systems—especially those in physical forms like robotics—that operate within its borders to ensure national security and data privacy.