The Hardware Ceiling Is Real
The AI boom’s dirty secret is that it is hitting hard physical limits, and the bottleneck starts far below the model layer. Christophe Fouquet, CEO of ASML, the Dutch monopoly on the extreme ultraviolet lithography machines essential for advanced chips, delivered a sobering reality check. He stated that despite a huge acceleration in manufacturing, the market will be supply constrained for the next two to five years. This means hyperscalers like Google, Microsoft, and Meta will not get all the chips they are paying for, no matter the price. Google Cloud COO Francis deSouza confirmed the staggering demand, noting its backlog nearly doubled in a single quarter to $460 billion. The message is clear: throwing more money at the problem won’t magically create more silicon. It is a physics problem, not a financial one.
The Energy Trap and Orbital Band-Aids
If chips are the first bottleneck, energy is the second, and it is far more intractable. DeSouza revealed that Google is seriously exploring orbital data centers as a response to terrestrial energy constraints, despite the fact that space is a vacuum which eliminates convection cooling, leaving only inefficient radiation as a heat sink. This is a desperate, costly fix. More pointedly, Qasar Younis of Applied Intuition argued that the true bottleneck for physical AI, like autonomous vehicles and defense drones, is not silicon but real world data that cannot be synthesized. Meanwhile, Eve Bodnia of Logical Intelligence is making a radical bet that the entire large language model architecture is wrong. Her energy based models (EBMs) at 200 million parameters claim to run thousands of times faster than trillion parameter LLMs by focusing on understanding rules rather than predicting tokens. She argues that for robotics and chip design, this is the only path that makes sense, suggesting the industry’s obsession with scale is a costly distraction.
Sovereignty, Control, and the Disappearing Entry Level
The geopolitical stakes are rising. Younis made a provocative claim that more countries possess nuclear weapons than can field a safe robotaxi, and every nation is demanding physical AI not be controlled by foreign powers. Fouquet countered that China’s progress, while real with DeepSeek, is fatally constrained by its inability to manufacture advanced chips due to the US export controls on EUV lithography. On the labor front, the panel offered a chillingly optimistic view. DeSouza hopes AI will solve neurological disease, but Younis bluntly noted that physical AI is not displacing workers in mining and agriculture; it is filling a void because no one wants those jobs. Shevelenko of Perplexity, which now sells a ‘digital worker’ that acts as a staff of 100, admitted that the entry level job is disappearing, but claimed the constraint for the future is now only one’s own curiosity and agency. That is a convenient belief for those selling the tools that are automating away the first rung of the career ladder.
Source: Techcrunch
