Five people who touch every layer of the AI supply chain sat down at the Milken Global Conference and delivered a sobering assessment. From ASML's chip manufacturing monopoly to Google Cloud's $462 billion backlog, the AI economy is hitting hard physical limits. Chips are scarce. Energy is insufficient. And one panelist argued the entire architecture underlying modern AI may be fundamentally wrong.
The Chip Bottleneck Will Last Years
ASML CEO Christophe Fouquet delivered the starkest warning. Despite a massive acceleration in chip manufacturing, he believes the market will remain supply-limited for three to five years. The hyperscalers — Google, Microsoft, Amazon, Meta — are not going to get all the chips they are paying for. Full stop.
Google Cloud COO Francis deSouza reinforced the point. Google Cloud crossed $20 billion in quarterly revenue last quarter. Its backlog nearly doubled in a single quarter to $462 billion. The demand is real. The capacity to meet it is not.
Applied Intuition CEO Qasar Younis raised a different kind of bottleneck. For physical AI — autonomous vehicles, drones, mining equipment — the constraint is not silicon. It is real-world data. No amount of synthetic simulation fully replaces sending machines into the physical world and watching what happens. That gap will persist for years.
Energy Is the Next Wall
If chips are the first bottleneck, energy is the one looming behind it. DeSouza confirmed that Google is exploring data centers in space as a serious response to energy constraints. Space offers more abundant energy. But cooling is harder in a vacuum. Radiation is the only way to shed heat — a much slower process than the liquid and air systems terrestrial data centers use.
Google's argument is that co-engineering its full AI stack — from custom TPU chips through to models and agents — delivers efficiency gains that companies buying off-the-shelf components cannot replicate. Running Gemini on TPUs is more energy efficient than any other configuration because chip designers know what the model needs before it ships.
Fouquet added a broader caution. Nothing can be priceless. The industry is investing extraordinary capital driven by strategic necessity. But more compute means more energy. And more energy has a price that someone will eventually have to pay.
A Challenger to the Entire LLM Paradigm
While the rest of the panel debated scale and infrastructure, Eve Bodnia offered something more provocative. Her startup Logical Intelligence is built on energy-based models — a class of AI that does not predict the next token in a sequence. Instead it tries to understand the rules underlying data, in a way she argues is closer to how the human brain works.
Her largest model runs on 200 million parameters. Leading language models use hundreds of billions. She claims her system runs thousands of times faster. More importantly, it can update its knowledge as data changes without retraining from scratch.
For chip design, robotics, and domains where a system needs to grasp physical rules rather than linguistic patterns, Bodnia argues energy-based models are the more natural fit. The AI field is beginning to ask whether scale alone — the approach that has defined the last three years — is sufficient. Bodnia's answer is no.
Agents Need Guardrails
Perplexity CBO Dimitry Shevelenko explained how the company has evolved from search into what it calls a digital worker. The new Perplexity Computer is designed not as a tool a knowledge worker uses but as staff that a knowledge worker directs.
The pitch raises obvious control questions. Shevelenko's answer is granularity. Enterprise administrators can specify which connectors and tools an AI agent can access and whether those permissions are read-only or read-write. When the agent takes actions, it presents a plan and asks for approval first.
Some users find the friction annoying. Shevelenko considers it essential — particularly after joining the board of Lazard, where he found himself sympathetic to the conservative instincts of a CISO protecting a 180-year-old brand built on trust. Granularity is the bedrock of good security hygiene.
Sovereignty and Physical AI
Younis offered the most geopolitically charged observation. Physical AI and national sovereignty are entangled in ways that purely digital AI never was. Autonomous vehicles, defense drones, mining equipment, agricultural machines — these manifest in the real world in ways governments cannot ignore.
Almost consistently, every country is saying they do not want physical intelligence controlled by another country operating inside their borders. Fewer nations, Younis noted, can currently field a robotaxi than possess nuclear weapons.
Fouquet framed China's position differently. DeepSeek's progress at the software layer is real. But without access to EUV lithography, Chinese chipmakers cannot manufacture the most advanced semiconductors. Models built on older hardware operate at a compounding disadvantage no matter how good the software. The US has the data, compute, chips, and talent. China does a good job on the top of the stack but is lacking elements below.
What It Means
The Milken panel painted a picture of an AI economy that is growing faster than the physical world can support it. Chips will be scarce for years. Energy constraints are pushing companies toward space. The fundamental architecture may need rethinking. And the geopolitical landscape is fragmenting the AI industry along national lines.
The optimism was there — deSouza pointed to diseases, climate, and infrastructure problems that AI could finally solve. Shevelenko argued that anyone with an AI agent now has a hundred staff on their team. But beneath the optimism, the five architects agreed on one thing: the wheels are starting to wobble. Whether the AI economy can build fast enough to stay ahead of its own ambitions is the defining question of the next three years.
