Chip shortages stretching five years out. A $460 billion backlog at Google Cloud. Data centers in orbit. Five executives who sit at different layers of the AI supply chain laid out where the strain is showing, and TechCrunch AI captured the full conversation from the Milken Global Conference in Beverly Hills.
The panel included Christophe Fouquet (CEO of ASML), Francis deSouza (COO of Google Cloud), Qasar Younis (CEO of Applied Intuition), Dimitry Shevelenko (CBO of Perplexity), and Eve Bodnia (founder of Logical Intelligence). Each one painted a piece of the same picture: the AI economy is hitting physical walls faster than most decks suggest.
The Silicon Wall Is Real
Fouquet runs the company that makes the EUV lithography machines every advanced chip depends on. His read: “For the next two, three, maybe five years, the market will be supply limited.” Translation for hyperscalers: Google, Microsoft, Amazon, and Meta will not get all the silicon they have paid for. Period.
DeSouza put numbers on the demand side. Google Cloud crossed $20 billion in quarterly revenue, growing 63%. The backlog (committed but undelivered revenue) nearly doubled in a single quarter, jumping from $250 billion to $460 billion. “The demand is real,” he said.
This is significant because it reframes the AI buildout as a multi-year supply problem, not a money problem. Capex announcements grab headlines. Lithography throughput decides who actually ships.
Energy Is the Next Bottleneck
If chips are constraint number one, power is constraint number two. DeSouza confirmed Google is genuinely exploring orbital data centers as a response to energy limits. The catch: space is a vacuum, so heat can only leave by radiation, which is slower and harder to engineer than the cooling systems used on Earth.
His fallback argument is integration. Co-engineering custom TPUs alongside Gemini gives more flops per watt than any off-the-shelf stack, he claims. Fouquet echoed the math from the manufacturing side: “Nothing can be priceless.” More compute means more energy, and energy has a price.
Physical AI Has a Different Bottleneck
Younis runs a $15 billion company building autonomy for cars, trucks, drones, and defense vehicles. His constraint isn’t silicon. It’s data you can only collect by putting machines into the real world. “You have to find it from the real world,” he said. Synthetic simulation does not close the gap, and won’t anytime soon.
That’s a useful corrective for anyone betting that physical AI will scale on the same curve as language models.
A Quiet Challenge to the LLM Paradigm
Bodnia is the contrarian on the panel. Her company, Logical Intelligence, builds energy-based models instead of next-token predictors. Her largest model has 200 million parameters (versus hundreds of billions in frontier LLMs) and she claims it runs thousands of times faster. It also updates as data changes rather than requiring full retraining.
Her pitch: “Language is a user interface between my brain and yours. The reasoning itself is not attached to any language.” Yann LeCun joined her technical research board as founding chair earlier this year, which adds weight to an argument the field is starting to take seriously: scale alone may not be enough.
Agents Become Staff
Shevelenko described Perplexity’s pivot from search engine to “digital worker.” Perplexity Computer is positioned as something a knowledge worker directs, not uses. “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?”
What Practitioners Should Take Away
Three practical reads from the conversation:
- Plan for compute scarcity. If you’re building anything that depends on frontier model access, assume capacity is rationed for the rest of the decade. Vertical integration with cloud providers becomes leverage.
- Watch the architecture debate. Energy-based models, world models, and other non-LLM approaches are gaining serious backing. Locking your stack to one paradigm is a risk.
- Treat agents as staff, not features. The interface shift from tools to workers changes hiring math, not just product roadmaps.
The boom is not slowing. It’s just running into the parts of reality that capital alone can’t fix. Full panel details at the original TechCrunch AI report.