If you want to stay ahead in the AI race, stop watching model benchmarks. Check chip supply contracts. Look at power infrastructure deals.
In May 2026, five architects of the AI economy gathered at the Milken Institute Global Conference. ASML's CEO. Google Cloud's COO. Perplexity's CBO. Applied Intuition's CEO. And the founder of a startup building AI in the exact opposite direction from LLMs. Their diagnosis converged on one point: the thing holding AI back isn't technology — it's physical constraints.
Wait, is this the real problem?
The Milken Institute Global Conference draws top names in finance and economics. This year's AI panel had a sharp topic: "Where are the wheels coming off the AI economy?"
The first shock came from ASML's CEO. ASML — the world's only supplier of EUV lithography machines — confirmed: "for the next two, three, maybe five years, the market will be supply limited". A single High-NA EUV machine costs roughly $370 million, fewer than 20 are made per year, and orders are already booked through late 2027.
Here's why that matters. No EUV means no cutting-edge AI chips. No AI chips means data centers don't matter how many you build. People debate whether the bottleneck is power or chips — but the real limit is the equipment that makes chips. The ceiling on AI chip supply is set by how many EUV tools ASML can deliver in a year.
Then there's the energy problem. Google Cloud COO Francis deSouza said revenue crossed $20 billion last quarter and the backlog nearly doubled from $250B to $460B in a single quarter. But there's not enough power on Earth to meet that demand. That's why Google is "seriously exploring" orbital data centers — in space, solar power runs 24/7 with no atmosphere blocking it and no local opposition to deal with.
Physical AI — autonomous vehicles, drones, agricultural robots — faces yet another bottleneck. Applied Intuition's CEO said real-world data collection is the actual constraint. Simulation data simply can't close the gap with real-world physics.
The most striking remarks came from Logical Intelligence founder Eve Bodnia. She's building energy-based models (EBMs) — a fundamentally different architecture from LLMs. Her largest model runs 200 million parameters, a fraction of leading LLMs, yet operates "thousands of times faster". With Yann LeCun as chair of the technical board, the company's argument is direct: LLMs are structurally incapable of genuine reasoning.
What have we been getting wrong?
The standard AI narrative has been: better models, more parameters, longer context. Milken 2026 told a different story.
| Common Perception | Reality (Milken 2026) | |
|---|---|---|
| Core advantage | Model performance, parameter count | EUV chip allocation, power contracts |
| Supply constraint | GPU wait times of a few weeks | EUV backlog fully booked through 2027 |
| Energy | Electricity cost issue | Physical infra limits, orbital data centers now under consideration |
| AI architecture | LLM is the standard, parameter race | EBMs solving identical tasks at 1/3,750 the cost |
| Geopolitics | Regulatory and ethics issues | Physical AI operates within borders → sovereignty becomes a real constraint |
Perplexity's Shevelenko described AI agents as "digital workers." For those agents to function inside enterprises, the key isn't raw capability — it's granular permission structures. "Granularity is the bedrock of good security hygiene," he said — something any team deploying AI agents right now needs to take seriously.
Younis's take on physical AI geopolitics was striking. Autonomous vehicles, drones, and mining equipment operate within visible, physical borders. Fewer nations can field a robotaxi than possess nuclear weapons — because you need infrastructure, legal frameworks, and data sovereignty simultaneously. Even China's strong software AI development is constrained below the model layer by EUV export controls.
The essentials: 5 bottlenecks and what to do
- Chip supply — already booked through 2027
No EUV lithography, no cutting-edge AI chips. High-NA EUV costs $370M per unit, fewer than 20 are produced annually. Unless you control a chip supply chain, your strategy should focus on maximizing efficiency with available compute. - Energy — no room left on Earth
The US grid interconnection queue exceeds total grid capacity. Analysis suggests 30–50% of planned 2026 data centers will slip to 2028. Google's orbital data center exploration is not a joke. Switching to energy-efficient models can reshape your cost structure right now. -
Real-world data — simulation can't close the gap
For physical AI (autonomous vehicles, robots), real-world data collection is the real bottleneck. Unlike software AI, failure cases are hard to repeatedly collect. If you're exploring this space, investing in a data collection pipeline is your highest-leverage move. - The LLM alternative — same task, 1/3,750 the cost
Not every problem needs a GPT-tier model. Logical Intelligence's EBM solved a complex reasoning task for $4; the same task cost ~$15,000 with a frontier LLM. The ability to match task to architecture is what actually determines cost structure. - Geopolitics — physical AI is a border issue
Cloud AI ignores borders; drones, autonomous vehicles, and robots don't. Local regulations, data sovereignty, and export controls apply differently by region. Where your data is processed and where your model runs are now strategic decisions.
The practical takeaway
With supply and energy constraints structurally raising AI costs, the team strategy that matters isn't "which model?" — it's "which task needs how much spend?" Architecture design beats model selection. That's the real competitive edge right now.
