Why AI Companies Are Running Into Chip Shortages

Five leaders from semiconductor manufacturing, self-driving cars, AI search, and quantum computing met at the Milken Global Conference in Beverly Hills to discuss a growing problem: there simply aren't enough chips to meet the explosive demand for AI systems.

The most direct warning came from Christophe Fouquet, CEO of ASML, a company that controls the machinery used to manufacture the world's most advanced chips. Fouquet predicted that chip shortages for AI will persist for two to five years — a significant timeframe that reflects how difficult it is to scale up production.

The shortage goes deeper than just factory capacity. ASML operates the only machines capable of making the most cutting-edge chip designs, and those designs require years to plan and build. Meanwhile, the companies training large AI models — Google, Microsoft, OpenAI, and others — are racing to buy every chip available. Each new, more powerful AI model needs more computational power to train, which means demand keeps rising while supply struggles to keep up.

Physical AI and Defense Contracts

Qasar Younis, who runs Applied Intuition, a company that builds AI systems for self-driving cars, delivery drones, and mining equipment, pointed out that the chip shortage affects more than just data centers. These physical AI applications need specialized chips, sensors, and components that face their own supply constraints.

Younis also noted that companies like his are increasingly turning to defense contracts — supplying technology to militaries and defense contractors — to maintain growth when consumer AI products face capacity limits. It's a pattern spreading across the industry: when chips are scarce, companies look for customers willing to pay more and commit for longer.

The Search-to-Agents Shift

Dimitry Shevelenko, chief business officer at Perplexity, explained how his company is evolving from a search engine into an "agents" company — software that can think through complex problems over longer conversations rather than just looking up facts. This shift demands different computing infrastructure than traditional search, and it means Perplexity is competing directly with much larger companies like Google and Microsoft for the same limited pool of chips.

Running these agent systems requires keeping track of longer conversations, handling different types of data (text, images, video), and working through multi-step reasoning — all of which demands significant computing power.

A Harder Question About Architecture

Eve Bodnia, a quantum physicist who started a company called Logical Intelligence, raised a more fundamental concern: maybe the real problem isn't that we need more chips, but that we're building AI systems the wrong way.

Today's AI systems rely on a specific approach called transformer architectures, which power everything from ChatGPT to image generators. Bodnia suggested that this approach might be inefficient, and that newer methods — including ideas borrowed from quantum computing — could do more with less computing power. If she's right, the hundreds of billions spent on AI chips could eventually look like a costly mistake.

This is worth flagging, though it's worth noting that architectural shifts in computing have historically taken decades to unfold. The mainframe-to-personal-computer transition didn't happen overnight, and neither will any move away from transformers if one happens at all. For now, companies have no choice but to keep building within the current system while keeping an eye on alternatives.

Unconventional Solutions

The panel also discussed more experimental ideas, like putting data centers in orbit around Earth. Satellites could tap into constant solar power and avoid some of the cooling and power challenges that ground-based data centers face. This sounds far-fetched, but we have seen this before: in the late 1990s, companies spent billions laying undersea fiber-optic cables to connect the internet globally, and what seemed like massive overinvestment turned out to be essential infrastructure.

What This Means Going Forward

The chip shortage appears to be creating a divide in the industry. Companies with strong relationships to chip makers — or enough money to secure supply — will be able to scale aggressively. Smaller companies may be forced to find creative ways to use less computing power or step back from their ambitions.

For organizations planning to deploy AI systems, it's important to understand that the timeline is longer than it was even a year ago. If you're hoping to run a custom AI application at scale, you should expect to wait months or years for the chips you need, and costs will be higher. A practical approach for many companies will be mixing cloud-based AI services with smaller, local systems that run on readily available hardware.

The panel's discussion made clear that the AI industry has outgrown its original infrastructure faster than anyone anticipated. While these supply constraints create real near-term problems, they're also pushing companies to think harder about efficiency and explore new ways to deploy AI systems — changes that could actually make the technology more practical and accessible in the long run.