The AI Infrastructure Arms Race: Building Faster Than Demand Can Catch Up

Why the AI boom is less about hype and more about timing mismatches, infrastructure slowdowns, and uncertain demand.

Not All Bubbles Pop—Some Just Miss the Mark

When we talk about tech bubbles, it’s easy to imagine catastrophic crashes. But in economic terms, a bubble is just a bet that overshoots. The wager might be grounded in something real—like AI—but if supply grows faster than demand, the bet unravels.

And in the case of AI, the bets are massive. But they’re also complicated, with timelines and dependencies that make the outcome far less binary than previous tech manias.

This isn’t about whether AI is real. It’s about whether our investments are sized—and timed—correctly.

The Core Problem: A Timeline Mismatch

AI software evolves at breakneck speed. But the hardware, energy, and real estate required to run it—like data centers—take years to build.

• AI models can be deployed or replaced in months.
• Data centers take 3–5 years to plan, construct, and connect to the grid.

That’s a critical mismatch. What looks like a smart infrastructure bet today may be out of sync with actual usage patterns by the time it comes online.

A Glut of Supply, A Shrug from Demand

Big Tech is betting huge on AI. Recent commitments include:

• $18B in credit for an Oracle-linked data center in New Mexico
• $300B in cloud contracts between Oracle and OpenAI
• A $500B AI infrastructure project (Stargate) involving SoftBank
• Meta’s $600B commitment to AI infrastructure over 3 years

But is demand keeping pace?

• A McKinsey survey found that while most companies are experimenting with AI, few are deploying it at scale.
• In many cases, AI is delivering localized efficiencies, not sweeping business transformation.
• Translation: Enterprise AI spending is still in a cautious, exploratory phase.

If you’re building data centers expecting a flood of AI usage, you may be early by years.

When Infrastructure Lags Innovation

Even if demand explodes, physical bottlenecks could delay or disrupt AI’s growth.

• Satya Nadella recently said he’s more concerned about lack of data center space than chip shortages.
• Some data centers sit idle, unable to deliver enough power to run modern AI chips like Nvidia’s H100s.
• Power grids and permitting systems haven’t caught up to AI’s energy needs.

In other words, you can buy all the chips you want, but if your site can’t handle the heat (literally), you’re stuck.

Betting Big Means Risking Big Bottlenecks

This doesn’t mean the AI boom is hollow. But the scale of investment—and the physical reality of infrastructure—makes it uniquely vulnerable to delays, misallocations, and unmet expectations.

Even if AI demand proves limitless, it won’t matter if the systems built to support it aren’t ready.

This is where bubbles form—not because the technology is fake, but because execution can’t keep up with ambition.

The Better Lens: Overcapacity, Not Collapse

Instead of thinking about the “AI bubble” as a crash waiting to happen, consider it a classic case of overcapacity:

• The bets being made today could oversupply the market by 2026 or 2027.
• The grid, construction, and cooling demands may lag behind AI software needs.
• Meanwhile, buyers (especially enterprises) may take longer to commit to large-scale deployments.

The result? Too much infrastructure, too soon. Not a collapse, but a correction.