Nearly half of the data centers planned for construction in 2026 have been cancelled or delayed, according to new industry reports. If you've been following the breathless AI infrastructure buildout announcements over the past two years, this should feel familiar: it's what happens when PowerPoint projections meet physics.

I've been in the startup trenches. I know what FOMO-driven investment looks like. And this wave of data center cancellations isn't a blip in the AI revolution - it's the market doing what markets eventually do: separating genuine demand from speculative fever dreams.

The numbers tell the story. Data center developers spent the last 18 months announcing massive facilities to support AI training and inference workloads. The pitch was simple: AI companies need compute, compute needs data centers, therefore build data centers and they will come. Except they're not coming. Or at least, not at the scale the projections suggested.

The reasons aren't mysterious if you've ever built infrastructure. Energy costs are crushing. Grid capacity is limited. And crucially, the actual compute demand from AI companies hasn't materialized at the rate the hype cycle predicted. Turns out training GPT-5 is expensive, but you only need to do it once. Inference is cheaper. And most "AI-powered" startups are just calling APIs to OpenAI or Anthropic, not spinning up their own massive compute clusters.

This is what a healthy market correction looks like. When I sold my fintech startup to Stripe, one of the things I learned was the difference between what investors want to hear and what actually scales. Data center developers sold the dream. Now they're facing the reality: you can't build speculative infrastructure at this scale and hope someone shows up to fill it.

The technology is real. The capabilities are impressive. But we've been here before with tech infrastructure buildouts - the fiber optic cable bubble in the late 90s laid dark fiber that sat unused for years. This feels similar. Too much infrastructure chasing demand that's growing, yes, but not at the exponential rate the presentations claimed.

Here's the thing that should worry AI companies more than cancelled data centers: energy costs aren't coming down. Grid capacity isn't magically expanding. If the economics don't work now with cheap money and high expectations, when exactly do they work?

Some of this is geographic. Building data centers in regions with expensive energy made sense when everyone assumed AI compute would command premium pricing forever. Now that companies are actually running the numbers on inference costs versus revenue, those facilities look less attractive.

The question isn't whether AI is real - it obviously is. The question is whether we actually need this much compute infrastructure. And increasingly, the answer appears to be: not yet, and maybe not ever at these levels.

This correction might actually be healthy for the industry. It forces AI companies to be honest about their compute needs. It pushes infrastructure developers to build based on actual demand rather than projected hockey sticks. And it reminds everyone that at some point, the physics and economics have to make sense.

The AI revolution is happening. But revolutions don't require building infrastructure for usage levels that may never materialize. Sometimes the smartest thing you can do is not build what you thought you needed.