Ed Zitron says planned AI data centers need $1.6 trillion to $3 trillion a year in demand, but the market has under $130 billion

Ed Zitron thinks the AI boom has a math problem so large it makes the dot-com bubble look quaint. On The Diary of a CEO, he put a number on it: planned AI data centers would need roughly $1.6 trillion to $3 trillion in annual demand, while the market today is not even at $130 billion.

That is the claim worth arguing about here. Not whether your cousin uses ChatGPT to rewrite a resignation email. Not whether a founder can make a pitch deck look less like it was assembled during turbulence. Zitron’s argument is that generative AI is being sold like a civilization-level upgrade while being financed like a casino with a power bill.

Right now, the demand we have for generative AI is predominantly subsidized.

Steven Bartlett does the obvious thing and reaches for the early internet analogy. Fair. Every tech skeptic eventually gets haunted by Paul Krugman’s fax-machine line. Bartlett points out that bubbles can burn most of the junk and still leave Amazon-shaped survivors behind. Zitron’s reply is that the analogy flatters AI too much. Fiber optic cable left over from the dot-com frenzy became useful as demand caught up. A GPU data center does not become a charming public good just because the PowerPoint was persuasive.

The fiber optic cable comparison is doing the heavy lifting

Zitron’s distinction is blunt: the dot-com overbuild left behind wires in the ground. The AI overbuild leaves behind buildings full of specialized chips that keep guzzling electricity. That matters because a data center is not a passive relic. It is a hungry machine. It needs power, cooling, maintenance, customers, and some way to turn demos into durable revenue.

there’s 190 gigawatts of data centers under in planning, don’t know about under construction. That works out at about 12 million a megawatt, that’s what like 1.6 trillion to 3 trillion dollars a year in annual demand you’d need for that.

This is where Zitron is at his most convincing. Even if you think he’s too allergic to AI hype, the cost stack is real. GPUs are expensive. Electricity is expensive. Inference is not a poetic concept, it is a meter running. If consumers love the product only when someone else eats the bill, that’s not product-market fit. That’s a free buffet with venture capital shrimp.

He also makes a useful language point: the term artificial intelligence lets companies bundle everything under one glowing umbrella. Protein folding, robotics, search ranking, autonomous systems, chatbots, code generators. All of it gets tossed into the same marketing cauldron. Then when someone criticizes large language models, a defender can point to medical research like a magician producing a dove.

The bubble is not the tool. It’s the promise.

The funnier and meaner version of Zitron’s case is that the biggest buyers may also be the biggest props. He argues that much of the apparent demand comes from a small loop of cloud giants and AI labs, with Microsoft, Amazon, Google, OpenAI, and Anthropic keeping the wheel spinning. That doesn’t make the technology fake. It makes the market weird.

they wouldn’t just have two unprofitable fail sons that they’re propping up with Christ, they’ve raised $217 billion just in 2026.

That line is vintage Zitron, half spreadsheet, half pub fight. It is also where skepticism should stay awake. His broad claim about subsidized demand is strong, but the more totalizing version, that the whole thing is basically a con, can flatten the genuine ways people use these tools. Bartlett’s fiancee using ChatGPT to write copy and make images for a small business is not imaginary value. It just may not be value that can support a trillion-dollar AI data centers land rush.

The coding discussion gives Zitron another lane: AI may not just be expensive, it may be making software worse. He connects AI-assisted coding to bug volume, outages, and developer complacency. Some of that is hard to separate from the general enshittification of big platforms. Google Search did not need ChatGPT to discover that more ads and worse answers could still print money.

Based on what Sam Altman has been saying for the last few years, clammy Sammy’s been promising the world saying this will replace software engineers.

That is the cleaner verdict: Zitron is less persuasive as a prophet of uselessness than as an auditor of overpromise. The tools can be useful and the bubble can still be grotesque. Your small business copy assistant can be real. So can the absurdity of building trillions in infrastructure on the assumption that everyone will soon pay full freight for autocomplete with vibes.

If Zitron is right, the listener’s stake is not whether to delete ChatGPT. It’s whether to believe every new power plant, every higher cloud bill, every worse search result, and every jittery software release is the necessary birth cost of the future. Maybe it is. But someone still has to pay the electric bill.