The views expressed here are my own and do not reflect those of any client, employer, or affiliated organization.
There is a version of the AI story that makes sense on its face. Intelligence requires computation. Computation requires hardware. Hardware requires infrastructure. Therefore, if you believe in the AI revolution, you build data centers. Lots of them. As fast as possible.
That logic has produced one of the most capital-intensive buildouts in the history of technology, financed by debt, premised on demand that doesn’t yet exist, dependent on supply chains that are quietly breaking. Ed Zitron, a tech analyst and writer who has been doing the unglamorous work of actually running the numbers, put it plainly earlier this year: “The data centers do not make sense. No one can actually square the math.”
He’s right. And the math is worth spelling out.
Building a modern AI data center costs roughly $44 million per megawatt — about $30 million in servers and GPUs, $14 million or more in construction. The debt financing most of these facilities is largely junk-rated. The customers expected to fill them are AI startups that are themselves unprofitable, paying for compute access with venture capital they are burning faster than they can raise it. OpenAI, the anchor tenant for much of the Stargate buildout, has reportedly committed to $300 billion in compute spend with Oracle — a company that, by Zitron’s analysis, doesn’t have the capacity to deliver it on the promised timeline, and a customer that doesn’t have the cash to pay for it. The gigawatt data center everyone is talking about has never been built by anyone. The power infrastructure alone takes two to four years, and that assumes everything goes smoothly.
None of this is going smoothly.
The hardware inside these facilities depreciates over six years at best and is technologically obsolete within one. A data center takes a year or two to build, meaning the chips inside are often already a generation behind when the doors open. Private equity is sinking tens of billions of dollars a quarter into infrastructure designed around GPU technology that will be outdated before it generates a dollar of revenue.
This is before you consider what the buildout actually runs on.
Semiconductor fabrication requires rare earth inputs — dysprosium, terbium, lutetium, yttrium — that flow almost entirely through Chinese processing facilities. It requires helium, a non-renewable gas used in dozens of fab process steps that cannot be substituted and whose primary export facility was physically struck and damaged during the Iran war earlier this year. It requires advanced chips manufactured almost exclusively in Taiwan, on an island whose strategic situation has materially worsened in the past several months. These are not hypothetical risks. They are supply chain vulnerabilities the industry is treating as background conditions rather than structural threats.
Meanwhile, the efficiency argument the industry ignores is being demonstrated in real time by a Chinese AI lab operating on export-controlled chips that US policy was specifically designed to deny them. DeepSeek’s architecture — activating only 37 billion of 671 billion parameters per token, compressing memory costs during inference, achieving frontier-competitive performance on hardware the US considered a meaningful handicap — is a proof of concept for the thesis that centralized compute accumulation is not the only path. It may not even be the right path. If you can get competitive intelligence from a fraction of the compute, the case for tens of billions in junk-rated data center debt becomes harder to make, not easier.
The rational infrastructure approach is not glamorous: algorithmic efficiency, expanded local deployment, targeted regulation of the most compute-intensive consumer applications that generate heat without much light. It doesn’t produce ribbon-cutting announcements or $500 billion Stargate headlines. It does produce a technology stack that isn’t one rare earth export license away from a supply shock.
The bubble isn’t a bug. It’s the inevitable outcome of building supply-side infrastructure for demand that was never validated, financed by parties with every incentive to keep the story going and none to ask whether the numbers work. Zitron’s prediction that it begins to unravel in 2026 is starting to look less contrarian by the month. All it takes, as he noted, is one or two facilities collapsing — a financing deal that falls through, a construction halt, an opening with no customers left. The cascade follows from there.
The AI era may well arrive. The data center buildout premised on its immediate and unlimited appetite almost certainly won’t survive the wait.
This is the first in a three-part series on AI infrastructure and its supply chain vulnerabilities. The next post examines how China turned a hardware disadvantage into strategic leverage.