AI's $500B Circular Debt: A Castle Built on Sand

Nvidia, OpenAI and the $500 billion accounting trick

There are 11,000 data centers operating globally, yet barely anyone talks about them. The figure appears in critical infrastructure maps, but not in headlines. Until the bill arrives. Not just the electricity bill: the financial one. Nvidia’s latest move to prop up OpenAI has named what many have been warning about for months: circular financing.

The energy hole no one wants to quantify

The consumption numbers are hard to swallow. A data center in Cantabria would have required as much electricity as the rest of the autonomous community, and the idea was shelved. Globally, the 11,000 centers would consume around a third of world business electricity. And real AI demand for chat, by the sharpest calculations, ranges from 350 megawatts to 1 gigawatt for all global inference. The Big Four (Amazon, Google, Microsoft, and Meta) plan to build capacity for 50 times more. It’s not a mystery; it’s a bet.

Circular financing: Nvidia backs OpenAI to pay Nvidia

Bloomberg uncovered the mechanism: Nvidia is in talks to provide up to $250 billion in financing guarantees to OpenAI. Why? So OpenAI can rent computing capacity at the massive Ohio data center, a 10-gigawatt, $500 billion project led by SoftBank. In other words: the chipmaker backs its main customer so that customer can buy chips from it. The debt is generated, circulates, and returns to the starting point. It works as long as everyone keeps playing the game.

The bubble with a 2008 taste

Some compare it to the 2002-2007 financial bubble: low interest rates, cheap money, and irrational conviction that prices will keep rising. Only this time, the collateral isn’t subprime mortgages, but debt guaranteed by companies paying each other. Skeptics talk of a bubble too big to fail and an establishment with unprecedented financial weapons. Regulators are starting to move: Britain’s Ofgem already charges between £237,500 and £712,500 per megawatt to curb speculative projects. Bad news if the ball is $500 billion.

The counterweight deflating the bubble

While the US piles up debt, China plays in another league. DeepSeek launches a model 100 times cheaper than Anthropic’s flagship, Kimi K3 hits US AIs, and Indian companies switch to Chinese open-source models: 95% of value for 10% of the cost. To top it off, spike neural networks (SNNs) promise to reduce LLM consumption by two orders of magnitude, and local hardware advances to bring powerful AI to the pocket. Even the Pentagon is not left out, with its DARPA program to coordinate 100,000 AI agents. If efficiency improves at this rate, why on earth build capacity for 50 times demand?

With circular debt, the geopolitical race, and energy efficiency all pressing at once, the uncomfortable question is who gets stuck with the bill. Optimists calculate that AI will pay back energy consumed many times over. Skeptics argue that when the music stops, the last to sign the guarantee will be left holding the megawatt. What if both are true at once?