Another session, another circular AI deal.

If this all implodes one day — you can almost see it: a year or two from now, one of the hyper-scalers suggests on an earnings call that demand for AI’s slowing or not panning out as expected, opening the proverbial floodgates — we’re going to remember the flurry of tie-ups announced over the past three months as an especially glaring example of the kind of self-referential absurdity which often typifies late-stage bubbles.

On Monday, Amazon said OpenAI will pay $38 billion over seven years to access Nvidia GPUs via AWS. “The rapid advancement of AI technology has created unprecedented demand for computing power,” Amazon declared, in a predictably hyperbolic press release. “As frontier model providers seek to push their models to new heights of intelligence, they are increasingly turning to AWS due to the performance, scale and security they can achieve.”

The announcement came hot on the heels of Amazon’s post-earnings rally. The rolling two-session gain for the stock, at ~15%, was the most pronounced in three years.

If you were inclined to suggest Amazon’s somehow an also-ran in the AI race, you should think twice, AWS chief Matt Garman suggested Monday. AWS, he said, will be “a backbone” for OpenAI’s ambitions, as Sam Altman “continues to push the boundaries of what’s possible.” (Never mind whether those boundaries should actually be pushed.)

As one habitually snarky market commentator (and no, I’m not talking about myself) noted a few weeks back, it’s easy to spend money you don’t have. OpenAI’s commitments are closing in on $1.5 trillion, and while a lot of that will come from investors both private and, eventually, public, one does have to wonder whether we’re pushing the limits in terms of positing the implausible.

Over the weekend, in “The AI Spending Question,” I noted that we’re nowhere near peak AI capex, and I was emphatic that investors understand as much and remain more than willing to fund those expenditures. I’m confident that’s the right assessment, but past a certain point, we’re conjuring castles in the air.

“Where are the AI revenues going to come from?” SocGen’s Andrew Lapthorne wondered aloud on Monday. “MSCI World constituent companies are already spending most of their cashflow on capex, dividends and buybacks and with a mere $44 billion to spare, corporate balance sheets are going to have to take the strain,” he added.

The table above’s from Goldman. Focus on the columns under “Revision to 2026 estimates.” For the hyper-scalers, 2026 net income estimates were revised up by 3% over the past two months, capex by 13%.

To reiterate from the above-linked article, everyone knows the mega-caps aren’t going to fund the entirety of the AI buildout with “extra” cash. That’s not the issue. The concern — my concern, anyway — is that this whole thing feels cartoonishly interdependent.

If you promise to spend $50 billion to access GPUs through my service, and I pledge to spend $50 billion buying more AI chips, and the AI chip company pledges to invest $50 billion in your equity, what is it we’re really doing? Something, I guess. To believers, it’s a virtuous circle. To skeptics, it feels uncomfortably like rehypothecation.

On Monday, Bloomberg said Alphabet’s looking to sell more than $22 billion in bonds following Meta’s mega-offering last week. Meanwhile, Microsoft committed to spending almost $10 billion over five years to access Nvidia-powered compute from an Australia-based provider and said it’ll spend nearly $8 billion investing in data centers and cloud capacity in the UAE.