Something without precedent is happening to the American economy, and most people are experiencing it only as a feeling – a vague sense that the numbers stopped making sense somewhere along the way. Here is the number that explains the feeling: the five largest AI data center companies are on track to spend around $900 billion in 2026, roughly what they spent in the previous three years combined. Not since the railroads, or the interstate highway system, has this much capital moved this fast into physical infrastructure.
I have been writing about artificial intelligence on The Cameron Journal since before ChatGPT was a household name – I was skeptical early, and I have not fully retired that skepticism. But what we are watching now is no longer just a technology story. It is a capital markets story, a macroeconomic story, and – if enough of the bet goes wrong – a financial stability story. This is my attempt to lay out the whole machine on the table.
The scale of the build-out
Start with the raw spending. US hyperscalers – Microsoft, Alphabet, Amazon, Meta – are on track to spend somewhere between $700 and $745 billion in 2026, six times their 2022 levels. Global AI capex is forecast to approach $900 billion this year and $5.5 trillion by 2030. Capital intensity at these companies has roughly doubled since ChatGPT launched; they now invest a larger share of revenue than the telecom operators ever did.
Where does it go? Chips, land, power, cooling, and the data centers themselves – concrete and copper as far as the spreadsheet can see. And the spending is pulling an entire ecosystem with it: utilities are expanding their own capex to feed the data centers, though not fast enough – power infrastructure is now the binding constraint in more markets than chip supply.
The debt problem nobody priced in
Here is where the story gets serious. The first phase of the AI build-out was funded by the richest companies on Earth out of their own cash flows. The current phase is being funded with debt – and not just ordinary bonds.
The numbers tell the story: AI-related bonds have gone from about 2% of the investment-grade market in 2023 to roughly 10% in 2026 – around $330 billion issued this year alone. Add the off-balance-sheet stuff – leases, SPVs, securitized data center paper – and by one S&P estimate there is another $675 billion in signed-but-unfunded lease obligations sitting out there. Analysts modeling a downside scenario put potential losses at $140 billion even with most of the exposure outside the regulated banking system.
The market is noticing. Credit default swaps on Nvidia – the company whose chips power all of this – hit their highest spreads on record this summer. Bond investors have started discriminating: paper tied to solid hyperscalers still prices fine, but debt tied to the “neoclouds” – the speculative smaller players renting out GPU capacity – is getting repriced hard. At least 47 large data center projects were delayed, withdrawn, or blocked this year. Banks that underwrote tens of billions in AI construction loans are now trying to hand them off.
The circular deals, explained
And then there is the part of the machine that would make a fraud attorney sit up straight in his chair: circular financing. Microsoft invests in OpenAI; OpenAI spends that money on Microsoft’s cloud. Nvidia is reported to have committed up to $100 billion to OpenAI, which OpenAI uses to buy chips and data center capacity from Oracle, and Oracle buys the chips – from Nvidia. The money goes around the loop and comes back as “demand.”
Now, some of this is defensible – in a capital-intensive industry, vendor financing accelerates legitimate build-out, and it has precedents in telecom and railroads (we’ll come back to how those ended). But when the buyer’s revenue comes substantially from the seller’s investment, reported demand is no longer demand. It is a company purchasing its own growth story. And indeed, a much-touted $250 billion Nvidia-OpenAI data center commitment reportedly came in around $105 billion when the details firmed up – a $145 billion haircut that tells you something about how firm “firm commitments” are.
Railroad economics, revisited
Every bubble has an infrastructure skeleton, and the AI build-out is no exception. The 1840s British railway mania left behind a rail network Britain still uses. The telecom fiber glut of 1999 buried most of the world’s long-haul fiber – fiber that became cheap and useful a decade later. The question for AI is never “is the technology real” – it clearly is – but “who is holding the bag when the financing meets reality?”
The honest answer, as of today: the hyperscalers can absorb almost any outcome. They have the balance sheets to hold these assets through a downturn – they are building shock absorbers in the classic sense, and I’ve written about why resilience beats efficiency in that kind of world. The risk is concentrated one ring further out: the neoclouds, the private credit funds, the data center REITs, and everyone holding the lease obligations. That is where a demand shortfall would actually break something.
What this means for you
If you invest: the era of buying anything with “AI” attached and being rewarded is ending. The market is moving from a demand-for-the-story phase to a who-pays-for-the-electricity phase. Watch the credit spreads, not the product launches.
If you run a business: don’t bet your operations on a neocloud’s cheapest-rate GPU contract – the provider that is borrowing heavily to survive is the provider that may not be there in two years. Contract with entities that can outlast the financing cycle.
And if you want the whole picture – capex flows, credit signals, chip supply chains, and what the build-out means for the real economy – that is exactly what I track at Story, Strategy & Signals, my intelligence briefing on Substack. The AI capex machine is the defining economic story of the decade, and the signals of how it resolves are already on the tape.
Related reading from the archive: my contrarian take in AI is not Intelligence and It’s Not Human, the employment angle in AI Job Replacement is Here, and the early warning in ChatGPT is Coming for Useless Jobs.
Discover more from The Cameron Journal
Subscribe to get the latest posts sent to your email.




