Every family knows the kitchen-table version of this debate. The roof needs replacing, the old one is leaking, and the contractor’s estimate is bigger than anything you’ve ever spent on the house. Do you borrow the money? The roof will protect everything underneath it for thirty years — that’s the argument for. The monthly payment will follow you around like a shadow — that’s the argument against. Everybody at the table agrees the roof is worth having. The fight is about who pays, how, and whether the math works.
That is, almost exactly, the debate consuming financial markets right now. Only the roof is the entire artificial intelligence future, the contractor’s estimate runs into the trillions, and the borrowers are the biggest companies on earth.
Start with the borrowing itself, because this week it reached a scale that’s hard to wrap your head around. Axios Markets reported that AI financing is getting bigger and more complicated, with S&P raising concerns that hyperscalers’ pristine credit ratings are at risk. Investors are so eager to fund the AI boom that companies are issuing a record surge of zero-coupon convertible bonds — borrowing, essentially, interest-free, in an environment where money is anything but free. Bond investors are getting twitchy over the scale of data-center borrowing. And in a development that should make every saver think, Big Tech’s borrowing binge is now giving Treasury bonds real competition: investors are increasingly choosing to lend to AI companies over the U.S. government, one more pressure pushing up Washington’s own borrowing costs.
A Morgan Stanley analysis put the number in perspective: roughly $3 trillion in AI-related infrastructure spending committed by seven Big Tech companies in off-balance-sheet commitments. Three trillion dollars that doesn’t fully show up the way ordinary debt does. There is even an effort underway to create a futures market for AI compute — contracts to hedge the price of processing power the way farmers hedge wheat. When Wall Street starts building a futures market for something, you know the spending has gotten enormous.
Oracle is the case study that brings it all home. This week the company reported a genuine blockbuster: adjusted earnings of $1.92 a share, revenue of $19.34 billion up 30%, cloud infrastructure sales up 121%, and a backlog of signed future business at $664 billion. And yet its shares were down more than 20% for the year before the report, because the company plans $90 to $95 billion in capital spending this year and S&P Global downgraded its credit rating in July over weak cash flow. That’s the whole debate in one company: the future looks extraordinary, and the bill is terrifying. One strategist’s summary was perfect — Oracle’s problem was never finding customers; it was proving that building ahead of demand was the right call.
The deals kept coming. Nvidia agreed to buy the open-source AI platform Hugging Face for $12.93 billion — about 86 times the company’s roughly $150 million in annualized revenue. That’s not a valuation; that’s a statement of belief about where the center of AI development will live. SpaceX signed an AI hosting agreement worth about $1.11 billion per month, starting December 1, disclosed by its CFO at the Goldman Sachs Communacopia conference — its fourth major compute deal in four months, after agreements with Anthropic, Google, and Reflection AI, and it underpins confidence in a $100 billion annual revenue run rate by year-end. SoftBank has built a $64.6 billion, 13%-plus stake in OpenAI. DeepSeek’s valuation climbed above $70 billion, with a shadow secondary market spawning around the fundraising frenzy. Two-year-old AI startup Sapien raised at a $180 million valuation led by Neo’s Ali Partovi, moving from financial-planning software into systems that link operational decisions directly to profits — its customer list includes Bayer and Carlex.
Then there’s the plumbing of the whole boom: the chips. ASML — which holds a 100% commercial monopoly on the extreme-ultraviolet lithography machines that make advanced semiconductors — saw Samsung and TSMC agree to adopt its next-generation high-NA EUV machines by 2028 and 2030. ASML’s shares are up 126% in the past year. And Axios reports “chipflation is here to stay”: memory chip prices are skyrocketing on AI demand with no end in sight, pushing up the cost of electronics and cloud services for everyone. That’s the part of the AI bill that shows up in your life directly — the laptop costs more, the phone costs more, the cloud storage costs more, because the machines that think are hungry for memory.
So who’s right — the borrowers or the worriers? The believers have serious voices. Yale economist Edward Yardeni, the man who called the 2020 market bottom, now puts 80% odds on the AI rally lasting into the early 2030s, arguing that profits are real, the economy is holding, and AI is delivering. His projection: the S&P 500 could hit 10,000 by the end of the decade. A PwC estimate carried by the Journal puts global AI infrastructure spending at more than $31 trillion between now and 2050 — nearly matching today’s entire U.S. GDP. Finimize’s retail investor survey found 70% of investors expect global markets to be higher in twelve months. The Anthropic IPO is the most-wanted private listing among retail investors, with 41% naming it.
But the week also carried the warnings, and they deserve equal weight. An Anthropic safety researcher who resigned this week told CNBC he believes there’s over a 10% chance AI could, in his words, “kill us all by the end of the decade.” However you weigh that claim — and reasonable people will weigh it very differently — it landed in the same week the money got bigger, which is worth sitting with. The buildout also isn’t spreading its rewards evenly: only 26% of new U.S. AI hires in 2025 were women, according to data Axios cited — the boom’s highest-paying jobs going overwhelmingly to men, the same imbalance as the old tech workforce. New money, old patterns.
It’s not enough to just ask whether AI is real. That’s the wrong question now — the earnings, the contracts, the $664 billion backlog all say something very real is being built. The right question is the kitchen-table one: who pays, how, and does the math work? Debt taken on for something that generates real returns is how every great expansion in history was financed. Debt taken on because everyone else is borrowing is how every great bust begins. The honest answer is that nobody knows yet which one this is — and anyone who tells you they do is selling something.
Here’s what I’d put my hope in. The discipline is starting to show up where it matters. Investors punished Oracle’s stock until it proved the backlog was real. S&P downgraded a credit rating when the cash flow didn’t justify the spending. A futures market for compute is, at its heart, a tool for pricing risk honestly instead of guessing. Markets are asking the second question — not “is AI exciting?” but “can you afford it?” — and that is exactly the question that keeps a boom from becoming a bubble.
For the rest of us, the practical takeaway is unglamorous and important. If you own the broad market, you own this bet — the S&P 500’s largest companies are the ones doing the borrowing. That’s neither a reason to panic nor a reason to pile in; it’s a reason to understand what you own. Check what your index fund’s top holdings are spending and why. And remember the kitchen table: the roof is worth borrowing for when the roof is real, the contractor is honest, and the payments fit the budget. The AI future is being built either way. Our job is to make sure we’re the ones asking whether the math works — before the bill comes due, not after.
