Borrowed Time
I’ve written before that currency is debt. Every dollar in your account is someone else’s liability. When a bank makes a loan, it doesn’t hand over someone’s savings. It types a number into the borrower’s account, and that number is new money.
So think of money as virtual tokens, or often referred to as credit. We can create as much as we like. The only constraint is that credit has to match real economic activity: goods made, services delivered, problems solved. Credit that matches activity is growth. Credit that doesn’t create loss, which has to be settled one way or another.
A loan is credit extended today against activity the borrower promises to generate before it’s repaid. AI is now being financed with credit on a scale we have never seen, and the interesting question isn’t whether AI is real. It’s whether the activity can show up to pay the debt.
The hurdle
For a loan to work, the asset it funds must earn more than it costs. Interest plus depreciation, as the asset wears out while you’re still paying for it.
For AI, most of the spend is on GPUs with a useful life of 3-6 years. At a five-year life, depreciation alone is 20% a year. Adding bond yields of 5%, roughly where the 10-year Treasury sits today, the hardware has to earn around 23% a year to break even.
Hyperscaler capex is expected to top $1 trillion next year, not counting projects financed off balance sheet through SPVs. Earning back $1 trillion of capex over five years at a 5% cost of capital takes roughly $460 billion of revenue a year at a 50% margin, and that’s for a single year’s spending. For scale, OpenAI’s annualized revenue was reported near $50 billion in September.
When there’s more credit than activity behind it. There are only two ways to settle:
- Creditors and owners pay. Default destroys the credit, and the loss stays with those who lent and invested.
- Everyone pays. A bailout funded by printed money, which causes inflation, or higher taxes, a deflationary fiscal squeeze.
Why returns fall
Every revolutionary technology starts with low-hanging fruit. The first uses are obvious, the payoff is huge, and returns sit far above the cost of debt.
Then three things push returns down:
- The easiest problems get solved first. Each new dollar goes to a slightly harder, more expensive one.
- A specific product or service also has decreasing value as supply becomes more abundant.
- As we increase scale, marginal cost increases. In AI, the bottleneck moved from GPUs to memory and packaging, then to power and grid connections, then to transformers and turbines.
Lenders typically underwrite based on the sector’s average return so far, which is backward-looking. A new dollar earns what the next project yields. Early on, the old and the new might yield similar returns. However, as the low-hanging fruit gets picked, they drift apart, and default rates start to rise.
That’s the debt cycle. The weakest, most leveraged players often fail. Their assets get sold at a fraction of the cost, and the new owner may finally earn a decent return at a lower cost basis. The technology survives. The creditors and shareholders don’t.
It happened with railroads in the 1840s and 1870s, with fiber in 2001. The technology was real, and the debt still defaulted.
AI is not an exception.
Value comes from people
It takes time for people and businesses to figure out how to use a new technology to create new things to either serve themselves or others. Finding product-market fit is an experiment. You ship something, wait, and then see the result. Do customers keep using it? Expand? Or just churn?
Each experiment has a feedback delay. Both enterprises and consumers need time to use a product to see if it works. You sometimes need a year of data to know whether a spike was real demand.
Throwing more resources doesn’t fix this. It buys more experiments in parallel, not faster ones. Run ten 30-day A/B tests at once, and you still wait 30 days for the answer. Worse, you can’t run unlimited experiments in parallel: there is a finite number of customers to test on, and experiments running at the same time contaminate each other.
RSI
The hottest idea in AI right now is recursive self-improvement (RSI): AI that improves its own inputs, with better chips, better code, and better models. If that works, the low-hanging fruit grows back, and returns never fall. But whether you call it RSI, AGI, or superintelligence, there are two problems it doesn’t address.
The first problem is physical. Every step of self-improvement needs more compute, which needs power plants, fabs, permits, and land. Recursion speeds up demand for physical things, which cannot be scaled up easily. But that’s the smaller problem.
The second and more principal problem is that the economy isn’t about machines. It’s about human desire and matching it with the right supply. People buy chips to train models, models to serve businesses, and businesses to sell to people. Everything besides consumer is derived demand. Machines improving machines is an intermediate step. Value is only validated when a consumer decides to spend money.
Businesses are, essentially, running giant experiments to find out what people like, and people vote with their wallets. RSI can shrink product development time, but not value-validation time, which often is more time-consuming. When production outpaces discovery, the debt starts to go bad.
What the market is saying
The market is flashing warning signs now. The 20 most valuable AI and chip stocks lost more than $1 trillion in a single week in mid-October. The bond market shows this more clearly. A bondholder’s payoff is lopsided: the best case is getting the money back with interest, the worst case is losing it. So credit investors mostly ignore the upside story and ask the hurdle question directly: will the cash flows cover the debt?
Record revenue and profit in hyperscalers doesn’t settle that question. Anthropic and OpenAI’s compute bill is a hyperscaler’s cloud revenue, which is a chipmaker’s data-center revenue, and some of it is funded by the hyperscaler’s own investment. Each layer books revenue from the one below; none of it proves a person at the end paid. Profit flatters too: chips are paid for upfront but depreciated over five or six years, so earnings lag the cost. Free cash flow, on the other hand, tells a much different story. Alphabet just grew revenue 25% and roughly doubled cloud revenue, and its free cash flow has deteriorated to negative $9.1 billion.
The cash flow gap is being filled with debt. New debt went from 9% of hyperscaler capex in fiscal 2024 to 32% on a trailing basis by mid-2026. Last quarter Alphabet raised $28.7 billion in notes and $12.4 billion by selling stock, and its long-term debt has gone from $46.5 billion to $125.1 billion this year.
Meanwhile, credit default swaps (CDS), the annual cost of insuring a company’s debt against default, quoted in basis points of the amount insured, have widened significantly:
- CoreWeave five-year CDS peaked at 1,167 basis points on October 22, up from 421 a year ago. Anything above 1,000 is generally considered distress, meaning it might go bankrupt.
- Oracle, one of the major hyperscalers, hit 315 basis points, up 311% in a year. That’s a range more typical of junk credit.
Spreads are widening fastest where debt is largest relative to proven cash flow: the neoclouds, then the most leveraged hyperscaler, and then the cash machine like Google and Meta, just by less.
Borrowed time
A loan is credit extended against activity that must show up before the debt comes due. AI makes the product faster than ever. It doesn’t make people decide any faster whether they want it, or shorten the time businesses need to collect enough data for a meaningful signal. Debt runs on a fixed calendar while discovery runs unpredictably and often slowly. When the debt comes due before the discovery pays off, someone has to absorb the difference.