Practicing Wisdom — Issue #19
A distillation of the most interesting things I explored, learned, and thought about.
1. What I Learned This Time
The Age of Borrowed Futures
Debt is one of the strangest human inventions because it lets us spend the future before it arrives.
At its simplest, a loan is a claim on cash that has not yet been earned. That is not inherently dangerous. Almost every productive civilization has depended on some version of this trick. Railroads were built before the passengers arrived and factories were financed before the goods were sold. Mortgages let people inhabit houses decades before they have paid for them.
The interesting question is not whether an economy uses debt - it is what kind of future people feel comfortable borrowing against. Right now, we seem extraordinarily comfortable borrowing against futures that are still mostly hypothetical.
The clearest example is AI. In a recent conversation, Ben Thompson describes the industry moving rapidly “down the capital curve”: first the enormous free cash flows of the hyperscalers, then corporate debt, then equity issuance, and now increasingly elaborate structures designed to tap pension funds, insurance capital, private credit and other pools of long-duration money. His concern is fundamentally a timing problem. The infrastructure is being paid for today; the revenues that justify it are expected tomorrow. If those two timelines fail to meet, you can believe completely in AI’s long-run importance and still experience a spectacular financial accident along the way, which is an important distinction.
We tend to frame bubbles as arguments about whether the underlying technology is “real,” but historically, that is often the wrong question. Railroads, the internet and fiber were real - the problem was that capital structures can fail long before technologies do.
Railroads had what Thompson calls a duration mismatch: enormous upfront costs funded by money that needed servicing long before the railroad generated its mature cash flows. Eventually the world did not run out of demand for railroads, it ran out of money with patience.
J.P. Morgan now estimates hyperscaler capital expenditure at roughly $697 billion in 2026. Meanwhile, AI-related hyperscaler bond issuance has reached about $220 billion so far this year, versus only $12.5 billion over the comparable period last year. Even high-quality tech borrowers are beginning to pay more: tech credit spreads have widened beyond the broader investment-grade market, and investors are demanding larger concessions to absorb new issuance. None of this says the investment is bad, it says that the balance sheet is beginning to carry a belief that the income statement has not yet proven.
Here is where debt becomes psychologically interesting.
Equity says, “This might work.”
Debt says, “This will work on schedule.”
An equity investor can endure delay. A lender eventually needs to be paid. Interest does not care that adoption is arriving one year late, that enterprise procurement is slow, or that the ultimate total addressable market remains enormous: debt converts optimism into a timetable.
That seems to describe much of the age we are living through. We are deeply impatient with waiting for cash flows to finance investment organically. Governments borrow against future tax revenues, corporations borrow against future AI profits, data-center developers borrow against future utilization and consumers borrow against future wages. Increasingly, the economy seems organized around the proposition that tomorrow will be rich enough to validate what we want to do today.
The psychology underneath that is fascinating. It is partly confidence, but perhaps also a kind of temporal entitlement: we have become accustomed to pulling prosperity forward. When everyone does this simultaneously, an uncomfortable thing happens: everybody’s future cash flows begin competing for the same pool of present-day savings.
The U.S. government is doing it too. The Treasury is financing persistent deficits at the same time corporations are issuing unprecedented amounts of long-duration debt. Marc Rubinstein’s piece captures the increasingly visible effort to manage that burden: shifting issuance toward the short end, encouraging new Treasury buyers such as stablecoin issuers, relaxing constraints on bank holdings, and now expanding buybacks of longer-duration bonds. His closing line is the important one: the risk is that fear itself becomes fundamental.
That has already become a market question. Long-term Treasury yields recently reached levels not seen since 2007, prompting the Treasury to double planned buybacks for some long-dated securities. Yet the intervention does nothing to reduce the deficit; it merely changes the plumbing through which it is financed. This means there are actually two giant duration mismatches occurring at once.
The private sector is borrowing against the future productivity of AI and the government is borrowing against the future taxable productivity of the entire economy. This is a good bet if tomorrow is sufficiently richer than today but remember - we need to not understand only if, but when. If capital becomes unwilling to wait the structures will weaken.
So, where would we see evidence that the promised cash is not coming quickly enough?
Here is where financing terms can help as a mark to market.
The first signal is the price of debt. If AI remains magnificent but lenders steadily demand wider spreads, larger new-issue concessions, more collateral and shorter maturities, the market is quietly saying: We still believe you, but less than we did six months ago. That appears to be starting already. Reuters reports that tech bonds now trade at wider spreads than the investment-grade market overall, with investors explicitly describing “indigestion” from the volume of issuance.
The second signal is who must guarantee whom. One of Thompson’s more interesting observations is that Nvidia is increasingly taking equity stakes, providing backstops and otherwise helping customers lower their cost of capital so those customers can continue buying GPUs. Risk has not disappeared; it has migrated back toward the vendor. S&P recently made this concrete by treating Nvidia’s residual-value guarantee supporting an AI infrastructure project as a debt adjustment. The guarantee is manageable for Nvidia today, but the important thing is the direction of travel: the supplier is beginning to help finance demand for its own product. Vendor financing is not automatically sinister, but it is one of the oldest yellow lights in capital-intensive industries. The moment a seller must increasingly finance the buyer, reported demand becomes harder to distinguish from financed demand.
The third signal is the migration from unsecured to structured capital. As long as investors happily lend to a company based on its general creditworthiness, the financing system is expressing confidence in the enterprise. When financing begins migrating into SPVs, asset-backed structures, GPU collateral, residual-value guarantees and bespoke project finance, capital is beginning to ask: Exactly what am I getting if the story fails? Again, this can simply be intelligent financial engineering. But the structure itself tells you where skepticism is emerging.
The fourth signal is asset values after the hype has left the showroom. The most important number in AI may eventually be the resale value of a three-year-old GPU. A surprising amount of current financing implicitly assumes that compute is a durable asset. Yet specialized AI data centers can have short customer contracts, fast hardware depreciation and significant re-leasing risk. If used GPU prices, data-center lease rates or renewal economics begin deteriorating, lenders will discover that the collateral beneath the debt is worth less than their underwriting assumed.
The fifth signal is equity issuance by companies that seemingly should not need it. Thompson highlights Google issuing equity as symbolically important. A company with one of the greatest cash-generating businesses in history is choosing to dilute shareholders rather than fund the entire AI buildout internally or through debt. That does not prove distress, but it tells you something about scale: even extraordinary cash flows may no longer be extraordinary relative to the capital being demanded.
And then there is perhaps the most important signal of all: whether the underlying users ever develop an economic reason to pay enough for what is being built. We spend enormous amounts of time measuring usage, engagement and model capability. Benn Stancil’s essay this week is a useful warning about treating metrics as if they were natural facts. A “user” is a construct. An algorithm is a set of human choices. Models themselves are not pristine objects discovered in nature; they are engineered distributions produced by countless subjective decisions.
The same caution should apply to AI economics: tokens are not revenue, revenue is not gross profit, bookings are not cash, capacity reservations are not utilization. Utilization is not necessarily an adequate return on invested capital.
The ultimate reality check is much more primitive: Does the productive value created by the machines generate enough cash to pay for the machines? Everything else is a model.
Perhaps that is the deepest connection between these pieces. Stancil argues that we forget how much of what appears objective is actually constructed. The same can happen in finance. A financing structure can make an investment appear self-sustaining for a very long time. Capital gets passed from one balance sheet to another; risk is securitized, guaranteed, refinanced or moved off the balance sheet. The system acquires an appearance of inevitability, but debt has a useful property: eventually someone needs to send cash.
That is why I suspect the end of this cycle, if it comes, will not initially look like everyone suddenly deciding AI was overhyped. It will look much duller.
A bond deal prices poorly.
A lender demands more collateral.
A GPU-backed facility gets refinanced at a much higher spread.
A data-center tenant declines to renew.
A hyperscaler slows its capital commitments.
A supplier provides more financing to preserve sales.
A private-credit fund quietly marks down a loan.
None of these events announces that the music has stopped, they are simply the first moments when somebody somewhere says, “I would still like to believe in your future. I am just no longer willing to lend against it at yesterday’s price.”
Sources Referenced
Winners & Losers in the AI Era — Invest Like the Best (link)
The Model is Man-made — Benn Stancil (link)
Great Scott — Net Interest (link)
2. Key Distillations
Debt is optimism with a maturity date.
Equity asks whether the future arrives. Debt asks whether it arrives on schedule.
When a vendor starts financing its customers, demand and credit begin wearing the same clothes.
Risk rarely disappears. Financial engineering mostly changes the address where it lives.
The first crack in a boom usually appears in the price of money, not the popularity of the story.
3. One Contrarian Viewpoint
The AI bubble can burst without AI being a bubble.
Most debates force an unnecessary binary: either AI produces extraordinary economic value and current spending is justified, or AI disappoints and the entire boom was irrational. There is a third possibility.
AI could be every bit as transformative as the bulls believe and still destroy enormous amounts of capital.
The mistake is assuming technological return and investor return must occur on the same timeline. Railroads transformed America while repeatedly bankrupting railroad investors. Fiber became essential infrastructure after much of the capital that financed it was wiped out.
A technology can be civilization-changing and badly financed at the same time.
The risk today may therefore be less “AI fails” than AI succeeds too slowly for its capital structure.
4. One Investable Idea
Follow the creditors.
Equity markets are built to tell stories while credit markets are built to survive them.
If I wanted one dashboard for this cycle, I would spend less time tracking benchmark scores and more time tracking the marginal financing terms of the AI ecosystem:
hyperscaler credit spreads and new-issue concessions;
secured versus unsecured borrowing;
GPU-backed loan spreads;
residual-value guarantees and vendor financing;
used GPU prices and lease renewal rates;
hyperscaler equity issuance;
data-center project refinancings;
and, above everything, the spread between the growth of AI infrastructure capital employed and the cash actually generated from AI services.
The interesting trade may eventually be less “long AI” or “short AI” than long the companies that can internally fund the transition and short the companies whose survival requires continuously cooperative capital markets.
That distinction barely matters while money is abundant; it becomes everything when money gets selective.
5. From the Archives: A Recall Highlight
“The market can tolerate uncertainty far longer than it can tolerate a broken financing mechanism.”
A good idea can survive a crash. A capital structure cannot survive indefinitely without cash.
