Big Tech Borrows, Carlson Finally Buys Bonds

How big the buildout is

Van Nieuwerburgh starts from companies' announced investment plans. He then discounts them, on the assumption that not every plan will go ahead. He compares the estimated AI buildout with earlier investment booms, such as railroads, in a table (shown below).

Taylor flags two caveats about that comparison. First, the AI figure is a forecast, while the others actually happened. Second, the time spans differ: the railroad boom, the second-largest figure in the table, lasted much longer than the few years the AI estimates cover. Even so, Taylor calls the AI estimates "striking."

The turning point: capex exceeds cash flow

• about $97 billion in 2020
• more than $400 billion in 2025
• a projected $800 billion-plus in 2026

Here is a simplified, hypothetical example of what that second option means. A tech company wants a data center but doesn't want the full cost and debt on its own books. A separate entity, such as a developer or a joint venture with an investment fund, builds and owns the facility. That entity raises its own debt, and the tech company signs a long-term lease. The tech company still depends on the building, but the debt sits with someone else.

Taylor ties this to an older puzzle. Through the 2010s, big US tech firms held more than $1 trillion in cash, and economists kept asking what they were saving it for. The AI buildout looks like the answer, and it has now used up more than the cash they generate each year.

Who supplies the outside money

The paper describes a mix of equity and debt:

The BIS study, by Puriya Abbassi, Iñaki Aldasoro and Sebastian Doerr (BIS Quarterly Review, September 2026), looks at that last group. Banks often want physical collateral before they lend. Private credit funds will lend against a borrower's expected cash flows and intangible assets, which suits tech firms. As bank regulation tightened, private credit grew to about $2.5 trillion in loans outstanding by 2025. Tech firms' share grew even faster. Their private credit borrowing rose from about $22 billion (22% of the total) in 2010 to over $1 trillion (44%) in 2025. So tech went from roughly a fifth of a small market to nearly half of a much larger one.

Why the funding source matters for risk

How the money is structured matters most if things go wrong. Investors in private credit funds usually commit their money for five to eight years and cannot withdraw it early. A bank, by contrast, lends out depositors' money, and depositors can ask for it back. That mismatch is what turns bad loans into bank runs and wider crises. If a private credit fund's loans go bad, its investors lose money. But Taylor says there is "much less reason to think" those losses would spread into the banking system.

That is the reassuring side. The worrying side is scale combined with interconnection. A market economy can usually absorb individual companies failing. But this investment is very large, and the risk is spread across a web of tech firms, developers, REITs, funds and lenders. That raises two open questions. Who takes the losses if AI revenue lags? And could those losses be big enough to hurt the broader economy? Neither paper, as Taylor summarizes them, gives a definitive answer.

Lessons from earlier booms

Taylor points to the dot-com bust as a precedent. US stocks fell about 50% between March 2000 and October 2002, many internet stocks fell much further, and the US had an eight-month recession in 2001. Companies vanished, but the internet did not, and the physical network built during the boom was later put to productive use.

He also borrows a question from the California gold rush: who got richer, the miners or the people selling them picks and shovels? Many miners found little, but every miner had to buy tools. Taylor argues that some of today's richly valued AI software companies will turn out to be unsuccessful miners. They will still have paid for the physical buildout either way.

He adds one piece of logic about the public debate. People worry that AI will be so cheap it causes mass unemployment, and also that its infrastructure will be ruinously expensive. AI, he notes, "will not be both cheap and expensive at the same time."

What remains open

The summary doesn't say how much of the investment sits in off-balance-sheet structures. It doesn't map which specific lenders are most exposed, and it doesn't estimate losses under a slowdown. The size figures rest on forecasts of plans that may change. What it does establish is the shift: AI's physical expansion is now financed partly with other people's money, through channels that are harder to see than a company's own spending.

The starting point: who pays for the machines

Fernández-Villaverde starts from a simple claim. Building out AI will take enormous investment, and that investment has to be paid for out of saving. Households and firms save more only if saving pays more, so the real interest rate has to rise a lot. In his benchmark scenario the ten-year real yield, meaning the inflation-adjusted return on ten-year government bonds, jumps from 5.1% to 13.6% the moment the transition begins.

To see how that works, here is an illustrative example, not taken from the paper. If r is 10% and g is 2%, debt of 100% of GDP becomes roughly 108% a year later. Interest adds 10 points and growth dilutes it by about 2. Holding the ratio steady would take a primary surplus of about 8% of GDP. Each year the gap stays open, it compounds.

The adopter: growth rises, but rates rise faster

The tax rate has to exceed the r − g gap because the tax itself slows growth: from 21.9% to 13.6% with unit elasticity, and from 11.7% to 9.1% at 0.2. It does nothing to lower the rate the Treasury pays, because in the model that rate is fixed by the machines' net return, which is set by the technology. The verdict: painful, but the boom survives and the debt gets paid.

The bystander: same rates, old growth

The second economy is a bystander like Spain. It is fully plugged into world capital markets but, as a deliberately extreme assumption, adopts no AI at all. It pays the world interest rate and keeps growing at its old 1.6% trend. Its capital income stays at about a third of GDP, and that GDP grows slowly, so there is no boom to tax.

The surplus it would need is 11% of GDP in 2030, 26% in 2035 and 44% forever under unit elasticity. In the low-substitutability case the figures he gives are 18% and 24%. On a balanced primary budget its debt doubles by 2033 and triples by 2035. "No tax system raises that," he writes.

He adds two things that make it worse:

• Front-loading. Long-term yields jump straight away. A Treasury that issues ten-year bonds locks in high rates sooner than the short-rate rollover assumes. It is the same bill, paid earlier.
• Capital flight. The bystander's own savings flow abroad to fund foreign machines, which pulls its GDP and tax base below trend.

One offsetting effect: the bystander's households can own other countries' machines and get rich on the returns, and their government can tax some of that income.

What to make of it

Fernández-Villaverde calls this "a benchmark, not a forecast." His real point is about the arithmetic: if growth arrives with an interest rate about twice as high, the boom makes debt worse, not better.

Cowen says he is "somewhat more optimistic," for three reasons. Zero adoption is an extreme assumption. Ideas are public goods, so Spain can benefit from American and Chinese AI and the science it produces without building it all itself. And the long-run supply of capital may be quite elastic, meaning saving could rise readily without needing such large rate increases.

Some questions stay open in the excerpt. It doesn't explain why the benchmark real yield starts at 5.1%, well above recent real rates in Europe or the US. It also leaves out how sensitive the results are to the IES of one-half, which drives the widening gap. The numbers in the 13–44% range rest on the model's assumptions, and the scenario is a stress test rather than a forecast.

A long-time stock investor buys bonds

Ben Carlson, a wealth manager who writes A Wealth of Common Sense, said on his podcast that he had bought bonds for the first time in his career. A 38-year-old reader with everything in broad stock index funds asked him why. His answer is less a market call than a way of deciding when bonds are worth owning and which kind to buy.

How bonds got here

Carlson starts with history. High rates in the early 1980s kicked off a 40-year bull market in bonds. When you start with high yields and rates keep falling as inflation cools, bond returns are very good. After the 2008 Great Financial Crisis, though, yields stayed very low for years, and they went lower still during the pandemic.

The other result is that yields are now the highest since 2007. Carlson points out that's roughly two decades in which bonds paid very little income, so a whole generation of investors has never seen yields this high.

What bonds are for

Carlson lists the jobs he thinks bonds do in a portfolio:

He explains why he skipped bonds when he was younger. They protected portfolios well in the 2008 crash, but he was in his 20s then, and falling stock prices just meant he could buy more shares cheaply. A long time horizon and steady savings were already doing what bonds would have done.

Duration: the key risk

So he sticks to intermediate duration or shorter, about 5–6 years at most. Within that range, how much duration he takes depends on the pay. For most of the 2020s, yields were too low to justify any duration risk. With higher rates over the last few years, he's willing to take "a little."

An illustrative example, not from the article: say a bond fund has a duration of 5 and yields 5%. If rates jump 1 point, it loses about 5% in price. But it keeps earning around 5% a year in interest, and new money gets reinvested at the higher rate. So in a bit over a year, the investor is roughly back to even. Back when the same fund yielded 1%, that same rate rise would have taken years to earn back. This is what Carlson means when he says higher yields offer "a margin of safety."

Why now

He gives two reasons. The first is that bond returns tend to follow starting yields, and his chart of the 10-year Treasury yield against its forward annualized return shows this. With yields where they are, he says, expected bond returns are the highest "since Apple launched its first iPhone."

He describes it as diversifying, not retreating. His retirement and brokerage accounts still hold plenty of stocks, and he sees the bonds as trimming risk after "one of the best 15 year runs in stock market history."

What he isn't saying

Carlson specifically says he isn't predicting a stock crash or a fall in interest rates. Yields could keep rising if inflation stays sticky, the government keeps spending heavily, or AI capital spending booms. If they do, his bonds will lose value for a while. He also notes that higher yields would mean higher expected returns afterward, which is the same breakeven logic as in the example above.

What's left open

The post doesn't say how big his bond position is compared with his stocks. It doesn't name the type of bonds or fund, and it doesn't compare Treasuries, municipal bonds, corporate bonds or CDs. It also doesn't take on the 2022 problem directly: bonds fail as a stock hedge when inflation is what's driving markets down. Carlson's list says bonds hedge "most" crashes, and that caveat is carrying more weight than the post admits. Topics like paying off a mortgage versus buying bonds were covered on his podcast, not in this post.

Today the index spans 11 sectors. By company count: industrials 83, financials 76, information technology 75, health care 60, consumer discretionary 47, consumer staples 34, utilities 31, real estate 30, materials 24, communication services 22 and energy 21. Those add up to slightly more than 500, because a few companies list more than one class of share. Carlson uses the comparison to argue that markets really are different now, mostly in investors' favor.

A broader index, but a top-heavy one

Carlson grants that today's index is more concentrated at the top. The S&P 500 is weighted by market value, so a handful of giant companies drive much of its movement. Still, he points out that the kinds of businesses it holds are far more varied than in 1960. In a footnote he adds that the index has more industrial companies (83) than tech companies (75). Tech dominates only because its companies are so much larger.

What investing used to cost

Carlson leans on a recent column by Jason Zweig, the Wall Street Journal's personal-finance writer, that lists how much worse the old market structure was:

How much the costs mattered (illustrative)

Here is a rough illustration, not from the article. Put $10,000 into a fund with an 8.5% load, and $850 goes to the salesperson before anything is invested. Ang's tax estimate also matters a lot over time. $10,000 compounding at 10% for 30 years grows to about $174,000. At 7% it grows to about $76,000. Add a 1%-plus annual fee on top of that, compared with a few hundredths of a percent for a broad index fund today, and the old-era investor kept a much smaller share of the market's return.

That's the basis for Carlson's main claim. Future stock returns may be lower before costs than in the past. But after costs, he argues, investors have never had it better.

Reg FD and the vanishing star manager

When knowledge itself was the edge

To show how little was once widely understood, Carlson cites Robin Wigglesworth's new book A Fabulous Debt. In college, Michael Milken read W. Braddock Hickman's study of corporate bonds from 1900 to 1943. Hickman found that lower-quality bonds sold at lower prices with higher yields and defaulted more often. Even so, the extra yield more than made up for the losses. Taking more credit risk paid off. One of Milken's colleagues at the investment bank Drexel later said maybe 2,000 people read the book and 1,999 ignored it. Milken built the high-yield ("junk") bond market on that insight and dominated it in the 1980s. As Carlson puts it, it "didn't end well for him": Milken later pleaded guilty to securities violations. Today, an edge like that would be priced in almost immediately.

What hasn't changed

Carlson's list of changes: more products and strategies, lower costs, and automated investing that makes it easier to stick to a plan. The constant is human nature. His warning is that more data breeds overconfidence. Knowing more about the past and present doesn't make the future easier to predict.

What the piece leaves open

This is a short column, not a study. Carlson doesn't put a number on how much net returns have improved, how big today's concentration risk is, or how valuations should be adjusted for the change in sector mix. The claim that Reg FD explains the shortage of star managers is his argument, not a tested result. The tax numbers rest on one researcher's estimate that leaves out state and local taxes.

Adam Grossman, founder of Mayport, a wealth-management firm that charges a fixed fee, starts from a simple point in a new HumbleDollar column. Nobody can predict markets. That gap gets filled by commentators with confident stories, and he argues the stories can persuade even when the data points the other way. He walks through two current debates, shows a strong case on each side, and ends with a process for investors in place of a forecast.

Debate one: is the stock market overpriced?

He adds a worry that CAPE may even understate the risk, because some AI companies' earnings may be inflated by money moving in circles. His main example: Nvidia has announced it will invest up to $100 billion in OpenAI, while OpenAI spends billions on Nvidia's chips. Nvidia has also invested in dozens of other AI firms that buy its hardware.

Here is a simplified illustration of the concern (not real figures). Company A invests $10 in Company B. Company B spends that $10 on A's products. A books $10 of revenue, but the money started as A's own cash. If much of the industry's demand works like this, reported earnings overstate demand from paying customers who are outside the loop. Grossman calls it an open question whether this revenue is sustainable or "a game of financial musical chairs." He says the look of these deals supports the argument that the market is "out over its skis."

Bears also like a piece of Wall Street lore. J.P. Morgan, asked the secret of his success, supposedly answered: "I made all my money selling too early." The lesson they draw is that caution pays.

Where the anecdote meets the data

His verdict: both sides "have a legitimate point." No single data set settles the future. Outcomes depend on interest rates, AI, elections, geopolitics and factors nobody is watching yet. So definitive predictions deserve little weight.

Textbooks call U.S. government bonds the "riskless asset." The reasoning is that the U.S. has the largest and most productive economy, and the Treasury has always been able to borrow what it needs. Federal debt passed $40 trillion a few months ago. Even so, the government's near-perfect credit rating has let Congress keep financing deficits of trillions of dollars a year.

Rising yields have other possible causes, such as expected inflation or a heavier supply of new bonds. Grossman offers fiscal worry as "one explanation," not as proof.

What he recommends instead

His formula has three parts:

For example, a reader whose plan depends mostly on U.S. stocks might stress-test a large drop in share prices. A reader holding long-term Treasuries as the "safe" part of the portfolio might test what rising yields do to bond prices, since bond prices fall when yields rise. The article doesn't give a model portfolio or explain how to run these tests, so the details are left to the reader.

Assessment

This is a framing piece, not new research. Its strength is showing that two respected-sounding arguments can point in opposite directions, and treating forecasts as inputs to a stress test rather than as signals to trade on. Its limits: the study it relies on gets one sentence, and the bond-yield explanation is offered without other possible causes. The advice to diversify is sound, but it doesn't say what diversification means when both stocks and supposedly safe bonds are the assets in question.