[Week 43 of 2026] Financing AI

Welcome back to Price and Prejudice with a few musings from Week 43 of 2026. Today we talk about who is paying for the AI buildout, and who ends up holding the bill if it does not work out.

Last week, CoreWeave – a company that buys AI chips and rents them out to the companies building AI models – had its worst stretch since going public. Its stock fell 15% on October 14, and the yield on its bonds due in 2031 rose to about 15.6%, from about 11.3% in mid-August (CNBC). The cost of insuring its debt against default, i.e., its credit default swap (CDS) spread, also hit a record of almost 12 percentage points a year on October 15. This is happening in the middle of what is probably the largest investment boom in decades, when demand for exactly what CoreWeave sells is enormous. So why would a company selling the hottest product in the economy start to look like a credit risk?

To understand why, it helps to start with a single data center. Suppose you want to build one. You need three things. First, you need a building, which is basically a large warehouse with very good cooling and will be useful for twenty-five years or more. Then you need power, meaning a connection to the grid and often new transmission lines and substations, which will last for decades. Finally, you need chips, which are by far the most expensive part, and which may be state-of-the-art for only a few years before something better comes along.

Money for these have to come from somewhere, and these three pieces invite very different kinds of money. For example, the building is literally real estate, and real estate has been financed with long-term mortgages for a very long time. The power hookup looks like a utility asset, and utilities finance those by borrowing cheaply and recovering the cost from their customers over decades.

On the other hand, the chips look like something else entirely, i.e., an expensive asset that loses value quickly and whose resale value depends on how fast the next generation arrives. As we know, the basic rule in finance is to match the life of the money to the life of the asset. Many of the interesting episodes historically come from stretching that rule. CoreWeave is also part of this trend, where itss most recent loan, for example, runs about five years and is secured by chips whose customer contracts run about three.

So who actually writes the checks? For a while the answer was simple. The large technology companies, (Microsoft, Alphabet, Amazon, Meta, and Oracle) paid for data centers out of the cash their existing businesses generate, which admittedly was a lot of cash. The problem is that capital spending has grown faster than that cash. Goldman Sachs estimates that these five companies spent about $405 billion in 2025, will spend about $750 billion this year, and about $1.2 trillion in 2027. In the second quarter, Alphabet’s capital spending exceeded its operating cash flow, giving it negative free cash flow for the first time since it went public in 2004.

Once you are spending more than you earn, you have to raise money from someone else. The first step is ordinary corporate bonds, which these companies can issue at low spreads because they are among the most creditworthy borrowers in the world. These companies have already taken this route: Amazon already sold about $54 billion of bonds in March, and Alphabet even sold a 100-year bond (which I covered in my class last spring).

The second step is to put the data center somewhere other than your own balance sheet. In a typical arrangement, a technology company sets up a joint venture with an investment firm, the joint venture borrows to build the data center, and the technology company leases it. Meta’s Hyperion campus in Louisiana is the best-known example. A separate entity borrowed $27.3 billion through bonds that run until 2049, Meta signed a lease with an initial term of only four years, and Meta guarantees part of the value of the property if it walks away. From the outside, Meta has a lease rather than $27 billion of debt, even though the debt exists only because of Meta. (You can see how people might be concerned).

The third step is private credit funds and securitization, where loans on data centers are packaged and sold to investors who want long-term, steady payments. Asset-backed securities backed by data centers have grown from about $4 billion outstanding in 2020 to about $61 billion this year. So this is also a popular trend.

Each step moves the risk a little further away from the company that actually wants the data center. That is not necessarily bad, since the whole point of a financial system is to move risk to whoever can best hold it. But it does make it harder to answer a simple question, which is who loses money if the data center turns out to be worth less than everyone assumed.

Then there is the circular part. Nvidia makes the chips, and it also took a $30 billion equity stake in OpenAI, which uses its money, in part, to pay for computing power. Some of that computing power is also rented from companies like CoreWeave, which buy Nvidia’s chips and borrow against them, and in which Nvidia both owns shares and has agreed to buy up to $6.3 billion of unsold capacity through 2032. Meanwhile, OpenAI has committed to buying about $300 billion of computing capacity from Oracle, which now reports a backlog of $664 billion. Once again, none of this is improper but it means that some of the revenue at each link in the chain comes from the next link’s investors rather than from end customers.

When you follow the money down, a lot of it rests on long-term promises from a handful of AI labs whose revenues are still much smaller than what they have committed to spend. OpenAI’s annualized revenue was about $50 billion at the end of September, according to an FT report last week that came in below the roughly $70 billion many had assumed, against computing commitments that run into the hundreds of billions of dollars over the next few years. CoreWeave sits right on top of this: it has about $50 billion of debt and other obligations, much of it secured by its chips, and its three largest customers, including Microsoft and OpenAI, accounted for 72% of its revenue in the second quarter. The selloff came after the OpenAI revenue report and some predictions that about 80% of companies like CoreWeave will not survive independently by 2030.

So it is not that demand for AI has collapsed. Rather, it so happens that CoreWeave is the layer with the most debt, the shortest-lived collateral, and the most concentrated customers.

There is also one wrinkle in the accounting. Because chips wear out or become obsolete quickly, how fast companies depreciate them matters a lot for reported profits. Microsoft and Alphabet extended the assumed useful life of their servers from four to six years a few years ago, while Amazon shortened part of its fleet from six to five years in 2025, citing the pace of change in AI. If chips are depreciated over six years but are only useful for four, earnings look better today and the loans backed by those chips look safer than they are.

It is worth stepping back from individual companies, because the financing question also has a macroeconomic side. Every dollar spent on a data center is a dollar that someone, somewhere, did not spend on something else. There are basically three ways for an economy to come up with the money: (1) Households can save more, which usually requires higher interest rates to persuade them; (2) Foreigners can send their savings, which shows up as a larger trade deficit and often a stronger dollar. (3) Or other businesses can invest less, because the AI buildout is outbidding them for money, equipment, and electricians. (Economists call the last one "crowding out" effect.)

As we discussed two weeks ago, real interest rates in the US have risen by about a percentage point this year, which is consistent with the first channel. But a point is not that much given the size of the buildout, which suggests the other two channels are doing a lot of the work. Foreign money is clearly part of it. Foreign investors bought a net $300 billion of long-term US securities in August alone, the US current account deficit widened to 3% of GDP in the second quarter, Abu Dhabi’s MGX closed a $49 billion AI fund in July, and SoftBank borrowed $40 billion to fund its stake in OpenAI. Crowding out is generally a bit harder to see in financial markets, but we can easier see it in the scramble for transformers, turbines, and electricians.

This means that some of the bill also lands on people who never (explicitly) agreed to finance AI. When utilities build new generation and transmission for data centers, part of the cost ends up in everyone’s rates. For example, the EIA expects average residential electricity prices to rise about 5% this year, and PJM, the grid operator for much of the eastern US, has cleared its capacity auction at the price cap for three years in a row. Taxpayers also pay a share through the tax code, since companies can deduct the cost of new equipment immediately. And every other borrower, including governments with large debts, now competes with the AI buildout for the same pool of savings.

Insurers are another quiet participant. The long-dated, steady payments from data-center bonds and leases are exactly what life insurers like to buy to match their long-dated liabilities. Moody’s estimates that US life insurers held about $85 billion of AI-related investments at the end of June, which excludes indirect exposure through lease-backed deals like the one behind Meta’s Hyperion campus. The insurance regulators are now reviewing the credit ratings on insurer-held data-center deals, which is probably a sign that the question of who holds the risk has reached them too.

An important question is who holds the claim when the chips wear out and the leases come due. It could be technology company shareholders, who are diversified and can absorb losses. It could be private credit funds and the insurers that buy their loans, for whom this kind of risk is newer. Or it could be electricity customers, who never chose to take it on. The CoreWeave episode is a nice reminder that collectively the market has started asking that very question and through financial markets, we are truly all connected with one another (for better or for worse).

Other Stuff

• Polymarket is losing ground to Kalshi, and its investors are pushing for changes, a follow-up to last week’s post on prediction markets.
• Airplanes are being scrapped earlier than expected, with an engine shortage making some jets worth more as spare parts than as aircraft.
• Municipal bond issuance is at multi-year highs, as states and cities borrow more at today’s higher yields.