Money Stuff: AAA AI

I think sometimes, in our age of artificial intelligence, about two opposing theories of asset pricing. One is the standard academic theory, which is that investors demand compensation for correlated returns. An asset that will pay back your money 98% of the time, but will give you zero in the 2% of cases where there’s a big market crash, is bad: It loses money at the worst possible time, the time when all your other assets lose money and you really need money. An asset that will pay back your money 98% of the time, but will give you zero in some random 2% of cases — when there’s a hurricane in the Caribbean, when the Bengals win the Super Bowl — is good: Sure it has risk, but that risk is uncorrelated to the rest of your risks; this asset probably performs well when you really need it, and loses money when you are otherwise fine. 

There is another popular folk theory in the financial industry, though, which is that investors minimize career risk and thus demand compensation for taking uncorrelated risks. An asset that pays back your money 98% of the time, but goes to zero in the 2% of cases when all of your competitors are also losing money, is fine. If everyone’s losing money at the same time, then (1) no one will blame you for losing money too and (2) your firm is probably blowing up anyway so owning the asset doesn’t change much. (“IBGYBG,” this theory is sometimes called.) An asset that pays back your money 98% of the time, but goes to zero in some random 2% of cases when everyone else is doing great, is embarrassing. If you buy that thing and it blows up on a balmy summer day, (1) you’re getting fired and (2) everyone else is getting a nice bonus and shaking their heads at your idiosyncratic mistake. “Why did she buy that thing that blew up,” they ask. “She should have bought normal stuff like the rest of us.”

There’s a classic paper rooted in the first theory, “Economic Catastrophe Bonds,” by Joshua Coval, Jakub Jurek and Erik Stafford. Here’s a 2007 working-paper version, and here’s the 2009 version in the American Economic Review. Some stuff happened between those dates. From the abstract:

The central insight of asset pricing is that a security’s value depends on both its distribution of payoffs across economic states and state prices. In fixed income markets, many investors focus exclusively on estimates of expected payoffs, such as credit ratings, without considering the state of the economy in which default is likely to occur. Such investors are likely to be attracted to securities whose payoffs resemble those of economic catastrophe bonds — bonds that default only under severe economic conditions. We show that many structured finance instruments can be characterized as economic catastrophe bonds, but offer far less compensation than alternatives with comparable payoff profiles. 

The “economic catastrophe bonds” they are referring to were, above all, AAA-rated mortgage-backed securities. Those securities had AAA ratings because they were, on some reasonable assumptions, unlikely to default. But if they did default, they’d definitely default at the worst possible time. From the paper:

Credit ratings describe a security’s expected payoffs in the form of its default likelihood and anticipated recovery value given default. However, because they contain no information about the state of the economy in which default occurs, they are insufficient for pricing. Nonetheless, in practice, many investors rely heavily upon credit ratings for pricing and risk assessment of fixed income securities, with large amounts of insurance and pension fund capital explicitly restricted to owning highly rated securities. In light of this behavior, the manufacturing of securities resembling economic catastrophe bonds emerges as the optimal mechanism for exploiting investors who rely on ratings for pricing. These securities will be the cheapest to deliver to investors demanding a given rating, but will trade at too high a price if valued based on rating-matched alternatives as opposed to proper risk-matched alternatives.

On the one hand, you know, pensions and insurers should not have paid high prices for AAA-rated stuff that was highly correlated to the housing market, and “manufacturing” that stuff was “the optimal mechanism for exploiting” those investors.

On the other hand, what else were they going to buy? There are a lot more assets that are correlated to the economy than not! If you’re a special smarty at a hedge fund, you can put all your money into hurricane bonds or Bengals Super Bowl bets and achieve steady returns with no correlation to the broader economy. But if you are, you know, the insurance industry as a whole, or if you are everyone’s retirement savings, you kind of have to own, like, the economy. You can’t have a whole economy that is uncorrelated to the economy. 

I have written a couple of times recently about banks making margin loans against stock in OpenAI and Anthropic. Margin lending against private stock is normally kind of risky and bespoke, but these deals seem different to me. Because they are, like, attractively correlated. I wrote:

“Borrowing against OpenAI to buy more OpenAI” is a decent description of the global economy right now. If OpenAI’s valuation collapses, then the banks that gave SoftBank this margin loan will be in bad trouble, but so will all the other banks. Might as well also do the margin loan.

Not everyone is literally making margin loans secured by OpenAI stock. But lots of people are buying investment-grade bonds backed by data centers, which ultimately depend on cash flows from, you know, AI. OpenAI is expected to provide $750 billion of those cash flows. Anthropic has committed $200 billion to Google, $100 billion to Amazon, etc. A lot of the economy, for the next few years, might flow through OpenAI and Anthropic. If the basic AI thesis works out, then (1) lots of companies will pay lots of money to use AI models, (2) that money will result in revenue for the big AI labs that build the models, (3) they will spend a lot of that revenue paying for computing power, (4) that will create cash flow for the data centers that provide the compute and (5) that will pay back the bonds.

For reasons — Anthropic and OpenAI are private companies, still quite young, and still losing money — the bonds are not issued directly by the AI labs, and they tend to look to the credit of big existing investment-grade public companies. (Either as direct issuers, or as providers of some guarantees for securitizations.) And you could easily imagine scenarios in which (1) things work out great for AI in general, and for data center bondholders but (2) things work out poorly for Anthropic or OpenAI or both. Maybe there is one winning model, and the winner gets huge revenue and needs all the capacity in all the data centers, so the data centers do fine, but the losing AI labs go to zero. Or maybe the AI revenue flows from customers to data centers, but the labs end up squeezed on both sides and keep only a tiny cut of that revenue. OpenAI and Anthropic are not perfectly correlated to “the AI trade” in general. [1]

Still there is some rough sense in which a lot of investment-grade data-center bond investors are basically buying indirect claims on OpenAI’s and Anthropic’s revenue. There is some rough sense in which the economy is increasingly the AI trade, and the AI trade is OpenAI and Anthropic, and where are you going to put your money if not in the economy? Anyway the Financial Times reports:

Anthropic and OpenAI’s bankers are lobbying for an investment-grade credit rating after their upcoming initial public offerings, a designation that would lower the borrowing costs for their ambitious AI infrastructure plans.

Morgan Stanley and Goldman Sachs have held talks with credit rating agencies in recent weeks on behalf of the two leading AI labs, as they look to gain access to the $11.7tn corporate bond market post-IPO, said people familiar with the matter.

Analysts at the rating agencies told the FT that bankers acting for Anthropic and OpenAI had argued that the two companies’ public listings would unlock vast amounts of liquidity and improve the health of their balance sheets.

“Wall Street is trying to minimise their overall debt impact by arguing that these two companies will soon be flush with liquidity,” said one senior credit analyst.

Achieving an investment-grade rating from Fitch, Moody’s and S&P soon after going public would be a remarkable feat for the two lossmaking AI labs, unlocking big benefits for the companies and their infrastructure partners including Oracle and Nvidia. …

Analysts at rating agencies are waiting to see the results of their IPOs before reaching a decision. The two companies remain unprofitable and have shown little sign of generating positive free cash flow. They also face growing risks, including the popularity of Chinese open-weight models.

“We still treat OpenAI and Anthropic as deep in speculative grade . . . they are in the red,” said another senior credit analyst.

I mean, sure they lose money, but what are investment-grade investors going to invest in if not the AI buildout? Maybe they can be graded on a curve.

Sure? Okay?

Defiance ETFs has submitted paperwork to the US Securities and Exchange Commission for a series of leveraged funds that would seek to double the moves of some individual stocks — over periods measured in just hours, rather than days. …

Existing two-times leveraged funds aim to deliver twice a stock’s move over a single trading day. Defiance’s proposed strategies would effectively restart that bet several times before the market closes.

The filing lists products tracking the hottest tech names including Meta Platforms Inc., Microsoft Corp., Nvidia Corp., Palantir Technologies Inc. and Tesla Inc. The proposed funds would use swaps or options to maintain roughly twice the underlying security’s exposure, rebalancing six times throughout the trading day instead of just once at the close, per the filing. If approved, the funds wouldn’t use a single price for each reset and would instead utilize a time-weighted average price, the paperwork says.

I … ? We have talked about levered exchange-traded funds before. Most 2x levered ETFs rebalance once a day. They aim to give you 200% of the return of the underlying stock for one day. If you hold them for two days, you do not generally get exactly 200% of the underlying stock’s two-day return: The ETF rebalances at the end of each day, buying more stock if it’s up and selling some if it’s down, so it does not perfectly replicate the experience of holding 200% of the stock for two days, or any longer period. Often this is bad: It’s called “volatility drag,” and it often means that people who hold the leveraged ETF for long periods get performance that is worse than 200% (or even 100%) of the underlying stock. 

What if you want exactly 200% of the performance of the underlying stock for some longer period, a week or a month? Well. You could, you know, go to your broker, take out a margin loan, buy $200 worth of stock with $100 of your own money, and get 200% of the stock’s performance for exactly the period — an hour, a day, a month, whatever — that you want. But as I often say, eventually every trade becomes an ETF, and so in fact there are “calendar reset leveraged ETFs” for that. You can buy an ETF that gives you, like, 200% of the performance of the S&P 500 for the month of September, that sort of thing.

The tradeoff is that, if you don’t buy that ETF on the first of the month, you don’t get exactly the right 200% return. The ETF doesn’t rebalance daily, so if you buy the September ETF on Sept. 9, you should expect to get a bit more than 200% of the S&P’s return for the rest of the month. (Because the S&P has gone down a bit, but the ETF has not rebalanced, leaving it slightly more levered than 2x.) The monthly ETF is designed to give you a 200% return if you buy at the beginning of the month and sell at the end, but if you buy in the middle it’s imprecise.

Someone could offer an ETF that gives you 200% of the return of some stock from Sept. 9 until Sept. 30, but so far no one does. Give it time!

Conversely, if you buy the daily ETF at 1 p.m. and sell it at 2 p.m., you might not get exactly 200% of the hourly return on the underlying stock. If the stock goes up in the morning, the ETF will be a bit underlevered by 1 p.m. [2] If you want exactly 200% of the hourly return, you are out of luck. Until now. Now you are in luck. A modest amount of luck. Now Defiance will sell you 200% of the hourly return. Why do you want that? That is the wrong question. The point is that this is a possible trade on an infinite list of possible trades, and eventually there will be an ETF for every trade on the infinite list of possible trades, and now there’s this one. I mean whatever here’s a reason you could want it: 

A trader expecting Nvidia to jump on a piece of news, for example, could buy the fund during one of its hourly periods, targeting roughly twice the stock’s move during that window rather than its move over the entire day. 

Sure?

James Seyffart, an ETF analyst at Bloomberg Intelligence, is somewhat skeptical.

“I’m not fully convinced this will offer completely differentiated exposures when compared to daily resetting products aside from very specific hourly periods around earnings or other announcements,” he said.

Right like. A 2x levered daily ETF does not give you precisely 2x the hourly returns of the underlying stock. But it gets pretty close. Do you need more precision? Why? Again, the point here is not to fill a need; it’s to fill every possible space on the map. But it’s weird. Earnings are normally announced before the open or after the close? Also, like, outside of those periods around earnings, do these ETFs just hibernate and have $0 of assets? Or are people doing weird arb trades between the hourly, daily and weekly ETFs? Every possible space on the map filled, just because it can be.

I mentioned a few weeks ago a reader’s pushback on my theory that every trade eventually becomes an ETF: Maybe that has been true so far, but in the not-too-distant future, retail brokerages’ agentic artificial intelligence tools will take over that job. Instead of saying “hmm I’d like to get 200% of the return of Nvidia for the next 70 minutes” and buying the 2x-Nvidia-1-to-2-p.m. ETF as the closest available proxy, you type “hmm I’d like to get 200% of the return of Nvidia for the next 70 minutes” into your brokerage’s chatbot and it goes and takes out a margin loan for you and trades the stock and gives you precisely the trade you want, for your own obscure reasons. It is fun to joke about an infinite-dimensional matrix of possible trades, with an ETF corresponding to each trade, but there is not actually an infinite number of ETFs. There is not literally an ETF for every possible trade idea you could think of; you can easily think of a trade idea that nobody has ever thought of before. If you want precisely that trade, you’ll need someone to build it specifically for you. In the past (present) that was (is) hard, and ETFs were (are) a reasonable substitute for the most popular, like, few thousand potential trades. In the future it will be easier and everyone will just get exactly the weird thing they want.

There are a lot of ETFs, though.

Okay here’s the basis trade:

1. The US government borrows a lot of money for long terms. It does this by selling 10-year or 30-year Treasury bonds: People give the government cash, and the government gives them the bonds. (The current administration is doing a lot more short-term borrowing, and relatively less long-term borrowing, but even so there are a lot of long-term Treasury bonds, and the interest rate that it pays on those bonds is of keen interest to the Treasury.)
2. Some investors — banks, insurance companies, pension funds, whatever — buy those long-term Treasury bonds, because they want to park their cash in long-term low-risk fixed-income investments. The government wants to borrow long-term, investors want to lend long-term, and everything works out.
3. Many investors, though, want something slightly different: They want the long-term interest-rate exposure of owning Treasury bonds, but they don’t want to park their cash in Treasury bonds. They want to invest their cash in corporate bonds, which pay higher interest rates, but they also want to own some Treasuries, to give them the exposure they want to long-term interest rates. [3]  But they’ve already invested their cash in the corporate bonds, so there’s no cash left over for the Treasuries. They want to own Treasuries without paying for them.
4. There is a way to own Treasuries without paying for them: Treasury futures. A Treasury futures contract is an unfunded bet on Treasury-bond interest rates: If you own a futures contract on $1 million of 10-year Treasuries, that’s like owning $1 million of 10-year Treasuries, but you don’t put up $1 million of cash. (You put up about $18,750 of cash, plus you pay more when Treasury prices go down and get money back when they go up.)
5. The government does not, however, sell Treasury futures. The government sells Treasury bonds. Unlike Treasury bonds, Treasury futures do not exist in nature.
6. That’s okay though. Big hedge funds are in the business of giving the market what it wants. A hedge fund will (1) borrow money (from money-market funds or other short-term cash investors in the repo market), (2) use the money to buy Treasury bonds and (3) sell you Treasury futures. The amount it receives on the Treasury bonds is slightly greater than the amount it pays out (to you on the futures and to its repo lenders), so it makes a bit of money by providing this plumbing service.

The hedge fund can borrow almost all of the money it uses to buy the Treasuries, so the trade is very leveraged; the hedge fund might buy $50 of Treasury bonds and sell $50 of futures for every $1 of capital that it puts up. Very leveraged trades make people nervous, and in fact this trade has occasionally blown up, so people get nervous about it.

What to do about it? We talked last year about a proposal from Anil Kashyap, Jeremy Stein, Jonathan Wallen and Joshua Younger, in which the Federal Reserve would be ready to take over basis trades from hedge funds if the basis trade blew up. (Instead of lending to banks against Treasury bonds, the idea was that the Fed would buy the Treasury bonds and sell the futures, taking over the whole position from hedge funds.)

Today Steven Williams proposes a fun simpler solution, with a full version on the web here and an FT Alphaville column here. The solution is: The Treasury should do the basis trade. Like, instead of selling Treasury bonds, [4] the US Treasury should sell Treasury futures. Williams calls it a “Deferred Settlement Auction”: “An investor competes for an allocation in the standard [Treasury] auction [of long-term bonds], wins at the clearing price, and takes delivery one month later rather than on the standard settlement date.” That’s a one-month future. And then the “DSA programme should be built around a seamless monthly roll,” so that each month investors extend delivery by one more month.

The investor has an unfunded bet, against the Treasury, on Treasury interest rates: The investor hasn’t put up (much) cash, but it has economic exposure to a long-term Treasury bond. If interest rates go up (bond prices go down), the investor has to pay Treasury some money at the next monthly roll; if rates go down (bond prices go up), Treasury has to pay the investor some money at the next roll. As a matter of interest-rate risk, the investor owns a long-term bond and Treasury has sold a long-term bond.

As a matter of cash, though, the investor hasn’t put up any cash and Treasury hasn’t gotten any. But that’s fine. In the normal basis trade, a hedge fund borrows money from money-market funds and other short-term cash investors, in the repo market, to buy the Treasury bonds backing the futures. In Williams’s trade, the Treasury borrows money from money-market funds and other short-term cash investors, in the Treasury bill market:

Because deferred settlement delays the receipt of auction proceeds, the Treasury must bridge this cash flow timing gap. It could do so by adjusting its Treasury bill issuance. Money market lenders who previously funded the basis trade naturally absorb this new supply. They simply shift their capital in a one-to-one substitution from repo to risk-free Treasury bills.

Like, in theory, Treasury could fund itself 100% by issuing short-term bills, but that would create a lot of interest-rate risk, and in practice Treasury likes to have a lot of long-term bonds in its funding mix. But Treasury could fund itself 100% by issuing short-term bills, and then get to its optimal duration mix by selling futures. If you issue $100 of bills and also sell $100 of 30-year futures, then that’s basically like selling $100 of 30-year bonds.

“Unlike Treasury bonds,” I wrote above, “Treasury futures do not exist in nature,” but that is just a contingent fact. If Treasury sold Treasury futures, then asset managers could buy Treasury futures from Treasury. They wouldn’t have to buy them from hedge funds. Which has two arguable advantages. First, right now, hedge funds make money by doing this trade, by manufacturing futures out of Treasury bonds. In Williams’s proposal, the Treasury would get that money instead: “Rather than flowing to the intermediary chain as compensation for a friction-heavy service, [the futures premium] would be partly internalised by the US government at near zero cost.” 

Second, I … guess this trade is less crisis-prone than the regular basis trade? Like, in the regular basis trade, if futures prices go up while bond prices go down, hedge funds blow up. In this trade, if that happens, the US government, I mean, probably doesn’t blow up? Though I suppose if this trade does blow up it’s much, much, much worse than a few hedge funds blowing up.

We talked a few months ago about a case brought by the US Department of Justice and 17 state attorneys general against a handful of big egg producers, accusing them of manipulating the benchmark price of eggs. The defendants settled the case by agreeing to (1) knock it off, (2) pay $3.3 million in penalties to the states and (3) deliver 53 million free eggs. I am sure this is not the first lawsuit that has ever been settled in eggs, but it’s the first I’ve heard of, and also surely the largest settlement ever paid in free eggs? I hope? Please email me with counterexamples. Anyway here’s a video posted by North Carolina Attorney General Jeff Jackson that shows him receiving some of the eggs. They’re for food banks, I mean. He didn’t eat them. That would be funnier.

Bessent Dares Traders to Bet Against Yen: ‘I Am the House Now.’ Millennium Joining Rokos With Plans for First Greek Outpost. How Greece is wooing hedge funds. States That Gave Data Centers Billions in Tax Breaks Are Now Ripping Up the Deals. Anthropic Researcher Quits Over ‘Out-of-Control’ AI Fears. “In past weeks I received almost daily WeChat messages and emails offering different DeepSeek [special-purpose vehicle] opportunities — and all are charging insane amounts of fees.” Where Does Spirit Go When It Dies? U.S. Gets Minority Stakes in Some Quantum Computing Companies in $300 Million CHIPS Act Funding Deal. India Lifts Ban on JPMorgan Unit in Index Manipulation Charge. Silver Lake to merge French software groups Cegid and Silae in €10bn deal. Husband and Wife Plan Sale of $37 Billion Credit Firm Palmer Square. AI Boom Wrecks Trump’s Plan for ‘Made-in-America’ Bitcoin Mining.

If you'd like to get Money Stuff in handy email form, right in your inbox, please subscribe at this link. Or you can subscribe to Money Stuff and other great Bloomberg newsletters here. Thanks!