Money Stuff: AI Builds Some Pyramids

The situation with artificial intelligence companies is:

1. The founders of AI companies place a high value on keeping control of their companies, not surrendering control to outside investors or boards of directors. Part of the reason for this is that all tech founders love control, but part of it is that AI founders believe they are creating a uniquely powerful and dangerous technology and don’t want it to fall into the wrong (BlackRock’s) hands. As Elon Musk put it: “I just don’t feel comfortable building a robot army here, and then being ousted because of some asinine recommendations from ISS and Glass Lewis, who have no freaking clue.”
2. AI companies need to raise a kajillion dollars of outside equity to build their robot armies.

This is a problem. The AI founders want to keep control of their companies, but traditionally control goes with stock ownership, and they need to raise so much money that they can’t really keep a majority of their companies’ shares.

There are technologies to address this problem. With SpaceX’s initial public offering last week, Elon Musk advanced the state of the art of this technology:

• He gave himself super-voting stock so that he can keep control even if he sells a majority of the company’s economic ownership;
• He separately gave himself the ability to appoint a majority of the board of directors; and
• He made it essentially impossible for shareholders to sue him if they don’t like his decisions.

Nobody has ever gone quite this far before, but on the other hand nobody has ever taken a giant existential-threat-to-humanity AI company public before. [1] I wonder a bit whether OpenAI and Anthropic, which are also thinking about going public, will take some pointers from Musk. Musk’s approach caused a certain amount of complaining, but it did not stop him from raising $75 billion from enthusiastic investors at a giant valuation, so I guess it works.

There is, however, an older technology. Back in the day (the 1920s) it was called “pyramiding,” and it worked like this:

1. You have a company, Company A. It’s worth $100. You own 100% of it. It needs to raise money.
2. You sell 30% of Company A to outside shareholders for $30.
3. Now you have raised $30, and you own 70% of the voting stock of the company.
4. You need more money.
5. You plop your 70% stake into a new holding company, Company B. You own 100% of Company B, and Company B’s only asset is 70% of the voting stock of Company A ($70 worth).
6. You sell 30% of Company B to outside shareholders for $21. 
7. Now you have raised $51, and you own 70% of Company B, which owns 70% of Company A. You own a 49% economic stake in the operating company (Company A). But Company B owns a majority of the votes in Company A, and you own a majority of the votes in Company B, so you still have control over Company A.
8. You need more money, you plop your Company B shares into Company C, and you sell 30% of Company C for $14.70. Now you have $65.70, but you still control 70% of Company C, which controls 70% of Company B, which controls 70% of Company A. You still have control despite only owning about one-third of the stock economically.
9. Company D, etc.

You can see why this technology has mostly been replaced by dual-class supervoting stock. This is sort of messy and irregular. It also requires raising money at lots of different holding companies, rather than at one main company, which creates a lot of paperwork and is bad for liquidity and for storytelling. 

Still, maybe it will make a comeback. The Information reports:

Chinese AI lab DeepSeek has closed its first funding round that raised more than 50 billion yuan ($7.4 billion) under an unusual deal structure, according to two people with direct knowledge of the matter.

The funding round, which values the large-language model developer at more than $50 billion, requires investors to put their capital into a limited partnership managed by DeepSeek CEO Liang Wenfeng, instead of DeepSeek itself, as a way for Liang to ensure he retains absolute control of the company, the people added. It also imposes a five-year lockup on all investors’ shares, during which they can’t sell their stakes, the people said. …

Under this structure, other external investors will not have voting rights at DeepSeek, though they will have access to privileged financial information and priority rights to invest in potential future fundings, typical in venture deals, the people said. …

Liang is a firm believer that AI should be open-source so the technology can benefit all, according to people close to him. To that end, Liang has been highly selective in choosing DeepSeek’s financial backers.

The goals of an AI founder are (1) to raise as much money from investors as possible and (2) to give those investors as little control as possible. In theory it might be hard to reconcile those goals with each other; in practice, so far, it’s pretty easy.

It would be fun to construct a prediction markets index. Collect the 100 biggest event contracts on Kalshi and Polymarket, weight them equally or by open interest or whatever, average their prices, and you get an index that tracks the overall level of prediction markets. Prediction-market investors could use the index to benchmark their own performance, and commentators could use it as a shorthand for how the market is doing. “Events have become more likely,” they could say, or “probabilities have gone down on profit-taking,” or “going into this week’s Fed meeting, traders expect things to happen less.” Just passive exposure to the market portfolio of probabilities.

Would the index be range-bound between 0 and 1, just an average of current contract prices? Or would there be some sort of mechanical roll process, where individual bets in the index would pay off 0 or 1 and then be rolled into new bets? In the latter case, would the overall level of the index go up over time? No, right? Or I guess it depends on your indexing rules? If the index was mostly high-probability bets (“bonds,” as they say), perhaps it would earn some expected return? (Those bets are arguably a sort of insurance, so you might expect to earn some premium.) Would the index be able to hold both “Yes” bets and “No” bets? It’s possible that an index of “No” bets could have positive expected returns. (Not gambling advice.) Fascinating questions of product design.

This is pretty stupid, but on the other hand I bet there will be an absolutely straight-faced launch of a prediction markets index in like two weeks. I really like the idea of television commentators saying “prediction markets were up today.”

There is not really a news hook for this; this is idle speculation and jokes and also I want to publish it now so that in two weeks I can say “I told you so.” But it was vaguely inspired by this story:

A single trader on Polymarket lost nearly $1 million when Cabo Verde fought Spain to a stunning draw on Monday in one of the most unlikely outcomes in recent World Cup history, powered by a 40-year-old goalkeeper who left the pitch in tears.

The losing bettor had wagered on what oddsmakers saw as a nearly certain Spanish win, according to trading records from Polymarket, one of the largest prediction market exchanges. …

The platform’s public trading records show that Spain’s odds of winning stood at an average of around 92% when the user — who went by the screen name “betoor619” — placed the massive unsuccessful wager. Records indicate the user opened the account last October, and had never won or lost more than $9,000 on an individual event before the Cabo Verde game.

Some traders place big bets on highly-likely outcomes in the hopes of capturing small profits on relatively low-risk bets. If Spain had won, betoor619 would have only made around $85,000 on an original bet of almost $1.1 million.

Yeah, I mean, people lose sports bets. The innovation of prediction markets is that now there are public regulated betting markets, and financial media can report on losing sports bets.

I will say, though, that “bet $1 million on an extreme favorite on Polymarket every month and, if you win, do it again” has almost the shape of a good financial product? Like, that pays a high bond-like return most of the time; it has some risk, but that risk is uncorrelated to other financial instruments. (It goes to zero when Spain draws with Cabo Verde.) Build an index, build a passive exchange-traded fund to track it, what is everyone waiting for? [2]

People keep sending me the Strategy thing. Strategy Inc. (formerly MicroStrategy) is a company that buys Bitcoins. Each Monday, it puts out a statement saying how many Bitcoins it bought the previous week. For a long time, Strategy never sold Bitcoins, but that became less tenable over time, and in the week ending May 31 it sold 32 Bitcoins.

Polymarket, the prediction market, had listed a contract on “MicroStrategy sells any Bitcoin by May 31, 2026?” Shortly after Strategy announced that it had sold Bitcoins in the week ended May 31, that contract resolved to “No.”

Why? Well, I mean, the short answer is that prediction markets are not a philosophical mechanism for achieving truth; they are a convenient consumer gambling platform. If you want to establish the truth of some question like “when did Strategy first sell Bitcoin,” or “when did US troops first enter Iran” for that matter, you might reach a tentative conclusion and be willing to revisit it whenever new facts come to light. Perhaps in 20 years newly declassified information will reveal that actually US special operations troops conducted secret ground missions in Iran in mid-March 2026. But that doesn’t work for Polymarket: Polymarket took some bets on “US forces enter Iran by March 31,” and it needed to resolve those bets in a timely fashion, and it hasn’t heard anything about any secret ground missions, so those bets resolved to “No.” If in 20 years you do archival research and find out that Polymarket’s resolution was wrong, and you go to Polymarket to inform them, they will stare at you and say “you are completely missing the point here.”

Of course, with Strategy, anyone who pays the slightest amount of attention knows that each Monday Strategy announces its Bitcoin purchases (or sales!) for the previous week, so the extremely obvious resolution mechanism would have been to wait for the Monday announcement. As it happens, June 1 was a Monday, so Polymarket could have accurately resolved “MicroStrategy sells any Bitcoin by May 31, 2026” by waiting one day. But it didn’t, out of, I don’t know, cussedness, or lazy consumer gambling platform design. The Wall Street Journal reports:

On June 1, the bitcoin-hoarding company Strategy disclosed that it had sold the digital currency during the previous week, sending shock waves through the crypto markets. [Hunter] Guo, a 20-year-old student at Kings College London, saw an opportunity to cash in. With a quirk of Polymarket’s rules keeping trading open, he bought thousands of betting contracts which said they would pay off if Strategy sold any bitcoin by May 31. He hoped to buy a Porsche with his winnings.

Then Polymarket released a note to provide “additional context.” The platform explained that Strategy’s disclosure didn’t count because it had come out too late. For Guo to win his bet, news of the bitcoin sale needed to have come out by 11:59 p.m. ET the previous night. The value of Guo’s contracts was wiped out within seconds. …

Other traders—including Polymarket veterans steeped in precedents set by previous disputes—argue the platform made the right call. They note that in the past, Polymarket has declined to consider information released after the cutoff date of a contract. If fresh information that trickles in late can be used to resolve bets, they might never be fully settled, these traders argue. ...

As a relative newcomer to Polymarket, Guo said he didn’t know about its arcane procedures for resolving disputes. All he knew was the plain language of how Polymarket described the Strategy bets on its app and website. 

“The rule asked whether Strategy sold bitcoin by May 31. Not whether the world discovered it by May 31,” he wrote in one tweet. “People arguing otherwise should read the rule again.”

Right, yes, the naive view is “this contract should resolve according to what actually happened,” whereas the sophisticated view is “we are running a market here and what matters is quick final resolution, not absolute truth.”

For a while, appraisal arbitrage had the reputation of offering asymmetric upside. In many places, when a public company is acquired for cash, shareholders who don’t like the deal price can go to court and ask for more money. This is called “appraisal”: The court will take evidence, decide how much the company was worth, and then order the acquirer to pay the disgruntled shareholders the appraised fair value of the company. Traditionally, the disgruntled shareholders will argue that the acquirer got a sweetheart deal and underpaid for the company, while the acquirer will argue that actually the deal was bitterly negotiated and the price was fair. And so traditionally the court would either (1) rule in favor of the shareholders and give them more than the deal price or (2) rule against the shareholders and give them the deal price. Either they’d get extra money, or they wouldn’t. But they’d never lose money: The worst realistic outcome was just getting the deal price anyway.

And so an appraisal arbitrage strategy of buying stock of merger targets shortly before the deal closed (at roughly the deal price) and suing for appraisal [3]  was pretty good: You made money if you won, and didn’t lose money if you lost. In Delaware, where most US public company mergers happen, things were even better, because the acquirer had to pay you interest, at an unusually high rate, until the lawsuit was resolved. If you lost the lawsuit, you got the deal price plus a nice fixed-income return; if you won you did even better. “It was a good steady fixed-income strategy with some big equity upside,” I once wrote.

This pleasant situation came to an end in roughly 2018, when Delaware courts, and acquirers, decided that sometimes the deal price was too high. If the target company is diligent and well-advised, it will negotiate a price for a merger that is higher than its fair value, by making the acquirer pay for synergies (value that is created by the merger, and is not part of the target’s standalone fair value) and/or by using a competitive auction to pressure the acquirer into overpaying. [4] Courts became more receptive to the efficient-markets notion that the pre-merger trading price of the target represented its fair value, and that appraisal arbitrageurs shouldn’t get any of the premium paid in a fairly negotiated merger. What used to be a trade with a lot of upside and little downside became much riskier, and so less popular, though you still see it sometimes.

That was a fairly US-centered story, but last week Bloomberg’s Sabrina Willmer and Yiqin Shen reported that the Cayman Islands are going through a similar evolution:

Dozens of hedge funds face having to repay hundreds of millions of dollars combined to a Chinese recruiting firm if a ruling is upheld on appeal that the company was overpriced.

The high-stakes legal fight over 51job Inc. could force the group, including Oasis Management Co., Pentwater Capital Management and Alpine Global Management, to return about $525 million that the company paid them in advance for their shares, according to people familiar with the matter.

The group of money managers, which also included Millennium Management and Man Group Plc, argued that the $61 per share paid to investors wasn’t sufficient after a private equity-led group took 51job private in 2022, a move that backfired after a judge ruled against them. …

[Appraisal] cases have largely succeeded in the Cayman Islands, with judges finding firms were worth more than the deal price and letting objecting money managers pocket the difference. But in a rare twist, Cayman Islands Justice David Doyle ruled in November that the Chinese firm was actually only worth $31.11 a share, in a particularly bad outcome for the hedge funds.

The hedge funds are appealing that ruling, in a case that’s set to be heard in July, the people said. ...

“This is new territory for these investors facing the downside risk of repayment from this sort of take-private action,” said Geoffrey Derrick, an asset recovery attorney at Boies Schiller Flexner. “It is going to significantly deter future appraisal arbitrage in the Cayman” Islands if the appeals court upholds the ruling.

The judge used the pre-merger trading price of the stock as the fair value of the company, so the arbitrageurs will have to pay back the merger premium. It is very pleasant to do trades with no downside risk, but eventually most of those trades go away.

In April, SpaceX struck a deal to buy Cursor, the AI coding company, for $60 billion. More or less. SpaceX couldn’t really buy Cursor in April: There was too much going on, it was in the middle of work on its initial public offering, and smushing another company into SpaceX would have delayed the IPO. So SpaceX and Cursor had an agreement to agree: SpaceX bought an option to buy Cursor, to be exercised shortly after the IPO, and everyone assumed that it would get exercised. Today it happened:

Early Tuesday, SpaceX formally agreed to buy Cursor in a deal that will entitle the startup’s investors to SpaceX stock. In doing so, Elon Musk is signaling his desire for SpaceX’s xAI to rapidly rebuild and catch up to rivals including Anthropic PBC and OpenAI that have capitalized on demand for artificial intelligence-powered coding tools in a way that his AI business hasn’t. …

As part of its acquisition announcement, SpaceX said it’s been working with Cursor to train a new AI model and aims to collaborate on developing more cutting-edge artificial intelligence capabilities.

The price of the deal was $60 billion, payable in SpaceX stock. In April, $60 billion was perhaps 3.3% of SpaceX’s market value. At noon today, $60 billion is perhaps 2.2% of SpaceX’s value: SpaceX went public at about a $1.77  trillion valuation, and has been ripping up from there; it was around $2.75 trillion at noon. If Cursor had gotten $60 billion of SpaceX stock in April, it would be worth something like $90 billion today. But it didn’t. It’s not even getting $60 billion of SpaceX stock now. It’s getting $60 billion of SpaceX stock sometime later this summer. [5] If Cursor’s founders and investors hold onto their SpaceX stock, they will participate in any future upside in SpaceX after this summer. But they missed out on the upside in the IPO: They agreed to sell themselves to SpaceX for SpaceX stock, while SpaceX was private, but they agreed to receive that stock after SpaceX was public and at the public valuation. [6]

This is perhaps another example of “private markets are the new public markets.” Traditionally, if you sell your company to a big private startup for stock, you expect to get the equity upside of that startup. By selling to a startup, you are betting on that startup. If you are part of the story of value creation in the initial public offering, you expect to earn the IPO pop. I mentioned in April that, back when Facebook was a private company, Mark Zuckerberg bought Instagram by (correctly) persuading Instagram’s founder that $1 billion of Facebook stock was really worth $2 billion. That was the sort of thing you could do in 2012: A big tech startup could buy a smaller tech startup for stock, by selling the smaller startup on the upside potential in the bigger startup’s stock.

Whereas when a public company buys another public company for stock, the value of the stock is the value of the stock. There’s a market price. “Really $60 billion dollars of our stock is worth $90 billion” is incoherent. SpaceX didn’t buy Cursor by offering it an equity stake in a promising private startup. It bought Cursor by offering it $60 billion at market prices.

“Excellence” has been redefined to mean “slop,” lol:

A KPMG report on how AI is being used by businesses across the world exaggerated adoption of the technology with bogus case studies that appear to have been based on AI hallucinations.

The October report, “Redefining excellence in the age of agentic AI”, made numerous false claims about the use of AI by organisations including the Swiss bank UBS, the UK’s National Health Service and the public transit groups Swiss Federal Railways and Transport for London. ...

Edward Tian, GPTZero chief executive, said error-riddled publications by the Big Four “poison the well of information”. KPMG’s findings have already been referenced by multiple industry publications as well as a big Czech newspaper, according to GPTZero’s research.

Big consultancies such as KPMG and EY are viewed as highly credible, so their reliance on false information “increases the risk of second-hand hallucinations”, Tian said.

It’s a fascinating recursive vulnerability: Modern AI relies on trusted human sources to understand the world, but humans increasingly rely on AI to understand the world. If the trusted human sources take whatever AI says at face value, then whatever AI says will just become fact. Who can dispute what both Claude and KPMG say is true? 

Elsewhere: “HR must manage AI bots as well as humans, says Accenture executive.” 

Justice Department Decision to Allow Paramount Deal Surprised Staff Investigators. Yum to Offload Struggling Pizza Hut Chain for $2.7 Billion. Hotel Owners Are Rebelling Against Marriott’s Loyalty Program. Bondholders Forced to Reckon With $4 Billion Tax Threat in Bitter Credit Feud. How the AI boom turned a buyout deal into one of history’s most lucrative. JPMorgan plans Chase expansion into Europe’s largest markets. Taiwan Frees Life Insurers to Deploy Capital in AI Projects. Private equity bosses warn of AI threat to bets on law and accountancy. Judge Dismisses Elon Musk’s xAI Trade-Secret Lawsuit Against OpenAI. Central banks repatriate gold as global insecurity rises. Fujitsu chair quits after claims of ‘woman-related’ improper conduct.  Both Parents Now Work Full Time in Most US Families. The Party Game Where Silicon Valley’s Elite Pretend to Kill Each Other. More Americans Dip Into Their 401(k) Savings for Emergency Funds. In New Luxury Kitchens, Everything Is Hidden — Even the Sink. The Infuriating Rise of the $8 Ice Cream Cone.

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!