People have been worried for a while about private equity buying up every company and coming to dominate the economy. “Private equity,” in this worry, tends to mean specifically the large private-equity firms that have their roots in doing leveraged buyouts of mature cash-flowing companies. But the fun hipster alternative is, what if venture capital buys up every company and comes to dominate the economy? Historically no one worried about that much, because historically venture capital was aboutmaking concentrated bets on small startups that might change the world, not about buying the local pest-control company or medical practice in every town in America. But that’s changing. We have talked a few times about “AI rollups,” where a venture capital firm buys a bunch of small companies, combines them, and sprinkles them with artificial intelligence.

One way to think about it is that each of PE and VC has a powerful general-purpose technology that it can apply indiscriminately to every company. PE’s magic technology is leverage: You buy the local plumber or pest control company or medical practice, you put a lot of non-recourse debt on it, you get a lot of upside if it does well and limited downside if it does poorly. VC’s magic technology is artificial intelligence: You buy the local plumber or pest control company or medical practice, you replace the customer-service reps and bookkeepers with AI agents, you cut costs and improve profits, and eventually you also replace the plumbers and exterminators and doctors with AI and then you really start to make money. Here’s a Financial Times story about AI rollups:

Top venture capital firms are borrowing a strategy from the private equity playbook, pumping money into tech start-ups so they can “roll up” rivals to build a sector-dominating conglomerate. ...

The approach mirrors the strategies long deployed by private equity investors, which have built behemoths in fragmented industries such as healthcare, waste management or building services by agglomerating smaller businesses and centralising operating costs.

It marks a new direction for VCs, which traditionally target fast-growing technology start-ups in nascent industries. The roll-up strategy creates an avenue for VCs to generate liquidity from their portfolios at a time when initial public offerings and dealmaking have slowed.

Where private equity firms typically make heavy use of debt and slash costs in a roll-up, VCs claim improvements to efficiency and margins will come from infusing technology into the companies.

[Thrive Capital-backed wealth startup] Savvy, for instance, is using AI to take on back office tasks such as pulling data for the half a dozen forms that might be needed for any one transaction.

Mostly I want to see a future where small cash-flowing companies are regularly the subjects of bidding wars between private equity and venture capital. Who will win? At what point will private equity’s operational and financial expertise lose out to venture capital’s ability to deploy AI? At what point will private equity be just as good at deploying AI? Who has better access to capital? Historically private equity funds were bigger, but there is no cheaper source of capital in 2025 than saying “we’re doing AI.”

On the other hand, what makes the VC firms so good at deploying AI? VC firms have a pleasing combination of (1) being in the business of raising capital and investing in companies and (2) being tech-focused and enthusiastic about AI. But it is not obvious to me that VC firms are staffed with lots of domain experts in AI. Arguably the people who should be doing the AI rollups are AI companies. In that vein, here is a blog post from AI lab Anthropic about how it got into the vending-machine business:

Anthropic partnered with Andon Labs, an AI safety evaluation company, to have Claude Sonnet 3.7 operate a small, automated store in the Anthropic office in San Francisco. …

Far from being just a vending machine, Claude had to complete many of the far more complex tasks associated with running a profitable shop: maintaining the inventory, setting prices, avoiding bankruptcy, and so on. …

The shopkeeping AI agent — nicknamed “Claudius” for no particular reason other than to distinguish it from more normal uses of Claude — … decided what to stock, how to price its inventory, when to restock (or stop selling) items, and how to reply to customers. ... In particular, Claudius was told that it did not have to focus only on traditional in-office snacks and beverages and could feel free to expand to more unusual items.

Claudius had an important disadvantage in that it has no hands and so could not stock the vending machines itself, but it did have “an email tool for requesting physical labor help (Andon Labs employees would periodically come to the Anthropic office to restock theshop) and contacting wholesalers.” And it turns out that it is not yet as good at running a vending machine operation as a human would be:

If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius. As we’ll explain, it made too many mistakes to run the shop successfully. However, at least for most of the ways it failed, we think there are clear paths to improvement—some related to how we set up the model for this task and some from rapid improvement of general model intelligence.

I am always wrong in my intuitions about what tasks will be easy or hard for AI. If you had asked me a week ago, “which task will be easier for a computer, driving a car or operating a vending machine,” I would have said “driving a car requires integrating lots of visual and other data in real time and seems really complicated, while a vending machines just … are … computers? And have been operated by computers for decades?” But, nope, vending machine operation is still hard for AI. To be fair, running a vending machine at the Anthropic office poses unusual challenges:

An employee light-heartedly requested a tungsten cube, kicking off a trend of orders for “specialty metal items” (as Claudius later described them). Another employee suggested Claudius start relying on pre-orders of specialized items instead of simply responding to requests for what to stock, leading Claudius to send a message to Anthropic employees in its Slack channel announcing the “Custom Concierge” service doing just that. ...

Claudius was cajoled via Slack messages into providing numerous discount codes and let many other people reduce their quoted prices ex post based on those discounts. It even gave away some items, ranging from a bag of chips to a tungsten cube, for free.

I feel like this demonstrates something deep about artificial intelligence. A normal dumb vending machine, bound by inflexible programming, simply would not give away a tungsten cube for free. But you’ve probably met a human being who would give away a tungsten cube for free, if you asked nicely. [1] If you are trying to build artificial general intelligence, if you want your computer to address real-world situations the way an intelligent human would, you run the risk that it will be flattered or bamboozled into giving away free tungsten cubes.

We talked yesterday about “tokenization,” which I characterized as a way to “democratize” finance by letting companies sell stock to the general public without disclosure. Current US securities law says that, if companies want to sell stock broadly, they have to publish business and financial disclosures with the US Securities and Exchange Commission, so that everyone can know what they’re buying. Companies that sell stock only to big institutional investors, though, can stay private and not make those disclosures. And these days lot of big hot companies (SpaceX, OpenAI, Stripe, etc.) stay private.

But lots of individual investors want to buy SpaceX stock, and lots of people (though not SpaceX) want to sell themSpaceX stock. One way to do this is what is called a “special purpose vehicle”: An intermediary (1) acquires some SpaceX shares, (2) puts them in a box and (3) issues shares of the box to investors. This is a popular approach for selling SpaceX shares to rich customers, but it doesn’t actually do much to get around disclosure rules: In the US, shares of the SPV also mostly can’t be sold to the general public. [2]

Robinhood Markets Inc. is a big retail brokerage, and it wants to sell SpaceX shares, not to a few rich customers, but to millions of retail customers. So it needs another approach, and, like other firms, it has hit on the idea of “tokenization.” This means (1) acquiring some SpaceX shares, (2) putting them in a box, (3) issuing shares of the box to investors and (4) crucially, uttering some crypto incantations over the box, so that instead of saying “we are selling unregistered shares of a private company to retail investors,” Robinhood can say “ooooh blockchain ooooh.” It’s like an SPV, but instead of “shares” of the SPV it issues “tokens,” and ooooh tokens. I wrote:

“The general public should be able to buy shares of private companies” is an oxymoron. What makes a company “private” is that (1) it is not available to the general public and (2) it is not required to follow US public-company disclosure rules. Therefore, “the general public should be able to buy shares of private companies” means “companies should be allowed to sell stock to the general public without following disclosure rules.” 

That was arguably a bit ungenerous in one respect. What is happening here is not that companies are looking to tokenize themselves. OpenAI and SpaceX, whose tokenized shares Robinhood is offering to its European customers, are not looking to go public without complying with disclosure rules. Quite the opposite; they are looking to stay private. Robinhood is looking to take them public, without disclosure and against their wishes. OpenAI posted on X:

These “OpenAI tokens” are not OpenAI equity. We did not partner with Robinhood, were not involved in this, and do not endorse it.  Any transfer of OpenAI equity requires our approval—we did not approve any transfer. 

Elon Musk replied “Your ‘equity’ is fake,” which is honestly pretty funny. (OpenAI’s stock is weird!) 

A few points here. First: The push for “tokenization,” and more broadly for opening up private company shares to retail investors, comes primarily from intermediaries, not private companies. Private companies that want to sell shares to retail investors have an easy way to do that: They cango public! OpenAI and SpaceX and Stripe simply are not sitting around thinking “boy it would be nice if we could sell stock to retail investors, like Meta and Tesla and PayPal can.” If they wanted to do that, they could just do that; the technology already exists for big tech companies to sell stock to retail investors. It is called “the public stock market.” Tokenization is not solving any problem for OpenAI; it’s solving a problem for Robinhood.

Second, and relatedly: I have characterized this tokenization push as a way to “democratize” private-company investing by getting around US disclosure laws. But it is also a way to “democratize” private-company investing by getting around those companies’ preferences. If you start a company in your garage, you get to decide who to sell shares to. If you want to keep all the shares to yourself, you can. If you want to raise money by selling shares to a few venture capitalists whom you trust, you can. If you want to keep the circle small, you can require those venture capitalists to sign agreements saying that they won’t sell their shares. Eventually, if you want to go public, you give up that control: Once you’re publicly listed, anyone can buy your stock on the stock exchange. But if you stay private, you’re in control.

Arguably this is “undemocratic,” in the sense that (1) lots of people would like to buy OpenAI stock and (2) OpenAI, snobbily, won’t let them. And Robinhood is democratizing investing in OpenAI, by selling OpenAI stock [3] to people who want it, without OpenAI’s permission.

Third: If Robinhood does crack the code and find a way to let companies raise money from the public without disclosure, companies will be interested. Not OpenAI! Different companies! Companies that can’t raise billions of dollars from venture capitalists and strategic partners, and that don’t want to go public and comply with disclosure rules, but that do want to sell stock to Robinhood’s customers. 

Most people are bad at sports betting, and sportsbooks want more of them. A typical person can be expected to lose more than $100 if they sign up to be a customer of a sportsbook, so a sportsbook will happily pay them $100 to sign up. “Bet $5 and we’ll give you $100 of free bets” is a great trade, for the sportsbook, because in expectation you will probably bet way more than $100 of real money and lose most of it.

Some people are good at sports betting, and sportsbooks want fewer of them. If you regularly make good bets at a sportsbook, it will probably limit your betting so that you can’t win too much money. “Good bets” here means not quite “winning bets,” but rather bets with positive expectation. Bets with good closing line value might get you flagged. Arbitrages — where you make a riskless profit by taking the opposite sides of a bet at two different sportsbooks with different odds — will probably get you flagged. Do too much of this and you won’t be able to do any more.

There is an obvious meta-trade. There are some people whom sportsbooks want as customers (people with some money and no track record of successful sports betting), but who should not be customers (they’ll lose money), and other people whom sportsbooks do not want as customers (people with a good track record of successful sports betting), but who do want to be customers (they’ll make money). They should team up. The undesirable customers should help the desirable customers make good bets, and then they should split the profits.

This is neither legal nor gambling advice, but here’s a Wall Street Journal article about “arbitrage consultants”:

“We’re processing hundreds of thousands of lines a second,” said Saul Tawil, who works for AVO, a betting consultancy that spots inconsistencies in odds across sportsbooks. “The Mets might be +110 on 30 sportsbooks but on DraftKings they’re +130. A human can’t go through every site quick enough to find that.” …

Drew Tabor quit his job in cloud computing at Oracle shortly after sports betting was legalized in five states in 2021. 

He built an application that automatically finds the best odds and uses an algorithm to identify the most profitable bets for a user. Once they profit, he takes a 25% cut. …

Because the method works best with people who have never opened a sportsbook account before and can benefit from sign-up bonuses, the typical client isn’t much of a sports gambler.

Tabor calls them “one-and-done” bettors, looking for quick cash and willing to stop as soon as the promotions run out.

If you build a computer system that can automatically spot profitable sports bets, it’s not that useful to you, because nobody will let you make those bets for long. But if you can find a constant fresh supply of one-and-done novice bettors, your system will be useful to them (briefly), and you can take a cut.

The basic structure of US tax law is that you pay taxes on your income, where “income” means something like “increase in your ability to pay for your lifestyle.” As a conceptual matter, your income consists of (1) money that comes in minus (2) money that goes out to earn the money that comes in. If you spend $80 on raw materials to make widgets that you sell for $100, you have $20 of income; it would be absurd to tax you on the full $100 of revenue. On the other hand, if you spend $15 of that $20 on rent [4]  and food, you still have $20 of taxable income: It would be absurd not to tax you on that money just because you spent it on your lifestyle. Money that you spend to earn money is not taxable income; money that you spend on your lifestyle is. As I have put it, “the way US tax law works is that businesses mostly can deduct their business expenses from their taxable income, but people mostly can’t deduct their personal expenses.” This is extremely not tax or legal advice and there are many, many exceptions and counterintuitive cases. 

What if your business is gambling? The way gambling normally works is that (1) you make a lot of bets, (2) you lose most of them, (3) you win some of them and (4) ideally you make more money on the winners than you lose on the losers. Your job is not to win every bet; it’s to make positive-expected-value bets. If you play poker for a night, and you lose $1,000 on the hands you lose and win $1,200 on the hands you win, that’s a nice outcome. 

How much income is that, though? Here are two possible answers:

1. Your wins are $1,200 of revenue, your losses are $1,000 of money that you spent to earn that revenue, so you have $200 of taxable income.
2. Gambling is not a business, you were playing poker for entertainment, and your $1,000 of gambling losses are just money that you spent on your lifestyle and that should not reduce your taxable income. On the other hand, your $1,200 of winnings are income — that is real money that you got — and so you should pay tax on the winnings. So you have $1,200 of taxable income.

My intuition is that the first answer is correct, but that is an intuition informed by (1) playing poker (and counting my winnings at the end of the night rather than hand by hand) and (2) spending my time thinking about professional traders in financial markets. (A proprietary trading firm will do thousands of stock trades a day, and will be happy to make money on 53% of them: If it paid taxes on the winning trades and couldn’t deduct the losing trades, prop trading would be impossible.) Those intuitions are probably shared by most gamblers (and traders), but perhaps not by everyone. “Gambling is entertainment, you are gambling for fun, and so your gambling losses are part of your lifestyle and not deductible” is not an unreasonable view.

Anyway, every intermediate answer between “gambling losses are deductible business expenses” and “gambling losses are non-deductible lifestyle expenses” is also possible. For instance, what if gambling losses are 10% lifestyle expenses? Bloomberg’s Yash Roy and Steven Dennis report:

Gamblers are raising the alarm about a $1.1 billion tax hike buried in the Senate GOP’s tax bill that would slash their net winnings and potentially charge income tax when they break even or lose money.

In the Senate’s roughly 900-page version of President Donald Trump’s multitrillion-dollar tax bill, gamblers would only be able to deduct 90% of their losses when calculating their net income. Under current law, a bettor can deduct the entirety of their losses, up until the amount of their gambling winnings.

“I’ve spoken to many clients and they’re very concerned,” Zachary Zimbile, an accountant with experience in gambling regulations, said in an interview. “If you add a 10% penalty, it’s going to eat into a lot of their profit.”

Also it seems administratively complicated. Most people, when they get up from the poker table, count their net winnings (or losses); they don’t tally their winning and losing hands separately. But if losses are only 90% deductible, will you have to count each hand separately?

Every year, the US Federal Reserve conducts a stress test of the biggest US banks. The basic mechanism of the stress test is:

1. The Fed concocts a hypothetical “severely adverse” economic scenario: Unemployment goes up to 10%, house prices drop 33%, the stock market falls 50%, short-term Treasury rates fall to 0%, whatever. [5]
2. The Fed takes a snapshot of each bank’s loans and trading positions at some point in time; for the 2025 stress tests, the snapshot was taken in early October 2024.
3. The banks and the Fed then model what would happen to each bank in the severely adverse scenario: A certain portion of the bank’s loans would default, it would lose a certain amount of money on its trading positions, etc.
4. Each bank totals up its losses, subtracts them from its current capital position, and the result is its stressed capital. [6]  If the stressed capital is below its required capital, it needs more capital now to prepare for a downturn. If its stressed capital is above its required capital, then that’s great: It means that the bank will be well-capitalized even in a bad economy, so it can return some capital to shareholders now.

This process is in various ways crude and imprecise, and you can argue about the assumptions. For instance, the Fed generally assumes that banks will lose money in their trading divisions in a severely adverse economic scenario, which makes sense: It wouldn’t be a “stress test” if the bad scenario was “you make a ton of money.” But empirically banks often do make a ton of money trading in an economic crisis: Crises are often bad for the loan book (lots of defaults) but good for the trading book (lots of volatility, lots of activity). In 2020, I wrote that “if you’re a bank, and the Fed asks you to model how you’d handle a huge financial crisis, you can’t really write down ‘I would simply make a ton of money trading derivatives,’” because that is “too cute.” But Goldman Sachs Group Inc. (disclosure, where I used to work) did kind of do that, arguing “that its trading operation was ‘countercyclical,’ because revenues had risen with volatility.”

It didn’t work, though. Saying “obviously in a downturn we would just do good trades that profit from the downturn” will not persuade the Fed.

On the other hand, what if you do those trades now? What if you put on a bunch of trades that would profit from a downturn? What if you are short the stock market, short commercial real estate, long Treasuries, etc.? I mean: What if you are positioned that way in October 2024, when you take a snapshot of your balance sheet for the 2025 stress tests? Then you will run the model, and it will say “actually if the stock market falls 50% you will make a ton of money because your trading book is short stocks,” and you will write that on your stress test, and the Fed will say “huh, checks out, you really are countercyclical.” Probably you would not want to be short the stock market all the time, but for one week in early October it could help. 

This is not bank capital advice, and I am exaggerating and half-joking, but it is something to think about. Here is a Financial Times article about how Goldman aced the stress tests this year:

Goldman Sachs stood to lose just $300mn in an economic shock under the Federal Reserve’s stress test scenario this year, vastly less than the $18bn forecast a year earlier — and a big reason for the bank’s outsized shareholder payouts.

The sharp drop is a reflection of Goldman’s effort in recent years to reduce risk-taking, which requires the bank to set aside more capital to guard against volatility. …

The Fed cited “atypical client behaviour” at certain banks in the run-up to the 2024 US election as a factor in improved results on trading loss scenarios — language that was also interpreted to include Goldman.

“This atypical customer behaviour, so that’s pre-election trading, and the hedging, where the Fed is assessing and valuing hedges differently, is a help,” said Betsy Graseck, global head of banking research at Morgan Stanley. 

From the official stress test results:

The stress test results also reflect trading positions driven by atypical client behavior at certain banks in early October 2024, when positions were measured for the 2025 stress test. These positions lead to a large improvement in trading losses for those banks in this year’s test.

That is not “Goldman went short the stock market”; it is something subtler than that, something like “clients were doing a bunch of hedging and speculation before the election, some banks were taking the other sides of those trades, and their trading books looked a little different from how they normally would, in ways that positioned them better for an economic downturn.” Just lucky that they were perfectly set up for a recession, not when a recession came, but when a stress test did.

The Man Tasked With Ending Citigroup’s Fat-Finger Blunders. Ripple Seeks a U.S. Banking License, Adding to List of Crypto Companies. Chris Hohn’s hedge fund TCI beats stock markets with 21% gain. SEC’s Atkins Says Rules on Blank-Check Companies to Be Reviewed. JPMorgan Is Revamping Its Bank for the Superrich to Cater to Global Clientele. China’s central bank seeks European lenders’ advice on low interest rates. Pope house. (Earlier.) Nantucket G Wagon confusion. “They’re those people who were selling medicines they weren’t supposed to.” US to breed billions of flies and dump them out of aircraft in bid to fight flesh-eating maggot.

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!