Money Stuff: Tricking the AI Investors

Podcasting note: This week’s episode of the Money Stuff podcast is out now (because of the holiday tomorrow). Katie and I discuss Susquehanna puts, sports ETFs and Stretch. You can listen to this week’s episode here.

Okay here’s an investor relations strategy:

1. You run a public company. Let’s say it’s called Money Stuff Inc.
2. Periodically, you have either good news or bad news.
3. When you have good news, you put out a press release saying “Money Stuff Inc. announces some good news,” and your stock goes up.
4. When you have bad news, you put out a press release saying “Money Stuff Inc. announces some bad news.” Except that, in this case, you don’t quite write “Money Stuff Inc.” You replace the “M,” “o” and “e” in “Money” with Unicode characters 041C, 043E and 0435, which are the Cyrillic letters “M,” “o” and “e.” Those Cyrillic letters look exactly the same, to the human eye, as the English [1] letters “M,” “o” and “e,” so your press release will look completely normal to any human who reads it.
5. Not everyone reading your press release is a human, though. To a computer, English “M” and Cyrillic “M” are totally different. Computer trading systems will read your press release as like “**n*y Stuff Inc. announces some bad news” and be confused. “What the heck is **n*y Stuff Inc., never heard of it,” they will say, and not take any further action. They won’t sell your stock on the bad news, so — because algorithmic traders are increasingly important to the stock market — your stock won’t go down.

From a human’s perspective, sometimes you announce good news and other times you announce bad news and you are always scrupulously honest. From a computer’s perspective, sometimes you announce good news and other times there’s some random nonsense in the feed. The computer only sees your good news, so your stock never goes down.

This is so stupid in so many ways, and also extremely not any sort of advice, but it is based on real academic research. Bloomberg’s Lu Wang reports:

At the University of Liechtenstein, Advije Rizvani, Giovanni Apruzzese and Pavel Laskov designed 10 LLM-based trading models to forecast share prices with sentiment analysis for a portfolio of stocks. All generated positive returns during a 14-month investment period through April 2025. Yet every model was fooled after researchers made subtle changes to financial news headlines that were barely noticeable to human readers, such as swapping letters for nearly identical-looking characters or embedding hidden text.

Here is the paper, “Adversarial News and Lost Profits: Manipulating Headlines in LLM-Driven Algorithmic Trading,” which really does involve “Unicode homoglyph substitution”:

A few characters in a stock name (e.g., A, e) can be replaced by visually indistinguishable Unicode counterparts (e.g., Cyrillic “A” А, “е” е). The manipulated headline appears to be unchanged to the human eye. When such headline is used in the [algorithmic trading system] pipeline, it may elicit wrong decisions in the stock mapping algorithm, determining which stock a given headline refers to, and thus “misroute” a headline.

“We consider an adversary with no direct access to an [algorithmic trading system] but able to alter stock-related news headlines on a single day,” they write, but isn’t the obvious “adversary” here the company itself? [2] We have, ridiculously, previously discussed the the idea of downplaying unpleasant corporate news by printing it in white text on a white background, and I wrote:

Honestly this all feels quaint already. Pretty soon all corporate disclosures will be read exclusively by computers, and concealing information by making it unreadable to humans won’t work. What you’ll want is to conceal it from the computers, which is harder.

Totally wrong! Unicode homoglyph substitution! Maybe it’s the future of stock promotion. 

We have talked a few times about the maximalist vision of artificial intelligence, which goes something like this:

1. In the future, AI robots will replace most human endeavor and nobody will have jobs anymore.
2. The problem of economics will be how to allocate all the value created by the robots.
3. In our current system of market capitalism, that value will be allocated to the owners of the robots
4. So you’d better own some robots.
5. Society should get out ahead of that problem by making sure that everyone owns the AI companies that will control the robots. Perhaps governments should nationalize the AI companies, or tax them extra to fund universal basic income. The more market-based approach is that the AI companies should be large and increasing components of the stock index, and everyone should buy index funds in their retirement accounts to own their share of the robots.

Now. You don’t have to believe this maximalist vision. You could think AI won’t be that big a deal, or it will create more opportunities for human labor, or it will wipe out humanity and thus render stock ownership irrelevant, or something else. Here I just want to make the narrow point that this particular maximalist vision is quite bullish for AI stock prices. [3] The message is “in the future, all of the value in the world will belong to AI companies, so you’d better buy AI stocks today.” [4]

I write sometimes that “nobody in history has ever been better at business negging than Sam Altman.” What I mean is that Altman has mastered the art of saying and doing facially bad things about OpenAI, so that people will think good things about OpenAI. Most notably, he used to spend a lot of time publicly worrying that AI might destroy humanity, which sounds bad when you put it like that, but which also might make you think “ooh this technology is really powerful, I’d better invest.” Similarly, here’s this:

OpenAI has discussed giving a 5 per cent stake to the US government as the $852bn AI start-up seeks to clear political obstacles by securing financial buy-in from the Trump administration.

Sam Altman, chief executive of the ChatGPT maker, has argued that giving the public a financial stake in the company is the best way to share the upside of AI and has suggested a stake of this size in early conversations with the administration, according to two people familiar with the talks. …

Altman and other OpenAI executives have suggested that each of America’s leading AI developers allot 5 per cent of their equity to a vehicle like the Alaska Permanent Fund, a sovereign fund that invests the state’s oil wealth into stocks and pays dividends to the state government and residents. ...

OpenAI and Anthropic have previously suggested in economic policy proposals that arrangements such as public or sovereign wealth funds may be required in future to distribute shares to the public.

In April, OpenAI proposed a “public wealth fund” that “provides every citizen — including those not invested in financial markets — with a stake in AI-driven economic growth”.

The OpenAI Foundation, the company’s non-profit arm, said in May that in an AI-led future, “society will likely need new approaches that give people durable stakes in the systems creating value”, pointing to public or sovereign wealth funds.

On the one hand, giving away 5% of your equity is in some nominal sense dilutive to your shareholders. On the other hand, giving away 5% of your equity is a small price to pay to persuade everyone that owning your equity is essential to the future of humanity. Then you sell the other 95%.

One way that I like to tell the story of private credit is:

1. In the olden days, insurance companies bought a lot of corporate bonds, clipped the coupons and held them to maturity. The insurance companies had predictable long-dated liabilities, and they matched them up with predictable long-dated liabilities from corporate bonds.
2. Then, like, Bill Gross invented bond trading: Instead of holding bonds to maturity, insurance companies could trade them to make better risk-adjusted returns.
3. In the beginning, this was lucrative, but ultimately the bond market just became more efficient and returns were competed down.
4. Then some insurance companies invented the opposite of bond trading: Instead of buying liquid traded bonds that pay relatively low returns, you can accept some illiquidity, offer borrowers a better experience (faster marketing, lower market risk, a single lender, a better relationship with that lender, less paperwork, etc.) and demand a higher interest rate. Then you hold the bonds until maturity and clip the coupons, which is fine, because you are an insurance company with predictable long-dated liabilities.

That is the core story, and then modern “private credit” introduces various inessential novelties. For instance: Fifty years ago, when Bill Gross was getting his start, the innovators in insurance company lending worked at insurance companies. In the 2020s, they work at “alternative asset managers,” which manage money for insurance companies and/or own insurance companies directly. And so now the high-profile way that insurance companies do illiquid buy-and-hold lending is through private credit funds, because it is cool and lucrative to manage a private credit fund and boring to run an insurance company.

But that is only approximately true, and in fact the boring old business of “insurance companies buy bonds and hold them” never entirely went away. Bloomberg’s Emily Graffeo reports that it’s having a bit of a renaissance:

A private bond market dating back more than a century is opening a new front in the trillion-dollar AI funding boom, allowing tech borrowers to sell debt directly to deep-pocketed insurance firms.

Borrowers are hunting for capital wherever they can to finance the vast sums needed for the build out of AI. Meanwhile, life insurers facing record demand for annuities are seeking longer-term, high-grade corporate bonds to finance those multi-decade liabilities.

Private bonds, where companies sell securities directly to groups of select institutional investors, are bridging the gap. Issuance hit roughly $81 billion this year through May, the most for the period in data going back to 2016, according to Private Placement Monitor. Industry participants say AI is fueling the surge.

“Especially with AI and data centers, there’s an insatiable need for capital,” said Sheel Patel, head of New York private credit at Mayer Brown. “The borrowers are increasingly more comfortable using the private market for larger financings, especially when they realize that certainty, flexibility and confidentiality are a priority in these transactions.” …

This shift comes as a swelling population of Americans over age 65 drives annuity sales to record levels. In fact, total annuity sales hit an all-time high of roughly $464 billion last year, according to the life insurance trade group LIMRA. To back these expanding long-term liabilities, insurance firms are scaling up their purchases of private corporate bonds.

In some ways the simplest possible financial story is: “Businesses need money to make long-term investments in physical capital that will increase their future profits. They borrow that money from insurance companies who have lots of money and need to plan for predictable long-term liabilities.” You can complicate that story with liquid public bond markets or private credit funds or whatever. But the simple story is that you want a steady income in retirement, AI data center rents can provide that steady income, and the financial industry will match you up with a data center and make that trade happen.

Elsewhere:

For the second straight quarter, two Blue Owl Capital Inc. private credit funds were hit with the industry’s largest redemption requests, forcing the manager to again cap withdrawals.

Investors in the roughly $34 billion Blue Owl Credit Income Corp., one of the largest in the industry, asked to pull 18.8% of shares, or $3.6 billion in the second quarter, according to an investor letter Thursday. That was down from $4.2 billion requested in the prior period from the fund known as OCIC.

The smaller Blue Owl Technology Income Corp. saw shareholders request 38.1%, or $1.1 billion, compared with $1.2 billion in the first quarter.

Disclosure: I have a small investment in BCIC, sorry sorry. “Insurance companies have predictable long-term liabilities and can commit to making illiquid long-term investments” is a good classic simple story; “retail investors probably won’t want their money back all at once so let’s put it in illiquid long-term investments” is also a story but is notably less good. It turns out this is not the best technology for matching your retirement plans to a data center.

Polymarket, the prediction market, offers five-minute binary options on Bitcoin. If Bitcoin trades at $62,000 at 9:30 a.m., you can put in a bet on whether it will be up or down at 9:35. If it trades above $62,000 at 9:35, the “Up” bets win $1 and the “Down” bets get $0; if it trades below $62,000 at 9:35, “Down” wins.

If you put a lot of money on “Up,” and at 9:34 and 55 seconds Bitcoin is trading at $61,999.99, you should probably go buy some Bitcoin. Your buying will push up the price, and if you can push the price up to $62,000.01, your Polymarket options will pay off. Then you can sell the Bitcoin you just bought. When you sell your Bitcoin, the price will probably go back down. You will lose money on the round-trip Bitcoin trades: You’ll buy on the way up, sell on the way down, and probably sell at a lower average price than you bought. But you will make money on the Polymarket trades: You bought “Up” at 50 cents or whatever, and it pays $1. Whether this is a good trade or not depends on the relative size and liquidity of those markets. If you can buy 10,000 Up contracts at 50 cents, you can make $5,000; if you can then move the Bitcoin price up by spending $4,000, you should. This is not any sort of advice at all, and this is obviously bad market manipulation, but, you know, we are talking about Bitcoin and Polymarket here. There are also five-minute Dogecoin binaries.

We talked yesterday about market manipulation in eggs. I wrote:

This is a familiar story and you probably know the ending. There’s a big market (egg producers selling eggs to supermarkets etc.), and there’s a small market (egg producers selling extra eggs to each other on an electronic exchange). The price in the small market determines the price in the big market. Participants in the small market are also participants in the big market. You can spend a little money in the small market to move the price, which can make you a lot of money in the big market.

“Big market” and “small market” there are imprecise terms. The market for Bitcoin is bigger, by volume, than the market for five-minute binary options, but binary options are discontinuous, so moving the price of Bitcoin by a little bit can have a big payoff in your binary options. The relevant question is whether you can spend a little money to move the price of Bitcoin in a way that makes you more money in the binary options market. When prediction markets were just a glimmer in the eye of libertarian economists, the answer to that question was “of course not, what are you even talking about.” But, as I wrote a few months ago, “now prediction markets exist and are a big business, so they are not only in the business of predicting reality: They’re also in the business of changing it.”

Here’s a paper by David Dai, Ruizhe Jia and Shihao Yu on “Settlement Manipulation in Prediction Markets”:

Empirically, we study Polymarket’s Bitcoin five-minute up/down contract: a binary claim that pays $1 if Bitcoin is higher at the close of a five-minute window than at its open, and $0 otherwise. … The contract’s launch on February 12, 2026 offers a clean natural experiment: no five-minute contract existed before that date, so we compare three regimes: no such contract (P1), the 15-minute and 4-hour contracts only (P2), and the five-minute contract live (P3).

As soon as it launched, a relatively small prediction market began redirecting large spot orders to the settlement seconds. In the final ten seconds before each close, trading on Binance spikes: the magnitude of net order flow jumps, and volatility rises with it. The timing is no accident: the spike appears only after the five-minute launch (about 50% above the pre-launch level), and is significantly attenuated at the fifteen-minute horizon. The spike is sharpest where a push is pivotal: in the roughly 6% of cycles whose contract price still implies a near-even outcome just before the close, the near-settlement order-flow jump is about 3.9 times that in the rest. The clearest sign that this is not information is the reversal: within ten seconds the price reverts, by about a quarter in the near-even cycles and a tenth in the others. Real information would persist; the price impact of a manipulative push reverts.

We have talked about prediction market manipulation before. There is a Polymarket contract on whether Jesus Christ will return before 2027, and there was another Polymarket contract on whether the Christ-will-return contract would trade above 5% during some one-hour window, and allegedly people bought the Christ-will-return contract in an effort to make the derivative contract pay out. I wrote: “What possible purpose could the ‘Jesus Christ return before 2027 Odds >5% February 17, 12-1 AM?’ contract have other than as a plaything for market manipulation?” I submit to you that if you lose money to market manipulators on the “Bitcoin up in the next five minutes” contract, you deserve it. Not legal advice or whatever.

Lots of investors have investment mandates with some sort of environmental, social or governance restriction. “Don’t invest in weapons companies” is a reasonably common one. Some companies make weapons and some don’t; weapons-restricted investors invest only in the ones that don’t. They have to pay attention, though, and update their holdings periodically. Sometimes a snack food company will start making weapons. Or a non-weapons company will acquire, or be acquired by, a weapons company. 

If you’re a public-company shareholder with a no-weapons mandate, this is not a big problem: You just sell the stocks that are now weapons stocks. If you’re a private-company shareholder, it’s harder: There’s no liquid market for your stock, so you are stuck. When there is no exit, you need to exercise voice: A no-weapons shareholder in a private company that starts getting curious about weapons, traditionally, has to tell the company’s executives “no, this is not what we signed up for.” But private markets are the new public markets, and now it’s easy to migrate no-weapons investors out of the shareholder base. The Financial Times reports:

German surveillance drone maker Quantum Systems will consider merging with kamikaze drone start-up Stark after using its latest $1.2bn funding round to part ways with investors opposed to investing in weapons.

Florian Seibel, Quantum co-founder and co-CEO, said the company used the funding round, which valued the defence technology group at about $8bn, to “clean up” its shareholder structure and allow it to develop lethal technologies for the first time.

Anyone who did not “feel comfortable with the potentially new alignment of the company had a chance here to exit”, he told the FT.

That could pave the way for moves into areas such as deep-strike missiles, he said, as well as a possible merger with Stark, which he also co-founded, to bring his “two babies back together”.

I like that Seibel loves drones indiscriminately but had to have separate killer and non-killer drone companies to appeal to two distinct sets of investors. But he has now been successful enough to kick out the non-killer investors and reunite his “two babies.”

I mean, you know:

Empery Digital Inc. (NASDAQ: EMPD) (the “Company” or “Empery Digital”) [Tuesday] announced that it has entered into a definitive agreement for a $65 million investment (the “Investment”) representing a 25% ownership into a private entity that is acquiring a strategically located Midwest facility to be converted into a state-of-the-art AI data center.

Empery Digital and Hunt Properties, Inc. have entered a strategic partnership to jointly originate, evaluate, and acquire powered land properties with secured tenants suitable for AI and high-performance computing data center development. The partnership will combine Hunt Properties’ decades of experience navigating utility interconnection processes, power procurement, and energy infrastructure development and their established network of relationships across the U.S. with Empery Digital’s public company platform, expertise in capital markets, and strong balance sheet, including its Bitcoin holdings, to execute a shared vision for AI infrastructure.

Etc. We have talked about Empery a couple of times before. As of early 2025, it was called Volcon Inc. and was in the business of selling electric bikes and golf carts. But in July 2025, a better opportunity came along, and Volcon became a digital asset treasury company, or DAT (and renamed itself Empery). That was the thing to do with a US public company, last summer: If you put $500 million of Bitcoin into a public company, the company would be worth $1 billion, so Empery did. 

That is no longer the thing to do with a US public company: Empery started trading at a discount to net asset value, and activists have been pushing it to sell its Bitcoin, buy back stock and close the discount. The DAT trade that looked so good last summer — “if you buy Bitcoin, your stock will trade at a premium” — no longer works. Never mind.

The thing to do with a a US public company in the summer of 2026 is of course AI infrastructure. People won’t pay a premium for DAT stock anymore, but they will pay a premium for AI infrastructure stock. Onward! 

SEC Probes Alleged Insider Trades That Cost Susquehanna. Trump Made $1 Billion on Crypto Deals While His Fans Lost a Fortune. Nvidia Says It Will Take a Cut of Some Customers’ Cloud Revenues. CoreWeave Junk Bonds Slide Further as Investors Question AI Boom. SpaceX Showed Investors Prototype of Elon Musk’s New AI Device. SpaceX Analyst Debut Set to Test $2.2 Trillion Valuation. One Leveraged ETF Is Reshaping Trading in World's Top AI Memory Stock. Bank of England to push ahead with plan to limit hedge fund leverage. Equity-Market Fundraising at Most Exuberant Since 2021, Mergermarket Says. Millennium Targets at Least $10 Billion in New Fundraising. The DeepMind trio who built a poker AI are now making money for quant hedge funds. UBS to trial US banking services in push for wealthy American clients. McKinsey Shakes Up Its Board After Scandals Over Past Work With Clients. Hot dog inflation: Americans brace for costliest ever July 4 parties. Whiskey Barrel-Backed Loans Are Plummeting in Value With Americans Drinking Less. The Wall Street Women Who Traded Finance Careers for Influencer Success. Ken Griffin Bought Out All 138 Condos in Miami Tower, One by One.

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!