13 Oct 2025 18:44 UTC - by Matt Levine
OpenAI Keeps Doing Deals
The essence of finance is time travel. In the future, your factory will make widgets and you will sell them profitably, so you can borrow money, today, to build the factory to make the widgets to sell to repay the money. If you go to a bank with an architect’s plans for the widget factory, and a business plan for making and selling the widgets, and a good track record of profitably manufacturing and selling widgets, the bank will be happy to lend you the money to build the factory. The bank has people who understand widgets; it can model up a widget factory pretty easily. It will figure out an appropriate leverage level for the factory and lend you that amount of money at, like, 8% interest.
Tech startups are an extreme version of this. In the future, human life will be utterly transformed by your invention: We will live on Mars or socialize in the metaverse or commute in self-driving flying cars or have omniscient robots in our pockets that can fulfill every human desire. Presumably, also, you can turn that into money. So you can raise money, today, to build the metaverse or cars or robots. I once wrote about Nikola Corp., the hydrogen-powered truck startup:
This is harder to model, though. How much money will you make selling the metaverse or the omniscient robots? A large amount, probably, but you are not going to have a detailed spreadsheet projecting your margins to a tenth of a percentage point. “Ehhh omniscient robots, come on,” you say to investors, and they say “yes of course” and give you money. But this will be based more on feel and enthusiasm and competitive dynamics than on detailed financial modeling. Also it will be equity. Investors will give you money now in exchange for a slice of the enormous profits of your world-changing vision, not for a fixed 8% return.
How much will they give you? Well, if you are very very very very good at pitching your vision, they might give you tens of billions of dollars. Global venture capital investment is something like $368 billion per year; getting 10% of that would be an astonishing feat.
What if you want, say, $1 trillion? That is harder. You can go to Masayoshi Son with your vision, and you can put your arm around his shoulder and gesture sweepingly into the distance and whisper “omniscient robots,” and he might say “yes, here is all the money I have,” but that’s still not $1 trillion. To raise $1 trillion, you need more than a compelling science-fiction vision of your world-changing product. You need a compelling science-fiction vision of world-changing financial engineering.
By this metric Sam Altman is surely the greatest tech founder in history. Sure sure sure sure sure chatbots, but look at this, man:
No! Wrong! He has shared tons of details of what his new financing tools will look like! He shares more every day! Here’s today’s:
The financing tool is, you go to Broadcom and you put your arm around their shoulder and you gesture sweepingly in the distance and whisper “omniscient robots” and they whisper “yesssss” and you say “we’ll need a few hundred billions dollars of chips and equipment from you” and they say “of course” and you say “good” and they say “do you have hundreds of billions of dollars” and you whisper “omniscient robots” again and they are enlightened. And then you announce the deal and Broadcom’s stock adds $150 billion of market capitalization and you’re like “see” and they’re like “yes” and you’re like “omniscient robots” and they’re like “I know right.” That is the financing tool! In some loose postmodern sense, OpenAI has borrowed hundreds of billions of dollars from Broadcom. You can buy hundreds of billions of dollars of equipment to build the robots to sell for money to pay for the equipment, because you’ve gotten everyone to believe.
Also:
If you owe the bank $100, that’s your problem. If you owe Broadcom $500 billion, that’s Broadcom’s problem. If you owe every big tech company hundreds of billions of dollars, that is their problem. Surely they’ll find a solution! Or you will. The money will figure itself out.
How? Well, in a world with insatiable demand for AI inference, probably you can raise quite a lot of money in debt to pay for the data centers. The bet here is something like “by the time we are actually building these data centers, they will be less like science-fiction speculation and more like widget factories, and people will be comfortable financing them based on cash-flow models.” But, for now, backstopping that theory is another theory, something like “OpenAI has a $500 billion equity value so surely it could raise more money to pay its bills.” The Financial Times notes:
Last week, we discussed OpenAI’s deal with Advanced Micro Devices Inc., in which OpenAI essentially spent some of that global tech rally on chips: OpenAI agreed to pay AMD tens of billions of dollars for chips, but it took back warrants on AMD’s stock; those warrants instantly became worth tens of billions of dollars as AMD’s stock rallied on news of the deal. The Broadcom deal does not have that sort of explicit monetization of OpenAI’s ability to add equity value to every company it touches. But does it implicitly monetize that ability?
I have previously written that “nobody in history has ever been better at, like, business negging than Sam Altman.” And:
Same with the computing costs. “The deals have surprised some competitors who have far more modest projections of their computing costs,” because he is better at this than they are. If you go around saying “I am going to build transformative AI efficiently,” how transformative can it be? If you go around saying “I am going to need 1,000 new nuclear plants to build my product,” everyone knows that it will be a big deal.
Schematically here is how crypto market structure works. I think Dogecoin will go up. You think Dogecoin will go down. Dogecoin is trading at, let’s say, $0.20. I go to a crypto exchange and make a bet on Dogecoin. I put down $100 for my bet, at 20-to-1 leverage; effectively, I have bought $2,000 worth of Dogecoin (10,000 coins) for $100. If Dogecoin goes up to $0.25, my position is worth $2,500; I have made $500 of profit on my $100 bet. If Dogecoin goes down to $0.19, my position is worth $0; I have lost my $100 bet, and my bet is closed out.
You, meanwhile, put down $100 for a 20-to-1 levered bet against Dogecoin. If Dogecoin goes down to $0.15, you make $500 of profit on your $100 bet. If Dogecoin goes up to $0.21, your position is worth $0; you have lost your $100 bet, and your bet is closed out.
Notice, though, that my $100 bet was on $2,000 worth of Dogecoins. How does the crypto exchange get $2,000 worth of Dogecoins, if I have only given it $100? One possible answer is “the crypto exchange borrows $1,900 from someone else to buy my $2,000 worth of Dogecoins,” but a moment’s reflection will tell you that that’s not the answer. The answer is that the crypto exchange doesn’t buy any Dogecoins. The crypto exchange is not in the business of buying Dogecoins; it is in the business of matching bets. I bet $100 that Dogecoin will go up, and you bet $100 that Dogecoin will go down, and the crypto exchange matches our bets against each other. If I win, you lose, and vice versa; there are no actual Dogecoins involved.
This is not the only crypto market structure: There are many exchanges where you can give the exchange $2,000 and get $2,000 worth of Dogecoin, which the exchange really holds in your account. There are probably some exchanges where you can give the exchange $1,000 and get $2,000 worth of Dogecoin, with the exchange finding someone else to finance the other $1,000 to actually buy the Dogecoin. But if you are getting 20-to-1 leverage on Dogecoin, nobody is buying any Dogecoin. You are making a pure derivative bet on the price of Dogecoin.
In crypto this sort of derivative bet is conventionally called a “perpetual futures contract,” or “perp.” In other markets it has other names. European stock traders might call it a “contract for difference.” Old-timey American stock traders would have called it “bucketing.”
I have simplified things by saying that you and I bet against each other. Actually in the first instance we bet against the exchange: I have a perp that will pay me if Dogecoin will go up, but I do not look to you for payment; I look to the exchange. The exchange will probably run a matched book (if I am long $2,000 of Dogecoin someone else has to be short $2,000 of Dogecoin), particularly if it is a decentralized exchange with no balance sheet of its own. But that is not a law of nature, and sometimes crypto exchanges are in effect long or short massive amounts of crypto bets against their customers. The famous example is of course FTX.
But let’s assume there’s a matched book: The only bets at the exchange are me betting that $2,000 worth of Dogecoin will go up, and you betting that it will go down. Let’s say Dogecoin (actual Dogecoin, traded on other, less leveraged exchanges) goes down to $0.17. In theory, I have lost $300 (my account has negative $200 in it) and you have made $300 (your account has positive $400 in it). But in practice my account balance on a decentralized crypto exchange can never be negative: I set up that account with a crypto wallet, not my name and credit card, and there’s no recourse to me. If my balance gets below zero, my account is liquidated but no one can ask me for more money. So really what happens is that my account has zero, your account has positive $400, and the exchange has the $200 of collateral ($100 each) that we posted. The exchange owes you $400 but only has $200. So the exchange calls you up and says “hey sorry actually your short Dogecoin bet is closed, and it was closed at $0.19. Here’s the $200 you made.” And if you are like “wait no Dogecoin is at $0.17,” that’s not the exchange’s problem.
This is called “auto-deleveraging.” Here is how Hyperliquid, a big decentralized exchange, describes it:
The people with winning positions socialize the losses. If you have a winning bet, someone has a losing bet. If the people with losing bets have lost more than they had, then you don’t get all of your winnings.
(Incidentally I once discussed this structure with Sam Bankman-Fried, of FTX, on Bloomberg’s Odd Lots podcast. He was quite dismissive of it, arguing that FTX had a much better liquidation system to prevent winning bettors from socializing losses. It did, until it didn’t.)
Anyway this is a bummer for you if you were betting that Dogecoin would go down: You were right, and you only got a portion of your winnings. It is more annoying if you weren’t doing that. What if you were an arbitrageur? You saw that traders on the decentralized exchange wanted to be long Dogecoin, so you did an arbitrage: You bought $2,000 of Dogecoin in real life, and you sold $2,000 of Dogecoin (via perp futures) on the decentralized exchange, effectively “manufacturing” the futures out of spot Dogecoin. The price of Dogecoin drops to $0.17, and:
The upshot is that when Dogecoin goes down by 5% (from $0.20 to $0.19), (1) gamblers who are long 20-to-1 levered perpetual Dogecoin futures get wiped out, (2) arbitrageurs who were short perpetual Dogecoin futures are auto-deleveraged and have to sell Dogecoin to remain hedged, (3) there’s lots of selling and no buying, (4) prices go down even more, (5) gamblers who were long 10-to-1 levered perpetual Dogecoin futures get wiped out, (6) etc.
Where is the limit? Well, in theory, at some level of liquidations, long-term fundamental value investors will see value. They will say “okay when Dogecoin fell from $0.20 to $0.18, that was one thing, but now Dogecoin is trading at just $0.15, which is well below its fundamental value.” So they will step in and buy Dogecoin, planning to make money when rationality sets in and the price of Dogecoin returns to its fundamental value. You see the problem here don’t you?
Anyway. There was a crypto crash last week. Bloomberg’s Muyao Shen and Olga Kharif reported:
Quite a lot of crypto traders were making levered perp bets on altcoins, crypto exchanges regularly offer 40 or 50 or 100 to 1 leverage, and so when prices move slightly all those bets go to zero. Also the people who were making levered bets against those altcoins had their own trouble, as the losing bettors got liquidated:
Right, yes, if a ton of crypto market structure consists of 40-to-1 levered bets, then you are going to have a lot of levered losing bettors blowing up, but you are also going to have a lot of weird results for the levered winning bettors. They can only get paid out of the losing bets.
Here is a more cynical take from John Hempton:
The idea is that, if you are a bucket shop that is systematically short crypto to your highly levered customers, an occasional 10% price drop can be very good for business: The customers all get wiped out and you keep their deposits. I don’t think that this explanation is necessary, just because standard crypto market structure explicitly buckets customer orders into matched books, but it is fun to think about.
In arguably related news, my Money Stuff podcast co-host Katie Greifeld reported last week:
If you have 50x levered Dogecoin, a 2% move will wipe you out. If you have 3x levered short AMD, a 38% move will wipe you out. Can’t really complain about that; you got the experience you paid for.
One simple model of the stock market is:
This model does not necessarily make economic sense. From a strictly rational economic perspective, you might say “I buy stocks to get long-term exposure to economic growth, and as the public stock market shrinks and becomes more mature, it provides a smaller share of future economic growth, so I should reduce my allocation to public stocks and buy more private equity or crypto or whatever.” But those are complicated conversations, and if you have a simple heuristic like “60% stocks,” it might be sticky even as the makeup of the stock market changes.
Victor Haghani and James White have a fun paper out along these lines on “The Impact of U.S. Stock Buybacks: Theory vs Practice,” noting that buybacks — which shrink the total size of the stock market — should have a levered impact on prices:
I should say that this is not just a story about buybacks; it is more broadly a story about shrinking public stock markets and “private markets are the new public markets.” There’s a trillion dollars of stock value — about 1.5% of the value of all publicly listed US stocks — just in OpenAI, SpaceX and Stripe; if those companies went public, they would more or less offset the $1.2 trillion of estimated buybacks this year. If the big private companies stay private and the big public companies keep buying back stock, then public stocks will get more scarce, which might make them more expensive.
What is the state of the art methodology for stealing quantitative trading secrets from your employer when you leave to take a new job at a different quantitative trading firm? Obviously, like, “email all of the trading algorithms to your Gmail” doesn’t work; your old employer tracks and probably blocks that. “Download them to a thumb drive,” or “print them on the office printer,” will also probably get you caught. “Pull up the trading algorithms on your work laptop over the weekend, and then use your personal iPad to take photographs of your laptop screen” is maybe somewhat better tradecraft. “Take the photographs using an iPad provided to you by your new employer” is I suppose a nice touch. This is not any sort of advice.
Here is a UK judicial ruling in a lawsuit brought by G-Research against a data scientist named Pierre Allain, who worked there from 2021 until 2025. “On Friday 21 March 2025, the defendant accepted a job offer from one of G-Research’s principal competitors, Citadel Securities LLC”; on Monday, March 24, he resigned from G-Research. Over the weekend, he did this:
His defense is that “the photographs were part of an aide memoire, and that he acted impulsively in taking them” and never shared them with his new employer.
From Tricolor to Saks, Bonds Are Now Crashing at Breakneck Speed. Bank deregulation set to unlock $2.6tn of Wall Street lending capacity. Thinking Machines Lab Co-Founder Departs for Meta. Warner Bros. Is Said to Rebuff Paramount Takeover Approach. Paramount Circling Warner Bros. Discovery After Rebuffed Approach. First Brands’ Founder Patrick James Steps Down as CEO. Jefferies Defends First Brands Deals, Says Losses Manageable. Brookfield to Acquire Remaining Oaktree Stake for $3 Billion. Silver Roars Higher as Short Squeeze Rocks the London Market. Kalshi Raises Funds at $5 Billion Valuation, to Expand Overseas. Senegal Raises $795 Million in Oversubscribed Regional Bond Sale. US Lifeline Buys Argentina Time as Investors Urge FX Overhaul. Warren Asks Hedge Fund Group If It Played Argentina Bailout Role. The Former Banker Betting on Unprofitable Business of Flood Insurance. The Real AI Risk is ‘Meh’ Technology That Takes Jobs and Annoys Us All. Amateur gold prospecting. NIL injury insurance.
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!
Read Original ArticleCategories rationale: The article primarily discusses the massive financing strategies employed by OpenAI (a major tech/AI player) to secure hardware, which is framed as a form of financial engineering that bypasses traditional lending. This relates strongly to Institutional Adoption (7) through the scale of deals being struck, Scalability (5) as it relates to funding massive growth, and Private Market (29) since OpenAI is a private entity funding its operations through non-traditional means that resemble private credit/equity dynamics. Level 2 choices reflect the focus on large-scale initiatives and the private nature of the funding.Characteristics justification: The sentiment is moderately negative (-0.4) because the text heavily critiques the 'science-fiction vision' financing methods, comparing them to speculative crypto derivatives (like Dogecoin perps) and highlighting risks like auto-deleveraging and the potential for massive counterparty risk if the underlying vision fails. The discussion of FTX's failure and the G-Research lawsuit adds a layer of negative context regarding financial misconduct and risk management. Uncertainty (0.6) is high due to the speculative nature of the financing being discussed (i.e., whether the AI vision will materialize into cash flow to cover the debt).Tag relevance: Tags focus on the key entities (OpenAI, Broadcom, G-Research), the core financial mechanism being analyzed (leveraged bets, perpetual futures, auto-deleveraging), and the broader context of private market financing (buybacks, private markets).asset-types: others
rwa: false
entropy: 0.75
sentiment: -0.4
staleness: 0.3
relevance: 0.85
uncertainty: 0.6
Brilliant breakdown of how Altman is essentially creating counterparty exposure at scale. The perpfutures analogy is spot on bc just like auto-deleveraging socializes losses when long bets blow up, these chip deals might socialize risk if OpenAI's revenue model dosent materialize. Saw somethign similar when hardware startups overcommit on manufacturing capacity.