The Great Unwind.

A young hedge fund maven lost 67% in July on his AI bets. But don’t worry for him, he’s still up 80% for the year. It’s not the volatility that deserves our scrutiny, it’s his investment thesis. This handsome fella Leopold Aschenbrenner runs Situational Awareness, a hedge fund named for a series of essays he wrote in 2024 under that banner. As if the circular funding fiasco that powers the AI trade isn’t incestuous enough, young Leo is a former OpenAI employee who’s marrying Anthropic’s chief of staff this weekend. And even though he was fired from OpenAI, is only 25 years old, had a stint at disgraced crypto investor Sam Bankman Fried’s foundation, and is a researcher by trade, these bona fides were enough to convince titans of Wall Street to invest in and provide leverage to his new hedge fund that went long on shovel trades—companies involved in construction, infrastructure and data centers like Micron and CoreWeave—and short on software and chips like Nvidia, Broadcom and Oracle.

So the investment thesis is fairly cynical when you stop to think about it. It’s basically betting that there will be a lot of money in infrastructure and little payout on the technology. For me, that pretty much says it all. And his thesis has worked out for the most part. Like I said, he’s still up 80% for the year. But what happened in July when certain names went upside down in a chaotic month for AI and software stocks illustrates how ridiculous this whole AI narrative has become and how Wall Street is loading up tinder on the bonfire.

Anyone got a light?

Hyperscalers or Hooversuckers?

Microsoft, Alphabet, Amazon, and Meta have told shareholders to expect roughly three-quarters of a trillion dollars in combined capital expenditure this year alone—nearly double what they spent just last year—and if you fold in Oracle’s borrowing binge, the industry’s total commitment for 2026 pushes past that mark entirely. AI capital spending has swallowed nearly the entirety of these companies’ operating cash flow, up from about a third of it just a few years ago. The promise, as sold to investors, is that this is generational infrastructure—the rails for the next economy. The next coming of the internet, supercomputer chip, railroad. You get the picture.

What’s actually happening is that a handful of enormous companies are reallocating the cash flow of the entire tech sector into a single, unproven bet, and asking everyone downstream to trust the payoff arrives before the bill does.

And “downstream” is the part that convinced Wall Street to invest $45 billion with young Leo who didn’t build his fund betting on ChatGPT getting smarter. He built it betting on cement, copper, and turbines. It’s worth $10 billion today, by the way. So either he’s early, or he’s about as good at this as Scott Bessent was when he ran his hedge fund into the ground.

Honestly, it’s not a terrible investment thesis. The real winners of the AI boom, so far, aren’t the software companies promising to reinvent your job—they’re the companies pouring concrete, running conduit, and welding steel. Data centers need land, power substations, cooling infrastructure, and skilled trades to build all of it, which means the AI trade has become a construction and heavy-industry trade as much as a technology one. That’s the wager behind “long shovels, short chips”—go long the industrial buildout, go short the technology it’s supposedly built to deliver. The purpose of this is not to critique his thesis or hold him out as the first domino of the impending collapse, it’s to show you how collapses occur.

Which brings us to the GDP report that dropped this week, because that construction spending isn’t just enriching contractors, it’s propping up the entire country’s growth numbers.

Second-quarter GDP came in at 1.5% annualized, a real slowdown from the first quarter, and it missed what economists were expecting. But bury yourself in the details and you’ll find business investment—driven in no small part by what the government itself flagged as “a surge in investment in artificial intelligence”—was one of the only bright spots keeping the number from being worse . Strip out the data-center buildout and this economy is coasting on very little. Construction spending on AI infrastructure has become a load-bearing pillar of U.S. GDP.

Now here’s the problem, and it’s the same problem every bubble eventually runs into: somebody has to actually pay for the thing being built, and the people footing the bill are getting cold feet. This week OpenAI quietly cut prices on two of its models, trimming one by a fifth and slashing the other by roughly 80%. Sam Altman framed it as a march toward better “price/intelligence,” but the honest read is that corporate customers finally noticed their AI bills and started asking what exactly they were getting for the money. The era analysts have taken to calling ”tokenmaxxing,”. This all-you-can-eat phase where companies told employees to run everything through AI without worrying about the meter is officially over, and even the industry’s own trade press is calling it dead on arrival.

Also, OpenAI isn’t cutting prices out of generosity, it’s cutting them because Chinese labs keep releasing frontier-class models at a fraction of the cost, closing the performance gap fast enough that American labs no longer have the luxury of charging monopoly rents. Worth noting, though: China’s AI giants are just as lost on the path to profitability as their American counterparts, burning through investor cash to buy market share with no clearer route to actually making money on any of it. Nobody in this race has figured out how to get paid. They’ve just figured out how to spend.

Enter Mark Zuckerberg, who has spent this year committing enormous sums of shareholder money to AI, with little to show for it in his stock price. Meta was one of the worst performers of the Magnificent Seven this year, and analysts keep asking the same uncomfortable question: where’s the return on all that CapEx? So instead of answering that question, Zuckerberg published an op-ed and gave interviews warning the world about the dangers of AI being “centralized” in a handful of companies, taking not-so-veiled shots at OpenAI and Anthropic for being too closed, too doom-obsessed, and too controlling of the technology’s future.

When you don’t have an answer for your own AI returns, you change the subject to how everyone else is doing AI wrong.

Meanwhile, over at Oracle, the numbers stopped being a story about growth and started being a story about survival. The company’s capital spending is now running at something like 174% of its operating cash flow, which is a fancy way of saying it is spending far more than it’s bringing in, funded almost entirely by other people’s money—and it’s reportedly gearing up to raise tens of billions more in debt just to keep pace with its data-center commitments.

Speaking of companies betting the house: SpaceX has fallen hard from its highs this summer, and short sellers have piled in aggressively, with bearish bets on the stock jumping from the single digits to nearly a third of the available float in just a matter of weeks. That’s not a coincidence, it’s timing. SpaceX’s post-IPO lockup begins releasing hundreds of millions of shares onto the market right after its next earnings report, with even more shares potentially unlocking if the stock had managed to hold well above its IPO price, which it hasn’t.

It gets more pointed still: some of the most profitable companies in the history of American business are posting negative free cash flow. Not slowing growth—negative cash flow, while still generating enormous profits from their actual, functioning businesses. That’s how much money AI infrastructure is eating. Reuters found that five of the largest hyperscalers are on pace to spend far more on CapEx through next year than they’ll generate in new operating cash flow over the same stretch, meaning for every dollar of new cash coming in the door, more than a dollar and a half is going out to build data centers.

Amazon’s free cash flow has been squeezed down to almost nothing, despite operating cash flow climbing sharply. These are not struggling startups. These are the most dominant, cash-generative businesses on the planet, and AI spending is turning their balance sheets inside out.

Where’s that money coming from, if not from operations? Increasingly, from debt, and increasingly, that debt isn’t landing the way it used to. When hyperscalers go to the bond market now, the reception has gotten noticeably chillier. Investor demand for these bond sales has thinned out considerably compared to earlier in the year, spreads have widened across nearly every maturity, and a large share of this year’s tech bond issuance is now trading in the secondary market at worse yields than when it was first sold—meaning the people who bought early are already sitting on losses. When the biggest, most creditworthy companies in the world start seeing their bond sales come in soft, that’s the market signaling less confidence in this story than the press releases suggest.

Instead of taking that hint, Wall Street’s response has been to lean further into circular financing deals—arrangements where the chipmaker funds the data center that buys the chips that generate the revenue that justifies the valuation that lets the chipmaker fund the next data center. Nvidia alone has now been party to hundreds of billions of dollars in these arrangements this year, including financing structures tied to OpenAI’s infrastructure and chip purchases, and the pattern has gotten so pronounced that both the International Monetary Fund and the Bank for International Settlements have flagged it as a systemic risk worth watching. Nvidia’s own credit default swaps—essentially insurance against Nvidia defaulting—spiked to a record this month on the back of this. When the market for insuring a company against failure gets more expensive at the exact moment that company is signing the most deals of its history, that says something about how institutional money is actually pricing the risk.

Now layer on top of all of this the quieter, more structural problem: duration. This year’s hyperscaler bond issuance hasn’t just been large—it’s been long. Amazon, Alphabet, Meta, and Oracle have already sold tens of billions more in bonds this year than they did in all of last year combined, and Wall Street expects that pace to keep accelerating into next year. These aren’t short-term notes. These are long-dated bonds that lock up investor capital for decades, which means every dollar a pension fund or insurance company commits to a 30-year Meta or Oracle bond is a dollar that isn’t available to lend to a hospital system, a manufacturer, or a mid-sized company trying to expand.

So it’s not just that AI has hoovered up the available R&D cash, depleted cash reserves of companies, and absorbed the lion’s share of equities investments, now they’re eating into the debt market. This sector isn’t just competing with other companies for capital anymore. It’s crowding them out.

The Unwind

So how does this all begin to unwind, and why might Leopold Aschenbrenner’s blown-up hedge fund be the canary in the coal mine?

Start with the Fed, which held rates steady this week despite an unusual three dissents from officials who wanted to hike—the first time that’s happened since 2016. Fed Chair Kevin Warsh said something that the markets hated. He said the bond market is already doing the Fed’s job for it, because yields have been climbing on both ends of the curve without the Fed lifting a finger. In other words, the bond market hates the AI-first economy.

Corporate bonds are going undersubscribed, and rates from the two-year all the way out to the 30-year are reaching a point of “fuck it,” meaning at some point, locking in a guaranteed return north of 5% for the next few decades starts looking a lot more attractive than riding a hyperscaler’s AI story that might not have a payoff for years, if ever. When that math flips for enough big money, it doesn’t trickle out of the AI trade. It rotates.

And if you want to know what that rotation looks like when it goes wrong, you already lived it in 2008. Back then it was mortgage-backed securities stacked into collateralized debt obligations (CDOs) stacked into synthetic swaps, layer upon layer of side bets on the underlying mortgages, until a wave of defaults in the actual homes triggered margin calls up a chain that had way more exposure than actual money behind it. Anything outside of the underlying investment, remember, is just a bet. Or a bet on a bet.

Today, the underlying asset is corporate bonds, and Wall Street’s money is stacked into this AI trade in nearly identical layers—direct equity stakes, collateralized loan obligations (CLOs) packed with data-center and chip-sector debt, hedge fund leverage of exactly the kind that just liquidated Situational Awareness overnight, private equity stakes in the infrastructure buildout, and private credit funds leveraged against all of it. Little Leo’s fund didn’t blow up because his thesis about shovels versus software was wrong. It blew up because his prime brokers made margin calls, forcing a fire sale of his entire public book in a single overnight block trade—a preview of what forced deleveraging looks like when it’s someone else’s turn.

That’s why the Treasury market is the thing to actually watch here, more than any single stock or any one hedge fund’s bad month. Rising rates make all of this debt more expensive to service at the exact moment the industry is issuing more of it than ever. If we get more cracks in private credit, or another margin-call whipsaw like we just watched with Situational Awareness, only bigger, the market gets spooked in a way that doesn’t stay contained to tech stocks. Because the doom loop in this case is that the exits are treasuries.

When the AI trade stops paying and the leverage starts unwinding, the money doesn’t look for the next hot data-center stock. It runs to the safest, most boring asset in the world, which pushes yields around even more, which makes the debt-fueled AI buildout even more expensive to sustain, which forces more of the trade to unwind. That’s the loop. And Leopold Aschenbrenner just showed everyone, in real time, exactly how fast it can spin.