The AI trade did not break; leverage did: what All-In's selloff debate gets right

The latest All-In episode treats the AI-stock selloff as a test of financial structure, not a verdict on artificial intelligence. Its argument is narrower—and more useful—than “the bubble is over”: a market can have real demand, real infrastructure spending and still produce violent losses when investors finance the trade with too much leverage. 1

The hosts describe a sharp fall across semiconductor names: the Philadelphia Semiconductor Index was down more than 20% over the month discussed, while Samsung and SK Hynix were also cited as having fallen sharply. They connect the move to a wider selloff in AI-linked equities, not to one isolated company. 1

The episode's central mechanism is leverage. It discusses reports that Leopold Aschenbrenner's Situational Awareness fund was running at roughly 3.5x to 4x leverage. In the account presented, around $30 billion of equity supported about $120 billion of positions. A 5% move against the portfolio therefore becomes roughly a 20% hit to equity before fees, hedges or other details are considered. A large enough decline triggers a margin call: the lender wants more collateral, and if the borrower cannot provide it, positions are sold. 1

That is why a correction can feed on itself. Forced selling pushes prices lower, which makes collateral thinner, which creates more forced selling. The person who made the original bet loses control of the exit. The hosts' practical lesson is blunt: an investor can be directionally right about AI and still be wiped out by the path the market takes to get there.

The distinction matters outside hedge funds. Any AI business whose plan depends on continuously rising asset prices, cheap financing or a narrow set of public comparables is exposed to the same feedback loop. “The technology is real” is not a financing plan.

The hosts also place the selloff in a higher-rate environment. When Treasury yields offer a relatively attractive alternative, investors demand more future growth to justify paying high multiples for semiconductor and infrastructure companies. Persistent inflation, government deficits and geopolitical risk can keep rates and risk premiums elevated. 1

This does not require a change in the long-term AI story. It only changes the price investors are willing to pay for that story today. A company can be on track to sell more compute five years from now and still lose half its market value if the discount rate, positioning or expectations change first.

The episode also points to Chinese open-source models and cheaper access to intelligence as another pressure on the market. If model capability becomes easier to obtain, some of the value currently assigned to scarce frontier access may move toward applications, distribution and infrastructure. That is a competitive question, not proof that demand disappears.

Despite the market damage, the hosts argue that the underlying AI capex thesis remains intact. Hyperscalers are still spending heavily on data centers, chips and networking because they expect future usage to justify the buildout. Their interpretation is that the selloff is mostly momentum and financing unwinding rather than evidence that customers have stopped wanting AI. 1

That view should be treated as an argument, not as a settled fact. The episode offers evidence for it—continued infrastructure spending and the persistence of model demand—but it does not resolve the harder question of returns. Spending can be real while the eventual economic surplus is smaller than investors assumed. The useful follow-up is not “bull or bear?” but “which layer captures the value after prices fall?”

This is also why the discussion of AI revenue and chip prices cannot be collapsed into one chart. Demand can grow while margins compress. Usage can expand while a model provider cuts prices. Compute can be scarce today while future supply catches up. A market correction may be pricing that transition rather than rejecting the technology.

The episode is skeptical of the frontier-lab employees' call to pace AI development. The hosts interpret it partly as virtue signaling, a form of regulatory capture or an attempt by leading labs to make themselves the gatekeepers of a dangerous technology. They prefer narrower safety reporting, self-regulation or more open competition to a broad pause. 1

There is a real tension in that criticism. The request to slow development may be self-interested, but that does not make every underlying safety concern false. Conversely, calling for caution does not prove that frontier labs would sacrifice their competitive position. The episode is strongest when it asks readers to distinguish the policy claim from the institutional incentive behind it.

The market lesson follows the same pattern. Do not read a forced liquidation as a clean referendum on AI fundamentals; do not read continued spending as proof that every AI valuation is sound. The update for an AI practitioner is to watch two separate curves: whether useful demand keeps expanding, and whether the financing structure around that demand can survive a less forgiving market.