This week, Situational Awareness — the AI fund run by a manager in his mid-twenties, up 439% through June — went from a $45 billion peak to a forced fire-sale of its entire public book to Citadel. Not because the AI bet was wrong, but because it was concentrated, levered 4×, and early — and the market prices early and wrong exactly the same. It's Archegos. It's LTCM. It's every account that is all-in, on margin, unhedged. Here is the physics of a blow-up — and the four instruments that stop it being yours.
Blow-ups feel like freak events. They aren't. Put the great forced liquidations on one tape and the same four ingredients appear every time — leverage, concentration, a hedge that wasn't, and a liquidity trap. The names change; the physics don't.
And this wasn't an isolated accident — it was the biggest casualty of a violent month:
Every one of them had a defensible thesis. Not one had a survivable structure.
An unlevered investor who is right-but-early simply waits. A levered one can't: a drawdown becomes a margin call, the margin call becomes forced selling, and forced selling into a falling market becomes the fall. At 4× gross, a ~25% move against the book erases 100% of the equity under it. The curve below is the shape of that doom loop — slow, then all at once.
The market can stay irrational longer than your margin account can stay funded.
None of these are exotic. They are the same forces whether the book is $45 billion or ₹45 lakh. The green line under each is the instrument that neutralises it.
At 4× gross, a 25% adverse move is a 100% wipe. Leverage converts a drawdown you would have survived into a margin call you can't answer — and hands the timing of your exit to someone else.
Memory, GPUs, data centres, power — "diversified across AI infrastructure" was one trade wearing five tickers. When the theme turned, everything fell together, 27–54% in a month.
The fund did have shorts — legacy software names like Adobe, the "losers" of AI. In the rout, the longs collapsed and the shorts rallied: both legs were the same momentum bet, and both lost at once.
Beside the levered public book sat a $5 billion private Anthropic stake — unsellable on margin-call day. So the liquid holdings were dumped into the worst bid of the month, and the illiquid stake is what survived.
The whole danger in one chart: the adverse move it takes to erase your capital, at each level of gross leverage. At 1× you need a catastrophe. At 4×, an ordinary bad month does it. At 20×, a headline.
Leverage is borrowing return from the future and posting your survival as collateral.
The cruellest arithmetic in markets: losses and gains are not symmetric. A 20% loss needs +25% to get even. A 50% loss needs +100%. A 78% loss — this month's example — needs +355%, roughly the fund's entire legendary run, just to get back to zero. Risk management isn't about avoiding losses; it's about keeping them recoverable.
You don't get the same road back. Below a certain depth, there is no road back.
In calm markets, correlations look low and a themed basket looks diversified. In a sell-off, correlations rush to 1 — everything you own becomes the same trade. The book that felt spread out was always a single directional wager. It's the oldest lesson on the tape: LTCM's "uncorrelated" trades converged in 1998; AI infrastructure's did this month.
This is not a hedge-fund disease. The identical structure now runs on Indian phones. The margin trading facility — broker-funded stock buying at up to ~4× your capital — just hit an all-time high: ₹1.44 lakh crore as of July 29, roughly six times its size three years ago. Point it at a themed basket — defence, railways, AI smallcaps — and you have rebuilt Situational Awareness in miniature: one theme, 4× gross, no hedge, minus the prime brokers who negotiate before liquidating.
And the concentration is harder to see here, because it hides across accounts: the momentum names in your demat, the same theme again in your F&O positions, the same factor once more in your mutual funds — each app showing you a flattering slice, no screen showing you the sum. SEBI's own arithmetic says how the leveraged version usually ends: nine in ten F&O traders lost money in FY25, ₹1.05 lakh crore between them.
You can assemble a blow-up from a phone now. What you can't see is that you've done it.
The institutions that survive these episodes aren't braver — they've put a number on the danger before the market forces one on them. Four of those numbers, and the question each answers:
The loss you shouldn't exceed on, say, 95% of days — computed across thousands of simulated paths, not a gut feel. Not a ceiling; a baseline. If your everyday VaR already scares you, you're too big before anything goes wrong.
Re-run today's book through real history — 2008, COVID, this month's AI rout. VaR covers the ordinary; stress tests cover the day that ends accounts. July 2026 was a stress scenario that arrived unannounced.
Exposure by theme, sector and factor — and how correlated the book really is under stress, not in calm. The honest answer is usually "less diversified than it looks."
Your real position is the sum across every broker and account — equity, F&O, funds — not the flattering slice in one app. Consolidated exposure is the only view that tells the truth about concentration and leverage.
Being right is a thesis.
Surviving long enough to collect on it
is risk management.
Aschenbrenner may yet be proven right about AI — it won't matter to the capital that was marked out at the bottom. The edge was never the idea. It's the structure that lets you hold the idea: the position size, the real correlation, the hedge that works under stress, the liquidity match, the stress test run in advance.
This is exactly the instrument panel Kautilya puts in front of you — Monte Carlo Value-at-Risk on your live portfolio, stress testing against real historical crashes, concentration, factor and correlation analytics, and consolidated cross-broker exposure so the sum of your bets is finally on one screen. Not signals, not tips: the instruments that show you the edge of the cliff while you can still step back.