On the last Wednesday of July, Nvidia vaporized hundreds of billions of dollars of market capitalization in a single session. The ticker kept printing. The algos kept rebalancing. And somewhere inside that noise, a small, newly rebranded AI-safety outfit learned the oldest lesson in finance: borrowed money amplifies conviction, but it also amplifies the bill.
The Financial Times reports that Situational Awareness — the organization formerly known as the Center for AI Safety, founded by Dan Hendrycks — has approached investors and lenders for support after borrowing amplified its losses during July's AI stock sell-off. Read that sentence again. The team built to price existential tail risk in artificial intelligence forgot to price the most mechanical tail risk in markets: the margin call.
This is not an AI story. This is a leverage story wearing an AI costume. And anyone sitting in crypto right now, thinking the AI trade's pain won't leak into our sandbox, is about to learn the same thing the hard way. The tape already told us part of the story in real time: a seven-session slide in the handful of stocks carrying the entire index, a volatility surface that went vertical, and a 'reasons to own it' narrative that went conspicuously quiet. The FT report just made it official.
The Backstory No One Wants to Tell
Situational Awareness occupies an odd corner of the cultural landscape. It began as CAIS, a serious-minded research nonprofit dedicated to AI risk analysis — the kind of outfit that publishes papers on model evaluations and gets summoned to Washington. In 2025, it rebranded. The name change signaled more than a logo refresh: it signaled posture. This was an organization that wanted to be on the ground, decoding what AI means for power, security, and capital allocation. The brand grew. The writing became a must-read. And somewhere along the way, the balance sheet developed a personality.
According to the FT, the fund borrowed money to gain exposure to the AI complex and carried that leverage through the July sell-off. The specifics of the structure matter less than the mechanism. A concentrated, levered bet on the biggest winner of the bull run is fine — right up until it isn't. And the 'isn't' arrived in a hurry.
Catalyst one: DeepSeek's R1 release questioned the assumption that compute scale is an unbreachable moat. Catalyst two: the Bank of Japan hiked on July 31, the carry trade that had been quietly financing global risk assets began unwinding, and by early August, the Nasdaq sat in correction territory. The sequence mattered. A speculative attack on the scaling narrative landed first; the carry-trade unwind followed nearly forty-eight hours later; by the time the volatility market repriced, everyone holding borrowed exposure had to make a choice — post more collateral, or sell into a tape with no bids left at the required size. There is a reason the phrase 'crash this week, cascade next month' exists. Weeks after the fact, the story breaks: a fund with a safety-first brand is the one ringing up investors and lenders.
I don't say this to mock. I say this because I've executed my own miniature version of the same stupidity. During the Terra collapse, I watched a heavily levered trader stare at his screen because his liquidation price was exactly where the bids had vanished. That trader, on a smaller stage, was me two years prior — and the only thing that saved me was on-chain transparency: I could see the pools, I could see the liquidations coming, and I could move before the calls landed. Situational Awareness didn't have that luxury. Nobody in TradFi gets to watch a margin call arrive in public. You just get the press release. Pain is just data you haven't decoded yet. In this case, the data was leverage.
The Leverage Ledger
Let's do the math the mission statement skipped. At 2x leverage, a 20% drawdown in the underlying asset eats 40% of equity. At 3x, that same drawdown eats 60%. The escape velocity back to breakeven is not symmetrical: after a 40% loss, you need a 67% gain just to get back to zero. After 60%, you need 150% — in a market that may spend months going sideways, which is the exact regime we're in now. That's not a comeback; it's a death sentence paid monthly through borrow costs.
Leverage doesn't just magnify losses — it converts volatility into permanent impairment. Consider two books with the same starting capital. The first is unlevered and sits through a 20% round-trip: down 20%, then up 25%, back to zero. The second runs 2x leverage through the same path: down 40%, then a 50% bounce in the underlying produces a 100% gain on remaining equity — barely clawing back. Now add the borrow cost, the funding drift, the fact that the bounce is rarely that clean — and the levered book ends the round-trip permanently behind. Volatility decay is the tax nobody quotes when they pitch 'structural AI growth.'
In crypto, we pay this tax explicitly, every eight hours, through funding rates. In equities, it hides inside borrow costs, margin rates, and the slow bleed of rolled swaps. Different labels, same leak. The only difference is that on-chain, at least you can see the leak coming. In TradFi, the leak is a feature of the plumbing — invisible until the basement floods. For anyone who has watched a liquidation cascade on a DeFi dashboard, this is table stakes. Yet the exact same mathematics wrecked a fund that writes about tail risk for a living. That disconnect is not a crime. It's a signal — and the signal is crowding.
Here, by mid-2025, the AI complex was the most crowded trade on Earth. Every allocation committee chased it, every momentum strategy piled in, every 'AI-everything' token followed. When crowding meets leverage, the exit dynamics go asymmetrical. On the way up, buying begets buying — green candles convince the cautious to join. On the way down, the order of exits is determined by who is forced to sell first, not who is smartest. Margin calls have no respect for thesis quality.
I got a front-row seat to that dynamic in my own backtesting. In 2024, after the ETF approvals, I ran about a thousand historical scenarios through a Python script, hunting for entry points that fired when institutional buying pressure spiked. The model kept flagging the same blind spot: forced sellers don't care about fair value — they care about the call. You can model 'value' until the candles burn out; the only number that matters in a cascade is the liquidation price of the marginal levered holder.
The Funding Stack
Here is the information gain most retail traders will miss. The FT story is not actually about the drawdown. It's about the funding stack beneath the narrative. It's about who provided the borrowed capital, what covenants they demanded, and what they get in exchange for not forcing a fire sale. When a levered fund approaches investors and lenders, the event stops being a mark-to-market loss and becomes a capital-structure negotiation. The equity layer eats the pain first, but the real action is what the debt layer extracts in exchange for patience.
This is the 2022 playbook. Every rescue in that cycle converted a panic into a power shift — the lenders got the terms, the distressed fund got a smaller upside, and the market got a permanent reminder that leverage always has a recursive price. In crypto, the same dynamics play out in public: when a whale's collateralized position wobbles, you can watch the negotiation in real time through the mempool and the oracle feeds. That transparency, the thing TradFi can't offer, is the only reason I survived 2022 with my skin attached. Situational Awareness gets no such window. The first time the market learns about its margin call is the story in the FT.
Now, the part that should scare crypto traders specifically is the correlation. During July's sell-off, AI-themed tokens did not escape. My own watchlists showed a rolling 30-day correlation between Nvidia's five-day returns and a basket of AI-linked crypto assets spiking above 0.7 — a level that would have seemed absurd at the start of 2025. A levered TradFi book in Silicon Valley can transmit pain into our sandbox through ETF flows, market-maker risk limits, and sentiment contagion. When the machine hits a margin call, the machine sells what it can, not what it wants. High-beta AI names in equities, AI narratives in crypto, tokenized equity products — all of it is collateral in the same trade. Market noise is just fear wearing a suit. The suit just got expensive.
The Contrarian Read
The contrarian take is not that leverage is evil. The contrarian take is that the AI trade was never 'safe' — and the safety framing is exactly what made it fragile. The market spent two years branding mega-cap AI names as bond-proxies, compounders, sleep-well-at-night holdings. That branding is what emboldened a nonprofit to behave like a macro hedge fund. When an asset gets labeled 'quality,' risk managers drop their guard. The same thing happens in crypto every cycle: 'ultra-sound money,' 'internet bond,' 'digital gold.' Labels are risk-management suppressants. They don't alter the physics of the margin call; they just make you less likely to calculate it.
The deeper blind spot is the irony. Situational Awareness built its reputation on pricing tail risk in artificial intelligence. But tail risk isn't an exotic event — it's the ordinary failure of assuming your position is safer than the market's structure allows. The candlestick doesn't lie, but your bias might. The healthiest position in any market is the one you can survive worst-case. The org forgot that. In a sideways tape, with the AI narrative trying to find a floor, that kind of forgetting tends to cascade. The market doesn't need you to believe the AI story; it needs you to know whether the person across from you is holding the same position with someone else's money.
The Takeaway
Trade the next margin call — don't theorize about it. Watch the funding stack of every leveraged narrative. If the AI trade's leveraged layer is being restructured, the path of least resistance is lower highs until the overhang clears. For the crypto side, watch for forced ETF redemptions and risk-limit cuts at market makers — they will show up in token liquidity before they show up in headlines.
For my part, I'll keep my AI-agent experiments small. When I deployed a sentiment-driven trading agent on a DEX in early 2026, the first thing I learned is that the model looked brilliant on a dry run and idiotic the moment I added leverage — because AI doesn't fear, and fear is the margin of safety. Removing fear from a position is how smart people turn losses into calamities.
The question that matters now isn't 'will AI survive?' It's 'who is still leveraged when the next catalyst lands?' Are you?