A fund up 439% lost 67% in a month: the AI trade's risk is financing, not the forecast

Leopold Aschenbrenner's Situational Awareness lost around 67% of its value in July after margin calls it could not meet forced it to sell its public equity portfolio ^[1]. The fund had returned 439% in the first half of 2026 on roughly 4x leverage against concentrated positions in AI chips and infrastructure, according to a paper on the collapse by RCK Analytics ^[2], which is the whole lesson: the forecast held and the financing did not.

Look at what was in the book. By the end of June, filings show close to half of the fund's US equity portfolio sat in names such as SanDisk and Micron ^[3]. The demand data behind that positioning kept improving through the drawdown. Counterpoint Research put enterprise solid-state drives at 48% of global NAND sales in the second quarter of 2026 as AI workloads moved from training toward inference ^[4]. TrendForce forecast NAND prices rising 10% to 15% quarter on quarter and DRAM up 13% to 18% ^[5]. Micron reported fiscal third-quarter revenue of $41.46bn against $9.30bn a year earlier ^[6], about 4.5 times the prior-year figure ^[7].

None of that settles a margin call. The fund's positions fell between 35% and 47% during July ^[8], and the Philadelphia Semiconductor Index dropped almost 30% from its June peak ^[9]. Three prime brokers, including Goldman and JPMorgan, called margin; the fund could not honour the calls and sold its entire public equities book to Citadel ^[10]. At 4x gross exposure, a 35% to 47% move against the portfolio is 140% to 188% of equity ^[11], so the realised loss of 67% implies leverage had already been cut or the book was liquidated before the full mark ran through it. The reported figures at least reconcile: a 439% first half followed by a 67% July leaves roughly 78% for the year ^[12], close to the 80% year-to-date figure circulated with Aschenbrenner's letter to limited partners on 31 July 2026 ^[13].

Aschenbrenner presented the loss as a risk event rather than a thesis failure, saying the fund is positioned "to fight another day" ^[14] and that "we embrace volatility. But it should never jeopardize the fund" ^[15]. He also said he would keep investing in public equities ^[13], and the firm put $400m into Source Foundry, a lithography start-up valued near $5bn in the one part of chip manufacturing ASML currently monopolises ^[16]. That is a liquid problem answered partly with an illiquid asset.

Supervisors were already pointing at this. The Bank of England's July 2026 Financial Stability Report flagged a "substantial increase in the use of leverage in equity markets", with emphasis on stretched valuations in a small number of AI-linked companies ^[17]. The Bank for International Settlements has used Archegos in 2021 and the UK liability-driven investment stress of 2022 to press non-bank funds to avoid exactly this outcome ^[18].

For anyone running crowded semiconductor or AI infrastructure exposure, the practical question is not whether memory pricing holds. It is whether your position survives a 30% index drawdown ^[9] on someone else's collateral terms. Watch prime broker haircuts on semiconductor collateral, whether TrendForce's price increases actually print ^[5], and whether other levered AI books surface in second-half reporting. The gap between a correct call and a solvent one was five weeks ^[1][2].