Nine Japanese filings — all dumped at once on Aug 12, over a month past Japan's 5-day deadline — reveal a position no 13F ever showed: Taiyo Yuden (Tokyo: 6976), one of the world's top MLCC makers. The fund rode it from 5.99% of the company to a 16.61% peak in three weeks — then the margin calls hit, and the same filings show it force-sold down to 4.41% by Aug 3. That residual stake — roughly $380M at today's price — still stands, making this one of the few positions that survived the unwind.
What it makes: multilayer ceramic capacitors — the rice-grain-sized parts that smooth power into every chip. ~71% of its ¥355bn FY3/2026 revenue is capacitors; the rest is inductors and filters. A modern GPU pulls hundreds of watts at ~1 volt with violent microsecond swings, so designers pack banks of high-capacitance MLCCs right at the package — AI servers carry several times the passive content of regular ones. Q1 sales mix: autos 28%, info-infrastructure/industrial (incl. AI servers) 27% and rising, smartphones 20%, IT devices 18%.
Why the stock halved anyway: three legs, none of them an earnings miss — a July 6–10 valuation unwind (the stock had run ~7.6x from January and topped 100x forward earnings), the July 29 Asia-wide AI/memory rout, and July 30–Aug 3, when this fund itself was force-selling ~12% of the company into a falling market. The Aug 5 earnings were actually good — guidance raised on AI-server demand — and the stock still gave back the pop the next day.
What to watch: whether the remaining 4.41% holds (any ≥1% change triggers a Japanese filing within ~5 business days); whether capacitor pricing keeps firming (Taiyo Yuden reprices Sep 1); and whether the >10% operating-margin promise from Q2 materializes. The risk to the whole thesis: MLCC demand is a derivative of AI capex — if the buildout slows, the #3 line item on the bill of materials slows with it. New capacity takes 18–24 months to build, which protects pricing now but adds supply into 2028.
Sources: EDINET large-shareholding filings (Aug 12), Taiyo Yuden Q1 FY3/2027 results deck (Aug 5), TrendForce, company IR. The ~$380M residual figure = 4.41% of the ¥1.39T market cap at ¥159/USD — our arithmetic, not a filing figure. Not investment advice.
What happened: after July's AI crash, margin calls from the fund's lending banks (Bank of America, Goldman Sachs, JPMorgan — per CNBC, Reuters and Bloomberg) forced Situational Awareness to unwind the bulk of its public stock positions. Citadel — Ken Griffin's hedge fund — bought the bulk in a single block trade reported at over $10B (WSJ first; Millennium and Jane Street reportedly bid too). Per an Aug 4 SEC filing, at least one position — Core Scientific — passed through the sale: the fund sold half of it (~$225M) in Aug 3 block trades and still holds 4.4%. Per Reuters (Aug 14), Jane Street was not just a bidder for the block — it was an investor in the fund, and took a July loss of roughly $15B partly tied to that stake, its first losing month since 2016 (Fortune, Aug 15: a Jane Street partner said the stake is “flat on the year, still up over the life of the bet”).
Rumor check: Citadel didn't "liquidate" the fund — the banks' margin calls forced the sale; Citadel was the buyer, which ended the forced selling. And how a fund this hedged got margin-called: per Business Insider the losses ran two-sided — the AI longs collapsed while its software shorts and semiconductor hedges also moved against it.
Where the book went (Aug 21): per Ken Griffin's first client letter on the deal (CNBC), Citadel has already unwound more than 80% of the risk it bought — via 100+ block trades totaling over $4B in market value, including single-session record blocks in 10 separate stocks. Citadel was a fast intermediary, not a long-term holder.
What remains: the fund is not dead — roughly $10B, including its unsold Anthropic stake (FT-valued around $5B), and it reportedly keeps trading. Size arc, as reported: ~$45B peak in early July → ~$20–24B at the sale → ~$10B after. July closed down −67%; the investor letter (via Bloomberg) put it plainly: "We took the steps that were necessary to fight another day." On Aug 4 came the first post-crisis move — a $400M investment in an undisclosed private company, on top of a $100M stake bought the prior month (~$500M total; Bloomberg, which lists Anthropic, Fluidstack and MatX among the retained private stakes). Aschenbrenner — up 439% net through June — had invited fresh capital from Aug 1 (July 24 letter, via the FT), but the raise did not materialize as hoped: per Bloomberg (Aug 7), despite a surge of inbound investor interest, the fund has told investors it is not accepting new capital for now.
Newly surfaced (Aug 12): Japanese filings revealed a position 13Fs never showed — Taiyo Yuden, bought up to 16.61% of the company pre-crisis, force-sold to 4.41% through the unwind, and still held at that level. Full breakdown in the green panel above.
What to watch: the Q2 13F (filed Aug 14) is now below — it is the June 30 book, before the sale, so it shows what he was holding when the margin calls hit, not what survived; further 13D/G amendments will show what did (the Aug 4 Core Scientific 13D/A was the first proof); any Japanese 変更報告書 on Taiyo Yuden (a ≥1% change triggers one within ~5 business days); and SharonAI — the resale registration went effective Aug 14 (424B3) and a 13G/A the same day reaffirms 19.9%: the gate is open, watch the Form 4s. The next 13F (Q3, ~Nov 14) is the first post-sale snapshot. The rows below are the June 30 book — history now, not holdings.
The June 30 book was $20.24B — +48% QoQ — but the mix flipped: Q1's $8.46B of puts on Nvidia, Broadcom, AMD, Oracle and the SMH ETF were all closed by quarter-end, and long common went from ~$3.9B to ~$20.2B (5x). 55% of it sat in two memory names — SanDisk and Micron — which then fell 47% and 29% in July (Bloomberg). Hedged in March, naked in June, margin-called in July: this is that book.
AI compute requires gigawatts of continuous power. By June 30 Leopold was long the unsexy backbone — and, above all, memory (the book as filed Jun 30 — force-sold July 30): SanDisk and Micron (55% of the book, common shares now, not calls), Bloom Energy (on-site fuel cells), TSMC and new STMicro, and neoclouds and miners pivoting to GPU hosting (Nebius, CoreWeave, Core Scientific, Applied Digital, Riot, IREN, CleanSpark). Thesis: the bottleneck is physical, not algorithmic.
In March the puts targeted names where consensus multiples had run too far: Nvidia, Broadcom, AMD, ASML, Oracle, TSM and the broader SMH ETF — $8.46B notional. By June 30 every one of them was closed; the only put left was $5M on Infosys. The bet was no longer these stocks have priced in too much — it was one-way long. That is how a fund that looked hedged in May could be margin-called in July: when AI hardware fell, there was nothing on the other side.
| Ticker | Company | Type | Value | % book | Shares / Contracts | Status | Why it's there |
|---|
Data on this page is from the Q2 2026 13F-HR (period ending 6/30/2026, filed 8/14/2026, accession 0000935836-26-000418); quarter-on-quarter deltas compare against the Q1 filing (3/31/2026, accession 0002045724-26-000008). Post-filing context: Bloomberg and CNBC (Aug 14) on the filing, Reuters (Aug 14) and Fortune (Aug 15) on Jane Street, SharonAI's 13G/A and 424B3 (Aug 14). Verify and refresh from:
To update: edit the inline DATA array in this file when the next 13F drops — roughly mid-November 2026 for Q3 2026, the first snapshot after the July 30 sale.
Read this before you continue — credibility is the whole point of this site.