Scion Asset Management (Michael Burry)
Michael Burry (founder; 'The Big Short')
The most famous AI-bubble short of the cycle: Burry's bet that AI hyperscaler/chip earnings are inflated by aggressive hardware-depreciation accounting, expressed via puts on Nvidia and Palantir before he wound the fund down to focus on his Substack.
- Wound down — No longer operating — capital returned and the vehicle closed. Analysis of a historical record.
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
- Partial 13F — The disclosed book is longs-only and ~45 days lagged — a slice, not the strategy.
Short positions
Q3 2025 13F: puts on 1,000,000 NVDA shares
Q3 2025 13F: puts on 5,000,000 PLTR shares; combined NVDA+PLTR put notional ~$1.1B
Recent moves
Disclosed the NVDA + PLTR puts in the Q3 2025 13F (Nov 2025), publicly sparred with Palantir CEO Alex Karp ('couldn't crack a simple 13F'), then de-registered Scion and launched the paywalled 'Cassandra Unchained' Substack (Nov 23 2025, ~$379/yr, 23k+ Day-1 subs) as his sole vehicle for the AI-bubble/depreciation thesis.
Loading disclosed positions…
Our take
Honest framing: Scion is no longer an operating registered fund — the iconic AI short now lives as a Substack, not a book. The depreciation-accounting argument is genuinely substantive (it's also Chanos's core point), but Burry's history of being early-and-loud means the call can be right in thesis and painful in timing.
The most misread fact here is the headline.
The Q3 2025 13F showed roughly $1.1bn of puts — about $912m notional on Palantir, $187m on Nvidia — because Form 13F requires options be reported at the underlying's market value, not premium paid. Burry later disclosed the real position: 50,000 contracts, $50 strike, 2027 expiry, bought at $1.84, or $9.2m of premium. A bet reported everywhere as billion-dollar was, in cash at risk, under 1% of that; the filing format manufactured the story. Second, deregistration reads as business logic, not capitulation. Cassandra Unchained crossed 300,044 subscribers 231 days after launch (12 Jul 2026) at $39/mo or $379/yr — a theoretical $113.7m annual gross, paid-vs-free split undisclosed. Even at modest conversion that plausibly exceeds a management fee on the book he ran, with no redemptions, lockups, drawdowns or 13F obligation. He traded a fee stream contingent on being right on time for one contingent only on being read — a real asymmetry that also removes the last public scoreboard.
Thesis
Scion's terminal position was that AI infrastructure earnings are an accounting artifact: hyperscalers depreciating Nvidia GPUs over 5-6 years when Burry argues economic life is 2-3, which he sized at roughly $176bn of understated depreciation and overstated profit across 2026-2028. Expressed via puts on Nvidia and Palantir in the Q3 2025 13F. The vehicle no longer exists — Scion's SEC adviser registration terminated 10 Nov 2025 — and the thesis now lives in a paywalled Substack, not a fund book.
Concentrated, thesis-driven shorts expressed through long-dated listed puts rather than stock borrow, alongside a small deep-value long book. Scion never disclosed AUM, so capacity is unknowable; the binding constraints were timing and option tenor, not size. Post-deregistration there is no fund strategy left to evaluate — it is a publishing business monetising the same research.
- 2026-07On 25-28 Jul 2026 Burry disclosed an active, expanded short book: more Nvidia, new or larger puts on Micron, Applied Materials, Caterpillar and SOXX, Tesla and Palantir kept, long Flutter and DraftKings.
Assessment
- The depreciation argument is specific and testable — named useful-life assumptions, a sized $176bn 2026-2028 impact, and identifiable issuers, not vague bubble rhetoric.
- Long-dated puts cap loss at premium — $9.2m on Palantir — the right structure for a short whose timing is genuinely unknowable.
- Impact is evidenced: Nvidia issued a memo to sell-side analysts in Nov 2025 rebutting his depreciation and circular-financing points by name.
- Publicly conceded in the 27 Oct 2025 investor letter that his valuations were out of sync with markets, before winding down.
- A 2026-2028 accounting thesis expressed in puts expiring 2027 still has mismatched clocks — the earnings reset can arrive after the option does.
- The counter-case is serious: GPU waterfall reuse into inference, multi-year customer contracts, and firm resale marks argue for longer economic life than 2-3 years.
- Deregistration ends 13F filing, so there is now no independent way to verify any position, entry, exit, or P&L — the record becomes self-reported.
- Being right on the accounting does not require the stocks to fall; earnings quality and price can diverge for years.
- Whether the NVDA and PLTR puts made or lost money has never been disclosed.
Record
Attribution is close to impossible and that is the honest finding. Scion's audited returns were never public, and the 13F lens is unusually bad here: 45 days lagged, long-only-plus-options, US equities only, and reporting options at notional — the exact distortion that produced the $1.1bn headline. What is knowable is context. Burry's reputation rests overwhelmingly on one correct, early 2005-2008 subprime trade that nearly failed on investor redemptions before it paid; the post-2008 record has no comparable public verification. On the AI trade the tape is mixed, not vindicating: Palantir closed 17 Jul 2026 at $132.38 against $177.75 at the start of the year, roughly -25%, and that follows a ~24% rally off its June low on the Nvidia sovereign-AI announcement, while Q1 2026 revenue grew ~85% to $1.63bn. That is multiple compression punctuated by a rebound, not the earnings-quality unwind the thesis called for.
- 2026-07Marked to 31 Jul 2026: Palantir closed $123.06, down 30.77% year to date from $177.75 — not the roughly 25% read off the 17 July print.
- 2026-07The June rebound was surrendered in late July 2026: -6% on 22 Jul to $124.53, -7.1% on the week to 25 Jul, $122.05 on 30 Jul, leaving Palantir near a 52-week low, about 41% below its $207.52 December high.
- 2026-07July 2026 split the tape rather than mixing it: the SOX fell 20.6%, its worst month since Oct 2008, and Micron 28.7%, while Nvidia — the named put leg — rose 0.3% on the month and 7.77% year to date.
Risks & fit
- Key-person is total — there is no institution, no team, no continuity, only one author.
- Zero transparency post-deregistration: no filings, no third-party verification of any claim about the book.
- Thesis risk: if useful-life extensions prove economically justified, the core argument fails outright.
- Publishing incentives reward a maintained bearish narrative more than a revised one — the exit ramp from a wrong call is narrower than in a fund.
- Being early is indistinguishable from being wrong for however long the position or the subscription lasts.
The depreciation thesis dies if, through 2027-2028, hyperscalers hold or extend GPU useful lives with no meaningful impairment charges while 2022-2023 GPU cohorts show durable utilisation and firm secondary-market pricing — i.e. the waterfall reuse case proves real and Nvidia's four-to-six-year defence holds. It is confirmed if a large operator takes a material write-down or shortens useful life, forcing a visible earnings reset across the group. The vehicle question is already settled: Scion cannot be assessed as a fund again unless it re-registers and resumes filing.
Useful as a case study in three things: how 13F option reporting at notional distorts headline position sizes; how a single-manager research franchise can be worth more as a subscription than as a fee-earning book; and how a substantively correct accounting critique can still be uninvestable on timing. It is not a fund that can be evaluated on an ongoing basis — Scion is deregistered and files nothing, so any future claim about its positioning is unverifiable by construction.
Fund fee terms were never disclosed and are now moot — the adviser registration terminated 10 Nov 2025. The successor vehicle is a Substack at $39/mo or $379/yr that reported 300,044 subscribers on 12 Jul 2026. Scion AUM was never disclosed.