Michael Burry’s Latest Portfolio: Inside His Bet Against the AI Debt Machine

Michael Burry is not simply betting that artificial intelligence will fail. His sharper claim is that parts of the AI boom are being financed as if demand can never disappoint, even while chips become more efficient, small models move onto local devices and data-centre builders sign years of commitments before the final customer has proved it can earn a return. 

That distinction matters. It turns a familiar “AI bubble” story into something more specific: a credit cycle wearing a technology costume.

Let’s break down Burry’s latest known positions, why he sold Alibaba, where he is short, what he is buying instead, and which trades the numbers actually support.

What is Michael Burry’s Latest Portfolio in 2026?

First, a necessary correction to almost every “complete Burry portfolio” headline online: no outsider can see his complete current portfolio.

Scion Asset Management’s last Form 13F covered September 30, 2025. Its SEC registration was terminated soon after. Burry now publishes selected trades through his paid Substack, Cassandra Unchained. There is no 2026 filing.

The honest solution is to use confidence labels.

Michael Burry’s latest publicly supported short positions

The six high-confidence shorts are the names still listed in the latest August 24 tracking of his Substack disclosures. Palantir and Caterpillar were clearly disclosed earlier in August, but were not included in that fresher list, so it would be careless to call their current size certain. The same reporting puts cash near 12% after the August 13 reshuffle.

Michael Burry’s latest publicly supported long positions

A “medium” does not mean the reporting was unreliable. It means no later public update confirms that Burry still owns the position today.

Michael Burry’s positions reported as closed or exited

This is why the widely repeated “$1.1 billion Nvidia and Palantir short” is misleading. The old 13F reported the notional value of shares underlying put options, not the premium Burry paid or his maximum loss. Burry later said the Palantir puts cost $9.2 million, not $912 million. Options and short sales are not interchangeable: a put can expire worthless, while a naked short can lose more than the original capital committed.

Michael Burry’s Real AI Thesis: Follow the Financing, Not the Chatbot

Burry’s argument has three layers.

First, he sees circular financing. One company funds an AI customer, that customer buys chips or cloud capacity from the same ecosystem, and the spending returns as reported revenue. This is not automatically fraudulent. The danger begins when investors treat financed demand as identical to cash demand from profitable end users.

Second, his “perimeter of debt” includes purchase commitments, uncommenced leases, guarantees, backstops and special-purpose vehicles that behave like debt even when they sit elsewhere in the filings. Think of a family with a small mortgage but ten years of school fees and car leases. The mortgage statement does not show the true fixed burden.

Third, efficiency may arrive before the industry earns back its infrastructure bill. A study of more than one million queries, over 20 models and eight accelerators found that small local models could handle 88.7% of the single-turn queries tested. “Intelligence per watt” improved 5.3-fold in two years, while hybrid local-cloud routing cut modelled energy, compute and cost by 60% to 80%. That does not prove cloud demand will collapse, but scarcity pricing should not be projected forever.

The sober institutional version of this warning comes from the Bank for International Settlements. Its June 2026 report said intense competition could fuel AI overinvestment, opaque financing was increasing, and corporate credit could reprice sharply if AI payoffs disappoint. Burry’s timetable is not tomorrow. His stated base case is 2028, while markets may react earlier.

Our Perimeter of Debt Scorecard

Here is a reusable way to test any AI-infrastructure stock. We score five risks from 1 to 5, where 5 is the most dangerous.

This table explains our central disagreement with Burry: his system-level concern is strongest where the debt and hardware sit, but Nvidia is the asset-light toll collector. If an overbuilt motorway fails, the leveraged road owner suffers before the company that sold the toll equipment.

Where is Michael Burry Short, and Does the Math Support Him?

Prices below are a market snapshot from August 25, 2026. They are reference points, not targets.

Nvidia (NVDA): the right question, but probably the wrong first short

Nvidia traded near $208, worth about $5.1 trillion and roughly 32 times trailing earnings. That valuation needs extraordinary execution. Yet the latest operating results were also extraordinary: quarterly revenue reached $81.6 billion, up 85% year on year; Data Center revenue grew 92% to $75.2 billion; non-GAAP gross margin was 75%; and management guided the next quarter to about $91 billion without assuming China data-centre compute revenue. Nvidia also held roughly $50 billion of cash and marketable debt securities against about $8.5 billion of debt.

Our view: Burry is right to question who ultimately funds GPU demand. We disagree with expressing that thesis most aggressively through Nvidia. The company has a huge software moat, supplier leverage, high margins and a net-cash balance sheet. If AI financing cracks, weaker neoclouds and leveraged capacity buyers should feel the pain first. Nvidia can still fall sharply because expectations are high, but it is the strongest business in the chain. For most investors, “avoid at any price” and “short” are completely different decisions.

Oracle (ORCL): Burry’s cleanest fundamental short

Oracle traded near $142 at about 26 times trailing earnings. Its growth is real: fiscal 2026 cloud revenue rose 39% to $34 billion, infrastructure revenue jumped 93% in the fourth quarter, and remaining performance obligations reached $638 billion. The funding mismatch is also real. Oracle produced $32 billion of operating cash flow, spent $55.7 billion on capital expenditure and therefore generated negative free cash flow of roughly $23.7 billion. It raised $43 billion of debt and $5 billion of equity in fiscal 2026, and expects another roughly $40 billion of debt and equity funding in fiscal 2027.

The simple ratio is uncomfortable:

Our view: We agree with Burry more here than on Nvidia. Customer prepayments and backlog reduce the risk, so Oracle is not broken. But if contracts are renegotiated, customers weaken, or hardware economics change, shareholders and lenders absorb the funding mismatch. The thesis is valid; the timing can remain brutal while cloud growth accelerates.

Nebius (NBIS): the highest-risk balance sheet in the group

Nebius reported second-quarter revenue of $582 million, up 454%, and adjusted EBITDA of $236 million. It also spent about $5.66 billion on property and intangibles and reported a $190 million GAAP loss from continuing operations. Long-term debt stood near $8.5 billion, while deferred revenue totalled roughly $6.0 billion.

Our view: This is the short where Burry’s financing argument bites hardest. Nebius must build far ahead of revenue, finance the gap, execute perfectly and hope today’s scarce-compute pricing survives. The counterpoint is that deferred revenue and EBITDA show genuine demand. Our verdict is “avoid as a long”, not “rush to short”, because a capacity-constrained stock growing revenue at 454% can squeeze bears without warning.

Micron (MU): cheap peak earnings can be expensive

Micron traded near $910 after an enormous run. Its latest quarter looked almost unreal: revenue of $41.5 billion, up 346% year on year, gross margin of 84.6%, operating cash flow of $25.4 billion and adjusted free cash flow of $18.3 billion. Management guided the next quarter to roughly $50 billion of revenue and an 86% gross margin.

This is the classic commodity trap. A low P/E at record margins often means the “E” is temporarily too high. Suppose annual revenue eventually normalises to $100 billion and net margin settles at 25%. That produces $25 billion of earnings, far below the current annualised pace. The valuation multiple would expand mechanically even if the share price did not move.

Our view: We agree with Burry’s cycle logic and disagree with treating current earnings as normal. We are less confident about timing. Tight memory supply, AI demand and long customer agreements can keep margins elevated longer than a short seller can remain comfortable. This is a good thesis with a dangerous clock.

SOXX and QQQ: better hedges than stock-specific predictions

The iShares Semiconductor ETF traded near $506 after a roughly 73% year-to-date return. It held 30 stocks, had a reported P/E of 64.4 and a three-year beta of 2.08. That makes SOXX a cleaner way to bet on sector-wide multiple and margin compression than trying to identify the one chip winner that fails.

QQQ puts are different. They are portfolio insurance against an expensive, concentrated growth index, not proof that every Nasdaq-100 business is overvalued. Rolling the puts to June 2027 and higher strikes increased their sensitivity while buying more time.

Our view: Of Burry’s bearish instruments, SOXX is conceptually cleaner than shorting Nvidia alone. QQQ puts make sense as a defined-loss hedge. Neither is a sensible copy trade for a reader who does not understand option premium decay. Being right in 2028 does not rescue an option that expires in 2027.

Palantir (PLTR): an excellent company with almost no valuation cushion

Palantir traded near $176, with a market value around $452 billion and a trailing P/E near 150. The business is not weak. Q2 revenue grew 93% to $1.94 billion, US commercial revenue rose 149%, adjusted free cash flow was $1.22 billion and cash plus Treasury securities reached $9.2 billion. Full-year adjusted free cash flow guidance was $4.5 billion to $4.7 billion.

At the $4.6 billion guidance midpoint, the market was paying roughly 98 times current free cash flow. Even if free cash flow tripled over five years, a 30-times exit multiple would imply an enterprise value around $414 billion, before adjusting for cash. That rough exercise shows how much future success is already required.

Our view: We agree with the valuation warning, not with extreme claims that the operating business is worth close to nothing. Palantir has cash, margins, customer growth and no financing problem resembling Nebius or Oracle. We would avoid chasing it at this valuation. We would also avoid an unhedged short against a company growing revenue 93%.

Caterpillar (CAT): the weakest link in Burry’s short book

Caterpillar traded near $811 at about 35 times earnings. It is exposed to the data-centre power boom, but it is not a pure AI company. Q2 sales rose 24% to $20.5 billion, adjusted operating margin was 21.9%, operating cash flow reached $4.4 billion and backlog grew across its three main segments.

Our view: We disagree. Caterpillar also serves construction, mining, energy and transport through a powerful dealer network. The bet needs several cycles to cool together, making it far less direct than the neocloud shorts.

Where is Michael Burry Long, and Which Stocks Look Attractive?

Lululemon (LULU): cheap, but only if earnings have found a floor

Lululemon traded near $123. One Substack reader reconstructed its weight at 17.4% of Burry’s tracked holdings, versus 9% for Molina. That is not an audited portfolio statement, but Burry replied that LULU was “screaming cheap”.

The attraction is obvious. Revenue still grew 4% to $2.5 billion in Q1, China grew 30%, and the company held about $1.5 billion of cash. The problem is equally clear: Americas revenue fell 3%, Americas comparable sales dropped 5%, gross margin contracted 410 basis points and operating margin fell from 18.5% to 11.2%. Full-year EPS guidance was cut to $10.95 to $11.15.

At the $11.05 guidance midpoint:

Our view: This is one of Burry’s better longs. At roughly 11 times guidance, the stock does not need a return to perfection. It does need US sales and margins to stop falling. We would build slowly, not all at once, and treat the next two earnings reports as an earnings-floor test.

Alibaba (BABA) to JD.com (JD): this is about capital allocation, not China

Burry moved his entire Alibaba position into JD.com. The trigger was Alibaba’s HK$80 billion, or roughly $10.2 billion, placement of 710 million new shares at HK$112.70, with all proceeds earmarked for full-stack AI. The issuance represents roughly 3.7% dilution. Burry said Alibaba would have to fall by half before he reconsidered it.

Alibaba’s operating story is not hopeless. Quarterly revenue rose 9%, AI cloud and compute revenue grew 45%, and cash plus other liquid investments stood near $69.9 billion. But net profit fell 75%, capital expenditure rose 75% to almost $10 billion and free cash flow was negative $6.6 billion. Alibaba chose to issue discounted shares despite a huge liquid balance sheet. That is the signal Burry rejected.

JD’s latest quarter was less glamorous but more shareholder-friendly. Revenue fell 2.9% to $51.1 billion, yet JD Retail operating margin reached 4.6%, quarterly free cash flow was $4.69 billion, and cash, restricted cash and short-term investments totalled roughly $34.6 billion. JD repurchased about 2.5% of its shares in the first half.

Here is the per-share contrast: Alibaba is adding roughly 3.7% to its share count while JD removed roughly 2.5% in six months. That is an approximate 6.2 percentage-point swing in ownership direction. It is not a valuation model, but it captures why two similar China exposures can produce very different per-share outcomes.

Our view: We agree with choosing JD over Alibaba today. We do not agree that BABA must fall 50% before becoming interesting. The mechanical dilution is only about 3.7%; the real concern is falling return on invested capital. If AI cloud margins and free cash flow recover, a blanket 50% hurdle will look too rigid. JD is our preferred China commerce exposure, but both deserve smaller position sizes because of competition, geopolitics and ADR risk.

Adobe, MercadoLibre and Zoetis: three very different quality longs

Adobe generated record Q2 revenue of $6.62 billion, 13% growth, $2.17 billion of operating cash flow and 8.5 million share repurchases. Full-year non-GAAP EPS guidance is $24.35 to $24.45. At $276, that is only about 11.3 times guided non-GAAP earnings. AI-first recurring revenue more than tripled and exceeded $500 million. The market is pricing Adobe as a melting ice cube while the reported numbers still look like a growing, cash-rich subscription franchise.

MercadoLibre’s Q2 revenue plus financial income grew 50% to $10.2 billion. Gross merchandise volume reached $22 billion and total payment volume $101 billion. The risk is inside the strength: its credit portfolio grew 75% to more than $16 billion. At roughly 53 times trailing earnings, investors must monitor credit losses as carefully as commerce growth.

Zoetis cut its full-year outlook after US companion-animal sales fell 11%, but livestock sales rose 23% and international revenue grew 8%. At roughly 12.4 times the midpoint of adjusted EPS guidance, the valuation finally compensates investors for slower growth and competition.

Our view: Adobe is our favourite long in the reported book because current cash earnings provide support while AI adoption creates upside. MercadoLibre is the highest-quality compounder but needs credit discipline and a long horizon. Zoetis is a sensible defensive recovery trade, not a fast-growth story.

Molina Healthcare (MOH): a 2027 recovery bet disguised as a cheap stock

Molina traded near $199. In Q2, premium revenue fell 6%, net income declined 76%, the medical-cost ratio worsened to 92.2% and adjusted EPS was $1.51. Management calls 2026 a trough year and guides to at least $5.25 of adjusted EPS.

At $199, the stock is roughly 38 times trough-year guidance. That is not statistically cheap. If normalised EPS returns to $12 to $14 and the market pays 14 times earnings, fair value would be roughly $168 to $196. The upside therefore needs a stronger recovery, acquisitions, or a higher multiple.

Our view: Burry may be right that politics and rate resets improve 2027. We would call MOH a watchlist recovery, not a bargain on current earnings. Medical-cost trends are the deciding variable.

Flutter (FLUT): a regulatory option with operating leverage

Flutter traded near $103 after Burry reportedly more than doubled his position at an average cost in the high $90s. Q2 revenue was $4.33 billion, but US revenue fell 6%, US adjusted EBITDA dropped 70% to $119 million, leverage was expected to rise to 4.3 times and full-year guidance was cut. International revenue grew 10%, providing the cushion.

Burry’s bet is political. Prediction markets currently compete with sportsbooks while enjoying a different federal and state regulatory structure. If that loophole narrows, FanDuel’s scale becomes more valuable. If it remains open, the US business can stay under pressure.

Our view: We prefer Flutter to the DraftKings position Burry sold because Flutter has international earnings and broader brands. It is still a high-risk turnaround with leverage, regulation and execution moving at the same time. This belongs in the speculative bucket, not the core compounder bucket.

Michael Burry’s remaining reported longs, in one decision table

PayPal’s Q1 adjusted free cash flow was $1.7 billion and Burry said he would not sell into the proposed $60.50 takeover price. Fannie and Freddie are a different animal: Treasury still holds senior preferred claims and warrants to buy 79.9% of the common stock for a nominal price, and common shareholders cannot receive dividends while the current capital structure remains. Fannie reported a positive $116.5 billion net worth, but its senior preferred liquidation preference had risen to $234.2 billion by June 30. That is political optionality, not a normal P/E trade.

Our Ranking: Where Michael Burry Looks Strongest and Weakest

A Framework Investors Can Reuse: Separate the Story From the Security

Burry’s portfolio contains one lesson more useful than any ticker.

1. Ask who is funding the customer

Revenue paid from a customer’s operating cash flow is higher quality than revenue financed by the vendor, a strategic investor or a special-purpose vehicle. Both count as revenue today. They do not carry the same default risk tomorrow.

2. Compare capex with operating cash flow

When capex stays above operating cash flow, the company must use cash reserves, debt, equity or customer prepayments. That can be rational during a buildout. It becomes fragile when several funding sources must remain open simultaneously.

3. Match asset life with technology life

A data-centre lease may run for 10 years while the chips inside it become commercially second-tier in three. Long funding against short-lived technology is the core mismatch.

4. Normalise the denominator

For cyclical companies, do not value the stock on peak earnings. For disrupted companies, do not value it on trough earnings. Micron looks cheap on record margins. Molina looks expensive on trough EPS. Both conclusions can reverse when earnings normalise.

5. Decide whether valuation means “avoid” or “short”

An expensive stock can remain expensive for years. A put option adds a deadline; a short sale adds theoretically unlimited loss. Apple can be “permacostly”, as Burry put it, without being a good short. That may be the most practical risk-management lesson in the entire portfolio.

Our Take

Burry is strongest when he follows the cash rather than the narrative. Oracle’s negative free cash flow, Nebius’s extraordinary capital intensity, Alibaba’s discounted share issuance and JD’s contrasting buybacks are hard numbers. His broader warning that AI demand quality matters is correct and under-covered.

Where he goes too far, in our view, is treating every beneficiary of the buildout as equally fragile. Nvidia is not Nebius. Palantir is not Oracle. One sells scarce picks and shovels with a net-cash balance sheet; another finances the mine; another sells software with little balance-sheet strain. A system can be overbuilt without every stock being an equally good short.

On the long side, the pattern is classic Burry: buy businesses facing low expectations and short businesses carrying heroic expectations. Adobe, JD and Lululemon express that idea best. Molina needs a larger earnings recovery than its “cheap” label suggests. Flutter is a smart regulatory bet but not a simple value stock. Fannie and Freddie are policy options masquerading as equities.

The durable takeaway is not “copy Michael Burry”. It is this: expectations are a hidden liability. When a company combines high expectations with heavy fixed commitments and outside financing, a small operating miss can become a large equity loss. When a company combines low expectations with cash generation and disciplined capital allocation, merely surviving can create upside.

That framework will remain useful long after Burry’s next trading post changes the tickers again.