November 14, 2025. The SEC’s EDGAR system posts Scion Asset Management’s third-quarter 13F. Michael Burry—the man who shorted subprime mortgage bonds before the 2008 collapse—has exited his positions in Microsoft and Oracle. Both closed entirely. Not trimmed. Not hedged. Gone.
The market’s response is the actual story. Microsoft closed roughly 2.5 percent above its September 30 mark. Oracle closed about 8 percent higher. No violent repricing. No "The Big Short 2: Electric Boogaloo" narrative. Just a quiet data point absorbed by passive flows that no longer read individual 13Fs.
I have worked in this industry long enough to distrust tidy narratives. In 2017, I spent six weeks manually auditing Kyber Network’s Solidity before its token event. I found three integer overflow vulnerabilities in the rate-calculation functions that automated scanners had missed. That experience taught me something that applies directly to reading Burry’s filing: what the document says and what the system actually executes are two different things.
The 13F is a photograph of a portfolio as of September 30. It was published November 14. The market traded that entire interval in real time, and its verdict was: Burry’s exit is a portfolio decision, not a systemic signal.
I think the market is half right. And in crypto—where the AI-crypto convergence trades on narrative velocity rather than cash flows—that half-rightness creates a specific, quantifiable risk.
Context: The 13F Is Not What You Think It Is
To parse the signal, you first have to parse the instrument. Form 13F is a quarterly report of long equity positions filed by institutional investment managers with more than $100 million in assets under management. It arrives up to 45 days after quarter end. It includes only long positions and option positions. It does not include short positions, cash, private investments, or intent. It is, at best, a delayed, lossy compression of a portfolio’s state.
The deadline for Q3 2025 was November 14. The market knew this date. The information contained in the filing—the fact that Burry liquidated MSFT and ORCL—was priced over six weeks of trading, during which major indices were hovering near their highs.
Now consider what these two tickers actually represent in the AI stack.
Microsoft is the closest thing to a pure AI commercialization play in the large-cap universe. It committed roughly $13 billion to OpenAI. It owns a substantial portion of OpenAI’s economics through its stake and the consumption arrangement that runs OpenAI workloads on Azure. Every OpenAI chatbot interaction, every API call from a third-party developer, travels through Microsoft’s cloud. Azure revenue growth was in the mid-30 percent range on a constant-currency basis in the June 2025 quarter. Microsoft’s total capital expenditure, including finance leases, crossed roughly $110 billion in fiscal 2025. It is, in effect, an AI infrastructure fund with the spending profile of a small sovereign state.
Oracle is a different bet. Oracle was a legacy database vendor that dragged itself into the cloud with OCI. Its turnaround thesis rests on signing massive multi-year AI data-center contracts. Oracle’s remaining performance obligations—contracted but not-yet-recognized revenue—reached approximately $375 billion in May 2025. In fiscal 2026, Oracle plans to roughly double capital expenditure to about $16 billion. It is acquiring land, securing next-day power capacity, and signing gigawatt-scale lease commitments.
Two very different businesses. One common attribute: both are leveraged to the continued expansion of AI capital expenditure. Burry did not sell "Microsoft" and "Oracle." He sold two versions of the identical thesis—that AI infrastructure spending can keep compounding faster than the revenues it generates, for years.
My 2022 work reverse-engineering Arbitrum One’s state-challenge mechanism took four months. I wrote a 40-page specification on the latency implications of optimistic proofs versus zero-knowledge alternatives. The lesson from that work was about information lag: the longer the dispute window, the more stale the state, and the wider the gap between what users believe and what the protocol can prove. A 13F has a 45-day dispute window. Burry’s filing is a fraud proof submitted late, for a state that has already moved on.
Core Analysis, Part A: Decoding the 13F’s Limits
A 13F is a block header committed to a chain you no longer sync. It proves the state at time T. It proves nothing about time T plus 45 days. In crypto, we would never trade fair value off a six-week-old header—that would be malpractice. On Wall Street, it is standard practice. That asymmetry is where the confusion begins.
First, the size problem. Scion Asset Management controls on the order of half a billion dollars in qualifying assets. Microsoft’s market capitalization at the time of the filing was roughly $3.4 trillion. Oracle’s was above $500 billion. Even a $100 million Scion position is a rounding error against a $3.4 trillion company—approximately 0.003 percent of the float. Microsoft regularly trades $10 to $20 billion in volume per day. A $100 million exit is less than one minute of average volume. When headlines say Burry "dumped" Microsoft, they are describing a whisper in a hurricane.
Second, the disclosure problem. The 13F does not show shorts, index puts, or derivatives that offset the long-equity exposure. It does not explain why a position was closed. Burry has spent the past two years warning about index-level leverage and has periodically held put options on broad equity indices. A simultaneous exit from two megacaps could be a funding exercise, a hedge adjustment, or a reallocation into a more concentrated thesis. The filing alone cannot distinguish between these.
Third, the hiding problem. Managers can request confidential treatment for positions whose disclosure would harm an active strategy. The SEC routinely grants this. Burry has used this mechanism historically. The 13F we see is not necessarily the 13F that exists. What appears to be a clean exit may be a smokescreen; what appears to be an exit may be a re-entry executed after the snapshot date.
Fourth, the context problem. We are not told what he bought. If Scion rotated out of MSFT and ORCL into, say, smaller software names, cash, or energy, the signal is sector rotation, not AI pessimism. Without the full book, the headline is noise.
The verifiable fact is thin: at close of business September 30, 2025, Scion held zero Microsoft and zero Oracle common stock. That is the entire proof. Verify the proof, ignore the hype. The hype—"Burry thinks AI is a bubble"—is not in the filing. It is in the newsroom.
Core Analysis, Part B: The Capex Math Nobody Wants to Do
Let me walk through the actual numbers, because this is where the technical story lives.
The four largest hyperscale spenders—Microsoft, Alphabet, Amazon, and Meta—combined for roughly $230-250 billion in capital expenditure in 2024. In 2025, that figure is trending toward $370-400 billion. That is a year-over-year increase of roughly 55 to 65 percent.
What was the revenue payoff? AWS grew in the high teens to low twenties. Azure grew in the low-to-mid thirties, decelerating through the year. Google Cloud grew around 35 percent from a smaller base. Meta’s AI spend is funded by advertising growth in the low twenties. Weight these together and the aggregate cloud-revenue growth of the four majors lands somewhere in the mid-twenties. The capex growth rate is more than double the revenue growth rate.
Here is the ratio that matters: incremental capex divided by incremental revenue. In 2024, each incremental dollar of cloud revenue required roughly $1.5 to $2 of new capex. In 2025, the marginal dollar of cloud revenue is consuming closer to $3 to $4 of capex. That is not a profitable expansion curve. That is a long-duration infrastructure buildout where the payoff sits years out—which means the payoff is a function of interest rates, utilization rates, and the cost of electricity, not of AI enthusiasm.
I built quantitative risk models during the DeFi Summer of 2020. I ran 10,000 Monte Carlo simulations on MakerDAO’s collateralized debt positions under a 50 percent drawdown scenario. The simulations correctly predicted the liquidation-cascade risk in heavily leveraged positions. The general lesson carries over: when leverage extends into duration-sensitive assets, the trigger is rarely the first shock. It is the second. For hyperscalers, the first shock is a missed earnings target. The second is a guided-down capex curve that ripples through NVIDIA’s order book, the power supply chain, and every token with "AI" in its name.
There is a legitimate accounting defense. Data-center capex depreciates over five to seven years. The capacity built today can generate revenue for a decade. Training loads are finite, but inference demand is the long game. If AI adoption follows a smartphone-like S-curve, today’s capex will look cheap in 2027. That is the bulls’ case, and it is not unreasonable.
The bear case is not that AI will fail. The bear case is that the pricing already assumes the S-curve, the buildout has no slack, and the cost of capital is no longer supportive. NVIDIA’s gross margins sit near 70 percent, which means the hyperscalers are paying a monopoly tax for the privilege of building capacity. The bond market is the marginal allocator of capital, not the equity market. If 10-year yields stay in the 3.5 to 4.5 percent range, the discount rate on a 10-year AI infrastructure project is punishing. Every year the cost of capital stays elevated, and the value of a 2035 inference-dollar shrinks. Burry, a man who reads 10-Ks the way other people read novels, does not need a machine-learning PhD to run that calculation.
Core Analysis, Part C: The Fragility Chain and the Power Bottleneck
Let me be precise about where the fragility sits. The AI infrastructure chain runs in a specific order: GPU orders, data-center leases, grid interconnection, power purchase agreements, cooling, utilization, and inference revenue. The bottleneck has shifted along this chain every year.
In 2024, the chokepoint was GPU supply—TSMC’s CoWoS advanced packaging and HBM memory. In 2025, the chokepoint moved to power. The US interconnection queue contains more than a terawatt of generation and storage projects waiting for grid approval. Average queue times, according to Lawrence Berkeley National Laboratory data, run four to five years. New data-center capacity that was contracted in 2024 is landing in a grid that cannot deliver electrons until 2029 or later.
That is why Microsoft struck a deal to restart Three Mile Island. That is why Google signed a small-modular-reactor agreement with Kairos Power. That is why Oracle is reportedly buying land with existing power contracts at a premium. The AI buildout has stopped being a semiconductor problem and become an electricity problem.
The consequence is that capacity expansion is inelastic. Hyperscalers cannot respond to demand by simply ordering more GPUs. They must secure power, which takes half a decade of permitting, transmission work, and fuel procurement. Inelastic supply in the face of uncertain demand creates violent price swings when the demand side wobbles. A single large customer—say, an AI lab that does not renew its $10 billion compute commitment—creates stranded assets at a scale the market has never priced.
The second chokepoint is inference utilization. Training loads are finite; inference demand is the long-term payoff. The hyperscalers know this, which is why they push multi-year compute commitments and why Oracle’s RPO backlog ballooned to $375 billion. These contracted revenue numbers look like strength. They are actually a hedge against the utilization question. If enterprise AI usage plateaus, those contracts become negotiation leverage rather than revenue.
Now bring this back to crypto. Decentralized physical infrastructure networks—DePIN—are a relative-value trade on exactly this bottleneck. If hyperscaler capacity expansion is throttled by power interconnects, then decentralized GPU marketplaces that tap existing, already-powered hardware become economically rational, not utopian. Akash Network, Render Network, and similar protocols aggregate idle GPUs that are already sitting in data centers with power. Burry’s exit from hyperscale is not automatically bearish for DePIN. It is a relative-value argument: the marginal watt is cheaper where the infrastructure already exists.
Core Analysis, Part D: Burry as an Indicator—Right, Early, and Often Wrong on Timing
Burry has a record that is worth treating as a data table rather than folklore.
- He identifies the subprime mortgage market’s structural leverage and buys credit default swaps. The collapse arrives in 2007-2008. He is right, but he spends two years paying carry costs and watching his investors rage-quit.
- He shuts down Scion, telling investors he does not want to live through another cycle of markets doing "irrational" things. He is, by his own admission, early and exhausted.
2020-2021. He warns on GameStop mania and takes a position against Tesla’s valuation. He is early. Tesla keeps rallying into 2021. He closes at a loss. The subsequent 2022 drawdown vindicates the direction, not the timing.
2023-2025. He repeatedly warns that index-level concentration and passive flows are creating a leveraged, fragile equity market. The market keeps grinding higher. His 13Fs come and go. He buys puts on broad indices. Some pay, some expire.
The correct statistical description: Burry is a left-tail detector. He finds structural imbalances where the funding basis can break. 2008 was a mortgage market running at 30-to-1 leverage. The AI capex cycle is a credit-funded, duration-heavy supercycle. If the funding stops, the unwind is asymmetric and fast. That is a genuine risk.
But a left-tail detector does not tell you when the tail arrives. The market can remain irrational longer than you can remain solvent—this is not a platitude, it is a cash-flow statement. The crypto version is identical. Flipping a high-beta token position in response to a single 13F headline is a shortcut to buying volatility at the wrong strike. Burry can be right and the trade still lose money. That is not a contradiction. It is the definition of tail-risk timing.
Core Analysis, Part E: The Crypto-AI Downstream
The AI-crypto complex is a sprawling set of claims stacked on a single macroeconomic assumption. Compute marketplaces, agent frameworks, data-provenance layers, and prediction markets all assert some version of "AI plus an accounting layer." The problem is that most of these tokens are priced as 100-year options on a three-year-old narrative.
Look at the valuation gradient. Microsoft, at a 25 to 30 times forward earnings multiple, is too expensive for Michael Burry. Oracle, trading at decade-high multiples of forward earnings, is also too expensive. If a 30-times-earnings global monopoly on enterprise software is "too rich" for a deep-value investor, what does that say about an AI-agent token with zero revenue, a task-execution testnet, and a circulating market cap in the nine figures?
It says the token has no earnings yield, no cash-flow comparison, and no floor except the next buyer’s order. That is exactly the kind of asset that re-rates violently when the narrative velocity slows. The AI narrative is not a set of code; it is a set of cash-flow promises. Verify the proof, ignore the hype.
I spent much of 2025 evaluating agent-interoperability and decentralized-identity projects against basic cryptographic verification standards. Across three major projects, four out of five failed simple checks: authentication keys controlled by centralized operators, no dispute windows, no audit trails, and no verifiable execution environment. None of these projects were fraudulent. They were premature. The identity layer—the thing that would let an autonomous agent transact without a human in the loop—was an afterthought.
Here is the de-risking filter I apply to any AI-crypto project in a capex-tightening environment. Proof of usage: does it have paying customers, not just testnet addresses? Render has real render jobs. Most agent frameworks do not. Proof of compute: is execution verifiable through ZKML or a trusted execution environment, or is it a screenshot of a dashboard? Proof of treasury: does the project hold stablecoin reserves, or is its $200 million "treasury" denominated in its own token? The projects that survive a narrative contraction are the ones that look like infrastructure. The ones that die look like story.
Code is law, but bugs are reality. The bug in the AI-crypto thesis is the assumption that token demand equals compute demand. Token prices are a function of speculation. Compute prices are a function of electrons, silicon, and power. The two can decouple for long stretches, but a capex slowdown forces the convergence. And when it converges, it converges fast.
The Contrarian Case: What If the Market Is Right?
The crowd’s reading of Burry’s exit is bearish AI. Let me offer four counter-readings, because the technical data supports at least one of them.
First, Burry is not an AI skeptic; he is a value investor who only buys cheap. He missed the decade-long FAANG run because he could not pay growth prices. That discipline keeps capital safe, but it also produces systematic underperformance in secular shifts. AI in late 2025 could still be "early," like the internet in 1997, rather than "peak," like the internet in 2000. The technical data on AI adoption is ambiguous enough that a value investor’s blanket dismissal is a signal about his style, not about the technology.
Second, the 13F is a nine-week-old snapshot. Between September 30 and now, Burry could have bought everything back. Position entries and exits at his scale take minutes of market flow. Without his Q4 filing—which will not publish until February 2026—you cannot conclude he is "out" of the AI trade. You can only conclude he was "out" on September 30. The distinction is material.
Third, the market’s non-reaction is itself an information aggregation machine. The post-filing price action—Microsoft up 2.5 percent, Oracle up 8 percent—is the collective judgment of every quant fund, index rebalancer, and options market maker in the world. Passive index flows now overwhelm individual discretionary flows by an order of magnitude. A $100 million exit against $15 billion in daily volume is not a signal; it is a rounding error. Market participants priced Burry’s position many weeks ago, during the 45-day window in which the market was trading the expectation of the filing. By November 14, there was nothing left to trade.
Fourth, and this is the counter-intuitive part that most analysts miss: a hyperscaler capex slowdown is not automatically bearish for equities. If Microsoft and Oracle redirect capital from data centers to buybacks and dividends, earnings per share improve. Capital discipline is bullish for mature cash generators. Burry’s value lens may be signaling a rotation within equities—out of AI capex and into capital returns, energy, healthcare, or financials—rather than a collapse of equities. That is a sector-neutral read of the same filing. The bearish narrative requires assuming he is correct about the macro, when his filing only proves he made a micro decision.
The asymmetry here is the real trade. In an environment where hyperscaler capacity growth is constrained by power, DePIN networks with live hardware and measurable utilization gain relative share. In that scenario, you sell the agent-framework tokens with no hardware and check the GPU-hour data of the compute networks. The market’s dismissal of the Burry signal is incomplete. It is correct about Microsoft’s price. It is wrong about the marginal effect on the long tail of AI-crypto assets, where leverage is higher and liquidity thinner.
My 2024 custody work on the Bitcoin ETF complex reinforces this. I found potential single points of failure in the key-management architectures of major ETF custodians, based on public documentation and prior industry incidents. The market did not react. Institutional adoption has its own lag and its own risk profile—and the same is true of 13F signals. Institutions barely model the 45-day lag in filings. They read the headline and move on. That gap between headline and mechanics is where edge lives.
Takeaway: What to Watch Instead
Do not watch Burry. Watch the data points that actually move the cash-flow chain.
First, Microsoft’s fiscal Q2 2026 earnings in late January. Listen for Azure growth rate and the "capacity-constrained" language. If Azure AI revenue growth slows and capex guidance ticks down, the trade is confirmed by fundamentals, not by a 13F.
Second, Oracle’s next RPO update. If the backlog continues to grow faster than revenue recognition, the buildout is still rolling. If RPO decelerates sharply, the lease commitments are unraveling.
Third, NVIDIA’s data-center revenue. That is the canary. All hyperscaler capex eventually flows through NVIDIA’s order book. A sequential decline in data-center revenue is the earliest hard signal of a slowdown.
Fourth, the US grid interconnection queue and electricity prices. The power bottleneck is the real cap on AI capacity expansion. Any policy shift that accelerates grid approvals becomes the bull case.
Fifth, for crypto specifically: DePIN utilization rates—actual GPU-hours sold, not token volume. Render, Akash, and similar networks publish or imply utilization. If utilization rises while hyperscaler capex guidance falls, the relative-value trade I outlined is confirmed.
The tracking framework is not glamorous. It is quarterly earnings language, RPO conversion, order-book momentum, and watt-hour accounting. That is where the proof lives. Read the filings, not the headlines. The difference between Burry’s September position and the market’s November price is the same difference in crypto between a token’s narrative and its transaction history. Verify the proof, ignore the hype.
Code is law, but bugs are reality. The bug in the AI trade is not in the code—it is the assumption that buildout growth can compound faster than its cost of capital indefinitely. Burry trades on that bug’s existence. The rest of us get to choose which block we verify: the headline, or the ledger. In a bear market, that choice is the whole game.
