The headline writes itself: Michael Burry shorts AI infrastructure. Scion Asset Management files a 13F, and two names leap out from the otherwise quiet row of holdings. Nebius. Oracle. The market panics. The content machine churns. The phrase "betting against AI infrastructure boom" goes viral. Everyone misses the point.
In 2018, I spent three months line-by-line auditing 0x Protocol v2. Seven integer overflow vulnerabilities. Zero praise from the community, zero PR value. But the code was wrong, and I knew it before the exploit landed. That discipline — read the source, ignore the story — has paid for my career longer than any single trade.
So let me read the source here. The 13F filing tells us three things. First, Burry holds put options on Oracle and Nebius. Second, the market interprets this as skepticism toward AI infrastructure. Third, the underlying reasoning is financial overvaluation and, crucially, a challenge to asset depreciation practices.
"Depreciation" is the word most news consumers will skip. It is an accountant's term. It is dry, non-cash, and buried in the footnotes. That is precisely why it is the true battleground.
The story everyone wants — Burry thinks AI is a bubble — is a fairy tale. The real trade is narrower and more brutal. Burry thinks both companies are overstating current earnings by stretching out the cost of hardware that will be obsolete before it is paid off.
We do not predict the storm; we short the rain.
This is a short on balance sheets. Not on technology. Let me break down why, and what it means for anyone holding AI-exposed portfolios — including the crypto-native side of this market, where the same accounting disease is running rampant.
Context: The Two Names and the Boom
Nebius Group. Formerly the global business of Yandex N.V. Re-listed on Nasdaq in 2024. Backed by NVIDIA — the company that prints the shovels for this gold rush. Nebius operates AI-native cloud infrastructure: GPU clusters, model training, inference workloads, and platform tooling. It has no legacy revenue to cushion the landing. Its valuation rests entirely on AI cloud growth.
Oracle. The 800-pound enterprise software gorilla. Databases, ERP, financial systems. All real. All cash-generative. Oracle Cloud Infrastructure has spent the last two years selling "AI superclusters": massive GPU deployments for companies like xAI that need compute at scope few can match. Oracle's narrative is that its AI backlog is enormous and expanding.
The market groups both companies because both are hostages to the same macro wager: that AI capital expenditure will continue to explode, and that those investments will generate sufficient future revenue. The deployment of that wager is visible in every earnings call. CFOs drone on about "capex intensity." Analysts nod. Nobody asks the question that matters: what is the useful life of that hardware, and whose books assume a longer life than physics allows?
The broader context is the AI infrastructure boom itself. Hyperscalers and specialized clouds are ordering GPUs by the gigawatt. Data center construction has become a global industry. NVIDIA's H100, H200, and Blackwell architectures sell out. Every operator — Oracle, Nebius, and dozens of others — races to lock down supply. The assumption embedded in every purchase order is that demand for AI compute will keep outpacing supply for the life of the asset.
That assumption was comfortable when the narrative was young. Now a famous short seller has looked at the two least cushioned versions of that assumption and concluded the insurance is cheap.
The financial mechanics matter more than the narrative. This is where the analysis splits from mainstream coverage, and where I need to be clear about evidence quality.
The original news wire is thin. It passes only four information points: the short positions, the AI infrastructure suspicion, the valuation concern, and the depreciation red flag. My confidence in each layer is weighted accordingly. The core fact is certain. The interpretation is inference grounded in public market knowledge. Treat the conclusions as probabilities, not prophecies.
Core: The Depreciation Game
The Accounting Lever
When a company buys a GPU cluster for $100 million, the whole amount does not hit the income statement in the year of purchase. Under accrual accounting, the hardware is capitalized as a fixed asset. Its cost is then spread over an estimated useful life through depreciation.
This is standard practice. The problem is the estimates. Assume five years of useful life and $100 million in GPUs: the annual depreciation charge is $20 million. Assume three years instead: $33 million. The difference — $13 million per year — flows directly to the bottom line as higher reported earnings.
Now scale that to a company spending billions, and you can manufacture profit out of an accounting policy.
That is not illegal. It is also not honest in an economic sense. The cash left the company when the purchase order was signed. The physical asset obsolesces at the speed of chip design, not the speed of GAAP. NVIDIA's architecture cycle is twelve to eighteen months. An H100 purchased in 2023 is functionally inferior to a 2025 Blackwell cluster on nearly every dimension: memory bandwidth, power efficiency, training throughput.
The fair value of that H100 cluster does not decay linearly over five years. It collapses in a step function the day the next architecture ships.
This mismatch is the core of Burry's thesis.
The Non-Cash Fallacy
When depreciation comes up on earnings calls, management has a reflex. "Depreciation is non-cash." Retail repeats the mantra. It is technically true and spiritually misleading.
The cash left the building when the hardware was ordered. Depreciation is not a fake expense; it is a deferred acknowledgment of a real one. Operating cash flow already reflects the outflow. Investors who ignore depreciation are claiming that an asset will retain its economic value, which is exactly the assumption under scrutiny.
I have watched this logic destroy portfolios in crypto. Yield farms advertise triple-digit APYs. Token emissions are called "incentives," not dilution. Users pile in, treating the subsidy as protocol value. Then emissions fade, TVL leaves, and the token resets to its economic floor. The subsidy was a deferral of consequences, not a source of alpha.
A generous depreciation policy is the same machine. It subsidizes current income at the expense of future accuracy. The bill arrives when the hardware is sold, impaired, or written off.
Leverage doesn't care about feelings. It also does not care about accounting fiction. Margin calls are denominated in cash, not EBITDA.
The Great GPU Overbuild
Here is the sector-level problem the short is targeting. Every major operator is building simultaneously. NVIDIA ramps production. New entrants order clusters. The supply curve is shifting right while the demand curve is still a hope drawn on a whiteboard.
The natural equilibrium for GPU capacity is not perpetual scarcity. It is surplus. That surplus will show up first in spot rental prices, then in contract renewals.
When I ran an NFT market-making bot in 2021, I learned that volatility without liquidity is a trap. I captured spread revenue for four months, then watched inventory draw down 60% when the market turned. The lesson was simple: the same asset class that rewards you in a bull market will punish you when the order book thins. GPU infrastructure is the same trade at a different scale. The revenue looks like spread income during a supply shortage. When supply arrives, the spread compresses, and the asset's carrying cost becomes the dominant term.
Burry is not predicting the end of AI. He is predicting the end of the supply shortage.
A Stress Test for AI Infrastructure
I built my early career auditing code. The same instinct applies to balance sheets. Here is the framework I use when I look at any AI infrastructure operator — public cloud, AI token project, or GPU lease fund.
- Depreciable life versus technology cycle. If a company depreciates hardware over five years while the chip vendor refreshes every eighteen months, the asset will be economically dead long before the books say so. The shorter the accounting life, the more conservative the policy.
- Cash conversion. Compare net income to free cash flow. If earnings consistently exceed operating cash flow minus capital expenditure, part of the "profit" is just the distance between accounting assumptions and economic reality. A widening gap is a warning flag.
- Capacity utilization. The best depreciation policy means nothing if servers sit idle. Ask what percentage of GPU capacity is contracted. Ask whether contracts are committed or cancellable. Utilization is the bridge between capex and revenue.
- Contracted revenue versus spot exposure. A five-year contract signed at peak pricing can justify a five-year depreciation life. Spot market rental cannot. When supply floods, spot prices decay, and the contracts written at peak rates become market anchors that drag everyone down.
Oracle and Nebius are both large-scale versions of this stress test. Oracle has a real cash cow underneath, which makes its accounting more credible but also more dangerous: the enterprise software business may be subsidizing AI capex before the AI cloud reaches sustainable profitability. Nebius has no such cushion. It is the purest expression of the wager.
The Davis Double-Kill Setup
Now the part that the "AI is a bubble" crowd gets wrong. Burry is not necessarily short AI adoption. He is short the intersection of narrative and accounting quality.
Oracle is the cleanest example. Its core enterprise software business trades at a reasonable multiple. But the AI narrative has pushed its valuation toward a growth premium that resembles a pure-play AI cloud. The supercluster deals are the hook. If those deals fill capacity and produce contracted cash flow, the stock grows into its multiple. If they stall — if customers delay deployment, if competitors undercut pricing, if utilization disappoints — the market re-rates the whole company based on the slower-growing software business.
That is a double kill. The AI premium evaporates while the core business's stability gets repriced as boredom.
Nebius is the mirror-image trade. High beta, low liquidity, all narrative. A short here is not a forensic analysis of a cash-rich giant. It is a leveraged bet on sentiment reversal. Burry chooses both because they represent the infrastructure trade's two extremes — one with a safety net, one without. The shared enemy is the depreciation line.
What My Own Track Record Tells Me
In 2020, I managed a $500,000 treasury for a synthetic asset protocol. I spotted the basis trade between Ethereum staking yields and liquid staking derivatives. I executed with leverage. I made 40% annualized before the market corrected. The lesson: efficiency in crypto markets is fleeting, and you must capture the window before the crowd discovers it.
The same is true of accounting-driven earnings. A too-generous depreciation policy is an inefficiency. The securities market will eventually discover it — through a short report, a whistleblower, an audit opinion, or a sudden impairment charge. The trade is not to predict the discovery. The trade is to recognize that the accounting is priced as if the assumptions were certain when they are anything but.
During the 2022 bear market, I transitioned to an options strategist role. I watched three major lenders collapse. I built structured credit protection strategies and generated alpha while the broader market bled. The lesson stuck: bear markets reward people who audit assumptions before they break. Burry's short is the same instinct at institutional scale.
The Crypto Parallel
Now the bridge to the industry where I spend most of my time. The AI infrastructure trade has a crypto-native twin, and it is equally infected.
AI-focused crypto projects — GPU marketplaces, decentralized inference networks, data provenance layers — carry the same accounting disease. Tokens are minted to fund GPU purchases. The GPU is capitalized as a "strategic asset," often with no disclosed depreciation policy. The project reports network growth in token terms, not in cash, because there is no cash. The market rewards the narrative of physical infrastructure while ignoring the decay of the physical equipment.
The pattern repeats monthly: a treasury buys hardware. Hardware produces token yield. Token price is the real product. When hardware becomes obsolete or the yield cannot cover electricity, the token collapses.
My Layer 2 work has taught me to be suspicious of over-weighted infrastructure claims. I have argued publicly that data availability layers are overhyped — 99% of rollups do not generate enough data to need dedicated DA solutions. The market spent billions building for demand that had not arrived. That is the same structural mistake as building AI data centers on five-year depreciation assumptions before customers exist. The infrastructure precedes the revenue, and the carrying cost compounds.
Burry's short is a warning to the broader capital allocation machine: when hardware outpaces revenue, someone eventually eats the depreciation. The question is who.
Contrarian: Where the Crowd Is Wrong
Let me be deliberately contrarian now. The crowd's response to this 13F is as wrong as the crowd usually is.
This Is Not a Short on AI
The single most misread element is the AI angle. Mainstream coverage frames Burry as a Luddite who wants to short the future. No. Burry is a financial analyst who identifies moments where the price of an asset assumes flawless execution. AI is real. The adoption curve is real. Enterprise cloud spending is real. What is not real — yet — is the assumption that every dollar spent on GPU infrastructure today will generate a competitive return before the hardware vanishes.
A short is not a statement about technology direction. It is a statement about price and timing. Oracle could dominate AI cloud and still be a bad stock at 35 times forward earnings. Nebius could become the leading AI-native platform and still lose 60% if the market re-rates its risk premium.
The crowd conflates technology with the ticker. That is a category error.
The 45-Day Delay Means the Trade Is Already Old
Another blind spot is the 13F's structural lag. Burry's positions are disclosed 45 days after the end of the quarter. The market is not reacting to a fresh trade. It is reacting to a stale snapshot. He may have built the position months ago. He may have already trimmed it. He may have added protection that cancels the entire directional exposure.
Retail traders who treat this as a live signal are acting on data with a built-in expiry. It is like reading an order book from 45 minutes ago and expecting the liquidity to still be there.
The professional response is not to copy the trade. It is to learn the reasoning and apply it to the current environment.
The Short's Own Weaknesses
A short is not a riskless arbitrage. Oracle has momentum: AI backlog, enterprise relationships, and a sales machine that can sign multi-billion-dollar commitments. If Oracle secures enough contracted capacity, its revenue growth will overwhelm the depreciation drag. The short thesis weakens with every new committed deal.
Nebius is a liquidity trap in the other direction. Low float, volatile price, potential for a single large buyer to squeeze the borrow. Shorting small-cap infrastructure plays is dangerous even when the accounting is clearly bad. The market can stay irrational longer than you can stay solvent — or liquid.
Burry knows this. He sizes accordingly. Anyone copying him without his balance sheet is an amateur.
The Opportunity the Crowd Misses
The real opportunity is not in the shorts. It is in the regeneration they trigger.
When a famous short seller scrutinizes depreciation practices, boards start sweating. Auditors ask harder questions. Investors demand better disclosures. This is a catalyst for accounting transparency. And transparency creates a premium for the companies that were already honest.
I have watched this cycle in DeFi. When a protocol is accused of manipulative token mechanics, the honest protocols stand out. Lending platforms with conservative collateral factors and proper stress testing gain market share. The same dynamic will hit AI infrastructure. Operators with conservative depreciation, disclosed utilization, and real cash flow will attract capital fleeing the narrative names.
The market is about to differentiate between two groups: those who are building for the AI era, and those who are just building.
The Burden of Proof
Here is the uncomfortable truth for both sides of the trade. The burden of proof has shifted.
Bulls who claim "AI capex will always pay off" must now explain why depreciation assumptions are conservative relative to the silicon cycle. Bears who claim "this is a bubble" must explain why actual customer demand and contracted backlog do not justify the current asset base.
Burry's short does not settle the debate. It reopens it on better terms. That is the only reliable information gain from a 13F filing: it forces the market to stop trusting narratives and start reading footnotes.
Takeaway: The Signals to Track
Leverage doesn't care about your thesis. Accounting cares about assumptions. The intersection of the two is where Michael Burry has placed his bet.
Track these signals over the next three to eighteen months.
The 13F filings. The next quarterly filing will show whether Burry added, trimmed, or exited. If he exits within two quarters, the trade was likely tactical. If he adds, the depreciation thesis is developing into a campaign.
Oracle's quarterly reports. Watch three lines: capital expenditure, depreciation, and cloud infrastructure revenue growth. The ratio of depreciation to revenue tells you whether the asset base is becoming a burden. If depreciation grows faster than cloud revenue, the accounting is eating the economics.
Nebius earnings and guidance. More important than revenue: utilization numbers, gross margin on GPU rental, and any disclosure of average asset life. A pure-play like Nebius cannot hide behind enterprise software. Its numbers will tell the truth quickly.
The AI compute spot market. GPU rental prices are the temperature gauge of the industry. If spot prices for H100 and Blackwell instances decline, the long-term contracts signed at premium rates become relics. If prices hold, the bear thesis is delayed.
Crypto AI projects with GPU treasuries. The same frameworks apply. Search for projects that disclose depreciation policies, publish utilization metrics, and conservatively estimate hardware life. They are rare. They will be rewarded.
The deeper lesson is one I learned auditing 0x v2 in 2018: the truth is not in the press release. It is in the bytecode. It is in the footnotes. It is in the assumptions that nobody reads because the narrative is more exciting than the math.
Michael Burry has published a weather report. He is early, maybe. He could be wrong, maybe. But the question he is forcing the market to ask — how fast does this hardware die, and whose books pretend it does not — is the correct question. Every AI infrastructure holder, public or crypto-native, needs an answer.
We do not predict the storm; we short the rain.
The rain is the accounting adjustment that follows the capex frenzy. It will come. The only open question is whether you are positioned for it, underwater from it, or dry on the other side.
Position accordingly.
