Depreciation Season: Burry's Short Targets the AI Infrastructure Accounting Gap

The 13F hit the SEC feed on a Tuesday afternoon, and the market took a beat to process it. Michael Burry — the man who called the subprime collapse — opened short positions against two AI cloud infrastructure names. Nebius and Oracle.

Not Nvidia. Not the entire sector. Two specific companies, two different balance sheets, one shared vulnerability. The numbers don't spell 'bubble'; they spell 'depreciation mismatch.'

This filing lands forty-five days late, as all quarterly 13F disclosures do. Burry's positions were already built while the rest of us were watching headline momentum. By the time the paperwork surfaces, the trade has matured. The real signal is not the short itself — it is the forensic reasoning the position implies.

Context

Nebius emerged from the implosion of Yandex N.V., re-domesticated in Amsterdam, and re-listed on Nasdaq in late 2024. It is the purest AI-cloud bet on the board: no legacy software cash cow, no diversified enterprise annuity. Just GPUs, data center capacity, and a growth story that NVIDIA's balance sheet helped endorse. Nebius is, in effect, an NVIDIA proxy with a tech startup's operational volatility.

Oracle occupies the opposite end of the scale. OCI — Oracle Cloud Infrastructure — became the AI hyperscaler narrative of 2024 and 2025. The company announced aggressive expansion plans. The market repriced Oracle as if its cloud division had inverted the enterprise software stalwart's entire multiple. The xAI order rumor reportedly landed near ten billion dollars. The equity priced in a future where Oracle's AI ambitions outgrow AWS-style dominance, not merely an incremental cloud share.

Two companies, different sizes, different cash flows, different investor bases. And Burry shorted both.

Core: The Depreciation Ledger

The pairing is coherent only when you examine the accounting trail.

Oracle's depreciation expense has ballooned in step with its AI data center buildout. Capital expenditures keep climbing, but the revenue curve lags the cost curve by several quarters — the classic infrastructure-investment front load. The question is not whether AI cloud revenue grows. It is whether the growth arrives before the depreciation treadmill makes the income statement run out of air.

Nebius has no legacy cushion. Its valuation sits entirely on future AI compute demand. Its cash burn is real, its revenue base is thin relative to capital commitments, and its stock carries a beta that amplifies every AI sentiment shift. A pure infrastructure bet with no margin of safety built in.

My own audit experience shapes how I read this. In 2024, I led a team building a dashboard that tracked five hundred institutional wallet clusters through the Spot Bitcoin ETF approval process. We analyzed $2.3 billion in pre-approval accumulation patterns. The one discipline I carried into that work and now into AI infrastructure analysis: verify the fixed asset assumptions before you trust the profit line. The numbers on the income statement are only as honest as the depreciation policy supporting them.

Here is the mechanical core of the trade.

When a cloud provider purchases an H100 GPU cluster, it capitalizes those servers as fixed assets. The typical useful life assumption runs three to five years. Straight-line depreciation spreads the cost across that window. The infrastructure boom of the past eighteen months — across Oracle, Nebius, Microsoft, Google — has generated billions in new fixed assets booked under these assumptions.

Meanwhile, NVIDIA's release cadence runs twelve to eighteen months per generation. H100 to H200. H200 to Blackwell. Each iteration delivers meaningfully better inference throughput per watt and per dollar. The financial assumption embedded in a five-year depreciation schedule is that the hardware retains economic utility for its full accounting life. The technical reality is that an H100 cluster loses competitive positioning the moment Blackwell enters volume production.

This is the depreciation gap. The gap between the accountant's schedule and the technologist's reality.

If a customer signs a long-term compute contract anchored to an H100 fleet, that contract's relative economics deteriorate as newer, denser hardware enters the market. The cloud provider remains obligated to operate the old fleet: power, cooling, maintenance, staffing — all costs that retain their weight while the revenue per workload shifts toward the newer nodes. The utilization rate determines whether the old hardware still earns its keep. If utilization drops as customers migrate to Blackwell-based capacity, the fixed asset becomes a financial anchor, not a revenue engine.

Burry did not need inside information to see this. He needed the depreciation footnote from the quarterly filing and a simple comparison against NVIDIA's product roadmap. The disconnection between those two documents is the entire trade.

Trace the outflow.

Oracle guided capital expenditures persistently upward across 2024. Free cash flow conversion wobbled as the data center buildout consumed operating cash. Nebius, meanwhile, burns capital to expand GPU-backed compute capacity across Europe and North America, with revenue that remains small relative to the cost base. Both companies are using the AI narrative to raise capital — for Oracle, the narrative is a supplement to a mature software business; for Nebius, it is the entire investment thesis.

The deeper issue is the range of judgment embedded in depreciation policy. Extending an asset's useful life by two years can add billions to reported operating income without altering a single physical fact. Shortening the same schedule to reflect actual technological obsolescence would slash reported profits and trigger a repricing. Companies always face the temptation to smooth the curve. Auditors rarely push back hard unless the gap becomes indefensible.

Floor broken. Liquidity drained — not from the market, but from the credibility of the accounting.

Oracle's AI revenue growth is real, but it runs against hyperscaler competitors with deeper integration ecosystems and pricing scale. AWS, Azure, and GCP hold the customer lock. Oracle's supercluster wins on speed of delivery and GPU availability, not on cost efficiency. Its depreciation burden, relative to its cloud revenue base, sits heavier than AWS's. Burry is not attacking the technology. He is attacking the distance between the valuation and the comparables.

Nebius trades as a leverage play on AI infrastructure sentiment. Its revenue is a fraction of its fixed asset base. The company embodies the speculative intensity of the current investment cycle. A high-beta balance sheet with a single narrative supporting it is precisely the kind of target a forensic short seller identifies.

Contrarian Angle

The market's simplest conclusion — 'the Big Short guy thinks AI is a bubble' — misreads the trade.

Look at what Burry did not do. He did not short NVIDIA, the primary beneficiary of the infrastructure buildout. He did not short the entire AI complex. He shorted two specific balance sheets with different operational risk profiles and one shared vulnerability: the gap between depreciation assumptions and technical obsolescence. That is a trade on accounting, not a referendum on the technology.

The 13F lag matters here. The disclosed positions were established in the prior quarter. The market has already moved. Oracle's stock may have already corrected partially since Burry entered the trade. The disclosure confirms the thesis; it does not reveal the entry price. Any retail trader attempting to piggyback on the disclosed position is entering a trade that is already several months old.

The more nuanced possibility: Burry may be partially correct in ways that create widespread damage without AI being a bubble. If his position triggers a broader scrutiny cycle — analysts auditing depreciation schedules, investors demanding shorter useful-life assumptions, regulators questioning aggressive accounting — then even fundamentally sound AI infrastructure names will be repriced. The collateral impact could exceed the direct impact on the shorts' targets.

Counterbalance this against the counterarguments. Oracle's enterprise software business generates substantial cash flow that funds its AI expansion, providing a cushion unavailable to pure-play startups. AWS demonstrated that cloud economics can generate durable returns once scale is reached, and Oracle's OCI — though smaller — has been growing revenue at significant rates. GPU supply has historically been tighter than the oversupply thesis implies. The short could be a targeted bet on financial narrative displacement rather than a prediction that AI infrastructure fundamentally collapses.

Takeaway

This is a signal, not a verdict. The trade entry window is closed; the disclosure is old news by definition. The fresh information will arrive in the next ninety days.

Watch Oracle's upcoming earnings for the relationship between depreciation expense in absolute terms and OCI revenue growth rate in the current quarter. If depreciation grows faster than cloud revenue for two consecutive quarters, the accounting gap is widening in real time. Watch Nebius's gross margin trajectory and cash burn per revenue dollar earned. Healthy infrastructure companies show the burn rate narrowing as revenue scales. A widening gap validates the short.

Watch institutional follow-through. If other large funds have built similar positions — and they will be revealed in the next 13F cycle — the AI infrastructure repricing becomes a coordination event, not a solo bet.

The data is live. The interpretation discipline is yours. Trace the outflow. The arbitrage window: Closed.