The Hook
The most valuable asset in the AI boom is not a GPU. It is a fiction called the depreciation schedule.
Michael Burry — the investor whose name is now synonymous with catching the 2008 subprime collapse — has opened short positions against two AI infrastructure companies: Oracle and Nebius. The headlines frame this as a bet against AI. That is wrong. The filing says something narrower and far more dangerous: financial valuations are stretched, and asset depreciation practices deserve a hard look.
Let me translate. In the accounting world, the moment a company buys a $40,000 GPU, it starts spreading that cost over years. Most AI cloud operators use a three-to-five-year lifespan. If those assumptions are wrong — if the GPU is economically obsolete in two years, or if rental prices collapse as GPU supply floods the market — the balance sheet becomes a warehouse of overstated assets. That is the exact point where a stock's narrative disconnects from its earnings.
The market doesn't care about your AI roadmap. It cares about whether the depreciation charge is eating more cash than revenue can replace. This is why Burry's short matters, and why it isn't what it looks like. The man who shorted mortgage-backed securities is not shorting artificial intelligence. He is shorting a specific accounting policy that lets AI infrastructure companies borrow against assets that may be repriced far sooner than their depreciation lines suggest.
Context: Two Targets, Two Fights
Two targets. Two different fights. The market is treating them the same because the ticker symbols follow the same AI infrastructure narrative. They do not share a story.
Nebius is the pure play. Born from the Yandex global carve-out, listed in late 2024, the company received a fresh balance sheet and a cloud-native AI mandate. It builds high-density GPU clusters, designs its own AI cloud platform, and leans heavily on its connection to NVIDIA — the chip giant participated in its private capital raise, cementing Nebius as the market's "responsible adult" among AI-native clouds. But Nebius is a small-float story stock. It does not have decades of enterprise software revenue to cushion a downturn. Its entire value proposition is built on the assumption that AI compute demand remains structurally scarce enough to justify premium rental pricing.
Oracle is the incumbents' AI gamble. Oracle Cloud Infrastructure, OCI, went from a laggard to a narrative monster in two years. The company signed a stream of massive AI contracts — reportedly including a $10 billion commitment from xAI — and repositioned its sales machine around "AI superclusters." Oracle's stock ran on the promise that its enterprise relationships, combined with NVIDIA GPU supply, would make OCI a legitimate challenger to AWS, Azure, and GCP.
The gap between the two narratives reveals Burry's strategy.
Shorting Nebius is a trade on valuation extension and retail momentum. The stock's story is built on future growth, and any negative signal — a delayed deployment, a missed utilization target, a financing dilution announcement — hits the multiple hard. Shorting Nebius via puts is a cost-controlled hedge against narrative fragility.
Shorting Oracle is a deeper position. Oracle's stock prices its entire portfolio, including legacy database and application businesses. To profit from an Oracle short, the AI-infrastructure narrative must reverse enough to outweigh the stability of the cash cow. This is not a trade. This is a thesis: the AI infrastructure capex supercycle is producing assets whose depreciation burden will outpace the cash flows they generate.
Both companies are spending at a scale that demands perfect execution. Oracle's capex run-rate has exploded into the tens of billions. Nebius is deploying capital as fast as its balance sheet allows. And neither has pricing power over the supply chain — they buy from NVIDIA, and they compete with Microsoft, Amazon, and Google for the same customers.
Why now? Because the market is at an inflection. Hyperscaler combined capex guidance crossed the $300 billion annual mark. Data center construction is running at industrial scale. The GPU supply that was the bottleneck in 2023 is becoming a wave in 2025. When supply catches up with demand, rental prices weaken. When rental prices weaken, the revenue side of the depreciation equation breaks.
Core: The Depreciation Weapon
We need to talk about the accounting weapon.
Under standard accounting rules, a GPU cluster is a fixed asset. It gets capitalized and depreciated over its "useful life." The useful life assumption is a judgment call — and that judgment call is load-bearing for some of the most valuable stocks on earth.
NVIDIA's flagship offerings have compressed the useful life of the previous generation. The H100 launched in 2022. The H200 was announced in 2023 and shipped in 2024. Blackwell arrived in late 2024. Each new architecture does not simply add performance. It makes the previous generation less competitive in the eyes of renters. A cluster of H100s is worth less today than it was a year ago — not because the hardware degraded, but because its opportunity cost in the market has risen.
Traditional servers have a ten-year depreciation runway because the hardware continues to serve undifferentiated workloads. GPUs are different. They are exposed to the fastest-moving technology cycle in compute history. The gap between accounting useful life and economic useful life is the quiet margin that Burry is targeting.
Let's put real numbers on this. Consider a GPU cloud operator deploying 10,000 H100 GPUs. If each GPU costs $30,000 including server integration, the hardware investment is $300 million. A four-year straight-line depreciation policy produces a $75 million annual depreciation charge. Now assume the operator signs contracts that rent this capacity for an average of $2.50 per GPU-hour at 70% utilization. Annual revenue reaches around $459 million. After direct energy and facility costs — typically 25-30% of revenue — the contribution margin is roughly $330 million. Depreciation of $75 million is covered. The model works.
But here is where the assault begins. Blackwell ramps, rental prices fall to $1.75 per GPU-hour, and utilization dips to 55% as newer clusters take priority. Annual revenue drops to $264 million. Contribution margin falls to $190 million. The depreciation charge is still $75 million. EBITDA compression is immediate — from roughly $255 million of headroom to $115 million. The stock trades on EBITDA, and the market reprices it violently.
Extend the same simulation to year four, when the company must either extend or replace that cluster. The replacement cycle arrives before the depreciation schedule ends. The logical response — refinancing via debt — only adds leverage to a balance sheet already carrying a cohort of under-earning fixed assets.
Based on my audit experience in the crypto mining sector, I have watched this exact playbook destroy leveraged balance sheets. Miners bought ASICs with three-to-four-year depreciation policies. The halving cycle, power price spikes, or a Bitcoin drawdown repriced those machines in months. Companies with positive book equity ended up impaired into insolvency. The AI cloud moment is the institutional-grade rerun of that lesson.
Oracle: The Incumbent's Trap
Oracle's financial structure makes it the more revealing target.
The company carries a significant legacy revenue base — databases, enterprise applications, and a services machine that generates high-margin, recurring cash flow. That base funds the AI experiment. But here is the structural asymmetry: the legacy business is stable, while the AI business is capital-hungry and marginally profitable at best. When a company with Oracle's history pivots its narrative toward cloud infrastructure, the market assigns value to the growth story, not the legacy anchor.
The accounting line that matters is the depreciation expense. In fiscal 2025, as Oracle converted billions of dollars into data centers and GPU clusters, depreciation climbed. The market cheered the revenue growth numbers without a full examination of how much of that growth was being consumed by the fixed-cost burden. If the depreciation charge grows faster than the cloud gross margin, the AI story starts consuming the cash cow rather than replacing it.
The relationship between contract booking and cash collection is another blind spot. Oracle's reported backlog includes large-scale AI cloud contracts with optimistic pricing assumptions. If GPU rental market rates fall — or if customers renegotiate as supply loosens — those backlog figures become forward-looking revenue mirages. EBITDA and backlog guidance may continue rising even as realized cash flow stagnates. That is exactly the kind of divergence that draws Burry's attention.
Then there is the financing architecture. AI infrastructure is increasingly debt-financed. Dedicated project vehicles, sale-leaseback structures, and asset-backed loans are funding GPU procurement. These structures are anchored to residual asset values. If the residual value assumption breaks — if the market determines that a three-year-old Blackwell-generation predecessor cannot retain 40% of its purchase price — lenders will tighten. The cost of capital rises for future deployments, and the balance sheet's leverage ratio jumps with it.
Here is the part most analysts miss: Oracle is not just exposed to its own capex. It has moved into the position of an AI-infrastructure bank. Customers do not pay the full cost of compute upfront. Oracle finances part of the deployment through multi-year commitments. When those customers are themselves AI-native startups with fragile funding, Oracle carries both the asset risk and the counterparty risk. A depreciation shock could become a loan impairment shock.
Nebius: The Amplified Beta
Nebius is the high-beta expression of the same trade. If Burry believed there were a deeply compelling, well-researched short thesis, he would be shorting Oracle. Nebius is an amplification trade — a smaller float, a more explosive upside during the AI boom, and a sharper downside if the financing environment shifts.
Nebius's balance sheet narrative is one of rapid deployment. The company raised capital at a premium valuation, gave itself a multiple-year runway, and committed to building out capacity ahead of demand. That is either a sign of confidence or the leverage that kills a pure play. In AI-native cloud, the fundamental question is utilization ramp time. A cluster that takes 18 months to reach 70% utilization loses money at the start and remains under pressure when depreciation charges begin.
The company's favorable narrative was reinforced by NVIDIA's involvement. But NVIDIA's investment should not be read as a guarantee of demand. It is a supply-chain alliance. NVIDIA wants its GPUs deployed, hyped, and proven. If Nebius cannot demonstrate sustainable rental demand, the alliance does not shield the income statement.
Nebius also carries a structural weakness relative to hyperscalers: it lacks the networking density, the global reach, and the enterprise sales ecosystem to match AWS or Azure on price and reliability. It is an infrastructure boutique in a market where scale is the ultimate survival skill. Not a technology problem — a unit economics problem. And unit economics are the theater where Burry's shorts operate.
The Stress-Test Simulation
I do not write about financial models from a distance. In my work building dashboards for on-chain liquidity and AI-agent trading strategies, I live inside these equations. So let me show the market what a Burry-style accounting stress test looks like. This is a compressed version of the model I built to evaluate AI infrastructure exposure.
Parameter set: - Hardware cost per GPU: $30,000 - Cluster size: 10,000 GPUs - Depreciation: straight-line, 4 years - Initial rental price: $2.50 per GPU-hour - Utilization at start: 70% - Price erosion: 15% per year - Utilization erosion: 5% per year starting year two - Direct operating costs: 25% of revenue
Year one: Revenue $459 million, opex $115 million, depreciation $75 million, EBITDA $269 million, EBITDA margin 58.6%.
Year three: Rental price $1.81, utilization 63%. Revenue $331 million, opex $83 million, depreciation $75 million, EBITDA $173 million, margin 52.3%. Still profitable — but margin compression is visible.
Year five: The model requires a refinancing/replacement event. Rental price $1.30, utilization 57%. Revenue $202 million, opex $50 million, depreciation $75 million, EBITDA $77 million — margin 38%. But by year six, the new cluster deployment requires $300 million of new capital. If the market values the asset base at a discount to book, the company raises capital at dilutive terms or stops expanding. Growth stalls.
The point is not the absolute numbers. The parameters are contestable. The point is the direction: every year of GPU-price erosion converts a seemingly safe depreciation schedule into a tightening vise. The market is pricing AI infrastructure companies as if rental prices remain sticky forever. Burry is betting that they do not.
Contrarian: The Transparency Trap
And now for the angle the market hasn't priced.
Everyone wants to know whether Burry is right. The more useful question: what if the short triggers the assumptions it depends on?
If scrutiny of depreciation forces Oracle or Nebius to accelerate their schedules — taking a one-time earnings charge to align accounting with physical reality — the near-term earnings hit would look like a bearish event. But a reset would actually improve future earnings quality and reduce the tail risk of a massive impairment. Burry's presence could force the very transparency that saves the stocks from a bigger crash later.
Compliance check: any forced adjustment to depreciation policy will require auditor review and potentially SEC comment letters. The moment a company signals a revision to useful life assumptions, expect a sharp one-time earnings reset followed by a clean baseline. This is the exact sequence that creates the most violent short-term moves — and the best long-term entries.
There is also the uncomfortable possibility that Burry is early. The AI capex cycle is driven by hyperscalers, not by the two targets of his short. Even if Oracle and Nebius collapse, the capital flowing to AI infrastructure can continue. The trade works for Burry, but it does not prove the collapse of the AI buildout. It proves that some operators are less resilient than others.
And there is another complicating factor: the 13F snapshot. The filing describes positions as of a specific date. Burry has repeatedly shown he adjusts quickly. A single snapshot is not a prophecy. The market is treating this public disclosure like the final word from a genius — but the genius has already moved on to the next position by the time you read this.
Takeaway: The Recalibration
The pivot is not a retreat; it is a recalibration.
In a sideways market — where every narrative trade gets chopped and where the daily move gives everyone whiplash — positioning matters more than direction. The AI infrastructure narrative has been one of the few trades that worked consistently for two years. Burry's short is the first institutional acknowledgment that the accounting foundation under that trade carries cracks large enough to bet against through options.
Speed is currency, but precision is the vault. The market rewards fast narratives; it rewards precise ones even more. Right now, the precise trade is not a simple copy of Burry's positions. It is a deeper analysis of depreciation lines, cash collection quality, and GPU residual values.
Watch the next Oracle earnings release. Watch whether Nebius reports utilization metrics rather than capacity commitments. Watch whether other major funds file AI-infrastructure shorts in the next 45 days. And when you see the inevitable headlines that "Burry is back," ask yourself one question: is he shorting the technology, or is he shorting the lie that the assets are worth more than the cash they generate?
That is a trade. This is the market.
