Michael Burry has doubled down on his shorts against Micron, Oracle, and Nebius, placing a high-profile bet that the artificial intelligence boom is barreling toward a cliff. His thesis rests on a familiar macro-accounting stick: if tech companies depreciate GPUs over five to six years when the hardware actually becomes obsolete in two or three, the entire neocloud sector is sitting on paper-thin profits. In Burry’s eyes, this is the dot-com bubble all over again, stretched out by creative accounting.
The problem with applying a classic short seller’s lens to this cycle is that it misses the physical realities of the current infrastructure squeeze. Earnings updates from Nebius and CoreWeave highlight pricing dynamics, contract terms, and credit underwriting that directly contradict the idea of rapid asset decay. Burry is treating high-performance GPUs like disposable consumer gadgets or application-specific Bitcoin miners, fundamentally miscalculating how compute, power, and cash flows behave in the real world
Real-Time Pricing Rejects the Crash Narrative
When an asset is truly on a fast track to obsolescence, market signals reflect it quickly. Resale values plummet, empty capacity builds up, and customers refuse to lock themselves into long commitments. In the GPU market, the exact opposite is taking place.
In its latest capacity auction, Nebius watched prices clear 15% above the highest levels it had ever recorded. From a previous baseline near $9.8 million per megawatt, short-notice hardware rentals surged to $40–50 million per megawatt—a four- to five-fold spike. A buyer does not pay a sky-high premium for temporary flexibility on an asset if they expect the underlying tech to be useless in 24 months. Pay-as-you-go buyers pay up precisely because the economic value generated by that compute remains massive.
The Ampere Reality Check
CoreWeave’s operational data offers an even tougher challenge to Burry’s thesis. CoreWeave recently secured a customer contract keeping Nvidia’s Ampere-generation A100 chips running all the way through 2029. The A100 launched in 2020. That deal locks in nine years of paid, productive life on a chip architecture that is already three generations behind Nvidia’s state-of-the-art silicon.
Why would anyone rent 2020 hardware in 2029? Two main reasons:
The Power Bottleneck: Modern flagship clusters require immense power densities and specialized liquid cooling. Thousands of standard air-cooled data centers simply cannot host newer, hotter racks without multi-million-dollar retrofits. Older GPUs living in air-cooled halls occupy scarce, grid-connected power that cannot be easily upgraded, giving them a durable demand floor.
Workload Migration: Frontier AI models get trained on the newest, fastest chips. But as models move from training to everyday inference, fine-tuning, and enterprise tasks, older GPUs take over. They don't get thrown away; they cascade down to simpler, cost-effective jobs
This workload shift isn't a new phenomenon. During the cloud buildouts of the 2010s, Amazon and Microsoft both extended server depreciation schedules from three years out to five or six. Critics back then called it accounting manipulation. It turned out to be an accurate reflection of physical reality—the hardware kept generating utility long after the original estimates expired.
Credit Markets Aren't Buying the Collapse
Burry’s thesis also hits a wall in the debt markets. When neoclouds buy GPU clusters, the initial three- to four-year customer contract usually pays off the debt used to finance the hardware. Once that loan is retired, the ongoing cost to run those chips drops mostly to electricity and basic maintenance. Any re-rental revenue after year four flows almost straight to the bottom line.
If lenders expected these chips to become valueless paperweights by year three, they would demand sky-high, junk-tier interest rates to offset default risk. Yet even through volatile credit conditions, specialized debt facilities for these platforms—and massive new capital frameworks, like Nvidia’s $500 billion GPU financing structure—continue to price at reasonable rates. The institutions whose entire survival depends on correctly pricing hardware default risk are looking at the exact same data and coming to the opposite conclusion as Burry.
Why the "Big Short" Mindset Is Missing the Mark
Michael Burry built his reputation finding structural cracks in over-leveraged, stagnant assets. But trying to apply that same framework to the AI buildout misunderstands how technology adoption curves scale.
Bitcoin ASICs become junk the moment a more efficient chip arrives because they can only perform one specific calculation. GPUs are flexible general-compute engines. As long as global demand for digital intelligence outstrips the physical availability of grid power, even legacy silicon remains a income-generating asset.
GPUs do not last forever, and intense physical strain from full-power training will eventually wear silicon down. But betting against the sector on the assumption that five-year accounting schedules are a fraud ignores real-world pricing power, multi-year contract renewals, and basic grid economics. Burry isn't catching an accounting scam he is fighting the structural realities of the physical grid
Comments