'Fingers already on the trigger': why is a Big Short investor betting against Nvidia?

Nvidia's record sales in the latest quarter give the market the impression of sustained demand for AI infrastructure - but also mask its fragility, writes investor Michael Burry, known as the prototypical "Downgrade Game" hero, on his Cassandra Unchained blog. Nvidia depends on a handful of major customers, but their purchases could be inflated by a temporary phase of AI testing and training, he believes. When that phase ends, orders for chips and external capacity could slow - and it's not just Nvidia itself that's vulnerable, but the entire infrastructure chain built in anticipation of constant demand, from chip assemblers to data centers and their lenders.

Burry is one of the staunchest critics of the market euphoria around AI that has fueled the securities of chip makers and other companies in the sector. He opened a short position in Nvidia at the end of 2025.

Nvidia's business is vulnerable due to "off the charts" customer concentration, writes an investor who has studied the tech giant's latest report. He points out that 64% of accounts receivable in the last quarter were formed by only the three largest customers. Such concentration, Burry says, threatens to put Nvidia under pressure even without a broad cooling of the AI chip market. In the dot-com era, even a less concentrated customer base didn't protect Cisco from being hit after demand reversed.

For example, if Microsoft, which Burry sees as Nvidia's likely largest customer, cuts capital spending on the company's chips by 20%, the company's revenue would fall by 4.2%. With Nvidia's capitalization above $5 trillion, $182 billion in future purchase commitments doesn't seem like much. But Burry points out the scale of the risk: the amount exceeds the company's annual operating cash flow, and 65% ($119 billion) of it is made up of commitments from just one customer.

In addition, Nvidia binds itself in advance with purchase commitments and dedicated TSMC lines to meet the expected demand for its chips. If large customers start to buy less, some of this capacity and commitment may become excessive. Therefore, customer concentration is not only dangerous for revenue: the production chain, which is designed for long-term high demand for AI infrastructure, also becomes less sustainable.

Burry believes that the market has not yet put much of that risk into Nvidia's stock price. The company's 50-day average trading volume has fallen to its lowest level since 1999, and put options remain relatively cheap - meaning investors, he believes, have little insurance against a possible drop.

But the problem Burry points to is broader than one company. The investor questions the very structure of the AI boom. He believes that today's chip and capacity purchases have a lot to do with the early stage of AI adoption, when companies are testing models, collecting data on how employees work with AI, and refining their own AI tools to be less dependent on external solutions.

After testing and data accumulation, compression will come, Burry believes, as companies reduce their reliance on third-party AI services and more accurately distribute the computing load. So for Burry, current purchases of chips, data centers and cloud capacity are not a reliable guide to future demand: they may reflect a peak in testing and data collection costs, after which external monetizable demand for AI capacity will begin to compress.

According to a Ropes & Gray estimate cited by Burry, the market will need about $275 billion over the next two years for data centers already under construction - to refinance construction loans.

As long as investors believe that capacity will be utilized for years, this can be done through bonds, private loans or securities backed by future data center revenues. But if utilization expectations weaken, refinancing construction loans will become more difficult: financing will become more expensive, and it will be harder to attract long-term capital. Projects whose debt model was designed for higher demand for AI computing will then come under pressure.

Burry calls this chain reaction the "whiplash effect": high demand first accelerates construction and debt financing, while when the market slows, those who have already invested in capacity expansion lose.

This article was AI-translated and verified by a human editor