The Big Short: Is Michael Burry right about the AI trade?

Note CoreWeave, which suffered a big hit this week after its earnings guidance disappointed, has done similar. In January 2023, CoreWeave extended the depreciation period for its GPUs from four years to six.

Data centre GPUs running AI workloads with utilisation rates of 60-70% have useful lifetimes of one to three years, according to a 2024 report.

CoreWeave’s chips are mainly from Nvidia’s Hopper generation, such as the H100, which were state-of-the-art in 2023 and 2024

Moreover, Nvidia and others keep coming up with new chips that outclass previous generations.

The new Blackwell chips boast 40 times the performance of Hopper, according to Nvidia chief executive Jensen Huang. These are in full production and will be outclassed again by the Rubin upgrade next year, expected to be three times more powerful. Rubin Ultra is expect to double performance again when it is released.

To understand Burry’s critique, one needs to grasp how depreciation works in this context, and why he thinks it’s being mis-handled.

Normal depreciation logic

When a company buys a capital asset (e.g., servers, data-centre hardware, AI chips), it doesn’t expense it all at once; instead it spreads (depreciates) the cost over its useful life.

If the useful life estimate is too long, depreciation expense is too low → reported earnings are higher (because less charge each year).

In contrast, if the useful life is realistically shorter—due to rapid tech obsolescence—then depreciation expense should be higher and earnings lower.

Burry’s argument in this context

AI infrastructure (chips, GPUs, servers) is evolving extremely rapidly (for example, new generations of AI chips from Nvidia Corporation keep appearing).

Hyperscaler companies (major cloud/AI infrastructure providers) are pouring enormous capex into AI hardware.

Yet they may still be depreciating equipment over 5-6 years (or longer) rather than a more realistic 2-3 years given replacement cycles.  

By extending the useful life, they under-charge depreciation, making earnings “look” better than the economics imply.

Burry estimates that across the big players this understatement inflation could amount to ~$176 bn of earnings between 2026-28.  

He views this not as just a minor accounting quirk, but as a structural mis‐valuation risk: if earnings get revised or hardware replacement becomes more frequent than assumed, the “earnings bubble” could burst.

Why it matters for valuations

Many AI/hyperscaler stocks are trading on high multiples, expecting strong profit growth and continued large capex leverage.

If those earnings are overstated (or if capex doesn’t convert into commensurate returns), then the current valuations may be unsupported.

In that sense, Burry sees a “fat tail” risk: the downside is larger than the consensus is discounting.

Burry’s reputation and why his view matters

Michael Burry made his name by correctly forecasting the US sub‐prime housing collapse and profiting from the fallout via his hedge fund Scion Asset Management.

Because of that track record, his contrarian moves attract attention: when he says “sometimes the only winning move is not to play”, it carries weight.  

The AI boom has become one of the dominant themes in global markets (hyperscaler infrastructure build-out, chip demand, AI software monetisation). Burry is saying there’s a bubble (or close to one) and he’s betting against parts of it.

On the broader market, he sees the AI beat-up as increasingly reminiscent of a bubble: large valuations, heavy hype, massive capex commitments, and potentially weaker returns once the cycle turns.

The capital cycle always ends with a shakeout: Let’s look at a similar period in the recent past – 2002/03 just after the dotcom bubble burst. In a separate post on X, Burry flags a passage noting that by 2002 less than 5% of US telecoms capacity was in use.

So far he’s taken out notional $1.1bn bets against Nvidia and Palantir and is hinting at more in this space.

The size of the disclosed short bets, especially on Palantir, signals strong conviction. The use of put options is consistent with a view of “big downside if things go wrong” rather than just a modest hedge.

These are two short bets disclosed already are not part of the hyperscaler depreciation thesis.

Burry however says to look out for more info on November 25th.

Now for the explainer bit...

The broad thesis is pretty clear - if the AI consumer/hyperscaler trade turns down, these positions benefit; if AI continues strong without disruption, the risk is that he is early or wrong.

As is typical with contrarian trades, timing matters—he may be early, and if markets stay frothy, the trade may underperform (or cost premium) for some time before payoff.

Assumptions

The useful life of AI-hardware assets is much shorter (2-3 yrs) than those companies use (5-6 yrs or more).

Depreciation is being intentionally or structurally underestimated, thereby overstating profits.

The AI investment boom will hit a turning point: either capex growth slows, utilisation falls, replacement cycles shorten, or returns disappoint.

The market will at some point recognise these issues and valuation multiples will compress (or growth expectations will be revised downward).

Risks / Counterarguments

Some old generation chips remain in use (for cloud, legacy workloads), which could justify longer asset lives. Indeed analysts cite that older Nvidia A100 chips (introduced 2020) are still contracted.  

AI may still deliver strong growth and returns, making the depreciation risk less relevant in the near term. If hyperscalers generate high incremental profits despite capex, then earnings may hold up.

Timing risk: Even if Burry is correct, the “realisation” event (e.g., earnings revision, capex slowdown, demand drop) may be years away. Investors shorting early may lose while being “right but early”.

Market sentiment: The AI trade is strong in the current momentum environment; even if valuations are high, momentum can carry the theme further.

Complexity in filings: Put options’ cost, strike, expiry matter a great deal. Not all puts translate into large gains unless correctly timed and executed.

Timing & tactical indicators to watch

Earnings guidance from hyperscalers (Meta, Amazon, Microsoft, Oracle) — look for signs of weaker-than-expected capex returns or higher depreciation/amortisation charges.

Capex disclosures: If hardware refresh cycles accelerate (2-3 years) or replacement needs increase, companies may raise depreciation expense.

Filing revisions / restatements: If companies revise their asset lives downwards, this could trigger recognition of “hidden” cost.

Stock price action of the shorted names (especially Nvidia and Palantir) — large downward moves or multiples re-rating would validate the trade thesis.

Broader AI hype vs fundamentals: If AI infrastructure build slows (e.g., fewer contracts, slower enterprise uptake), then hyperscaler earnings may disappoint.

This material is marketing content and should not be regarded as investment advice. Trading financial instruments carries risks and historic performance is not a guarantee of future results.
The instrument(s) referenced in this content may be issued by a partner, from whom Saxo receives promotional fees, payment or retrocessions. While Saxo may receive compensation from these partnerships, all content is created with the aim of providing clients with valuable information and options.