The GPU Depreciation Debate Behind This Week's AI Dip

#A Rough Week Reopens An Old Argument

AI chip stocks sold off sharply on Monday. On September 14, 2026, Nvidia fell more than 3%, Broadcom dropped nearly 5%, and AMD slid over 4%, dragging the semiconductor sector down sharply as investors weighed fresh doubts about how fast AI development, and AI spending, should really be moving. It is the kind of selloff that makes a quieter, more technical argument suddenly feel urgent. How long do the chips behind all that spending actually last?

That argument belongs to investor Michael Burry, best known for calling the 2008 housing crash. Since late 2025 he has argued that hyperscalers are flattering their earnings by depreciating Nvidia GPUs over far longer periods than the hardware actually lasts.

#The Case For Faster Depreciation

The servers and computing equipment that include Nvidia GPUs are generally depreciated by major cloud companies over roughly five to six years. Burry says the real economic life is closer to two to three years, given how fast Nvidia ships new chip generations, each one two to three times more efficient than the last.

Part of the bearish case rests on technological obsolescence. Newer GPU generations can deliver substantially more performance per unit of power and cost, potentially reducing the economic value of older hardware even while it remains usable.

Run that gap through hyperscaler balance sheets and Burry estimates it adds up to roughly $176 billion of understated depreciation, and overstated profit, across the industry between 2026 and 2028. He has put specific numbers on individual companies too, estimating Oracle's earnings could be overstated by close to 27% and Meta's by around 21% by 2028. He has backed the view with short positions on Nvidia and Palantir.

#Nvidia And Its Customers Push Back

Nvidia and its largest cloud customers have pushed back on that premise. CoreWeave's CEO has said H100 capacity coming off expired contracts was rebooked at around 95% of the original contract price, and Nvidia CEO Jensen Huang has pointed to H100 rental rates that jumped 22% in a single month to $3.28 an hour, evidence he says shows the chips are earning more, not less, as they age. Huang has also highlighted a CoreWeave contract to rent Nvidia's older A100 chips all the way through 2029, nine years after that chip first launched.

The two sides are looking at useful life from different angles. Burry argues that rapid advances in newer GPUs can make older chips economically obsolete faster than current depreciation schedules assume. Nvidia and CoreWeave counter that older GPUs remain heavily used and can still command strong rental prices years after launch. A chip can fall substantially in value and still generate plenty of rental income, so neither measure alone settles how quickly it should be depreciated.

#What This Means For Investors

For a retail investor, the practical point is simple. Some of the reported profit associated with AI infrastructure depends on accounting estimates about how quickly the cost of that equipment is recognized through depreciation. Depreciation schedules are a matter of judgment under accounting rules, not a fixed fact, and companies have moved them in both directions before. Meta extended some useful lives in 2025, Amazon shortened others the same year.

Nobody knows yet which side of this argument will prove right. If older GPUs continue generating strong rental income for years, that would support Nvidia's case that longer useful-life assumptions are reasonable. If it does not, and this week's selloff is a reminder that sentiment can turn quickly, future reported AI earnings could come under pressure if companies shorten useful-life assumptions or recognize impairments as older hardware loses economic value faster than expected. Either way, it is worth watching depreciation assumptions the next time a hyperscaler reports earnings, not just the headline profit number.