AI bears have plenty to complain about.
Valuations. Massive capital spending. Circular financing. Questions about whether AI revenue will ever justify the hundreds of billions being poured into data centers.
But one of the more compelling arguments is much simpler…
Companies like Microsoft, Meta, Amazon, and others are spending staggering amounts on servers and data center infrastructure. And all that spending has triggered an increasingly important question:
How long will all that hardware actually be worth something?
Michael Burry—the investor famous for predicting the housing crash—put that question squarely in front of Wall Street late last year.
His argument was simple.
Companies are depreciating much of their server and networking equipment over five or six years. But Nvidia is releasing increasingly powerful chips at a much faster pace.
Burry argued that the economic life of AI equipment could therefore be closer to two or three years—and estimated Big Tech could understate depreciation by roughly $176 billion between 2026 and 2028 as a result.
Simply put, if Burry is right, the AI boom could be considerably less profitable than corporate earnings make it appear.
But we’re finally getting more real-world evidence about what happens to these expensive AI systems as they age.
And so far, that evidence is making the bear case considerably harder to prove.
Why depreciation matters so much
Depreciation is straightforward.
Say a company buys a server for $60,000. It generally doesn’t record that entire $60,000 as an expense on day one. Instead, it spreads the cost across the period it expects to use the equipment.
If it expects that server to last six years, that’s roughly $10,000 in depreciation expense per year.
If the server is really only economically useful for three years, that’s closer to $20,000 per year.
Same server, same cash expense… very different impact on reported earnings.
And Big Tech is making some consequential assumptions about those useful lives.
Microsoft recently extended the useful life of those assets from four years to six, saying software improvements and advances in technology allowed it to operate them more efficiently for longer.
Meta extended the estimated useful lives of most of its servers and network assets to 5.5 years beginning in 2025. That single accounting change reduced its 2025 depreciation expense by $2.92 billion and increased net income by $2.59 billion.
Oracle also increased the estimated useful lives of its servers and networking equipment from five years to six.
The investment case is simple: The longer these assets remain economically productive, the more reasonable those depreciation schedules look.
If they’re rapidly becoming obsolete, the opposite is true.
And with Nvidia constantly releasing faster, more efficient chips, the bearish argument makes intuitive sense.
Why would anyone want a six-year-old AI chip when something dramatically better is available?
Well, we’re starting to get an answer.
Six-year-old AI hardware is still finding buyers
Earlier this week, CNBC interviewed Shelly Li, CEO of Sprout, a company that manages the lifecycle of technology equipment after its original owner is finished with it.
Sprout refurbishes, resells, redeploys, and eventually recycles hardware for some of the world’s largest technology companies.
In other words, Li gets a firsthand look at what actually happens to AI hardware when it gets old.
And the numbers are surprising.
According to Li, Sprout has processed roughly 46,000 systems this year, including around 13,000 pieces of AI hardware.
Among the equipment increasingly coming through its doors are Nvidia A100 systems dating back to 2020.
Sprout has launched an entire business around refurbishing AI servers, offering certified systems containing Nvidia A100, H100, and even H200 GPUs to companies that need compute capacity.
Li told CNBC the industry is discovering that GPUs can last six, seven, eight, even nine years—far longer than many originally expected.
Moreover, Li says some hardware around three-and-a-half years into its life can still retain roughly 60%–70% of its value.
And Sprout isn’t the only evidence we’re seeing.
CoreWeave recently disclosed it has signed contracts to keep Nvidia A100 GPUs, which debuted back in 2020, in service through 2029.
That means customers are contracting for a generation of AI hardware roughly nine years after it first hit the market.
Suddenly, the assumption that every new Nvidia chip makes the previous one economically irrelevant looks much harder to defend.
Cutting-edge and obsolete are two very different things
Every new generation of Nvidia hardware will almost certainly be better than the last… But that doesn’t mean the previous generation suddenly stops being useful.
The latest and most powerful GPUs can go toward the jobs where performance matters most—training frontier AI models and running the most demanding workloads.
Older GPUs can move down the stack.
They can still handle inference… smaller AI models… internal corporate applications… less demanding workloads… and countless other computing tasks where having the absolute fastest chip matters far less than having economical access to compute.
We’ve seen this dynamic across technology for decades.
The latest iPhone makes last year’s model less desirable to someone who always wants the fastest device. But millions of people can still use the older phone perfectly well—and at a price that makes more sense for what they need.
Economic usefulness—not whether something newer exists—is what ultimately determines whether an asset can continue producing value.
There’s a huge reason old GPUs remain valuable
There’s also a structural reason older GPUs may retain their value longer than expected: The AI industry doesn’t have enough computing power to throw perfectly usable hardware away.
Demand for computing capacity continues to grow faster than the industry can build everything required to provide it.
And increasingly, the biggest constraint is the ability to power them. Data centers need electricity… grid connections… transformers… cooling equipment… networking… backup generation… and enormous amounts of supporting infrastructure.
Li emphasized that point during the CNBC interview, and it’s something we’ve been hammering for months.
You can manufacture another million GPUs… But you still need someplace to plug them in.
And that constraint could actually extend the economic life of existing AI hardware.
New Nvidia systems can demand dramatically more power and more sophisticated cooling than older A100 infrastructure. That can require major upgrades before a data center can accommodate them.
Meanwhile, existing A100 systems can keep generating revenue inside infrastructure that was already designed to support them.
That’s one reason CoreWeave says pricing for older GPU generations remains strong.
Simply put, the shortage of AI infrastructure may be making yesterday’s technology more valuable for longer.
And that changes the economics of the entire AI boom
None of this proves Big Tech’s AI spending will generate adequate returns.
That’s still the trillion-dollar question.
Microsoft, Meta, Amazon, Alphabet, Oracle, and their peers can buy hardware that lasts six years and still earn terrible returns if AI doesn’t ultimately generate enough revenue.
But useful life is an important part of that equation.
The bearish argument says today’s extraordinary AI capital spending could be even more expensive than it appears because companies will need to rip out and replace obsolete hardware faster than their accounting suggests.
The evidence we’re seeing now points toward a different possibility. Instead of becoming stranded assets every time Nvidia launches a new chip, older AI systems may simply move down the computing stack—continuing to generate revenue from inference and other less demanding workloads for years.
Meanwhile, shortages of power and data center capacity give companies another reason to squeeze every bit of productive life from the infrastructure they’ve already installed.
The bottom line: The AI bears can no longer simply point to Nvidia’s rapid product cycle and assume yesterday’s GPUs become economically irrelevant when the next generation arrives.
And if Big Tech can keep these incredibly expensive systems productive for five, six, or even more years, the economics behind the AI infrastructure boom could be considerably stronger than the bears expect.
This is exactly the kind of second-order AI trend we’re constantly watching at Curzio Research.
The biggest opportunities don’t always come from buying the most obvious name. They often come from understanding where the money flows next—into power, infrastructure, networking, and the companies helping keep the AI buildout running.
Want to find out where the money is headed next? Tune into Wall Street Unplugged.
