The useful lives of machines

Dan Bouk writes a short history of the most powerful idea in accounting: depreciation.

The historian of “things clouded in shrouds of boringness” explains that the bookkeeping convention was first conceived as a kind of actuarial table for machines.

In 1905, he says, a German official was among the first to apply the idea of a “survival curve” to industrial goods — estimating how long wooden telegraph poles were likely to remain useful in the same way insurance companies estimated how long their policyholders were likely to remain alive. 

The practice was slow to catch on, probably because accountants had so little data to work with. Railroad companies, for example, couldn’t possibly keep track of how long their millions of railroad ties remained useful (especially as they were made of all different types of wood and used in all different climates).

Still, companies were at least starting to think about it. Bouk quotes a railroad engineer warning that even perfectly functional equipment “is burning up as surely as is the coal which is shoveled under boilers.” 

Soon, everyone was forced to think harder. In 1909, the Supreme Court ruled that public utilities had to be guaranteed a profit, and that the profit had to account for depreciation.

“Before coming to the question of profit at all,” the majority opinion in Knoxville v. Knoxville Water Company read, “the company is entitled to earn enough to provide not only for current repairs, but for making good the depreciation of plant and replacing deteriorated portions thereof."

In other words, the cost of water in Knoxville had to be high enough to account for the finite lifespan of the infrastructure required to provide it. 

Bouk believes the idea had significance far beyond Tennessee: “It is powerful and important that governments affirm the necessity of reinvesting in productive capacity.”

For our material welfare, yes — and maybe our spiritual welfare, too. “There is something utopian in this arrangement,” Bouk adds. “It is an argument for a society’s collective responsibility to respond to decline and decay.”

And what could be more important than that?

More prosaically, the Knoxville decision prompted companies to record and document the decay of their industrial property. 

This gave accountants the data they needed to forecast the rate of decline of a company’s capital stock, just as actuaries had long done for people’s lives. 

The actuarial science of machines has preoccupied both accountants and investors ever since — never more so than in the age of AI.

The art of accounting

In December, Michael Burry accused the hyperscalers of using depreciation schedules to cook the books and mislead investors: “Understating depreciation by extending useful life of assets artificially boosts earnings,” he wrote on X, “one of the more common frauds of the modern era.”

The hyperscalers generally depreciate the GPUs they buy from Nvidia over the course of five or six years — their best guess for how long it will be until the chips either melt from overuse or become too antiquated to be of value.

Burry argued they would be obsolete in just two or three years — a reasonable assumption to make given how quickly AI tech has been advancing.

Even one of the companies he accused of book-cooking seemed to agree, at least directionally. Last year, Amazon shortened its estimate of the useful life of some of its AI hardware from six to five years due to “an increased pace of technology development.”

That may have been too conservative.

A deep-dive report by Azeem Azhar finds that global revenue in AI compute hit $25 billion in the first quarter of 2026 (ex China), exceeding the $21 billion the industry recorded in depreciation costs. 

In other words, the industry’s immense spending on AI capex — the spending that Burry said was fraudulently uneconomic — “just about clears the depreciation hurdle,” Azhar told Bloomberg.

So where did Burry go wrong?

It turns out the chips he thought would be hopelessly outdated in two or three years are still highly valuable after four. The Nvidia H100, for example, has retained nearly 80% of its original value from way back in 2022:

Even better, a report on secondary market prices for used GPUs suggests that the A100 — a six-year-old chip — has retained most of its original value, too.

This could change. Azhar cautions that profit margins are thin, that his data excludes financing costs, and that supply might catch up with demand.  

“The economics are holding,” he concludes. “But the margin for error is narrow.”

The market seems similarly wary.

Hyperscalers have vastly underperformed semiconductor makers in recent months, which suggests that investors believe that making and selling GPUs is a better business than buying and using them.

On the other hand, a new report from VanEck Digital Assets suggests the opposite: among bitcoin miners transitioning to AI infrastructure, the report says, investors assign higher valuations to companies that own and operate GPUs than to those that simply lease out space in their data centers.

In other words, hyperscaler investors appear more worried about depreciation schedules than investors in bitcoin miners.

This is unsurprising, because depreciation schedules have always been more art than science.

“They believed that good decisions can be informed by numbers,” Bouk wrote of the earliest practitioners of depreciation, “but ultimately require judgment.”

More than a century later, not even the AIs can tell you how long they’ll live.

— Byron Gilliam