The artificial intelligence (AI) boom is in full swing, which means lots and lots of comparisons to the late-'90s dot-com boom and bust. One such investor who was around at that time is short-seller Jim Chanos, who became famous by shorting Enron before it crashed in 2001.

Chanos still sells short through his investment advisory Kynikos Associates, and recently took to X, formerly known as Twitter, to make a bearish case for one of this year's hottest IPOs, AI "neocloud" CoreWeave (CRWV -3.17%).

So, should CoreWeave's shareholders be worried? The bear case is worth noting, but CoreWeave management also has a serious argument to refute it.

What is the useful life of a GPU?

Last week on X, Chanos questioned whether CoreWeave was appropriately accounting for the depreciation of its AI GPUs, which are pretty much entirely Nvidia (NVDA 0.94%) chips. He wrote:

Chanos is basically saying that CoreWeave may be giving a "generous" useful life to its GPU chips, plus of course the ancillary infrastructure of buildings, networking, and memory associated with them.

For instance, if the useful lives of its GPUs were only half what CoreWeave estimates at, say, 3.75 years, that would double depreciation expense. And if depreciation expense were doubled, CoreWeave wouldn't have made an adjusted (non-GAAP) operating income of $200 million last quarter -- it would have rather had a $400 million operating loss.

In other words, if CoreWeave is overestimating the useful lives of GPUs, its current growth may only end up leading to greater and greater losses, not profits. It would akin to a bank growing its loan book a lot, only to have those loans go sour a few years out. Another example would be if an insurance company underestimates losses, it could show strong growth and profits for a while, until those losses came back to bite it later on.

Thus, it's perhaps not surprising that Chanos actually calls CoreWeave not a technology company but rather a "financial company" with "short-lived assets that truly depreciate."

Is CoreWeave being too generous with useful lives?

There's a good case to be made that CoreWeave may, in fact, be too generous estimating the useful lives of its GPUs. After all, Nvidia recently moved from introducing a new chip architecture every two years to every year. That means each year, Nvidia will have a newer, faster GPU architecture. The more rapid introduction could in fact cause older generations to become obsolete more quickly.

Furthermore, CoreWeave admits in its SEC filings that overestimating the useful life of its chips is a big risk to its business model. In the "risks" section of its recent quarterly report, CoreWeave says:

We continually work to upgrade and enhance our platform, solutions, and services in response to customer demand and to keep up with technological changes. Part of this process entails cycling out outdated components of our infrastructure and replacing them with the latest technology available. This requires us to make certain estimates with respect to the useful life of the components of our infrastructure and to maximize the value of the components of our infrastructure, including our GPUs, to the fullest extent possible. We cannot guarantee that our estimates will be accurate or that our attempts at maximizing value will be successful. Any changes to the significant assumptions underlying our estimates or to the estimates of our components' useful lives, or any inability to redeploy components of our existing infrastructure to extend past their contracted life could significantly affect our business, operating results, financial condition, and prospects.

Back before its March IPO, CoreWeave stated in its S-1 registration statement it believes the useful life of its infrastructure exceeded that of its typical contract, which range from two to five years. CoreWeave also stated it has structured its contracts with an average typical "payback period" of 2.5 years. That means if CoreWeave spends $1 billion on a data center, CoreWeave believes it will earn $1 billion in EBITDA from that data center over two and a half years.

Thus, profitability beyond that 2.5 year payback period depends on the "useful life" of the infrastructure. For instance, if the useful life of the infrastructure is 2.5 years instead of 7.5 years, then CoreWeave doesn't make any money, essentially. And since CoreWeave uses a lot of debt to fund that infrastructure buildout, it has additional interest costs to bear as well on top of that.

Lit up blue servers in a data center.

Image source: Getty Images.

But CoreWeave has a good defense -- for now, at least

In its S-1, CoreWeave maintained that it has successfully rerented older Nvidia A100 GPUs in new contracts with either the same or new counterparties after their initial contracts were up. Management also noted that older GPUs may be attractive for inference-related workloads or lower-intensity training for smaller models in the future.

On its recent second-quarter conference call, management continued to claim it has been able to continue renting A100s and H100s for inference and other lower-intensity jobs, saying specifically that older chips falling off-contract "are still being recontracted out."

The short case doesn't appear to be at hand... at least not yet

While training the latest and greatest large AI models will typically require the latest and greatest chips, it seems that inference and lower-cost training workloads offer CoreWeave plenty of opportunity to recontract its older chips at this point. Therefore, it doesn't seem that Chanos' theory holds up -- at least not yet.

However, if Nvidia's more rapid pace of chip introduction means older chips go out of style more quickly, it could potentially be a problem. And of course if there were to be more competition and oversupply across clouds and neoclouds, maybe with competition in the inference market, that could put pricing pressure on renewal contracts, which could eventually become a problem.

So CoreWeave's near future appears bright and intact amid AI hypergrowth; however investors should be aware the company will have to maintain or find new use cases for older chips as it grows, and will need to continue to do this in perpetuity.