Nvidia CEO Jensen Huang made a joke this week that his biggest customers probably won’t find funny. “I said before that when Blackwell starts shipping in volume, you couldn’t give Hoppers away,” he said at Nvidia’s big AI conference Tuesday. “There are circumstances where Hopper is fine,” he added. “Not many.”
He was referring to Nvidia’s latest AI chip-and-server package, Blackwell, which is notably superior to the previous version, Hopper, released in 2022. Major cloud providers, including Amazon, Microsoft, and Google, purchase vast amounts of these GPU systems to train and operate the large models driving the generative AI revolution. Meta has also invested heavily in GPUs in recent years.
While these companies should celebrate the enhanced capabilities of Blackwell, a challenge arises:
AI Obsolescence
As new technologies evolve rapidly, earlier versions can quickly become obsolete or at least markedly less useful. This depreciation can make these assets less valuable, forcing large cloud firms to make adjustments. This process, known as depreciation, involves reducing asset values over time to account for factors like wear and tear and eventual obsolescence. The quicker the depreciation, the more substantial the impact on earnings.
Ross Sandler, a leading tech analyst at Barclays, cautionary advised investors last Friday that major cloud companies and Meta would likely need to make such adjustments, which could significantly impact profits. “Hyperscalers are likely overstating earnings,” he wrote, although Google and Meta did not respond to inquiries from Business Insider regarding the matter on Friday, and Microsoft chose not to comment.
Amazon Takes the Plunge First
Consider Amazon Web Services, the largest cloud provider, which became the first to face this issue. CFO Brian Olsavsky noted during Amazon’s earnings call last month that the company “observed an increased pace of technology development, particularly in AI and machine learning.” Consequently, he announced they would reduce the useful life of a portion of their servers and networking equipment from six years to five beginning in January 2025, resulting in an anticipated $700 million hit to operating income this year.
Furthermore, Olsavsky mentioned that Amazon had “early-retired” some of its servers and networking equipment, a move that led to “accelerated depreciation” costing approximately $920 million. The company expects this will further lower operating income in 2025 by around $600 million.
A Much Larger Problem for Others
Sandler included in his research a compelling chart detailing the cost of renting H100 GPUs, based on Nvidia’s older Hopper architecture. As the new Blackwell GPUs have become more available, rental prices have significantly decreased.
He indicated that this issue could be much more pronounced for companies like Meta and Google, as he estimated a one-year reduction in the useful life of Meta’s servers could raise depreciation in 2026 by over $5 billion, matching the decline in operating income. A similar adjustment at Google might result in a $3.5 billion drop in operating profits, according to Sandler.
However, there’s an essential caveat: Just because one major cloud provider has made this adjustment doesn’t ensure others will follow suit. Some firms might design their AI data centers in a way that extends the lifespan of Nvidia GPU systems or mitigates their obsolescence.
The Time Has Come
As the generative AI surge gained momentum in the summer of 2023, Bernstein analysts raised concerns regarding this depreciation. “All those Nvidia GPUs have to be going somewhere. And just how quickly do these newer servers depreciate? We’ve heard some worrying timetables,” they expressed in a note to investors. Analyst Mark Shmulik communicated with my colleague Eugene Kim, stating, “I’d imagine the tech companies are paying close attention to GPU useful life, but I wouldn’t expect anyone to change their depreciation timetables just yet.” Now, that time seems to have arrived.
“`
