As global tech giants prepare to invest $1 trillion in AI data centers over the next five years, a seemingly mundane accounting question has become a focal point for management and investors: How should the depreciation period for GPUs be determined? This technical issue actually lies at the heart of corporate profitability and investment return calculations.
Infrastructure leaders like Google, Oracle, and Microsoft estimate their servers can last up to six years. However, Microsoft's latest annual report reveals a wide depreciation range of two to six years for computing equipment. The market is now debating whether NVIDIA's high-value GPUs might face accelerated depreciation due to rapid technological obsolescence.
Depreciation timelines directly impact financial performance. Longer asset lifespans allow companies to spread costs over more years, softening profit impacts. But as a relatively new asset class, AI GPUs lack sufficient historical data, creating uncertainty for investors and lenders financing massive AI infrastructure.
Market concerns are already visible. CoreWeave shares have plunged 57% from June highs, while Oracle has dropped 34% since September 2023 peaks, reflecting growing investor skepticism about AI overinvestment.
**The GPU Depreciation Puzzle: Limited Historical Data**
AI chip depreciation presents unique challenges. NVIDIA's first data center AI processors launched around 2018, with the current AI boom starting after ChatGPT's late 2022 debut. Since then, NVIDIA's data center revenue skyrocketed from $15 billion to $115 billion as of January this year.
Unlike heavy industrial equipment with decades of usage history, GPUs lack comparable lifespan records. "Is it three, five, or seven years? From a financing perspective, that's a massive difference," said Haim Zaltzman, Vice Chair of Emerging Companies at Latham & Watkins, who works on GPU financing deals.
Depreciation—the accounting practice of allocating hardware costs over expected useful life—has grown increasingly crucial as companies try to forecast how long their hundreds of thousands of NVIDIA GPUs will remain functional or valuable.
**Three-Year Lifespan Reality?**
Divergent views emerge on GPU value retention. CoreWeave, which leases GPUs to clients, adopted a six-year depreciation cycle in 2023. CEO Michael Intrator told CNBC this week the company is taking a "data-driven" approach to evaluate GPU longevity, noting that its 2020-era A100 chips remain fully booked. Even returned 2022 H100 chips were immediately re-leased at 95% of original value.
However, CoreWeave shares plunged 16% post-earnings due to third-party data center delays affecting guidance. Notably, short-seller Michael Burry recently revealed bearish positions against NVIDIA and Palantir, arguing tech giants overestimate AI chip lifespans. Burry suggests server equipment realistically lasts just two to three years, meaning companies may be overstating earnings.
**Accelerated Innovation Adds Pressure**
Analysts cite multiple reasons AI chips might depreciate within six years: physical wear, damage, or economic obsolescence as new GPUs emerge. NVIDIA CEO Jensen Huang hinted at this dynamic when joking about predecessor Hopper chips losing value after Blackwell's launch: "When Blackwell ramps, nobody will want Hopper," Huang said at NVIDIA's March AI conference. "There are some cases where Hopper is okay, but not many."
NVIDIA now releases new AI chips annually versus its previous two-year cycle—a pattern matched by rival AMD. Amazon recently shortened some server lifespans from six to five years, citing accelerated AI/ML advancements, while other cloud providers are extending GPU lifespan estimates.
**Tech Giants' Countermeasures**
Despite massive AI infrastructure plans, Microsoft CEO Satya Nadella emphasized diversifying chip purchases to avoid overcommitting to single generations. "The biggest competitor to any new NVIDIA chip is its previous generation," Nadella noted, adding that accelerated migration cycles are a key consideration: "I don't want to carry four or five years of depreciation on one generation."
Depreciation experts like Emrydia Consulting's Dustin Madsen explain that tech's rapid evolution often upends initial estimates. Auditors scrutinize assumptions around obsolescence, maintenance, historical analogs, and engineering analyses. "You must convince auditors your proposed lifespan reflects reality," Madsen said. "They examine all factors—like engineering data suggesting six-year viability—at extremely granular levels."