AI Hardware Depreciation: The Risk Nobody Models
Hyperscalers spending hundreds of billions on AI infrastructure are using depreciation schedules of 5–6 years for GPUs and supporting equipment. This is longer than the historical useful life of comparable hardware, and longer than the architectural cadence of NVIDIA’s GPU releases. Every year of useful life that gets revised downward translates directly to billions of dollars in higher annual depreciation expense — and lower reported earnings. The depreciation question is the single most underappreciated risk in the AI infrastructure investment story.
The Math Behind the Question
A $100 billion CapEx program depreciated over 6 years adds $16.7 billion in annual expense. The same program depreciated over 4 years adds $25 billion. The $8.3 billion difference is pure earnings impact, before any change in revenue. For hyperscalers reporting in tens of billions of operating income, this is material — and the affected companies (Microsoft, Google, Amazon, Meta) sit at the center of most growth portfolios.
Why 5–6 Years Looks Aggressive
NVIDIA releases new GPU architectures roughly every two years. Each generation delivers 2–5x performance improvement per watt. A four-year-old GPU competes against new hardware that is 4–10x more efficient. At what point does keeping older hardware in production cost more in power and rack space than replacing it? The depreciation schedule answers one version of that question; the physical economics may answer it differently.
The CoreWeave Model
Specialty cloud providers — CoreWeave, Nebius, Lambda — lease GPU capacity with shorter contracts. Their financial models assume aggressive utilization in the first 2–3 years, with declining yield after. This is closer to economic reality than a 6-year straight-line depreciation. The gap between specialty providers’ models and hyperscaler depreciation schedules is the underwriting risk — and it is visible in financial disclosures if anyone bothers to look.
When the Question Becomes Visible
A depreciation schedule revision is a non-cash earnings hit but a confidence shock. The market typically does not price this risk until an explicit accounting change is announced. Catalyst windows: annual 10-K filings, auditor concerns, peer-group revisions. A single hyperscaler revising its schedule downward would force the industry to follow — and the market reaction would be immediate.
Investment Implications
Hyperscaler equity multiples partially depend on the durability of current earnings, which depend on current depreciation assumptions. AI infrastructure suppliers benefit from any depreciation revision that accelerates replacement cycles — shorter useful life means more frequent buying. The longer the current depreciation regime holds, the more violent the eventual revision becomes. This is a tail risk for hyperscaler equities and a tail benefit for AI hardware suppliers.