An Nvidia GPU can keep working for years and still become economically outdated much sooner.
That distinction matters as Microsoft, Alphabet, Amazon and Meta pour enormous amounts of capital into AI infrastructure. Investors tend to focus on the headline cost of chips and data centers, but depreciation determines how much of that investment eventually becomes an expense on the income statement.
Depreciation spreads the cost of equipment over its expected useful life. A company that spends $1 billion on servers normally does not record the entire amount as an expense immediately. Instead, the cost is recognized over several years.
The problem with AI hardware is that physical life and economic life can be very different.
Why AI GPUs Can Lose Value Quickly
A GPU does not have to stop functioning to become less valuable.
New generations can train models faster, consume less electricity and provide more computing power per dollar. That means an older accelerator may remain technically usable while becoming increasingly expensive to operate compared with newer hardware.
Nvidia's rapid product cycle makes that especially important. The company has moved from Hopper to Blackwell while continuing to develop newer architectures, increasing the risk that expensive data-center hardware becomes economically outdated faster than traditional servers.
Nvidia outlines its current product and technology strategy through its investor relations materials.
The issue also connects directly to whether unprecedented corporate AI spending actually produces attractive returns. Our guide to AI investment returns explains how investors can use free cash flow, margins and return on invested capital to judge whether those expenditures are paying off.
Why Depreciation Can Change Big Tech Earnings
Imagine a company spends $6 billion on AI servers and expects them to remain useful for six years.
Very simply, that could mean roughly $1 billion of annual depreciation before considering timing and other accounting details.
If management instead concludes the equipment will remain economically useful for only four years, the annual expense could rise substantially.
The important point is that changing the useful-life assumption can affect reported profits without changing the cash that was originally spent.
Large cloud companies disclose these assumptions in their financial statements. Investors can review Microsoft's depreciation policies through its investor relations filings, Alphabet's disclosures through Alphabet Investor Relations and Amazon's filings through Amazon Investor Relations.
That does not mean older GPUs suddenly become worthless. Companies can move them from demanding model-training workloads into inference, internal applications or less compute-intensive tasks.
That second life can extend their economic usefulness.
The Number Investors Should Really Watch
The biggest question is therefore not simply how much Big Tech spends on GPUs.
It is whether the revenue and cash flow produced by those GPUs grow fast enough to compensate for depreciation, electricity costs and rapid technological replacement.
That matters more as Nvidia's AI-chip revenue grows. Coinpaper's recent Nvidia earnings coverage shows the extraordinary scale of current accelerator demand.
Hardware cost is also only part of the equation. Modern AI data centers consume enormous amounts of electricity, making energy availability another increasingly important constraint. Coinpaper's explainer on AI power markets examines why electricity economics may ultimately become as important as the GPUs themselves.