Accounting 101: A Crash Course in Depreciation
Before we dive into the deep end, let’s get one term straight: depreciation. Imagine you buy a delivery truck for your business for $60,000. You don't record a single $60,000 expense in the first year. Instead, accounting rules say you should spread that
cost over the 'useful life' of the truck. If you decide the truck will last six years, you'd record a $10,000 depreciation expense each year. This is intended to match the cost of the asset with the revenue it helps generate over time. Now, replace 'delivery truck' with 'multi-billion-dollar fleet of AI servers packed with GPUs.' The principle is the same, but the stakes are much, much higher. The key variable here is 'useful life.' How long you decide the asset will last directly changes how much expense you book each year. A longer useful life means lower annual depreciation, which, in turn, makes your reported profits look higher in the near term.
The Six-Year Standard
For years, the tech world was accelerating. New chips came out, and old ones became obsolete. Hyperscalers like Microsoft, Google (Alphabet), and Amazon (AWS) used to depreciate their servers over relatively short periods, like three or four years. But then, a coordinated shift happened. Starting around 2022, the giants of the cloud began to extend the useful life of their servers. Microsoft announced it was increasing the estimated useful lives of its server and network equipment from four years to six. Alphabet also moved to a six-year schedule for servers. Amazon followed a similar path, though its disclosures are more complex, with different schedules for servers and networking gear. The justification was that investments in software and technology had made the hardware last longer. This change had a massive, immediate impact. For Microsoft, it was projected to boost operating income by $3.7 billion in the first full year alone. For Alphabet, a similar change added billions to its bottom line.
The 'Mistake': Do AI GPUs Really Last Six Years?
This is where the 'mistake' in the headline comes in. While a six-year lifespan might make sense for a standard server running predictable, lower-intensity tasks, it’s a questionable assumption for the high-end GPUs powering the AI arms race. These specialized chips are the workhorses of AI, running at near-constant high utilization rates that subject them to intense thermal and electrical stress. Some technical analyses and even comments from industry insiders suggest the practical physical lifespan of a heavily used data center GPU is closer to one to three years. Beyond physical wear, there's the blistering pace of innovation. NVIDIA, the dominant supplier, is on a rapid release cycle. A GPU that is state-of-the-art today might be economically obsolete for cutting-edge AI training in just a couple of years. Interestingly, Amazon recently acknowledged this reality by shortening the useful life for a subset of its servers back down from six to five years, explicitly citing the 'increased pace of technology development, particularly in artificial intelligence and machine learning'. This move runs counter to the broader trend and adds weight to the argument that a six-year schedule for AI hardware is overly optimistic.
How This Distorts the Earnings Picture
When a company like Microsoft applies a six-year useful life to a fleet of GPUs that may only be physically or economically viable for three, it creates a disconnect between accounting reality and operational reality. In the short term, earnings are inflated because the annual depreciation charge is artificially low. It makes the colossal capital expenditures on AI seem more manageable and more immediately profitable. Investors looking at the headline earnings-per-share number might get a rosier picture than is warranted. The true cost of keeping the AI infrastructure at the cutting edge is smoothed over and pushed into the future. It’s not that the cost disappears—it just shows up on the income statement much more slowly. This accounting choice can mask the true, staggering replacement cost required to stay competitive in AI, where having the latest and greatest hardware is a key advantage. While older GPUs can be repurposed for less intensive tasks like inference, the most valuable and demanding workloads require the newest chips.
