Big Tech’s AI boom faces a quiet threat from server depreciation math

The AI arms race is usually measured in chips, gigawatts and eye-watering capital budgets. A Reuters Breakingviews column published Oct 29 argues the real pressure point is far less glamorous: a single accounting assumption about how long servers last.

That assumption, known as useful life, decides how quickly companies like Alphabet, Amazon, Meta and Microsoft expense their data-center hardware. Shorten it, and profits shrink. Stretch it, and earnings look healthier, at least on paper.

Four giants, four different clocks

Servers work the same way on a corporate income statement. The longer a company assumes its hardware stays useful, the smaller the annual expense, and the larger the reported profit.

Amazon went the conservative route. Starting January 1, 2025, it cut the useful life of certain servers and networking equipment from six years to five, pointing to faster technological obsolescence driven by AI and machine learning.

That decision carries a price tag. Amazon expects it to add $1.4 billion to its 2025 depreciation expense and reduce net income by $1 billion.

Meta took the opposite path. It extended the estimated useful life of most of its servers and network assets to 5.5 years, up from a prior range of 4 to 5 years.

The effect runs in Meta’s favor. The change is projected to lower 2025 depreciation by $2.92 billion and lift net income by $2.59 billion, or $1.00 per share.

Alphabet and Microsoft have largely stayed put. Alphabet generally applies a six-year life to its servers, while Microsoft uses a flexible range of 2 to 6 years depending on the type of asset.

The $36 billion sensitivity test

Breakingviews ran the numbers on what happens if the optimists are wrong. Its estimate: if all four companies shortened their useful-life assumptions by just one year, combined net income could fall by approximately $36 billion in 2027.

The valuation hit is even larger. Applying current valuation multiples, the column calculates a theoretical market-cap impact of nearly $800 billion.

The scale of spending explains why the stakes are so high. Breakingviews notes that outlays on the chatbot boom are heading into the trillions of dollars, with data centers consuming electricity now measured in gigawatts.

Meanwhile, the bills have not fully hit income statements yet. Cumulative AI-related capital expenditures across the group have reached hundreds of billions, and capex has run at about 3.4 to 5.5 times depreciation charges in recent quarters.

Why server lifespans became a debate

Amazon explicitly tied its shorter schedule to the pace of AI and machine learning innovation, which can make expensive gear less competitive well before it physically wears out.

Meta’s extension reflects a different read. Its longer schedule implies confidence that its servers and network assets will remain productive for longer than it previously assumed.

What this means for investors

The most immediate takeaway is that reported earnings across Big Tech are not perfectly comparable. Amazon absorbed a $1 billion net income hit from its change, while Meta booked a $2.59 billion boost from moving in the other direction.

Because capex has been running several times higher than depreciation, a growing wave of expense is already baked in for future years regardless of policy choices.

Alphabet’s six-year assumption and Microsoft’s 2-to-6-year range will draw particular scrutiny. If either follows Amazon’s lead, the $36 billion sensitivity estimate stops being a thought experiment and starts showing up in actual earnings.