Big tech just reported earnings. The headlines were uniformly positive.
Meta: record revenue. Microsoft: Azure growth reaccelerating. Google: advertising holding firm. The narrative on Wall Street is clear — AI is working, and the numbers prove it.
But there is a variable buried inside every one of those earnings reports that almost nobody is talking about. It doesn't make the headlines. It doesn't show up in the summary tables. And it is quietly inflating reported profits by hundreds of billions of dollars across the sector.
It's called useful-life depreciation — and understanding it is the difference between reading big tech earnings and actually understanding them.
What depreciation actually does
When Microsoft spends $80 billion building AI data centers, that money doesn't hit the income statement all at once. Instead, the company spreads that cost over the "useful life" of the asset — the number of years it expects that hardware to remain productive.
If a server costs $1 million and has a useful life of 4 years, you book $250,000 of depreciation per year. If you extend that useful life to 6 years, the annual charge drops to $167,000. Your reported profit just increased by $83,000 per year — without a single dollar of additional revenue, without a single operational improvement.
The useful-life assumption is not neutral. It is one of the most consequential decisions a CFO makes.
And right now, across Meta, Microsoft, Google, and Amazon, those decisions are moving in one direction: longer lives, lower charges, higher reported earnings.
What the companies have done
Meta has extended server depreciation schedules twice in three years. From 4 years to 5 years in 2023. Then to 5.5 years for AI servers in 2025. The impact of the 2025 change alone: a $2.9 billion reduction in depreciation expense — equivalent to nearly 4% of pre-tax profits. That is not operational improvement. That is an accounting adjustment.
Microsoft extended the useful lives of its server and networking equipment from 4 years to 6 years back in 2022. The effect was significant: the extension added approximately $3.7 billion to Microsoft's 2023 profit. Microsoft has not revised this schedule since — despite now spending $80 billion annually on AI infrastructure that includes GPU-dense hardware operating on very different refresh cycles than traditional servers.
Amazon did something unusual: it moved in the opposite direction. AWS shortened its AI training server lifespan from 6 years to 5 years in January 2025, taking a $920 million one-time charge and accepting a $700 million drag on 2025 operating income. Amazon is the exception. And that exception matters.
Google maintains server useful lives of up to 6 years — a figure that critics argue significantly overstates the economic reality of GPU-dense AI infrastructure.
The $176 billion problem
Bank of America estimates that AI capex will drag EBIT margins by 1.6 percentage points across the major hyperscalers in 2026. That is the conservative view.
The more aggressive estimate comes from analysts who have modeled the actual depreciation math: between 2026 and 2028, depreciation across hyperscalers may be understated by approximately $176 billion, causing reported profits to be overstated by more than 20%.
That is not a rounding error. That is a structural distortion that changes how you should read every EPS beat from these companies.
The GPU obsolescence problem
Here is the core tension that makes this more than an accounting debate.
Traditional servers — the kind Meta and Microsoft built their businesses on — genuinely last 5 to 6 years. They run stable workloads. The hardware depreciates at roughly the same rate as its economic usefulness.
AI servers are different. They are dominated by GPUs. And GPUs follow NVIDIA's product cycle.
NVIDIA has shipped H100, H200, Blackwell, and Rubin within a 36-month window. Each new generation delivers 2 to 3 times the performance of its predecessor. That means an H100 cluster purchased in 2023 is not economically equivalent to a Blackwell cluster in 2026 — it is materially inferior for training frontier models, and increasingly uncompetitive for inference workloads where latency and efficiency determine margin.
If the real economic useful life of an AI GPU cluster is 2 to 3 years — which the technology cycle implies — and companies are booking 5-to-6-year lives on their accounting schedules, the gap between reported earnings and economic reality is substantial.
Company by company: the numbers
| Company | Stated useful life | Est. annual EPS inflation | 2026 AI capex |
|---|---|---|---|
| Meta | 5.5 years (AI servers) | ~$2.9B / year | $64–72B |
| Microsoft | 6 years | ~$3.7B / year | $80B+ |
| Google / Alphabet | Up to 6 years | ~$2.5B / year | $75B |
| Amazon (AWS) | 5 years (revised down) | Minimal — adjusted | $105B |
Amazon's decision to shorten its depreciation schedule deserves specific recognition. It is the only hyperscaler that has moved toward recognizing the real economic life of AI hardware. The $700 million hit to 2025 operating income was the honest accounting decision — and it penalized Amazon's reported numbers relative to peers who have not made the same adjustment.
What FCF tells you that GAAP earnings don't
This is where the analysis becomes actionable.
Depreciation is a non-cash charge. It reduces reported earnings but does not affect free cash flow. This means that for companies where depreciation is being systematically extended and understated, the gap between GAAP earnings and free cash flow is the signal.
When a company's FCF significantly exceeds reported GAAP net income, depreciation methodology is usually part of the explanation. When that gap is growing year over year — which it is at Meta and Microsoft — the accounting tailwind is compounding.
This does not mean these are bad investments. It means the EPS numbers you are reading are not a clean signal of underlying business performance. And when you use those EPS numbers to calculate P/E ratios and PEG ratios, you are potentially anchoring on an inflated denominator.
At Primus Pilus Capital, this is why we weight FCF and EBITDA more heavily than GAAP EPS for big tech positions — and why we treat EPS growth rates from companies with aggressive depreciation assumptions with a margin of adjustment.
The risk that nobody is pricing
There is a scenario that the market is not discussing: a depreciation cliff.
If and when AI GPU hardware reaches the end of its stated useful life at scale — approximately 2028 for assets purchased in 2022–2023 — companies will face two options. Either retire the hardware and take replacement costs through capex (which compresses FCF). Or take accelerated impairment charges (which hits reported earnings).
Amazon's decision to shorten useful lives now is a preemptive move against exactly this scenario. Companies that have extended useful lives are deferring the recognition of hardware obsolescence — not eliminating it.
The question is not whether the depreciation cliff arrives. It is when. And whether it lands in a quarter where the rest of the business can absorb it.
What to actually measure
For big tech AI positions, the metrics that matter:
Free cash flow yield — FCF divided by market cap. For META and MSFT at current prices, this remains attractive despite the depreciation distortion, because cash generation is genuinely strong.
EBITDA margin trajectory — Operating leverage before depreciation. If EBITDA margins are expanding while GAAP margins are flat, that is a sign of real operational improvement being masked by depreciation drag. If EBITDA margins are flat while GAAP margins expand — accounting is doing the work.
Capex as % of revenue — The companies spending the most on AI infrastructure are creating the largest future depreciation obligations. This is not a reason to avoid them — it is a reason to model the earnings differently.
Revenue per dollar of capex — The number that ultimately validates the AI thesis. If $600B+ of combined hyperscaler capex in 2026 does not translate into accelerating revenue by 2027–2028, the depreciation cliff arrives with no earnings growth to offset it.
The bottom line
Meta and Microsoft are genuinely strong businesses with real AI monetization beginning to appear in their results. The depreciation methodology doesn't change that.
What it changes is how you read the EPS numbers. When Meta reports a beat and attributes it to operating leverage, ask how much of that leverage is accounting versus business performance. When Microsoft extends a depreciation schedule again, note the dollar impact before celebrating the earnings growth.
The $176 billion is not fraud. It is accounting within accepted GAAP standards. But accepted standards and economic reality are different things.
At Primus Pilus Capital, we build our models on economic reality. When the two diverge, that is where the most important investment decisions live.