Six-Year Servers, One-Year Chips: The AI Data Centre GPU Depreciation Gap
Microsoft, Alphabet, Amazon and Meta all stretched server lives to five or six years just as Nvidia went annual. One of them has already reversed course. The gap between book life and economic life is the quiet subsidy under AI earnings, and somebody is holding it.
The four companies spending most on AI infrastructure all depreciate their servers over five to six years. Nvidia now ships a new accelerator generation every year. That is AI data centre GPU depreciation in two sentences: the accounting runs on a slow clock, the product cadence runs on a fast one, and the difference is currently being booked as profit. The interesting questions are how wide the gap has become, and who is holding the paper when it closes.
The assumptions are not hidden. They sit in the filings, along with what each change did to reported earnings. Microsoft extended the depreciable life of its servers and network equipment from four years to six starting in its 2023 fiscal year, disclosing that the change would add about $3.7bn to that year's operating income (the language is in Microsoft's 10-K filings). Alphabet followed in January 2023, moving servers from four years to six and estimating a reduction of roughly $3.4bn in 2023 depreciation (Alphabet's annual reports carry the estimate). Meta stretched certain servers and network assets to five and a half years from January 2025, worth about $2.9bn in reduced 2025 depreciation (Meta's filings). Amazon went the same way, to six years in January 2024 for about $900m of operating income benefit, then did the more interesting thing: in February 2025 it cut a subset of servers and network equipment back to five years, took an accelerated charge of about $920m for kit retired early, and guided to roughly $700m less operating income in 2025, citing the pace of technology development, above all in AI (Amazon's 10-K disclosures).
Read that sequence for what it is. Every change until Amazon's reversal ran in the earnings-flattering direction, and the extensions were granted just as fleet composition shifted from CPUs, which genuinely do age gracefully, towards GPUs, which do not. A single year of assumed life is worth billions of dollars of reported operating income across these four companies. Amazon is so far the only one to mark the assumption to reality, and it blinked on precisely the AI-driven portion of its estate. None of this alleges anything improper: estimates are estimates, and auditors signed all of them. It is an observation about where the burden of proof now sits.
How fast does AI data centre GPU depreciation really run?
Book lives are opinions; rental prices are facts. The cleanest public proxy for what a GPU fleet can earn is what the open market pays to rent one, and that curve looks nothing like a six-year straight line. H100 rental rates fell from around $8 an hour at the 2023 shortage peak to under $2 by late 2024, a decline of roughly three quarters in about eighteen months, and most of it happened before Blackwell shipped in volume. The cadence behind that decay is public too: Hopper in 2022, Blackwell in 2024, Blackwell Ultra in 2025, with Rubin scheduled for 2026. Every fleet deployed today meets a superior competitor within about a year of arriving.
The mechanism that turns falling rents into a shutdown date is brutally simple. A paid-off accelerator earns for as long as the market price of its output exceeds its electricity and hosting bill. As illustration rather than forecast: a cluster drawing a megawatt costs roughly £700,000 a year in electricity at 8p per kilowatt-hour, a cheap rate by current UK standards, and that bill never falls. What a megawatt can earn does fall, because compute is priced against the best machine on the market, not against yours. Once the going rate drops below the old fleet's power cost, the rational owner switches it off, however much demand exists. Economic life is therefore set by the efficiency of the next chip generation and the local price of power, not by the physical durability of the silicon, which runs to a decade or more.
Who holds the paper?
The gap lands very differently depending on the balance sheet. When Microsoft or Alphabet under-depreciates, the eventual correction hits shareholders through earnings: survivable, and at their margins nearly invisible. The GPU rental specialists are another matter. CoreWeave depreciates its computing equipment over six years (the assumption is disclosed in its S-1) while borrowing against the GPUs themselves; its business is the depreciating asset, with no software or advertising franchise wrapped round it. If six years is generous for Microsoft, which can refill a hall with its own workloads, it is heroic for a rental business whose product is the thing decaying. And the further the asset travels from its user, into leases, GPU-backed loans and securitised structures, the more the residual-value assumption becomes someone else's problem: the lender's, the lessor's, the insurer's. Mispriced depreciation at a hyperscaler is a quality-of-earnings quibble. Mispriced depreciation in a leveraged rental fleet is a default.
The comfort blanket usually produced at this point is fibre: capital overshot in 2000, the glass sat dark, patient buyers lit it later with better lasers and it went on to carry the modern internet. The analogy fails on physics. Fibre held its value because the glass was never the bottleneck; the intelligence sat at the ends of the strand and could be upgraded. A GPU has no upgradable ends. The chip is the endpoint, and improvement arrives as a replacement, not a retrofit. A buyer of a superseded fleet is buying the decaying rental stream described above, and the market has already shown how fast that stream decays.
What would change my mind
Three things, none fanciful, none currently the base case. A liquid secondary market clearing superseded accelerators at a healthy fraction of book value, sustained across a generation change, would show the decay running slower than the rental curve implies. A chip generation that fails to deliver a real efficiency gain would stretch every fleet's life at once. And durable power scarcity would prop up old silicon by starving the new: Janus Henderson estimates that of 157GW of data-centre capacity announced for 2030, only about 84.7GW will actually be installed, and if new capacity cannot get grid connections, the alternative to an inefficient GPU is no GPU at all. Note what that last one means, though: the bull case for residual value then depends on the industry failing to build, which is not the case any of the financing documents were written on.
The practical instruction follows. If you lend against, lease, insure or invest in this class of asset, model economic life at roughly one chip generation, value salvage as the discounted earnings above the power-cost floor, and treat any book life beyond that as a claim requiring evidence: clearing prices, not convention. The asymmetry does the rest of the argument for you. Model conservatively and turn out wrong, and you have lent too cautiously, a bounded error. Adopt the six-year book and turn out wrong, and the losses arrive correlated across debt, leases and insurance, all triggered by the same product launch. That stress test belongs in technical strategy work before the first contract is signed, and before that in the more basic question of whether you should be building at all.
Questions people ask
What useful life do Microsoft, Amazon, Alphabet and Meta assume for their servers?
As of their most recent annual reports, Microsoft and Alphabet depreciate servers over six years, Meta uses five and a half years for certain server and network assets, and Amazon uses six years with a subset cut back to five in 2025. Each assumption, and the earnings impact of every change to it, is disclosed in the companies' 10-K filings, which is the only way outsiders can track them at all.
Why did Amazon shorten the useful life of some of its servers in 2025?
Amazon told investors in early 2025 that the increasing pace of technology development, particularly in AI, justified cutting a subset of its servers and network equipment from six years back to five. The change reduced 2025 operating income by roughly $700m and came with an accelerated charge of about $920m for equipment retired early, making Amazon the first hyperscaler to move the assumption in the conservative direction.
Can old AI GPUs be redeployed for other workloads once they are superseded?
Yes, and for a while it pays: a paid-off accelerator has no capital cost to recover, so it can serve smaller models and batch inference profitably wherever the going rate exceeds its electricity and hosting bill. The catch is that each new chip generation pushes the going rate down while the old fleet's power bill stays flat, so the profitable niche shrinks with every launch and survives longest where power is unusually cheap.
Related
- Your Cloud Provider Is Financing Its Own AI Revenue. Do the Financial Due Diligence
- The Announcement Is Not the Deal: AI Infrastructure's Announced vs Contracted Gap
- The Target-Less Number: Glucose Sensors Are Turning Fitness Trackers into Health Anxiety Machines
- Data & Strategy
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.