How Long Does an AI Chip Really Last?

One ratio decides the life of the machine

Revenue per kilowatt-hour, not years on a schedule

Accounting depreciation spreads the cost of a server evenly over a useful life that management chooses. Economics asks a different question: how much cash can this machine still earn, and when does it stop being worth the electricity? For an AI accelerator the cleanest single measure is the ratio of revenue per kilowatt-hour to the price of a kilowatt-hour.

A chip earns its rent in its first years and then loses ground, because each new generation produces more output per kilowatt-hour and the market reprices older chips against it. Two thresholds fall out of the arithmetic.

The life itself follows from one line:

Starting at 32x with a floor of 3.7x, a five-year life requires rental prices for that chip to fall no faster than about 35% a year. That single threshold is the whole debate.

Same machine, three pricing worlds

The model places one current-generation accelerator (GB300 class) on one megawatt of capacity and runs it year by year. Inputs: 2.58 kW per GPU all-in at the wall, about $60,000 per GPU, US industrial power at $90.3 per MWh, 80% of hours billed.

Read across the rows. Under scarcity, the chip pays back in two years and could run for a decade. Under efficiency pricing, it still pays back, barely, but is finished in four years. Under the long-term contract, the provider earns almost nothing above its cost of capital: the contract shifts the obsolescence risk to the buyer, who keeps paying the fixed price after the market price has fallen below it.

What that means for the depreciation schedule

Economic depreciation is the fall in what the machine is worth, which is the fall in the cash it can still earn. It is front-loaded in both regimes. The difference is how front-loaded.

The evidence points to scarcity now, efficiency later

If efficiency pricing arrives, about a fifth of earnings is depreciation not yet taken

We restated each company's server spending, historical from filings and forward from our base-case capex paths, using the economic schedules above in place of the company's straight line. Only the server share of capex is affected; buildings and power gear keep their long lives. The figures are the change in earnings per share as a share of our base-case EPS, before any separate write-down from prices falling below cost.

Under scarcity pricing, the book lives are close to right, and Amazon's five years is if anything conservative. Under efficiency pricing, every company's straight line is too slow by a similar margin, because the spending wave of 2025 to 2027 is written off over four years instead of five or six. The gap is largest in 2027 and 2028 and narrows after, when the early vintages are fully written off under both methods. The order of magnitude, not the decimal, is the point: these are model estimates resting on the assumptions listed below.

What would tell us the regime is changing

What this model does not know

• Estimates. Per-GPU cost EST is anchored on disclosed per-megawatt GPU capex, not a unit price. All-in power per GPU EST includes network and a 1.25 PUE; the GB300 rack power is not published. Operating cost at 12% of revenue, 80% billed utilization, building and power gear at $12M per MW over 20 years, and a 10% cost of capital are all estimates.
• Prices. The launch price uses independent-cloud on-demand rates. The large platforms' list prices are two to two-and-a-half times higher, and their effective prices after discounts are not disclosed; if they realize more, every ratio here is higher and lives are longer.
• Resale. The model gives a retired chip no resale value. An active secondary market for used accelerators would lengthen effective life.
• Server shares. Amazon (42%) and Meta (43%) come from disclosed property schedules; Alphabet (50%) and Microsoft (65%) are estimates, because neither breaks out servers on a comparable line. EPS effects scale with this share.
• Base case. The EPS percentages are measured against our own base-case forecasts, not consensus.