GPU economic life and depreciation

For a company that owns GPUs, depreciation is the business. The chips are the largest cost it carries, and they are paid for up front, in cash, before they earn anything. Whether the business makes money depends on one question more than any other: how many years does a GPU keep earning enough to repay what it cost, before a newer generation makes it uncompetitive?

The accounts answer that question one way, and the rental market answers it another. This piece explains both, how our GPU Economic Life Tracker reads the market's answer from rental prices, and why the gap between the two matters to investors and lenders.

Useful life in the accounts

Every company that owns servers states a useful life for them in the notes to its accounts, and depreciates them in equal amounts over that life. The choice is an estimate made by management and reviewed by the auditors, and it can change. When a company lengthens the life it uses, its annual depreciation falls and its reported profit rises, though its cash is unchanged. When it shortens the life, the reverse happens.

For the largest owners of GPUs, the useful life is one of the most consequential numbers in the accounts. A small change, applied to a very large fleet, moves reported earnings a long way. That is why investors read the note on property and equipment closely, and why a change in estimate draws attention whenever it comes.

Economic life: how long a chip earns

Economic life is a different idea. It is the period over which a GPU's rents, the money it earns by the hour or under contract, justify the cost of owning and running it. A GPU does not stop working when a newer one arrives; it keeps running for years. What changes is what customers will pay for it. As each generation arrives, the rent for the one before tends to fall, and with it the value of the remaining years of the older chip.

The market offers several signals of economic life: the trend in rental prices for each generation, the prices of second-hand GPUs, and the terms on which lenders will finance them. Lenders tend to underwrite GPUs on shorter lives than the accounts use, which is itself a view on economic life.

How the tracker reads it

Our GPU Economic Life Tracker turns stored rental-price history into an implied economic life for each GPU generation. The method is public and versioned.

• The rent. For each generation, from the A100 to the B300, it reads the weekly on-demand rent across providers, counting a week only when enough providers list that generation.
• The trend. It fits the trailing year of weekly rents as a constant annual rate of change, with an interval around it that allows for the way weekly prices follow one another.
• The residual curve. If rents fall at a constant rate, a GPU's value, which is its remaining rent, falls at the same rate. The tracker shows the share of today's rent left after each of the next several years.
• The implied life. It converts the trend into the straight-line life that loses half its value in the time the rent takes to halve. The two ends of the interval give a shorter and a longer life, so the range is shown, not just a point.
• The comparison. It reads the useful lives that Microsoft, Alphabet, Amazon, Meta, Oracle, CoreWeave, Nebius and IREN state for servers and GPUs in their latest annual reports, using the reports' own tagged data. For each, it shows how much of the benchmark H100's rent would be left when the stated life runs out, and how annual depreciation per billion dollars of servers would change if the implied life replaced the stated one.

Today's neocloud on-demand price for an H100 reads

When rents rise

The method has a deliberate stop. If a generation's rent did not fall over the trailing year, the tracker shows neither a residual curve nor an implied life, and says so. A rise cannot be extrapolated into a decline, and inventing one would turn a forecast into a fact.

That case is not hypothetical. Through much of 2026 H100 rents firmed rather than fell, against the widely held expectation of steady deflation, as the AI Infrastructure Primer documents. When that happens, the honest answer is that the rental market is not yet showing the decline that a short accounting life assumes, and the tracker reports exactly that.

Why the gap matters

If a GPU's economic life is shorter than the useful life in the accounts, reported profits are overstated year by year, and the difference comes back later as a write-down or as a loss on disposal. If it is longer, the accounts are conservative and the fleet keeps earning after it is fully depreciated.

The gap matters beyond earnings. Adjusted EBITDA, the measure many GPU owners lead with, leaves depreciation out entirely, so it flatters a business whose main cost is depreciation. Lenders size their loans to the years in which a fleet can repay them, which is why GPU-backed financing leans on customer contracts rather than on the value of the chips. And for the hyperscalers, whose server fleets are among the largest assets they own, a change in stated life moves reported profit materially, without any change in cash.

The limits

The tracker is careful about what it cannot see. One year of weekly history shows the market's recent direction, not a full replacement cycle. The rental data mixes form factors and regions. The rents are on-demand list prices, while contract prices are lower and move more slowly. A floor set by the cost of electricity was considered and left out, because at today's power prices it sits far below any rent and never binds.

Where to see it

The tracker's table appears each month in the Neoclouds Monthly, for Plus and Pro subscribers. Its method, inputs and caveats are public on the GPU Economic Life Tracker page. For the arithmetic of what a GPU earns in the first place, see GPU-hour economics.

Sources

• Compute Terminal GPU Economic Life Tracker, method version 0.9 (23 October 2026); instrument page.
• GetDeploying GPU price dataset, CC BY 4.0: weekly on-demand provider medians by generation.
• Microsoft, Alphabet, Amazon, Meta, Oracle, CoreWeave, Nebius and IREN: latest annual reports (10-K and 20-F, SEC EDGAR), stated useful lives of servers and GPUs.
• AI Infrastructure Primer 2026, Compute Terminal: "Pricing, supply and the October 2026 state of play" and "Financing the buildout".