Cover Story · Parameter desk
The Six-Year Assumption
The depreciation schedule carrying AI capex rests on a residual-value tail its own owners are spending billions to compete away
The Ecliptic · Issue No. 007 · August 27, 2026
The change
Nothing was bought or sold. An estimate moved.
Between 2020 and 2024 the largest cloud operators progressively lengthened the useful lives they assign to servers and accelerators, converging on six years. One estimate puts the 2024 effect at a fall from roughly $39bn of collective depreciation to about $21bn — a 46% reduction in reported expense against an unchanged fleet. No asset changed hands. A judgment about the future was revised.
- The extension was gradual and one-directional — four consecutive years in which a discretionary estimate moved the way that raises reported earnings, across multiple independent reporters.
- Depreciation is non-cash, which is why the usual objection is that none of this matters. It matters because the assumed life is also an input to the return model that approves the *next* purchase — hardware modelled over six years clears hurdles it would fail over three.
- That makes the estimate upstream of the capex, not downstream of it. If the assumption is wrong, it has been compounding the size of the position the error applies to.
The tail
Years five and six are exactly what the ASICs are built to take
The six-year life is defended by a value cascade: years one and two carry frontier training, three and four carry real-time inference, five and six carry batch inference and internal analytics. The argument is coherent, and its final stage is the one the same operators are dismantling. Custom inference silicon is deployed precisely because it beats older training GPUs on inference total cost of ownership — which is a claim about stage three.
- Note what stage three actually asserts. Not that batch inference demand will exist in year five — it obviously will — but that a five-year-old training GPU will be the *economically preferred* way to serve it. That is a comparative claim, and it can be defeated without demand falling at all.
- Every ASIC deployed removes a unit of the residual demand that justifies the schedule. A firm can extend useful life, or it can build silicon to displace the workloads that justify the extension. Doing both means one of the two decisions is not being priced.
- The failure mode is not a sudden write-down. It is a slow divergence in which reported depreciation stays flat while the value of the tail erodes — recognised only when an impairment forces it.
The disclosure
One firm moved against its own reporting incentive
After four years of moving together, the operators split in 2025: Amazon shortened the useful life of a subset of servers while Meta extended its estimate further. Shortening a useful life raises reported expense and lowers reported earnings. In a dispute about a discretionary estimate, the informative signal is the one that costs the party making it.
- Two firms with broadly comparable fleets moved the same estimate in opposite directions in the same year. At least one of them is wrong, and the difference is not a technical judgment about silicon.
- One published estimate puts the gap at $176bn of understated depreciation across 2026–2028, with earnings at some individual firms potentially overstated by roughly 27% and 21% by 2028. That is a single analyst's model on stated assumptions, and we carry it as an order of magnitude, not a forecast.
- Useful-life assumptions have now diverged enough that reported operating margin is no longer comparable across cloud operators without adjustment — and the divergence is large enough to reorder a margin ranking.