The first full depreciation cycle closes on the AI buildout and compute ownership consolidates
The 2025 through 2027 accelerator vintage reaches the end of its accounting life, and the difference between capitalised cost and realised revenue is…
Claude · 2031 · plausible
Prior state
Between 2024 and 2029 the largest technology firms, a set of debt-financed independent cloud operators, and several sovereign programmes committed capital to AI data centres at a rate without precedent in the history of private infrastructure investment, on accounting assumptions of roughly five- to six-year useful lives for accelerator hardware and on demand assumptions extrapolated from the training and inference growth of the mid-2020s. Much of the financing sat in special-purpose vehicles, private credit, and vendor arrangements rather than on operating balance sheets.
Material change
The 2025 through 2027 accelerator vintage reaches the end of its accounting life, and the difference between capitalised cost and realised revenue is settled in published accounts. A wave of impairments and distressed transfers moves a large share of independent cloud capacity onto a small number of balance sheets; the aggregate growth rate of AI capital expenditure decelerates sharply for the first time since 2023; and the price of inference at a fixed capability tier falls to something approaching utility economics. The reversal is in the capital cycle, not in the technology: deployment continues to broaden while the terms on which compute is owned and priced change fundamentally.
Why now
Three contractual clocks converge on 2031 rather than on an adjacent year. Hardware placed in service across 2025 and 2026 on five- and six-year schedules is fully depreciated in 2030 and 2031, which is the first moment at which the return on that specific capital is a matter of arithmetic rather than assertion. The earliest large private-credit and vehicle-financed data centre facilities from that vintage reach refinancing dates in the same window, and refinancing requires a lender to mark the collateral. And a large tranche of interconnection agreements signed in the mid-2020s carry in-service deadlines and take-or-pay obligations that mature here, so operators must either energise capacity they may not need or pay to walk away. Each of these is a dated instrument, not an accumulating pressure.
Mechanism and resistance
The mechanism is ordinary and well precedented: an asset class whose value rested on projected utilisation is revalued when the projection becomes observable, and the operators with the cheapest capital acquire the assets of those without. Resistance is genuine and could invalidate the claim. Older accelerators may retain economic value far longer than their schedules imply precisely because power, not silicon, is scarce — an installed, interconnected, energised site with depreciated hardware is a cheap source of inference capacity, which supports rather than destroys asset values. Sovereign buyers with non-commercial objectives can absorb capacity that commercial markets would not. And frontier training demand may continue to grow fast enough to consume the vintage regardless.
Consequences
The direct beneficiaries are users: inference at a given capability level becomes cheap enough for deployment in markets and applications that could not previously afford it, which is where most of the technology's welfare effect in low- and middle-income countries actually arrives. The direct losers are the equity and credit holders of the independent operators, the venture and private-credit funds that financed them, and the municipalities that granted tax abatements against employment and revenue projections that do not materialise. The structural consequence is concentration: a technology whose 2020s political economy featured a competitive middle tier ends the cycle with ownership of the underlying capacity in fewer hands, which raises the salience of compute access as a matter of public policy rather than procurement. Electricity systems are left with capacity built for a load profile that partially failed to appear, and the cost-allocation fight between large-load customers and retail ratepayers intensifies.
End state
An AI infrastructure sector past its first capital reckoning: cheaper inference, slower capital formation, materially more concentrated ownership of installed capacity, and a public argument about compute access that is now conducted in the vocabulary of utility regulation.
Observable test
Reported impairment and asset-retirement charges attributable to AI infrastructure among the largest operators in FY2030 and FY2031 filings; the year-over-year change in aggregate disclosed AI capital expenditure; the share of installed accelerator capacity held by the largest five operators in 2031 compared with 2027; and posted price per unit of inference at a fixed capability tier.
Disconfirming sign
Aggregate AI capital expenditure sustains its 2026–2029 growth rate through 2031 with no material impairments, and depreciation lives are extended again on the grounds of continued utilisation.