The AI Industry Has a Depreciation Problem

The artificial intelligence industry is engaged in an exceptionally large capital investment programme. Technology companies are building data centres, buying accelerators, securing power and constructing networks. In May 2026, TrendForce forecast combined capital expenditure of approximately $830 billion for nine major cloud service providers. That is a forecast of their combined expenditure, not a measurement of spending exclusively on AI chips, but it indicates the scale of the commitments being made.^[1]

The central debate has largely focused on demand. Will artificial intelligence generate enough revenue to justify the expenditure? Will enterprise adoption grow quickly enough? Are companies overbuilding capacity, or will demand continue to exceed supply? Beneath these questions sits another that may ultimately be just as important: how long will the assets being built today remain economically valuable?

This is not merely an accounting question. It goes to the economics of the infrastructure itself. A conventional data centre combines buildings, power systems and cooling equipment with computing hardware that may need replacement on a different schedule. Treating the facility as one asset with one useful life obscures the different things an investor is actually financing.

Artificial intelligence makes that distinction especially consequential. A building or grid connection may remain useful across several generations of equipment, while the accelerators producing its revenue face a much faster rate of technological change. An investment programme that looks like the construction of long-lived infrastructure may therefore contain a substantial recurring replacement obligation.

The buildout could involve substantial recurring renewal rather than a one-off investment followed by decades of unchanged operation. That need not be a poor business: an asset can earn an adequate return over a relatively short life. The depreciation problem examined here is uncertainty about that life and the resulting financial assumptions, not a demonstrated finding that companies across the industry are understating depreciation.

Depreciation is often treated as a technicality because the charge does not itself require a cash payment in the year it appears in earnings. The payment for the asset and the recognition of its cost occur on different schedules. Under IFRS, IAS 16 provides for systematic allocation of the depreciable amount over useful life; the exercise is an allocation of cost, not an annual attempt to reproduce the asset's resale price.^[2][3]

An illustrative example shows why the estimate matters. With straight-line depreciation, no residual value and a six-year useful life, a £6 million server installation would generate an annual depreciation charge of £1 million for a full year of use. If its remaining productive life later proves substantially shorter, the future expense profile and possibly its recoverable carrying value need reconsideration. That does not establish that the original estimate was unreasonable when made.

Accounting rules do not allow useful lives to become permanent assumptions simply because management once selected them. For companies reporting under IFRS, useful lives and residual values must be reviewed at least at each financial year-end, with changes treated as changes in estimates. The separate question of impairment also matters. American technology groups generally report under US accounting rules, so an IFRS illustration should not be mistaken for a description of every company's precise policy.^[3]

The economic significance of a useful-life estimate is visible in companies' own disclosures. Microsoft reported in 2022 that it would extend the estimated lives of server and network equipment from four years to six, citing operating efficiencies and technological advances. At the time, it estimated that the change would increase fiscal 2023 operating income by $3.7 billion. This historical example demonstrates the sensitivity of reported earnings to the assumption; it is not evidence that the change was improper.^[4]

The present question is whether AI hardware will remain productive for the periods assumed. A newer accelerator does not automatically make an older one worthless. Different workloads require different amounts of memory, bandwidth and computational capacity, and software improvements can extend the usefulness of equipment already installed. Conversely, hardware that still functions can become much less attractive if its operating cost is high relative to available alternatives.

The Bank of England set out both sides of this uncertainty in its July 2026 Financial Stability Report. Demand for older chips and current shortages can support longer lives, while rapid innovation and efficiency gains can shorten them. It also identified potential fragility when debt outlasts the assets it finances. Importantly, it said most AI-related debt to date had financed buildings and facilities rather than the chips inside them.^[5]

That qualification matters. A twenty-year loan against a building is not economically identical to a twenty-year loan against a particular generation of accelerator. A data centre may support several generations of equipment, and cash flows from those successive generations may service the same financing. But somebody still has to fund the replacement equipment, and the facility must remain suitable for operating it.

Adaptability therefore becomes part of the investment case. A new generation may require different power density, cooling or networking. An owner cannot assume that replacing the accelerators will be the only expense involved in keeping a facility competitive. The Bank also warns that facilities unable to support newer chips could lose economic value, affecting collateral and debt-service capacity.^[5]

An implication is that strong demand alone cannot settle the question of investment quality. An asset may be heavily used and still produce a disappointing return if the operator has to replace it sooner than expected or undertake expensive modifications to continue selling a competitive service.

Suppose an operator builds a GPU cluster and achieves high utilisation for three years. A subsequent generation may reduce the price customers will pay for the older service. Whether the original investment earned an adequate lifetime return depends on its acquisition cost and the cash flows over its life. Whether to keep operating it now is a separate, forward-looking comparison of expected revenue, operating expenditure and the opportunity cost of using the equipment and capacity.

Economic depreciation is different from physical deterioration, and neither is identical to the accounting allocation of cost. A working chip can lose market value while continuing to generate positive operating cash flows. Recovering its original purchase price does not itself make its next period of electricity or maintenance cheaper. Past payback informs the investment record, not the incremental economics of operating it tomorrow.

This distinction becomes important when considering the timing of the investment boom. Newly purchased equipment enters the accounts before its full cost has passed through earnings. A period of rapidly rising capital expenditure can therefore be followed by rising depreciation charges even if purchases subsequently stabilise. That timing is a normal feature of capital accounting, not evidence of concealed losses.

The harder question is whether the schedules reflect the benefits the assets will actually generate. Where estimates remain reasonable, depreciation records the consumption of those benefits over time. Where assumptions change, companies may need to revise useful lives or recognise impairment. Investors need to distinguish those possibilities from the separate issue of how much cash will be required to maintain competitive capacity.

A company can grow revenue while earning an inadequate return on the capital needed to provide its service. But replacement spending alone does not establish that outcome. Investment analysis must compare expected receipts with the relevant initial and subsequent expenditure, financing assumptions and required return, without treating every replacement as a loss or counting the same capital cost twice.

Financing extends the consequences beyond the operator. Debt can be backed by equipment, facilities, customer contracts, guarantees or a combination of these. Each structure places different parties between a disappointing operating outcome and a lender's loss. The relevant question is not simply how much debt exists, but what cash flows and protections support it.

For chips specifically, the Bank of England cites a JPMorgan estimate of more than $2 trillion in funding needs over five years and notes the difficulty of matching uncertain chip lives with investors' preferred debt maturities. This is a funding estimate and a risk assessment, not a prediction that the loans will fail.^[5]

An infrastructure owner could manage that risk through faster amortisation, conservative residual values, strong customer commitments or sufficient equity to finance replacements. A customer might accept some of the risk through a long-term contract. A lender might rely primarily on the creditworthiness of a guarantor rather than on selling the chips if something goes wrong. These arrangements do not remove technological risk; they determine where it sits.

Understanding that allocation will matter more as the market becomes more complex. A contract described as long-term may offer less protection than its headline duration suggests if it allows termination, depends on performance conditions or requires repeated investment to deliver the promised service. Equally, a short-lived asset can be financed sensibly when the repayment schedule and operating cash flows reflect its likely life.

The industry faces a peculiar interaction between efficiency and replacement. Better hardware is good news for users when it reduces the cost of inference. For the owner of installed equipment, the same improvement can reduce the prices its existing assets command. The technological progress supporting the industry's expansion can therefore also accelerate the loss of value in parts of its capital base.

There is no requirement that this becomes a crisis. Demand can support older equipment, software can improve its performance and different workloads can favour different configurations. Continuing to operate an existing cluster may be attractive when its future receipts exceed the relevant operating and opportunity costs. Replacing it may be preferable when the expected benefits justify the new outlay; either decision can be rational regardless of whether the original purchase has already paid for itself.

Amazon's P4 offering illustrates the need for caution about instant-obsolescence claims. AWS introduced A100-based P4 instances in 2020 and continues to list that family commercially. Availability does not establish the profitability or utilisation of every older installation, but it shows why a new hardware generation should not be treated as automatically ending the previous one's useful life.^[6][7]

The conclusion is not that current depreciation schedules are necessarily wrong. It is that useful life, residual value and the cost of replacement have become central to evaluating an AI infrastructure business. These assumptions deserve the same attention as customer growth, capacity announcements and revenue forecasts.

An investor needs to distinguish expenditure that expands productive capacity from expenditure that maintains it, while recognising that one project may do both. Replacing an old cluster with a more capable one can simultaneously preserve competitiveness and increase output. A headline capital expenditure figure cannot explain how much of the spending represents each function.

The same scrutiny applies to facilities. A site that can accommodate several generations of higher-density equipment may retain value differently from one requiring substantial reconstruction. Access to power may remain useful even if a particular hardware configuration does not. Separating the assets is more informative than assigning a single depreciation story to an entire data centre.

These judgements remain uncertain because the economic life of today's equipment depends partly on technology, demand and prices that have not yet emerged. Companies must make long-term commitments without knowing exactly how future models will use computing resources or how quickly competing systems will improve. The uncertainty is part of the business, not a reason to pretend that either very short or very long useful lives are already established facts.

The winners of the infrastructure boom may not simply be the companies that build the most compute. They may be those that build adaptable facilities, match financing to the assets and cash flows actually available, and recover enough capital before technological change alters the economics again.

The AI investment debate is often framed around whether demand will justify the spending. A complete answer also requires an estimate of how long each asset will contribute, how it can be adapted and when replacement is worthwhile. Short asset lives can support good investments, and long accounting lives do not establish good ones. The depreciation problem is the need to make those assumptions explicit and keep testing them as the technology and market change.