Networking cables plugged into a patch panel, showcasing data center connectivity.
Technology

AI compute useful life is now the build-versus-rent call

As ordinary companies approve owned GPU infrastructure for the first time, the assumed life of that hardware, not its purchase price, is the number that decides whether the business case clears the hurdle rate.

The depreciation argument that began with hyperscaler filings has moved down-market, and it has arrived in a form most finance teams have not priced before. When a company buys its own GPUs rather than renting capacity, the AI compute useful life it assigns becomes the hinge of the entire business case. The hardware invoice is knowable. The life estimate is a judgment the CFO signs, and the auditor revisits it every year.

Why the life estimate decides the business case

Rent and build are not two versions of the same decision. Renting inference and training capacity keeps AI cost in operating expense: variable, cancellable, and visible to the board each quarter as a run-rate. Building converts the same workload into a fixed asset whose reported annual cost is a function of one assumption. Spread a cluster over six years instead of three and annual depreciation roughly halves. Nothing about the physical hardware changed. Only the estimate did.

That is why the life assumption, not the price per accelerator, is the swing variable in most proposals landing on CFO desks right now. CFO.com has reported that companies expect to roughly double AI spending in 2026, and agentic AI is being named a top finance priority at the CFO level, which is precisely what drives sustained inference volume and makes owned capacity look attractive. The arithmetic only works, though, at a life the finance chief is willing to defend.

The hardware invoice is knowable. The life estimate is a judgment the CFO signs, and the auditor revisits it every year.

What the hyperscaler disclosures actually signal

Large cloud providers have spent the past several years extending server useful lives into the five-to-six-year range, a change that lifted reported operating income by reducing periodic depreciation. More recently the direction has reversed in at least one case: Amazon moved a subset of servers back to a five-year life and flagged a material operating income impact. Finance teams should pull the current property and equipment footnotes from the latest 10-K filings rather than rely on secondhand summaries, because the disclosed lives and the scope of equipment they cover differ meaningfully across filers.

The reason this matters for a mid-market build case is reputational as much as technical. Short-seller commentary, including public criticism from Michael Burry, pushed the question of understated depreciation into mainstream financial press. Legal and audit commentary has followed, with the National Law Review among the outlets framing GPU useful lives as an earnings-risk issue. Any CFO adopting a six-year life for AI hardware in 2026 is adopting an assumption that is already under external scrutiny at firms with far deeper engineering telemetry.

The accounting does not let the estimate sit still

Both ASC 360 and IAS 16 require useful life and residual value to be reassessed at least annually. A build decision therefore does not lock in a number. It schedules an annual conversation. If utilization drops because a newer accelerator generation makes the installed fleet uneconomic for frontier workloads, the question is not only whether to shorten the remaining life prospectively but whether an impairment trigger exists at all.

Residual value deserves more attention than it typically gets. A business case that assumes meaningful resale value for three-year-old accelerators is making a claim about a secondary market that may be thin exactly when the company needs it. Model the residual at zero as a sensitivity. If the project only clears the hurdle rate with a generous salvage assumption plus a stretched life, the proposal is leaning on two estimates at once.

What to require in every AI infrastructure proposal

The practical control is procedural, not technical. Require every owned-compute proposal to present three life cases - three-year, five-year and six-year - with the resulting annual depreciation, return and payback for each. Require an explicit residual value assumption, stated separately rather than buried in the terminal value. Require a named owner for the annual reassessment, so the estimate has an accountable person attached rather than defaulting to the controller's year-end checklist.

Require a cloud baseline too. The board should see what the identical workload costs rented over the same horizon, including egress and commitment discounts, so the comparison is capacity-to-capacity rather than capex-to-nothing. Finally, budget the people. Datarails research cited by CFO Dive found nearly one in three finance job postings now call for AI skills, and owned infrastructure carries operating headcount that a managed cloud contract absorbs. That cost belongs in the build column, not in a footnote.

Where the hurdle rate meets the audit file

The uncomfortable position is the one in the middle. Stretch the life to make the internal rate of return work and the company inherits impairment exposure plus an audit discussion that is already playing out publicly at larger filers. Shorten it to a defensible three years and many proposals simply will not clear the hurdle rate, which is itself useful information: it suggests the workload should be rented until volume and model stability justify ownership.

The discipline worth importing from the hyperscaler debate is disclosure hygiene. Document the evidence behind the chosen life, including utilization data, vendor refresh cycles and internal replacement history. If the number later moves, the file should show a reasoned estimate updated with new facts rather than an assumption reverse-engineered to clear an approval threshold.

Key takeaways

  • AI compute useful life, not hardware price, is the swing variable in owned-infrastructure business cases.
  • Require three-year, five-year and six-year life cases plus an explicit residual value in every proposal.
  • ASC 360 and IAS 16 mandate annual reassessment, so a build decision schedules a recurring audit conversation.
  • Always present a cloud baseline for the same workload, including the headcount an owned fleet requires.
  • Verify hyperscaler useful-life disclosures directly in the latest 10-K property and equipment footnotes.