Your AI CapEx Plan Has a Governance Gap Nobody's Watching

There's a moment that keeps happening inside enterprises right now, and it rarely makes it into a boardroom deck.

Finance builds a five-year depreciation schedule for a new GPU cluster. Procurement signs a multi-year cloud commitment to lock in pricing.

The technical team, meanwhile, is quietly aware that the chip generation they just bought will be obsolete for frontier workloads in eighteen months, and that the model they built the whole pipeline around might not be the one anyone's using by the time the asset is half depreciated.

Nobody's lying. Nobody's incompetent. Each function is doing its job correctly, using its own clock. Finance uses an accounting clock. Procurement uses a contract clock. Technical teams use an obsolescence clock. And right now, in most organizations, nobody owns the job of reconciling the three.

That gap has a name: Stranded AI Asset Risk — the risk that infrastructure, licenses, or custom-built AI capability become economically dead well before they become accounting-dead, with no governance mechanism designed to catch it before the CapEx is committed.

Most AI governance frameworks in circulation right now are built for a different problem entirely — model risk, bias, safety, regulatory compliance. Those matter, but they assume the AI is already built and deployed, and they ask "is this safe and fair to use?"

They don't ask the earlier, more expensive question: "will this asset still be worth anything by the time we've paid for it?"

That's a strategy and governance question, and right now it's falling into a crack between finance, procurement, and technical leadership.

This isn't a theory. The hyperscalers have been quietly running this experiment in real time, and the paper trail is public.

Between 2020 and 2023, Amazon, Microsoft, and Google all extended the assumed useful life of their servers and networking equipment — moving in stages from three or four years out to a uniform six-year schedule.

Microsoft's CFO told investors in Q4 FY2022 that the company was extending depreciable life on cloud infrastructure from four years to six; the disclosed benefit to reported earnings was roughly $3.7 billion in FY2023 alone.

Google made a similar move around the same time. The effect across the group was to meaningfully lower reported depreciation expense and lift operating income, at a moment when AI CapEx was about to explode.

Then, in 2025, something interesting happened: Amazon reversed course.

Effective January 1, 2025, it shortened the useful life of a subset of its servers and networking equipment from six years back to five — citing, in its own words, the accelerating pace of technology development in AI and machine learning. That single change cut net income by roughly $162 million in Q1 2025 and by nearly $300 million in Q3 2025 alone.

Meanwhile, Meta moved in the opposite direction, extending useful life on most of its servers and network assets even further. Three companies, buying largely the same Nvidia chips, running the same category of workload, made three different bets about how long the asset would remain useful, and those bets show up directly in reported earnings.

This is the tell. If the world's most sophisticated finance organizations, sitting on more operational data about their own hardware than anyone else on earth, can't agree on the answer, the assumption baked into your own capex plan is very likely someone's best guess — not a fact.

The investor Michael Burry made this the center of a public argument in late 2025, accusing hyperscalers of understating depreciation by as much as $176 billion between 2026 and 2028 and inflating reported operating income at some companies by more than 20% above what he considers economic reality.

Reasonable people dispute his numbers. What's harder to dispute is that the debate exists at all, in public, among people paid to know.

If the depreciation debate feels like an accounting footnote, the credit markets have made it a balance-sheet-shaping issue. CoreWeave — the largest of the pure-play NeoCloud GPU providers — has built its growth almost entirely on debt collateralized directly by Nvidia chips.

It started with a $2.3 billion facility in 2023; by 2026 its total debt exceeded $20 billion, some of it structured well enough to earn investment-grade ratings from Moody's and DBRS, backed by long-term contracts with customers like Meta.

Here's the detail worth sitting with: analysts covering that debt note that a high-end Nvidia GPU can lose roughly half its resale value within three years, even as it sits inside a loan structure rated as investment grade. The rating isn't really vouching for the hardware. It's vouching for the customer contract wrapped around the hardware.

Strip the contract away, and the chips underneath are depreciating in the real world on a clock that has nothing to do with the loan's amortization schedule.

That's a neocloud problem today. But the same logic applies, quietly, inside any enterprise that's financing, leasing, or long-term-contracting AI infrastructure and treating the underlying asset's economic life as settled because someone else's depreciation schedule says so.

The hardware story is the cleanest version of Stranded Asset Risk because it shows up in dollar figures. But the same failure mode shows up just as often — and arguably more expensively — in the software and pipeline layer, and here the data is even starker.

MIT's Media Lab published research in mid-2025 analyzing more than 300 publicly disclosed enterprise AI initiatives alongside dozens of executive interviews.

The finding: roughly 95% of generative AI pilots inside enterprises were producing no measurable financial return. Only about 5% had made it into production with real, trackable value. The gap wasn't explained by model quality. It was explained by how the organization integrated the tool into its actual workflows, and whether anyone was accountable for whether it was working.

Put that next to the hardware depreciation story and the pattern comes into focus.

Enterprises are stranding two kinds of assets at once: physical infrastructure that ages out faster than it's being depreciated, and pilot programs, custom models, and internal tooling that never clear the bar to become a durable asset in the first place but that absorbed real budget, real headcount, and real opportunity cost getting there.

Both failures trace back to the same root cause: nobody was assigned to ask, before the money moved, "what is the actual useful life of what we're about to build or buy, and who signed off on that assumption?"

It's worth being specific about why this doesn't get caught by the structures most companies already have.

1. Finance owns the depreciation schedule, but not the technology roadmap. They apply a standard useful-life assumption, often inherited from general IT equipment policy or benchmarked against what hyperscalers disclose, without a mechanism to test it against the actual pace of change in the specific AI stack being purchased.

2. Procurement owns the contract terms, but not the obsolescence risk. A three-year cloud commitment negotiated for favorable pricing looks like a win on the day it's signed. Whether it's still a win eighteen months later, once a materially better and cheaper option exists, is a question procurement isn't set up to keep asking after the ink dries.

3. Technical and AI teams see the obsolescence curve most clearly, but don't think in capEx terms. Engineers and data science leads know exactly how fast the ground is shifting under a given chip generation or model family. That knowledge rarely gets translated into the language finance and the board actually use to make capital decisions.

4. AI governance committees, where they exist, are almost universally scoped to model risk — bias, safety, hallucination, regulatory exposure — not asset lifecycle risk. It's simply not the question they were chartered to ask.

Each function is rational and competent within its own lane. The stranded asset risk lives entirely in the white space between the lanes, which is exactly why it's invisible until an impairment charge, a credit downgrade, or a board question forces it into view.

This is meant to be usable, not just descriptive. Here's a structure for building the missing governance layer.

1. Build a stranded asset risk map. Take every material AI-related capital commitment — compute, licenses, custom models, data pipelines — and score each one on two axes: the accounting or contractual horizon (how long you're committed to paying for it) against the realistic economic horizon (how long it will actually deliver differentiated value before a materially better or cheaper alternative makes it obsolete).

Anything where the accounting horizon is meaningfully longer than the economic horizon is a stranded asset candidate, and it should be visible to whoever owns capital allocation, not buried in a footnote.

2. Add a decision gate before CapEx approval, not after. Before any significant AI infrastructure or platform commitment gets signed off, require an explicit answer to one question: what happens to this asset in 18 months, and who is accountable for that assumption being right?

This doesn't need to be a heavy process. It needs to exist and be documented, so the assumption is owned by a named person rather than inherited by default.

3. Run an assumptions audit on what you've already committed to. Pull your own organization's depreciation schedules and contract terms for AI-related spend, and compare them honestly against what you know about the pace of change in that specific category — chip generations, model releases, vendor roadmaps.

You're not trying to replicate hyperscaler-level precision. You're trying to find out whether your own assumptions were inherited from a standard IT policy that was never designed for this asset class.

4. Assign clear ownership. This risk needs a home. It shouldn't sit exclusively with finance, which can't see the technical obsolescence curve, or exclusively with technical teams, who don't naturally think in capital terms.

The workable model is a small cross-functional review (finance, procurement, and technical/AI leadership) with a named owner accountable for flagging stranded asset risk before it's locked in, not after.

5. Track leading indicators, not just lagging ones. Don't wait for an impairment charge to find out an assumption was wrong. Watch the signals that move first: secondary market pricing for the hardware you've bought, vendor roadmap announcements that shorten the realistic life of your current stack, and, just as importantly, internal utilization data on what you already have. Infrastructure sitting idle is the earliest and quietest signal that an asset is stranding.

None of this requires predicting exactly how fast AI hardware or models will age. Nobody can do that reliably, including the hyperscalers, who disagree with each other in their own filings. What it requires is making the assumption visible and owned, instead of letting it default to whatever number finance inherited from a policy written for a different era of computing.

The organizations that get burned by this won't be the ones that made an aggressive bet and got it wrong. They'll be the ones that never realized a bet was being made at all.