Signal Scanner · ARTIFICIAL INTELLIGENCE & AUTOMATION

AI’s Reported Profits Rest on an Accounting Guess About How Long GPUs Last

The AI-bubble debate fixates on demand. The quieter fragility sits on the books: hyperscaler earnings increasingly depend on the useful life assigned to a GPU, a discretionary estimate a single reset could cut by roughly a fifth across 2026-2028.

The consensus argument about AI has narrowed to one question: is it a bubble? The Bank for International Settlements gave a chapter of its 2026 annual report to the risk that a trillion-dollar buildout ends the way the dot-com boom did (Bank for International Settlements, 28/06/2026). Investors argue over demand and whether the capex pays off. Beneath it sits a smaller, technical number that decides how profitable the buildout looks today: the useful life a hyperscaler assigns to a graphics chip before it is written off. Stretch that number and reported profit rises; shorten it toward the real replacement cycle and a large slice of AI-era earnings disappears. The weak signal is that this accounting choice, not demand, is the nearer fault line.

Signal Identification

This is an emerging inflection in earnings quality, not a capability story. The four-to-six-year depreciation schedules hyperscalers apply to AI servers have drifted from the two-to-three-year cadence at which frontier GPUs become uneconomic. As the gap widens and the asset base compounds, a move to shorter lives would land as a sudden, large cut to reported profit.

Time horizon: 1–3 years (divergence already visible in FY2025-FY2026 filings; a reset most likely across the 2026-2028 reporting cycles) Plausibility band: Medium Geographic / Jurisdictional Scope: Primary: the United States, where hyperscaler filings, GAAP useful-life discretion and most AI capex sit. Spillover: global equity and credit markets exposed to US mega-cap AI names and AI private credit. Sectors exposed: Cloud hyperscalers and AI labs; semiconductors; data-centre developers and neoclouds; private-credit and leveraged lenders; equity investors; audit committees.

What’s Changing

Goldman Sachs Global Institute put a name to the fault line in May. Of the four supply-side assumptions that set the scale of the buildout, it found the economic useful life of AI silicon to be the single most influential (Goldman Sachs Global Institute, 01/05/2026). Data-centre shells depreciate over about 20 years; the accelerators inside them turn over far faster. Because silicon is such a large share of the spend, moving assumed life between four and six years swings annual depreciation by hundreds of billions of dollars. Its caution: the books can show orderly depreciation while the hardware goes obsolete far faster.

The divergence is already in the filings. Meta lifted the useful life of its servers and network assets from four years to five and a half, cutting 2025 depreciation by $2.9 billion even as its capex climbed from $72.2 billion toward a projected $135 billion (Harvard Business School, 01/06/2026). Microsoft, Google and Oracle made similar extensions between 2022 and 2024. Then, in February 2025, Amazon went the other way, shortening a subset of servers back to five years, citing the pace of AI development and booking a $920 million accelerated-depreciation charge (Footnote Brief, 21/05/2026). Same chips, opposite calls.

Assumed useful life of AI servers and accelerators (years)

Peers (MSFT, Google, Oracle) 6.0 Meta (from 2025) 5.5 Amazon (server subset) 5.0 Replacement-cycle view 2–3

Assumed depreciable life versus the replacement-cycle view of when frontier GPUs become uneconomic. Source basis: Harvard Business School (01/06/2026); Footnote Brief (21/05/2026).

Disruption Pathway

The pathway runs in three stages. First, divergence, now: one hyperscaler already depreciates part of its fleet faster than its peers, and each FY2026 filing reopens the question. Second, a trigger, plausibly 2026-2027: a second hyperscaler matching Amazon, a sharp fall in secondary GPU rental rates, or an SEC comment would force the others to defend or shorten their schedules. Third, the reset, 2026-2028: a shift toward a four-year or shorter basis. A four-year counterfactual implies roughly $200 billion of suppressed depreciation over three years, close to a prominent short-seller’s $176 billion estimate (Footnote Brief, 21/05/2026).

The stress does not stay on the income statement. AI capex now runs ahead of the hyperscalers’ cash flow, pushing them toward debt and opaque circular financing among chipmakers, labs and clouds (Bank for International Settlements, 28/06/2026). Because computing infrastructure has roughly doubled to about 1.5% of US GDP and become the leading driver of private-investment growth (Epoch AI, 05/06/2026), a profit revision that cools capex is a macroeconomic event, not a footnote. Two adaptations follow: investors and lenders should underwrite AI-era earnings on a normalised depreciation basis, not the reported figure; and data-centre financiers should price the risk that the collateral ages faster than the debt.

Why This Matters

For boards, CFOs and investors, useful life has quietly become a governance question. The IMF’s April stability report warned that leverage financing AI investment, combined with rapid capital-obsolescence risk, raises uncertainty around future earnings and valuations (IMF, 14/04/2026). That is the mechanism here: a defensible accounting estimate, applied across a compounding asset base, is holding up reported profit for the mega-caps that dominate the index. Audit committees should be able to show why their assumption survives Amazon’s reversal and falling secondary GPU prices; investors should know how much of their AI exposure is earnings and how much is a schedule.

Decision-action posture for this signal: Prepare — the divergence is in the filings now but a reset has not yet been forced; treat a second hyperscaler shortening its schedule, or a sharp drop in GPU rental rates, as the trigger to move to Decide.

Counter-Argument

The strongest objection is that the longer schedules are right. Hyperscalers run tiered fleets: frontier chips train for a year or two, then cascade to inference, and some 2017-era hardware still earns revenue. Goldman notes trailing-edge A100 and H100 rental prices have stayed high enough to imply five-to-six-year lives (Goldman Sachs Global Institute, 01/05/2026), validating the prevailing treatment. On that reading Amazon is being conservative on one cohort, not flagging an industry-wide error.

The cascade was real for the pre-AI fleet. It is unproven for frontier training silicon whose value collapses the moment a materially better generation ships on an annual cadence. And the objection need not be fully wrong to bite: even Amazon’s own five-year basis, applied across the group, implies tens of billions in suppressed depreciation. The live question is not whether reported earnings are flattered but by how much.

Implications

This is a durable repricing of earnings quality, not an accounting spat. The inflection window is the 2026-2028 reporting cycles, when the compounding asset base makes the depreciation gap too large to ignore and the FY2026 filings test whether Amazon’s reversal spreads. Investors that mark AI earnings to a normalised useful life will be positioned; those anchored to reported profit, and the valuations and capex plans built on it, are exposed if the schedules shorten. Whether or not a bubble bursts, the buildout’s profitability rests on one estimate more than the headline numbers admit.

Early Indicators to Monitor

Disconfirming Signals

Strategic Questions

Keywords

GPU depreciation; useful life; hyperscaler earnings; AI capex; ASC 360; capital obsolescence; AI buildout; data centre economics; earnings quality; AI private credit; AI bubble; accounting estimates

Bibliography

Source tiers: Tier 1, governments, regulators and intergovernmental bodies. Tier 2, think-tanks, academic institutes, major consultancies and quality data providers. Tier 3, quality journalism and specialist trade press. Tier 4, vendor, company and practitioner sources, used only as directional corroboration.

This Signal Scan is one component of a Shaping Tomorrow intelligence cycle on Artificial Intelligence & Automation. Every monitored topic also includes monthly Change Trackers that assess what is materially shifting in momentum, and a quarterly Decision Intelligence briefing that converts the cycle into board-level strategic implications. Cycles may also close with Action Triggers, an end-of-cycle synthesis that reads across the three formats and surfaces patterns no single deliverable carries on its own. Continuous intelligence on this topic is available as a Single-Topic or Multi-Topic programme.

Discuss a Tailored Programme

Prepared by Shaping Tomorrow: 11 July 2026