The Assumption That Decides $48 Billion

Combined 2026 capex guidance across the four largest hyperscalers climbed to $720-745 billion following earnings calls through mid-year, then was revised upward by Moody's to $785 billion after Q1 results. That figure is up 77% from roughly $410 billion in 2025. Every one of those revisions landed as a press release, a conference call, a wire story. None of them mentioned the assumption doing more to shape reported profitability than the spend itself: how many years those assets will stay in service.

FASB chair Rich Jones added a research project on 6 April 2026 to monitor data infrastructure investments, defined as data centers, the financing behind them and their use in a company's business and related contractual arrangements. The project sits in the research phase; no timeline has been given for whether it moves to the technical agenda. But the fact that it exists at all tells the auditor what the board already knows: the hyperscalers extended the useful lives of their servers and networking equipment, and capitalised GPUs can be deployed to generate revenue before significant depreciation expenses have to be recorded.

The rule requires a useful life estimate that reflects how long a company expects to use the equipment. On 1 January 2025, Amazon and Meta made opposite accounting decisions about the same kind of asset under the same technological conditions. Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning". Meta went the other way, increasing the estimated useful life of certain servers and network assets to 5.5 years, effective beginning fiscal year 2025.

A $10 billion GPU fleet placed in service today, depreciated straight-line with zero salvage value, generates $2.50 billion in annual depreciation at a 4-year life and $1.67 billion at a 6-year life. That $0.83 billion per year difference flows through operating income and EPS even if unit economics and cash spend are identical. Meta added $2.9 billion in earnings by extending server life. Amazon cut it. Two of the most sophisticated capital allocators in the world produced judgments diverging by eighteen months on the same fleet, with billions of dollars of operating income riding on the difference.

The variance is legal. Both companies hired auditors who signed the financials. The standard does not prescribe a number; it asks for a supportable estimate. But when the Big Four are blessing 5-year, 5.5-year, and 6-year assumptions in consecutive 10-Ks covering the same asset class during the same quarters, the choice is no longer an engineering judgment. It is a P&L lever.

What the Auditor Can See

In 2020, the four major US cloud providers depreciated servers over three to four years. Then they all began to extend those lives, citing improvements in server durability and utilization. Microsoft now uses 6 years, extended from 4 years; Alphabet 6 years, extended from 4 years; Amazon 5 years, shortened from 6 years in January 2025; Meta 5.5 years, extended from 4 years in steps. Those are GAAP useful lives. The tax class life remains 5 years regardless of what a company reports on its financial statements.

AI accelerators have a useful life typically estimated at four to six years, bounded by physical degradation on one side and economic obsolescence on the other, as each new generation delivers step-change improvements in performance. Meta's Llama 3 405B training study implied an annualized failure rate of approximately 9% for high-utilization training clusters. Yet federal tax law treats them as 5-year property and many large operators report them on 5-to-6-year depreciation schedules for financial accounting, creating a gap between real replacement cost and reported earnings.

The auditor reviewing capex against depreciation sees a pattern. Meta's depreciation expense increased from $11.0 billion in fiscal 2023 to $15.2 billion in 2024 and to $22.0 billion in 2025, according to its annual report. The increase reflects infrastructure that has already been built and placed in service, rather than spending that is still to come. But the rise in depreciation lags the rise in capex by the time it takes to commission the asset and by the average life assumption applied. A six-year life spreads the charge; a four-year life pulls it forward. The divergence means earnings comparability across the peer set is gone.

Microsoft's 10-K provides only a useful-life range of two to six years for equipment. Satya Nadella has acknowledged the timing mismatch publicly, telling an interviewer in late 2025 that the company was deliberately spacing AI chip purchases to avoid being "stuck with four or five years of depreciation on one generation". That is a direct admission from the operator that the depreciation schedule does not match the replacement decision. The quote is in the public domain. The auditor reads it and then signs a 10-K that assumes the chips will be used across the span the P&L requires.

The Regulator's Constraint

Accounting treatment related to the way big tech firms are financing data center projects is drawing scrutiny. While companies are generally required to describe major risks to their business, no specific disclosure requirement applies to power-purchase agreements that tech firms are initiating to achieve their AI goals. Meta secured billions in financing to build a big data center in Louisiana without taking on the debt itself, with such deals generally allowing companies to push the risks associated with the AI boom onto lesser known companies.

The securities regulator can ask for more disclosure. It cannot tell a board which useful life to pick. The standard already requires periodic review of useful lives and salvage values. Proposed GASB guidance emphasizes the requirement for governments to perform a periodic review of the estimated useful lives and salvage values used in the calculation of depreciation expense for infrastructure assets. Public companies follow FASB, not GASB, but the principle is identical.

The constraint is this: the accounting allows judgment, and the auditor's role is to test whether the judgment is reasonable, not whether it is the same as the competitor's. A divergence between Amazon at 5 years and Alphabet at 6 years is not evidence of error. It is evidence that the standard gives preparers discretion and that the market has not yet forced convergence. The regulator who wants consistency either needs a prescriptive rule or needs to conclude that existing disclosures are insufficient for investors to compare like with like. Neither path is fast.

AI infrastructure combines long-lived buildings and power systems with hardware whose economic life may be much shorter and more uncertain. This raises basic questions about depreciation, collateral values, lease duration, refinancing risk, and the appropriate maturity of debt. It also creates incentives to separate infrastructure-intensive assets from high-growth operating companies in order to preserve reported capital intensity and valuation multiples. Research is needed to determine whether current accounting and performance measures adequately capture the economic cost of the capital employed.

That passage is from a Brookings BPEA conference draft on financing the AI buildout, dated 24-25 September 2026. It was written for economists and policymakers. The conclusion is the same one the securities regulator reaches when the S-1 queue fills with SPVs whose sole tenant is a hyperscaler and whose debt has been carved out of the parent: the accounting is not catching the structure.

What Keeps Moving

Microsoft expects to spend roughly $190 billion in calendar 2026, Meta has lifted its range to $125-145 billion, and Alphabet has nudged its outlook to $180-190 billion. Together with Amazon's spending, analysts tracking the four companies peg the combined total at up to $725 billion for the year. Microsoft CFO Amy Hood said on the earnings call: "For calendar year 2026, we expect to invest roughly $190 billion in capital expenditures, which includes approximately $25 billion from the impact of higher component pricing." Meta CFO Susan Li said, the capex increase "reflects our expectations for higher component pricing this year and to a lesser extent, additional data center costs to support future year capacity".

Amazon planned to invest approximately $200 billion in cash capex in 2026, with the majority supporting AI and AWS. Capex guidance was raised to approximately $220 billion in cash capex in 2026, with the increase driven by higher costs of memory. Component inflation is real. It shows up in the cash flow statement and in the press release. Depreciation is a non-cash charge. It does not change the amount leaving the bank account, but it changes the profit the company reports and the margin the analyst models.

A debt-funded expansion commits to a multi-year spend trajectory regardless of short-term demand signals. Q1 2026 earnings confirmed the guidance increases precisely because the hyperscalers had already locked in financing commitments and supply contracts that made cutting back more expensive than pushing forward. The equipment has been ordered. The power has been contracted. The useful life will be set when the asset goes live. The choice made in that moment will determine how much of the $725 billion gets expensed this year, next year, or the year after, without changing a single chip that was bought.

If you sit on a board, especially the audit committee, ask management what useful life is assumed for AI accelerators specifically, apart from general-purpose servers and buildings, and what evidence supports five or six years in a market moving on an annual cadence. That is the question this issue exists to put in your hands.

The director gets that question from an activist or a newsletter, asks it in the closed session, and receives an answer the CFO has already rehearsed with the auditor. The answer will cite utilization rates, workload migration, cascade strategies that move GPUs from training to inference to batch processing. All of that is true. None of it resolves the fact that same hardware, same month, opposite conclusions.

The Number That Compounds

The issue is not the absolute level of any single estimate. A 5.5-year life is not obviously wrong for a fleet that includes general-purpose servers, storage, and network fabric. The issue is the rate of change and the direction of travel. Every hyperscaler lengthened server lives between 2020 and 2024. Then Amazon reversed course in January 2025, the same month Meta moved the other way. Amazon stated in Q4 2024: "We completed our most recent servers and networking equipment useful life study and are changing the useful lives of a subset of our servers and networking equipment, effective 1 January 2025, from six years to five years. For those assets included in Property and equipment, net as of 31 December 2024, whose useful life will change from six years to five years, we anticipate a decrease in 2025 operating income of approximately $0.7 billion. We expect to continue to acquire more of these server and networking assets in 2025".

That is not a rounding error. The $0.7 billion refers only to the equipment already on the books, not the $220 billion of new equipment coming. Apply a one-year difference across that base and the swing is material to any income statement.

The regulator cannot compel convergence. The auditor cannot refuse to sign a judgment that sits within the range the standard permits. But FASB's research project exists because the divergence is now large enough, public enough, and frequent enough that the board is being asked whether the guidance is still fit. Staffers will eventually present information to the FASB board to determine whether to add items to the FASB technical agenda, according to a news release. The release didn't give a timeline for when that may happen.

In the interim, the assumption that decides $48 billion in aggregate reported depreciation across the sector is being set one company at a time, by the same management teams whose compensation is tied to the operating income the depreciation directly reduces. The feedback loop is obvious. The remedy is not.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.