The Useful Life of a Useful Life

On January 1, 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.” Combined with the carry-through from $920 million of accelerated depreciation Amazon had booked in the fourth quarter of 2024 for related equipment retired early, the total impact on AWS operating profit in 2025 came to approximately $1.3 billion.¹ On the same effective date, Meta extended the useful life of certain servers and network assets from a prior range of four to five years up to five and a half years, reducing 2025 depreciation expense by approximately $2.9 billion, a figure the company’s full-year 2025 results confirmed flowed through as a decrease in the depreciation growth rate.² Two firms, the same chips coming out of the same supply chain, hit by the same pace of generational change from NVIDIA — and the accounting moved in opposite directions, with a combined swing in reported operating income on the order of $4.2 billion.

Both decisions are correct under GAAP. Useful life is meant to reflect each company’s expected use of the asset, and two firms operating similar hardware can arrive at different, fully compliant estimates if their workloads, cooling architectures, or replacement strategies differ. That is the framework working as designed. But the magnitude of the divergence matters. Two decades ago, two large enterprises looking at the same fleet of servers might have produced useful-life judgments within six months of each other on a roughly five-year asset, and the dollar consequence would have been small enough to live inside ordinary accounting noise. In 2025, 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 range of reasonable judgment has widened, and it has widened because the same chip is simultaneously subject to three different useful-life judgments — one set by NVIDIA’s eighteen-to-twenty-four-month architecture cadence, one set by the workload migration to newer generations, and one set by the supply constraint that keeps the older generation in service even after the newer one is available. The single useful-life number on the balance sheet has to hold all three at once. Amazon and Meta moved the number in opposite directions because each was responding to a different one of the three.

Three Judgments, One Number

Depreciation is a timing assumption. The cost of a long-lived asset is recognized as expense over the period the asset is expected to produce revenue, and the central judgment is how long that period is. The framework was built for asset classes where one number can carry that judgment. A delivery truck is depreciated over seven years because the operator plans to use it for roughly seven years and the truck performs roughly the same way across that period. An airliner is depreciated over twenty-five years because the operator plans to fly it for roughly twenty-five years and the airframe holds its operating profile across the cycle. The accountant and the operator are looking at the same horizon.

AI infrastructure does not work this way, and the three judgments named in the opening operate on the same fleet at the same time.

The first judgment is the one NVIDIA forces on the operator. NVIDIA designs the graphics processing units, or GPUs, that do the bulk of the computation inside an AI data center. New GPU architectures are now shipping every eighteen to twenty-four months and each generation delivers two to three times the performance per watt of the previous one.³ From the perspective of frontier model training, the H100 that was the most valuable chip in the data center in 2023 is no longer the chip an operator would buy for that workload in 2026 — the H200 and B200 generations that followed have displaced it. The chip still runs. It does not earn at the rate a frontier-grade asset is supposed to earn, because the workload that pays frontier rates has moved on.

The second judgment is the one the workload migration forces on the same chip. The H100 that no longer earns at frontier rates is not retired. It cascades down to inference — the act of running a trained model to answer queries, rather than training it — and to lower-margin compute, and it generates revenue at those rates for years. CoreWeave’s CEO has cited a specific case in which H100s from a 2022 contract, returning to the market after the original contract expired, immediately rebooked at roughly 95 percent of their original prices — evidence, in his framing, that previous-generation chips retain substantial value as inference demand absorbs them off the training frontier.⁴ The accountant looking at the workload-cascade revenue curve sees a long-lived asset with a defensible six-year useful life. The accountant looking at the frontier-revenue curve sees an asset whose economic life ended at twenty-four months.

The third judgment is the one the supply constraint forces on everyone. The B200 that would, in an unconstrained market, allow the operator to retire the H100 from inference workloads and run the entire fleet on current-generation silicon is not available in those volumes. The constraint is not in the GPU design but in two components further up the supply chain — high-bandwidth memory (HBM), the stacked DRAM that feeds the GPU at the speeds modern models require, and CoWoS, an advanced packaging process at TSMC that bonds the GPU and the memory onto a single substrate. Both are bottlenecked, and TSMC’s most advanced manufacturing capacity is also shared with the high-performance central processing units, or CPUs, that orchestrate the GPUs inside each server. As AI workloads have shifted from pure training to agentic and inference-heavy patterns, Intel disclosed on its first-quarter 2026 earnings call that the CPU-to-GPU ratio inside AI data centers has already tightened from roughly one CPU per eight GPUs to one per four, with one-to-one projected for agentic workloads — squeezing the same supply chain from a second direction.⁵ The H100 stays in service longer than its successor’s release date would otherwise imply, because the successor cannot displace it at the speed of the order book.

A defender of current accounting would say none of this matters. Depreciation is meant to track cash generation across the period the asset earns, not technological relevance, and the H100 keeps earning whether or not it remains the frontier chip. That is a serious objection and it has the virtue of describing how the framework was built. The problem is that the H100 does not earn the same cash across its life. It generates frontier-rate revenue for roughly two years, cascade-rate revenue for several more, and is held in service beyond that by a supply constraint that is unrelated to anything intrinsic to the chip. The CFO defending the current schedule is implicitly assuming a flat or smoothly declining revenue curve. The actual curve is stepped, and the steps correspond to economic phases that pay radically different rates. Straight-line depreciation across that curve is not tracking cash generation. It is averaging across regimes the framework cannot see.

These three economic realities are compressed into a single accounting judgment. A useful life of twenty-four months captures the first. A useful life of six years captures the second. A useful life that extends further still captures the third. The single number on the books has to be one of these, or some compromise across them, and the operator chooses which compromise to make. Amazon’s choice to shorten by twelve months leans toward the first judgment. Meta’s choice to extend by six months leans toward the second. Both are responding to the same fleet under the same conditions, and the framework gives them no way to reflect all three forces simultaneously. Individual useful-life changes at the major hyperscalers between 2022 and 2025 have moved billions of dollars of operating income across reported periods, with no change in the underlying cash flow.⁶

What the Allocation Hides

The framework problem does not stop at the useful-life number. It extends into how depreciation is allocated across the reporting segments that the market actually watches. The hyperscalers do not run separate fleets of GPUs for separate businesses. They run one fleet, centrally managed, and the depreciation flowing from that fleet is allocated across segments by internal rules that the firms do not fully disclose. The rules were designed for a more homogeneous era. They are now being asked to allocate the cost of the largest and most heterogeneous capital base any of these companies has ever held.

Alphabet states the principle in its 10-K. Technical infrastructure is managed centrally at the consolidated level and the associated costs, including depreciation, are allocated to operating segments as a service cost generally based on usage or headcount.⁷ A TPU — Google’s custom AI accelerator, an alternative to NVIDIA GPUs for certain workloads — may support Search, Gemini, and Vertex AI workloads through the same shared infrastructure. The depreciation on that TPU is split across Google Services, where Search and the consumer Gemini app live, and Google Cloud, where the enterprise AI services live. The allocation method is described as usage-based but the precise weighting is not disclosed. The reader of Google Cloud’s segment margin cannot see how much of the segment’s reported depreciation reflects TPUs used for cloud workloads versus the share of central infrastructure assigned to the segment by an allocation rule.

Amazon’s structure is similar in principle and cleaner in disclosure. The 10-Q is explicit that technology infrastructure assets are allocated among segments based on usage with the majority allocated to AWS.⁸ When Amazon shortened useful life by twelve months in 2025, the disclosed impact was that the change “primarily impacted our AWS segment.” But “primarily” is not “exclusively.” A portion of that $700 million flowed into other segments through the same usage-based rule, and the reader cannot see exactly how much. AWS is the largest single consumer of Amazon’s compute infrastructure but it is not the only one.

Microsoft is the most opaque of the three. The company’s 10-K provides only a useful-life range of two to six years for equipment.⁹ 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”¹⁰ — a direct admission from the operator that the depreciation schedule does not match the replacement decision. The allocation rule that determines whether this mismatch surfaces in Intelligent Cloud, in Microsoft 365 Copilot, or elsewhere is not disclosed.

The familiar complaint is that analysts cannot audit segment allocations. That has always been true. The newer and more uncomfortable observation is that management may not be able to interpret them perfectly either. Internal product-level economics — the numbers used to decide which workloads earn their cost of capital, which products to price up or sunset, which lines to invest in — use the same allocation rules as external segment reporting. If those rules are compressing three contradictory useful-life judgments into one number and then redistributing that number across segments by undisclosed weights, the internal product economics inherit every distortion the external segment economics inherit. The firms making the largest capital allocation decisions in the history of the private economy are doing so partly on signals their own framework can no longer reliably produce. Alphabet’s first-quarter 2026 depreciation expense rose forty-four percent year over year, from $4.5 billion to $6.5 billion, while capital expenditure more than doubled, from $17.2 billion to $35.7 billion.¹¹ The consolidated numbers flow into the income statement immediately. They flow into segment margins, and from there into internal product economics, through allocation rules that no analyst can audit and that management itself may struggle to interpret cleanly.

Depreciation, Reconsidered

The cloud segment margin is the most-watched number in the most-watched business at four of the five largest companies in the world. Investors price the stock on it. Boards approve capital plans on it. Analysts build models on it. Internally, the same number, broken down by product line, tells the operator which workloads are earning their cost of capital and which are not. The number does a great deal of work.

The framework producing the number was designed for a different business. Server depreciation policy was built when servers were a homogeneous class of asset, replaced on a predictable cycle, generating revenue across a relatively flat profile, and small enough as a share of total operating cost that the timing assumption mattered less than the cash. Each of those conditions has changed. The fleet is heterogeneous, with multiple generations of accelerator running in parallel and a tightening ratio of CPU to GPU. The replacement decision is decoupled from the depreciation schedule, set by NVIDIA’s eighteen-month cadence rather than the operator’s five-year books. The revenue profile is no longer flat — the same chip produces frontier-rate revenue for two years, cascade-rate revenue for several more, and is held in service beyond that by a supply constraint that operates independently of the chip’s own characteristics. And the depreciation expense is no longer a small share of operating cost. In a November 2025 conversation with Satya Nadella, the interviewer framed the depreciating asset as approximately seventy-five percent of the total cost of ownership of a data center across its life — a figure Nadella accepted without contesting in the discussion that followed.¹² The timing assumption is now the cost structure.

The Amazon-Meta divergence shows that the framework permits a $4.2 billion swing in opposite directions on identical underlying physics, and both firms are correct under GAAP. Michael Burry’s public argument since November 2025 — that the hyperscalers are systematically inflating earnings by extending useful lives — is a debate about whether the numbers chosen inside the framework are conservative or aggressive.¹³ The deeper question is whether any single useful life number can produce a reliable signal when the conditions the framework was designed for have changed across every dimension that matters. The profitability signal that guides hundreds of billions of dollars of AI capital allocation may be partially disconnected from the economic reality of the assets producing it.

The first signs will appear in impairment charges, in accelerated-depreciation disclosures, and in the unprompted useful-life revisions that began with Amazon in January 2025. Until those signs accumulate, the number on the page is the best the framework can produce. The deeper question, and the one the next few earnings cycles will begin to answer, is whether depreciation remains a useful proxy for asset consumption when the asset itself moves through multiple economic lives before it reaches the end of its accounting one.