The Invisible Write-Down

How Big Tech is carrying hundreds of billions in AI chips at prices the market stopped believing years ago

Liquidity Desk   |   Equity Deep Dive   |   April 2026

There is a number that does not appear on any earnings call slide deck. It lives quietly in the footnotes of annual reports, buried under accounting policy disclosures few analysts bother to read. That number is the gap between what the largest technology companies in the world claim their AI chips are worth and what those same chips would actually fetch in the market today.

This is not a story about fraud. It is a story about accounting discretion, the pace of technological change, and what happens when an industry collectively decides that the useful life of a $30,000 GPU is six years, while Nvidia ships a new generation every eighteen months.

Understanding it matters whether you own Microsoft, Alphabet, Meta, Amazon, or any company downstream in the AI infrastructure chain. It matters because depreciation is not a cash expense. It is a non-cash charge that flows through the income statement and directly shapes reported earnings per share. Extend the assumed life of an asset and earnings rise. Compress it and earnings fall. At the scale of capital deployed in 2025 alone, the difference is measured in tens of billions of dollars.

Part I: The Mechanism

What Depreciation Actually Does

When a company purchases a physical asset, generally accepted accounting principles do not allow the full cost to hit the income statement in the year of purchase. Instead, the cost is spread over the asset's expected useful life through annual depreciation charges.

The logic is sound. A server that generates revenue for five years should have its cost allocated across those five years, not front-loaded into year one. The problem arises when management sets the expected useful life and that estimate turns out to be considerably longer than the asset's economic life.

Consider a simple example built on actual 2025 capital expenditure. The five largest hyperscalers spent approximately $400 billion in 2025 on infrastructure, the majority directed at AI servers, GPUs, and networking. The annual depreciation charge depends almost entirely on the assumed useful life:

The difference between a three-year and six-year assumption on a $400 billion asset base is approximately $66 billion per year in reported expenses. That is not a rounding error. That is a number that moves earnings per share, price-to-earnings multiples, and analyst price targets across the entire sector.

"Even a small change of several months in the depreciation policies can change earnings in a given quarter by billions." - Olga Usvyatsky, Bloomberg, November 2025

Part II: The Scale of the Bet

2025 Actuals and 2026 Commitments

To appreciate the stakes, it helps to look at the actual numbers. The five major hyperscalers spent approximately $400 billion on capital expenditure in 2025, up roughly 60% from 2024. In early 2026, each of the four largest announced plans to increase spending again, dramatically:

The combined 2026 guidance from the Big Four alone exceeds $630 billion, with some analyst estimates placing the total including Oracle above $750 billion. For context, that is more than double what they spent just two years earlier, and it is happening while every company in the group simultaneously reports compressed or negative free cash flow.

Amazon, which spent $131 billion in 2025, is projected by Morgan Stanley to run negative free cash flow of nearly $17 billion in 2026 on its $200 billion commitment. Meta's free cash flow is expected by Barclays analysts to fall by close to 90% as capital expenditure approaches 54% of revenue. These are not normal technology company metrics. They resemble the capital structure of a utility or an industrial operator.

Every dollar of this capital expenditure lands on the balance sheet as a PP&E asset. And every one of those assets will be depreciated over the assumed useful life the company has chosen. The choice of that life is therefore not a minor accounting detail. It is a variable that determines whether these companies report the earnings their valuations imply.

Part III: The Great Extension

How the Industry Moved in Lockstep

The story of how hyperscalers arrived at six-year useful lives for AI hardware did not happen in isolation and did not happen by accident.

In January 2020, Amazon was the first to move. AWS extended the depreciation life of its servers from three to four years, citing operational data showing servers remained productive beyond initial assumptions. The reasoning was defensible. Moore's Law was slowing. At Amazon's scale, older hardware could still serve long-tail workloads, analytics jobs, and general-purpose compute.

What followed was a sequence of coordinated revisions across the industry, timed with striking precision to the years of peak AI capital expenditure:

• Amazon: 3 yrs - 4 yrs (2020) - 5 yrs (2022) - 6 yrs (2024) - 5 yrs again (early 2025, explicitly citing AI pace)
• Alphabet: 3-4 yrs - 6 yrs (2023). Quantified impact: approximately $3.9B annual depreciation reduction.
• Microsoft: 3 yrs - 4 yrs (2021) - 6 yrs (2022). Quantified impact: approximately $3.7B annual depreciation reduction.
• Meta: 3 yrs (2020) - incremental extensions - 5.5 yrs (2025). Quantified impact: approximately $2.9B annual depreciation reduction.
• Oracle: 4 yrs - 6 yrs. Independent analysts estimate operating income overstated by more than 20% as a result.

The collective impact across the group is estimated at approximately $18 billion in annual depreciation savings, cutting reported depreciation from roughly $39 billion to $21 billion on the same asset base. That is a 46% reduction in a non-cash expense flowing directly to the bottom line, achieved through a change in accounting estimates rather than a change in operations.

What gives analysts pause is the coordination. When several large companies arrive at the same conclusion within a short window, shortly before peak AI capital expenditure begins, the question is whether the accounting reflects genuine operational insight or whether one company's revision gave the others permission to do something that was financially convenient.

Part IV: The Divergence

Amazon Breaks from the Pack

The most revealing moment in this episode came not from an extension but from a reversal.

In February 2025, Amazon announced it was shortening the useful life of a subset of its servers and networking equipment from six years back to five years, effective January 1, 2025. The stated reason was explicit: the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.

In the same quarter, Meta extended its useful life assumptions to 5.5 years and booked a $2.9 billion reduction in depreciation expense. Two large technology companies, operating the same hardware, under the same technological conditions, reached opposite conclusions. Amazon also early-retired certain equipment in Q4 2024, taking a $920 million accelerated depreciation charge.

Amazon shortened lives and took impairment charges, citing AI pace. Meta lengthened them. Same chips. Same year. Different earnings.

This divergence is the clearest signal available to investors. It confirms that useful life is a management estimate, not a physical fact, and that different teams can look at identical technology cycles and reach dramatically different accounting conclusions depending on how much earnings pressure they face.

Part V: The Balance Sheet Question

Book Value, Market Value, and What the Rules Say

This is where the accounting gets important to understand precisely, because it is easy to misstate the problem.

Under both GAAP and IFRS, physical assets held for use in operations are recorded at historical cost minus accumulated depreciation. They are not marked to market. A server purchased at $30,000 and depreciated over six years sits on the balance sheet at $25,000 after year one regardless of what an identical server would sell for in the secondary market that day. This is not a defect in the accounting. It is the intended design for assets a company intends to use, not sell.

So the balance sheet, as presented, is not wrong under the rules. The recorded value of AI hardware at the large hyperscalers reflects management's best estimate of remaining useful life applied to historical cost. If a company genuinely believes its H100 fleet will serve productive workloads for six years, carrying it at depreciated historical cost is the correct accounting treatment.

The issue arises when the gap between book value and economic reality becomes large enough to trigger a different rule: impairment testing.

When the Rules Require a Write-Down

Under ASC 360 (GAAP) and IAS 36 (IFRS), a company must test long-lived assets for impairment whenever events or circumstances suggest the carrying amount may not be recoverable. Specifically, if the undiscounted future cash flows expected from an asset fall below its book value, the asset must be written down to fair value.

This is where the depreciation debate becomes a balance sheet debate. If a company refreshes its H100 fleet in year three because Blackwell and Rubin have made those chips economically obsolete for frontier workloads, it cannot simply walk away. It must test the retired assets for impairment and, if the secondary market value is below the remaining book value, record the difference as a loss on the income statement.

Amazon demonstrated this in Q4 2024. The $920 million accelerated depreciation charge was exactly this mechanism: hardware retired before the end of its assumed useful life, written down because it was no longer generating cash flows at the level implied by its book value. The charge was real. It hit earnings. And it was a direct consequence of AI hardware refresh cycles moving faster than the six-year depreciation schedule anticipated.

The latent risk on the balance sheets of companies maintaining six-year lives is therefore not that the balance sheet is misstated today. It is that the balance sheet contains a deferred impairment obligation. If refresh cycles accelerate, as Amazon's management explicitly acknowledged when it cited the AI pace as the reason for its reversal, a portion of the hardware currently carried at book value will eventually require either accelerated depreciation or outright impairment charges.

Secondary Market as an Early Warning

The secondary market for GPU rentals provides an imperfect but real-time signal of economic value. Rental rates for H100 capacity have fallen approximately 70% from their 2024 peaks. A unit renting at $30 per GPU-hour in early 2024 was available for well under $10 by late 2025.

This does not mean H100s are worthless. The hardware still runs. CoreWeave reports its older fleets remain booked at rental rates well above zero. Google has noted that its oldest custom TPUs remain at near full utilization. The value cascade argument, where hardware migrates from frontier training to inference to batch analytics over its life, has empirical support.

What it does mean is that the marginal revenue per GPU-hour at the frontier has collapsed, and that the economics of the assets as originally purchased, at peak 2023 to 2024 prices, for cutting-edge training workloads, do not match the economics of those same assets as they age. Whether the impairment test triggers depends on whether management's cash flow projections for aging hardware are realistic. That is a judgment call that external analysts cannot make with precision. But they can note the direction of the evidence.

Part VI: The Scorecard

Where Each Major Player Stands

Amazon: Revised and More Defensible

Amazon's decision to reverse course in early 2025 and take the $920 million early retirement charge makes it the most conservative in the group. By shortening useful lives and recognizing accelerated depreciation proactively, AWS is aligning its accounting with the observed pace of AI hardware refresh. Combined with its substantial underlying cash generation from the broader AWS business, this represents the most credible accounting posture in the group, even as its 2026 free cash flow turns negative on the weight of $200 billion in planned capital expenditure.

Alphabet: Strong Cash, Aggressive Assumptions

Alphabet extended to six years and reduced annual depreciation by approximately $3.9 billion. With 2025 capital expenditure at approximately $91 billion and 2026 guidance of $175 to $185 billion, the asset base being added to the balance sheet under six-year lives is growing rapidly. Free cash flow remains positive and the company has significant financial flexibility, but the impairment exposure grows proportionally with each year of elevated spending.

Microsoft: Large Base, Aware Management

Microsoft carries roughly $130 billion in server and networking PP&E under six-year lives, with 2025 capital expenditure of approximately $90 billion and 2026 guidance of $120 billion or more. The company has deployed over 485,000 Hopper GPUs plus 100,000 Blackwell Ultra units. CEO Satya Nadella's public comment that he did not want to be locked into multi-generational depreciation schedules suggests management awareness of the gap between accounting lives and hardware refresh reality, even if the financial statements do not yet reflect it.

Meta: Highest Capital Intensity, Most Compressed FCF

Meta's decision to extend to 5.5 years while Amazon was shortening is the starkest divergence in the group. With capital expenditure approaching 54% of revenue in 2026 and free cash flow projected to fall by close to 90%, Meta has the least room to absorb a future downward revision in useful life estimates. The $2.9 billion annual depreciation reduction booked through the extension is not negligible relative to the company's earnings trajectory. Independent analysts consistently flag Meta as the clearest earnings quality concern in the hyperscaler group.

Oracle: Leverage Amplifies the Stakes

Oracle presents the highest risk profile for a reason that goes beyond depreciation. The company is simultaneously running aggressive six-year useful life assumptions and financing its data center expansion through debt at a pace Moody's has flagged as significant. CDS spreads have moved to levels not seen since 2008. Capital expenditure as a percentage of revenue approaches 86% under 2026 guidance. In this context, the depreciation question is not merely an earnings quality issue. It is a question about the adequacy of asset coverage for the debt being issued against those assets.

Part VII: What It Means for Investors

The Questions Worth Asking

This analysis does not argue that the hyperscalers are committing accounting fraud. GAAP and IFRS explicitly permit management to set useful life estimates and revise them as conditions change. The extensions are disclosed in SEC filings. Auditors have signed off on them. The accounting is legal.

What it does argue is that the earnings quality of companies with large and growing AI hardware asset bases deserves more scrutiny than headline EPS numbers suggest. And that the balance sheets of those companies contain a latent impairment exposure whose size depends on how quickly management chooses to acknowledge the gap between accounting lives and economic lives.

If you own any of the companies discussed in this piece, a small set of questions is worth keeping in mind:

• What is the company's hardware refresh cycle in practice? If Blackwell is replacing Hopper clusters before the six-year depreciation clock expires, the accounting life is divorced from operational reality.
• Is management acknowledging the technology pace? Amazon's reversal and Nadella's public comments are positive signals. A company that never revises its useful life assumptions regardless of Nvidia's release cadence deserves skepticism.
• What does secondary market pricing imply? A 70% decline in H100 rental rates since 2024 peaks is an early warning signal worth tracking. It does not trigger impairment automatically, but it informs the direction of cash flow projections.
• What is the free cash flow position going into 2026? Companies with strong FCF can absorb a future write-down without structural damage. Those already at zero or negative FCF cannot.
• What is the debt structure? For highly leveraged operators, the depreciation question is not just about earnings. It is about the collateral supporting their debt obligations.

Burry's estimate that the industry faces $176 billion in cumulative understated depreciation charges between 2026 and 2028 may be aggressive. But the direction of the argument is sound. The companies spending $600 to $700 billion on AI infrastructure in 2026 are adding to balance sheets that already carry large positions in hardware whose secondary market value has declined sharply, under accounting assumptions that the most operationally sophisticated company in the group has already partially reversed.

The five hyperscalers together plan to add roughly $2 trillion in AI-related assets to their balance sheets by 2030. If those assets depreciate at approximately 20% per year in economic terms, the implied annual depreciation exposure exceeds the combined profits these companies reported in 2025. That is not a prediction. It is an arithmetic observation about the scale of the commitment being made and the assumptions required for it to be sustainable.

Conclusion

The AI infrastructure build-out is real. The capital being deployed is enormous. The demand for compute capacity is not manufactured. None of that is in dispute.

What is in dispute is whether the earnings generated by that infrastructure are being measured accurately, and whether the assets carrying that infrastructure are being held on balance sheets at values that reflect the pace at which the technology is actually evolving.

The balance sheet, under current accounting rules, is not wrong. A GPU held for operational use is carried at historical cost minus depreciation, not at market value. That is the design. But inside that design is a conditional: if and when the asset is retired before the end of its assumed useful life, or if and when future cash flows fall below book value, the gap between accounting assumption and economic reality must be recognized. Amazon showed what that looks like in practice.

The invisible write-down is not a scandal waiting to happen. It is a deferred accounting reconciliation, the size of which depends on decisions about hardware refresh that have not yet been made, at a scale of capital commitment that has no modern precedent in the technology industry.

The investors who navigate it best will be the ones who look past headline EPS, read the PP&E footnotes, and ask what it would cost to run the depreciation math at three years instead of six.

The number is there if you know where to look.

This article is published by Liquidity Desk for informational and educational purposes only. It does not constitute financial advice, investment advice, or a recommendation to buy or sell any security. All data sourced from SEC filings, publicly available earnings reports, and analyst research as of April 2026. Past performance is not indicative of future results.