The Most Important Number in the AI Boom Is a Guess Filed With the SEC

There is a number in every hyperscaler's financial statement that is not a record of anything. It is a forecast — about the future, made by the company, disclosed in a footnote, and used to decide how much profit the company gets to report.

The number is the estimated useful life of its servers.

Amazon has changed that estimate three times in four years. Each change moved billions of dollars between one year's earnings and another's, and the sequence is the most legible evidence we have of something the AI buildout has not yet had to prove: that the hardware it is buying will still be worth something when the debt matures.

Two extensions, then a reversal

Effective January 1, 2022, Amazon extended the estimated useful life of its servers from four years to five and its networking equipment from five to six. Its 10-K attributed the longer lives to "continuous improvements in our hardware, software, and data center designs." The change cut depreciation and amortization expense by $3.6 billion for the year and improved Amazon's reported net loss by $2.8 billion.

Effective January 1, 2024, the company extended servers again, from five years to six. That change reduced depreciation by $3.2 billion and benefited net income by $2.5 billion — about 23 cents a share.

Two extensions. Roughly $5.3 billion in cumulative benefit to reported results — $2.8 billion of it reducing a net loss in 2022, $2.5 billion of it adding to net income in 2024. Hudson Labs' compilation quotes each 10-K verbatim.

Then, effective January 1, 2025, Amazon reversed part of it. It shortened the estimated life of a subset of servers and networking equipment from six years back to five. The 10-K gave the reason directly: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning."

The reversal cost the company $1.4 billion in additional depreciation and $1.0 billion in net income for the year, mostly in AWS.

Read the sequence again. The same company, applying the same accounting standard, decided twice that its hardware would last longer — and then decided that AI was making it wear out sooner. Both decisions were legal. Both were disclosed. Both were, on the evidence available, defensible. And together they determine what AWS reported as profit in four consecutive years.

The scaled consequence is now visible in the segment disclosures. AWS depreciation and amortization ran $2.8 billion in the first quarter of 2024 and $7.3 billion in the first quarter of 2026.

Why the number is unverifiable by design

Depreciation exists to charge the cost of an asset against the years in which it earns revenue. It is a matching device. It is also, unavoidably, a projection — and GAAP treats a change in estimated useful life as a change in accounting estimate, applied prospectively, which means the effect lands in the current period and no prior year is restated.

Nothing about that is improper. It is what the standard requires when facts change. But it has a consequence worth stating plainly: the estimate can move by billions of dollars without any cash changing hands, without any prior figure being corrected, and without any external party being able to check it.

A revenue figure can be traced to a contract. A cash figure can be traced to a bank. The useful life of a GPU can be traced to nothing, because the remaining value of the asset depends on what Nvidia ships next, on whether the buyer can find workloads to run on it, and on whether anyone wants to rent it in year four.

There is a secondary market, and it does not agree with itself. Published 2026 estimates of used H100 resale prices run from roughly $15,000 to $28,000 per card, with other sources putting the range at $18,000–$22,000 and the same index pegging retention at 75–85% of value through 24 months in service. Those are not rounding differences. That spread is what a market looks like before it has found a price.

The order of magnitude is what matters, not any single quote. Across four quarters through March 2026 the four largest US hyperscalers purchased about $433.9 billion of property and equipment against roughly $149 billion of reported depreciation — a gap that is, definitionally, the future. Calendar 2026 capex guidance for the same group sits in the $720–745 billion range, with one tally at $725 billion. When the residual value of an asset is genuinely uncertain, so is the depreciation schedule of the new one — and the schedule is the number the earnings depend on.

The vendor has an opinion, and it is long

On August 10, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish what it called independent compute financing platforms, intended to mobilize more than $500 billion of third-party capital for AI infrastructure.

Read as a marketing document — which is what it is — it makes an argument about asset life. Nvidia compute, it says, "provides the lowest token cost, highest revenue and longest life." Jensen Huang, in the same document: the platform is "continuously improved through CUDA software — extending its useful life and improving its economics over time."

That is the supplier publicly opining on the depreciation schedule of the thing it sells. Which is not an accusation of anything. It is the natural position of a vendor, and it may well be correct. But it belongs in the record next to the accounting, because the party with the strongest interest in a long assumed life is now also the party supplying the asset and convening the capital that buys it.

The financing now runs through the same asset class

Here the steelman deserves a hearing, because it is strong.

Independent underwriting is exactly how this kind of risk should be distributed. Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR are the institutions with the duration and the analytical capacity to hold long-lived infrastructure exposure; it is what they exist to do. Moving half a trillion dollars of compute financing onto their books, rather than onto the vendor's, is a genuine improvement over the precedent.

And the precedent is Lucent. In the late 1990s Lucent Technologies lent money to its own customers so they could buy Lucent equipment — booking the sale immediately and the credit risk later. When the telecom bubble broke, the customers defaulted and the revenue proved to have been financed by the seller. Lucent's fiscal 2001 annual report documents the aftermath; the telecoms crash took the equipment vendors and most of their customers with it. The failure mode was not that the technology was worthless. It was that the vendor's balance sheet and the customer's ability to pay were the same balance sheet.

Nvidia is not doing that. It is convening third parties instead of lending its own book, and the platforms are described as independent, financing the buildout "across NVIDIA's ecosystem." That is a real difference and it should be said.

Where the argument narrows is on the second word. The pools of capital are for NVIDIA customers, built around NVIDIA infrastructure; the release's own framing of the asset is a single reference architecture, one CUDA ecosystem, one "deep global ecosystem of developers, customers and offtakers." So the independence is in the underwriting, not in the exposure: six institutions, one asset class, one vendor roadmap. Correlated exposure across six independent underwriters is not diversification. And whatever useful life they underwrite is a life the vendor is simultaneously, publicly, arguing is long.

What this has to do with agents

The estimate only works if the assets are used. An idle accelerator depreciates on schedule regardless.

So the depreciation assumption creates a standing obligation to find workloads — and the largest source of new demand for inference at scale is agent deployment. Agents are, structurally, the cheapest producers of value on that compute, because the value they produce is captured by whoever runs them.

The useful-life estimate flows into reported earnings. Reported earnings flow into the debt covenants and the equity story. The pressure to justify the estimate therefore becomes pressure to deploy — at whatever terms are on offer.

Nobody designed the terms of agent deployment around a depreciation schedule. But the schedule is one of the inputs into how much that deployment has to earn, and therefore into how hard the people setting those terms need to bargain.

That last step is an inference, not an observation, and it is the part of this argument I would most want contested.

What I don't know

A few things, stated plainly.

The Harvard Business School material on this subject is paywalled, and I worked from published abstracts: "Meta: Accounting for AI Data Center Depreciation" (Heese and Korganbekova, June 2026), which examines "the role depreciation assumptions play in shaping reported earnings"; "Meta Platforms: Accounting for the AI Arms Race" (Srinivasan, Shin and Kim, March 2026), which describes a $27 billion off-balance-sheet data center lease, accounting decisions that "reduced depreciation expense by billions," and an auditor's decision to flag the structure as a critical audit matter; and "Circular AI Deals: Strategic Flywheel or Fragile Stack?" (Korganbekova, Serafeim and Pinckney, May 2026), listed on the HBS faculty page.

I could not open Amazon's 10-K text directly for this piece. The filing language quoted above comes from Hudson Labs' compilation, which names the source documents; the primary source is the filing, and the quotations should be checked against it before this goes further.

The used-GPU price sources disagree with each other and I have reported the disagreement rather than resolving it, because the disagreement is the finding.

And management may simply be right. Estimating the service life of a machine that has not finished being obsoleted is genuinely hard, and nothing in the record I have described suggests misconduct. Changing an estimate when the facts change is what the standard asks a company to do.

The question is not whether Amazon acted in bad faith. It is who is allowed to know whether the estimate is right — and what happens to the financing if it isn't.

Sources

• NVIDIA, "NVIDIA Partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital," press release, August 10, 2026. https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital
• Hudson Labs, "Amazon Server Depreciation (AMZN)" — compilation of Amazon 10-K FY2022, FY2024, FY2025 and 10-Q disclosures, quoted verbatim. https://www.hudson-labs.com/research/amazon-server-depreciation-amzn
• DeepQuarry, "Amazon revises server lifespan amid AI depreciation debate" — treatment of change-in-accounting-estimate under GAAP. https://deepquarry.substack.com/p/amazon-revises-server-lifespan-amid
• Heese, J. and Korganbekova, M., "Meta: Accounting for AI Data Center Depreciation," Harvard Business School case 126-034, June 2026 (abstract; full case paywalled). https://hbsp.harvard.edu/product/126034-PDF-ENG
• Srinivasan, S., Shin, S. and Kim, J., "Meta Platforms: Accounting for the AI Arms Race," Harvard Business School case 126-070, March 2026 (abstract; full case paywalled). https://store.hbr.org/product/meta-platforms-accounting-for-the-ai-arms-race/126070
• Korganbekova, M., Serafeim, G. and Pinckney, C., "Circular AI Deals: Strategic Flywheel or Fragile Stack?," Harvard Business School, May 2026 (via HBS faculty listing). https://www.hbs.edu/faculty/Pages/profile.aspx?facId=1563739
• Silicon Analysts, "Hyperscaler AI Capex and the Depreciation Wall, 2026" — capex versus reported depreciation, four quarters through March 2026. https://siliconanalysts.com/analysis/hyperscaler-ai-capex-depreciation-wall-2026
• TMT Finance, "2026 hyperscaler capex tops US$700bn — analysis." https://www.tmtfinance.com/intel/2026-hyperscaler-capex-tops-us700bn-analysis
• AI Weekly, "Amazon, Microsoft, Alphabet, Meta plan $725B AI capex in 2026." https://aiweekly.co/alerts/amazon-microsoft-alphabet-meta-plan-725b-ai-capex-in-2026
• ServerBuyBack, "GPU Resale Value Index" — used-accelerator pricing and retention through 24 months in service, 2026. https://serverbuyback.com/resources/gpu-resale-value-index/
• GPUnEx, "Buying and Selling GPUs in 2026" — used H100 price range. https://www.gpunex.com/blog/buying-selling-gpus-2026/
• Lucent Technologies, Form 10-K for fiscal year 2001, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1006240/000095011702003045/a33915.htm
• Wikipedia, "Telecoms crash" (background on the 2001 telecom equipment collapse). https://en.wikipedia.org/wiki/Telecoms_crash