Amazon's ROIC Dropped From 28% to 22%. The 10-K Explains Why.

Amazon's fourth quarter of 2025 was excellent by every traditional measure. Revenue grew 12.4% to $213.4 billion. Operating margins expanded to 11.7%. AWS reaccelerated to 24% growth. Operating cash flow hit $139.5 billion for the full year — up 20%.

Then Amazon announced $200 billion in 2026 capital expenditures, the largest single-year commitment in corporate history. The stock dropped 11%.

The selloff wasn't about the quarter. It was about a single question: what happens to returns on invested capital when you double an already-massive infrastructure program?

The numbers that matter most are the ones that don't fit together. Operating cash flow is $139.5 billion. Free cash flow is $7.7 billion. The gap — $131.8 billion — is the capex. For every dollar Amazon generates in operating cash, it reinvests 94 cents. This isn't a company returning capital to shareholders. This is a company transforming its balance sheet at a pace that has no historical precedent in American business.

Will this capital eventually earn returns that justify the transformation? Or is Amazon building infrastructure faster than the market can absorb it?

To answer that, we need to look inside the return itself.

Why Amazon's ROIC Declined 6 Points — And It's Not About Margins

Amazon's return on invested capital fell from 28.1% to 21.7% over the past year — a 6.4 percentage point decline. In most companies, a decline of that magnitude would signal serious operational trouble: pricing power eroding, costs spiraling, competitive position weakening.

But when you decompose the return into its two constituent parts — profit margin and capital turnover — something unexpected emerges.

NOPAT margin went from 9.30% to 8.97%. That's a decline of 0.33 percentage points. Barely perceptible. The operating engine is earning roughly the same return on each dollar of revenue that it earned a year ago.

Capital turnover went from 3.02x to 2.42x. That's a collapse of 0.60x — and it accounts for 86% of the entire ROIC decline.

For investors who aren't familiar with the framework: DuPont decomposition is a method that breaks return on invested capital into two multiplied components — profit margin (how much you earn per dollar of revenue) and capital turnover (how much revenue you generate per dollar of invested capital). When ROIC declines, this decomposition tells you whether the problem is profitability, capital efficiency, or both.

In Amazon's case, the answer is unambiguous. The operations are fine. The balance sheet is absorbing capital at a rate that revenue cannot match.

The quarterly series tells a richer story. Q3 2025 stands out — ROIC dropped to 17.0% on a 7.3% NOPAT margin. That quarter absorbed $4.3 billion in one-time charges: a $2.5 billion FTC lawsuit settlement and $1.8 billion in severance costs. When those charges cleared in Q4, the margin snapped back to 9.51% and ROIC recovered to 23.9%.

This is important evidence. The operating engine is resilient. When you remove the noise, margins hold. The question isn't whether Amazon can run its business profitably — it clearly can. The question is whether the balance sheet transformation overwhelms the P&L.

The FY2025 10-K filing shows exactly where the capital went.

$59.3 billion in servers and networking equipment added in a single year. To put that in perspective: that's more than the annual revenue of all but roughly 50 companies in the entire S&P 500.

$71.7 billion in infrastructure that isn't finished yet — not generating revenue, not being depreciated, not earning any return. This is pure invested capital drag on ROIC, and it grew by $25.1 billion in a single year.

If the turnover compression were limited to Amazon's lower-margin retail segments, it might be less concerning. Retail has always been a capital-intensive, thin-margin business. But the 10-K reveals the compression is happening at the segment level where it matters most:

AWS's property and equipment grew 72% in a single year — from $110.7 billion to $190.1 billion. And despite 20% revenue growth, AWS operating margin compressed 160 basis points, from 37.0% to 35.4%. Even the highest-margin, fastest-growing segment in Amazon's portfolio cannot outrun the capital being deployed into it.

This is what a generational infrastructure buildout looks like from inside the financial statements. Margins hold. Revenue grows. And the balance sheet swells faster than both.

Amazon's Incremental Returns Are Falling — But the Raw Numbers Lie

If overall ROIC tells you the average return on all invested capital, ROIIC — return on incremental invested capital — tells you what each new dollar of investment generates. It's the more granular question: is the marginal capital being deployed efficiently?

Over the past four quarters, trailing twelve-month ROIIC has declined steadily: 33% → 24% → 12% → 4.3%.

On the surface, this looks catastrophic — an 87% collapse in incremental returns. But surface-level analysis is exactly what we're trying to avoid. Two adjustments reveal a more honest picture.

First, one-time charges. The $4.3 billion in FTC settlement and severance costs in Q3 2025 aren't recurring operational expenses — they're legal and restructuring costs that depress the numerator of the ROIIC calculation for the trailing twelve months. Removing them, adjusted ROIIC rises to approximately 7.3%. Still a significant decline. Not catastrophic.

Second, the FY2024 tax anomaly. Amazon's effective tax rate in FY2024 was 13.5% — anomalously low, likely due to international tax structuring or one-time tax benefits. In FY2025, the effective rate was 19.6%, much closer to a normalized level. The low FY2024 rate inflated that year's after-tax earnings (NOPAT) by roughly $5 billion, which means the starting point of the ROIIC calculation — the baseline against which FY2025 is measured — was artificially high.

The difference between "incremental returns collapsed 87%" and "incremental returns compressed 64% during a generational infrastructure buildout" is the difference between alarm and analysis. The first framing suggests something is broken. The second frames a demanding but potentially rational investment cycle.

The structural question that the data cannot yet resolve: Amazon's three-year trailing ROIIC average is 41.8%. The current adjusted run rate is 7-12%. Is this compression cyclical — the normal pattern of a buildout where capacity gets deployed before demand fills it — or structural, reflecting permanently lower returns because AI infrastructure doesn't monetize the way cloud compute did in the 2010s?

The honest answer is that four quarters of data can't distinguish between these two explanations. What the data can do is identify the specific mechanisms that will determine the answer — and the most important one is in the depreciation schedule.

Amazon Reversed Its Server Depreciation. Meta Went the Opposite Way.

Buried in the property and equipment footnotes of Amazon's FY2025 10-K is a two-sentence disclosure that tells an extraordinary story about how Amazon views the economic life of its AI infrastructure:

Read that again carefully. In 2024, Amazon extended server useful life from five years to six — the same direction that Meta, Alphabet, and Microsoft all moved during the same period. The logic was straightforward: modern servers last longer than their predecessors. Extending useful lives reduced annual depreciation expense and boosted reported margins. It looked like an industry consensus.

Then, in January 2025, Amazon reversed course. But only for a subset — the servers and networking equipment dedicated to AI and machine learning workloads. Those went back to five-year lives. The financial impact: approximately $1.4 billion in additional annual depreciation expense.

This is a remarkably revealing decision. Amazon is implicitly telling the market something its peers are not: AI-specific hardware — the GPUs, custom accelerators, and high-bandwidth networking that power generative AI — may obsolete faster than general-purpose servers. Custom silicon evolves rapidly. Training infrastructure that is state-of-the-art in 2025 may be economically uncompetitive by 2028.

The contrast with Meta is striking — and is the clearest illustration of the depreciation divergence reshaping hyperscaler earnings quality. In the same period, Meta extended GPU useful lives from 4.5 to 5.5 years, reducing its depreciation expense by $2.9 billion. Both companies run Nvidia's latest GPU accelerators. Both are making generational bets on AI infrastructure. Amazon shortened useful lives. Meta extended them. Same hardware. Opposite accounting conclusions. One of them is wrong — or at least, more aggressive than the other.

Amazon simultaneously extended the useful life of heavy equipment — data center buildings, power infrastructure, cooling systems — from a flat 10 years to a range of 10 to 13 years. This partially offsets the server shortening, reducing depreciation by approximately $0.6 billion. The net impact of all useful life changes: roughly $0.8 billion in additional depreciation expense.

The pattern is intellectually honest: conservative on the compute hardware that faces rapid technological obsolescence, lenient on the physical infrastructure that surrounds it. A data center shell will be useful for decades. The GPUs inside it may not last half that long. Amazon's accounting reflects this distinction. It also creates a near-term margin penalty that its peers have chosen not to absorb.

The quarterly depreciation data from MetricDuck's pipeline shows the wave is already building. FY2025 D&A expense by quarter: Q1 $14.3 billion, Q2 $15.5 billion, Q3 $16.8 billion, Q4 $19.5 billion. That's 36% growth within a single fiscal year — and Q4's $19.5 billion run rate annualizes to $78 billion, already well above the $65.8 billion full-year figure.

Year	D&A (Est.)	Growth vs. FY2025	Operating Margin Headwind
FY2025	$65.8B	Baseline	—
FY2026E	~$80-85B	+22-29%	~180-240 bps
FY2027E	~$94-100B	+43-52%	~350-430 bps

At a blended useful life of approximately seven years, $200 billion in 2026 capex generates roughly $28-29 billion per year in new depreciation expense. By FY2027, total D&A could approach $100 billion — approximately 12% of revenue at current growth rates. For context: Amazon's entire FY2025 operating income was $68.6 billion. The depreciation from this buildout alone could consume a significant portion of future operating income growth.

There is one partial offset. The "One Big Beautiful Bill Act" provides 100% accelerated depreciation for qualifying capital equipment, saving Amazon an estimated $4 billion per year in cash taxes. But this is exclusively a cash benefit — it does not reduce GAAP depreciation expense. Investors focused on reported margins see the full headwind. Cash flow investors see partial relief. The divergence between Amazon's GAAP ROIC (21.7%) and its Cash ROIC (47.1%) is this phenomenon in real-time — and it will widen as the wave builds.

One final piece of context: Amazon's capex-to-depreciation ratio is 2.00x — the most conservative among major hyperscalers. Meta's is 3.95x. Alphabet's is 4.33x. This means depreciation catches up to capital expenditures faster at Amazon than anywhere else, which is why the margin headwind materializes sooner and hits harder.

The depreciation wave is building regardless of what happens to revenue. But this raises the next logical question: if returns are compressing and costs are rising, can Amazon adjust course?

$440 Billion in Commitments Amazon Can't Walk Away From

The 10-K's commitments and contingencies footnote answers this directly.

$440 billion. Contractually locked in.

Two line items deserve particular attention.

$96.4 billion in leases not yet commenced. These are data centers and facilities that Amazon has contracted to build or lease but has not yet put into service. They represent years of committed infrastructure spending that will flow into the balance sheet as the facilities come online — each one adding to invested capital, entering the depreciation schedule, and deepening the capital turnover compression described above. These commitments are, for practical purposes, irreversible. You don't walk away from a signed data center lease.

$71.7 billion in construction in progress (up $25.1 billion from the prior year). This is infrastructure currently being built that generates zero revenue and incurs zero depreciation. When these assets enter service, they simultaneously add to the depreciation wave and the invested capital base. Every dollar of construction in progress is a future drag on ROIC that has already been committed but has not yet appeared in the return calculations.

The 10-K also reveals a shift in how Amazon funds this expansion:

Total long-term debt grew from $58.0 billion to $68.8 billion. Amazon has shifted from a net debt repayer to a net debt issuer. The balance sheet remains strong — $21.3 billion in net cash, 35x interest coverage — but the trajectory matters. This is a company that is now borrowing to build, not just reinvesting operating cash flow.

What the Math Requires

With $200 billion in planned 2026 capex, invested capital is expected to grow to approximately $460-490 billion. Here's what different revenue and margin scenarios produce:

To maintain 20% ROIC at approximately $460 billion in invested capital, Amazon would need revenue of roughly $907 billion — a 26.5% growth rate at current margins. That's more than double its current growth trajectory. The only realistic path to maintaining 20% ROIC is simultaneous revenue acceleration and margin expansion — possible, but demanding.

Every scenario in the table produces ROIC in the 17-18% range. Still above the estimated cost of capital (~10%). Still value-creating. But the margin of safety has compressed from comfortable to thin.

The tension between GAAP and cash metrics sharpens the analysis. GAAP ROIC is 21.7%. Cash ROIC — which uses operating cash flow instead of NOPAT — is 47.1%. The gap exists because operating cash flow ($139.5 billion) dwarfs NOPAT ($64.3 billion); depreciation is a massive non-cash charge that reduces GAAP earnings but doesn't consume cash. As the depreciation wave builds, this divergence will widen. Investors who focus on cash flow will see a healthy, improving company. Investors who focus on GAAP returns will see steady deterioration. Both are measuring real things — they're just measuring different things.

Six Metrics That Will Tell You If $200 Billion Is Working

The data doesn't resolve whether Amazon's infrastructure buildout will earn adequate returns. What it does provide is a set of specific, measurable thresholds that will signal whether the math is working.

Capital turnover stabilization. This is the single most important metric. Invested capital turnover is 2.42x and declining at a rate of -0.51x per quarter (8-quarter trend). It needs to hold above 2.0x. If it breaks below 2.0x at current margins, ROIC falls under 18%. Monitor this quarterly on Amazon's ROIC page.

Incremental return recovery. Adjusted ROIIC should begin climbing back toward 15% or higher within two to three quarters if the infrastructure is being efficiently deployed. If ROIIC remains below 10% for three or more consecutive quarters, the investment cycle is likely structurally over-extended — capacity is being built faster than demand can absorb it.

Depreciation versus revenue growth. Depreciation expense grew 24.5% in FY2025 while revenue grew 12.4%. If this gap persists or widens for three or more consecutive quarters, margin compression is structural rather than cyclical. Revenue growth needs to close the gap for margins to stabilize.

Construction in progress conversion. $71.7 billion in CIP must eventually start generating revenue. Track the quarterly flow from construction in progress into placed-in-service property and equipment. If CIP continues growing without corresponding revenue acceleration, the turnover problem deepens.

Q1 2026 earnings noise. The first quarter of 2026 will be unusually noisy. Approximately $15 billion in non-cash Anthropic mark-to-market gains will inflate headline earnings. Amazon Leo — mentioned for the first time in an SEC filing — adds roughly $1 billion in incremental costs. The first quarter of accelerated FY2026 capex will also begin to appear. Look through the headline numbers.

AI product disclosure. Amazon committed $200 billion to infrastructure but named zero specific AI products in the FY2025 10-K. No Rufus. No Trainium. No Graviton. No Project Rainier. The filing describes AI and machine learning capabilities in entirely generic terms. Until Amazon provides product-level revenue attribution or capital return metrics, investors are underwriting a capacity bet — not a monetization story. Watch whether future earnings calls or filings begin quantifying what the $200 billion is actually building.

Track capital turnover trends across hyperscalers on our ROIC Screener.

---

The Bull Case Is Concrete

The bull case for Amazon's infrastructure buildout doesn't rest on hope — it rests on specific, measurable evidence. AWS backlog stands at $244 billion with a 4.1-year weighted average remaining life, providing a floor of committed future revenue. Advertising revenue reached $68.6 billion in FY2025 — up 22% — making it Amazon's third-largest and highest-margin revenue stream, one that rides on top of existing retail infrastructure with minimal incremental capital. Amazon has built and monetized infrastructure at massive scale before: the 2015-2018 fulfillment center expansion was widely criticized at the time but ultimately produced the logistics moat that underpins today's retail dominance. And the 42% three-year ROIIC average suggests the capital base has historically earned returns far above cost of capital. The question is not whether Amazon can deploy capital effectively — its track record is among the best in corporate history. The question is whether this cycle's returns will converge toward the historical average or settle at the current compressed level.

The data doesn't say Amazon is making a mistake. Amazon's operating engine is healthy — margins held, cash generation is exceptionally strong, AWS continues to grow at 20%, and the company has deployed capital at massive scale before and earned extraordinary returns. What the data says is that the margin of safety between current returns and cost of capital is compressing, and the commitments driving that compression are irreversible.

At $200 billion in capex, the investment case rests on believing that Amazon's 42% three-year ROIIC average reflects the long-term earning power of this infrastructure — not the current 7-12% adjusted run rate. That belief may prove correct. Amazon's track record of patient capital deployment is among the best in corporate history.

But it is a belief, not a certainty. And there is one thing the 10-K makes clear that no earnings call or analyst day has addressed: Amazon provides no quantitative framework — no ROIC targets, no segment-level capital return metrics, no product-level AI revenue disclosure — for investors to evaluate whether $200 billion is working.

The numbers exist. Amazon chooses not to share them.

Until they do, investors are left to build the framework themselves.

Methodology

Data Sources

Filing	Company	Form	Period	Filed	CIK
Amazon FY 2025 10-K	Amazon.com, Inc.	10-K	FY 2025	Feb 6, 2026	0001018724

All financial metrics extracted via MetricDuck's SEC filing analysis pipeline. Quarterly metrics computed from 10-Q filings for Q1-Q3 2025 and the FY2025 10-K.

ROIC is calculated as NOPAT (Net Operating Profit After Tax) divided by Invested Capital. DuPont decomposition breaks ROIC into NOPAT Margin multiplied by Invested Capital Turnover. ROIIC (Return on Incremental Invested Capital) measures the change in NOPAT divided by the change in Invested Capital over trailing twelve months. TREND8 represents the linear regression slope of the metric over the trailing 8 quarters, measuring the rate of acceleration or deceleration. Cash ROIC substitutes operating cash flow for NOPAT, isolating the impact of non-cash charges like depreciation. Forward ROIC scenarios use estimated invested capital based on projected capex and historical IC-to-capex relationships.

Limitations

• This analysis is based on publicly available SEC filings and cannot assess non-public information, management intentions, or forward-looking outcomes.
• ROIC calculations use reported figures; actual economic returns may differ based on depreciation policy choices and accounting estimates.
• Forward ROIC scenarios assume proportional invested capital growth based on stated capex guidance and may not reflect actual capital deployment timing.
• Segment-level ROIC is not available; AWS capital efficiency is inferred from segment-level PP&E and operating income disclosures.

Disclaimer: This analysis is for informational purposes only and does not constitute investment advice, a recommendation to buy or sell securities, or an offer or solicitation of any transaction. Past performance is not indicative of future results. Always conduct your own due diligence before making investment decisions.

MetricDuck Research does not hold positions in AMZN, META, or GOOGL. This analysis is based solely on publicly available SEC filings and is provided for educational purposes.