Follow the Dollar: Who Captures the Economics of AI Capex

Your superday is Thursday, and one question is almost certain: is AI capex a bubble? Three sources give you three different spending numbers, and Nvidia just reported a 75.0% gross margin. Without your own math, you can only repeat headlines.

Today, suppliers capture most of the measurable profit from AI capex. Nvidia reported a 75.0% gross margin on $96.2 billion of revenue in the quarter ended July 26, 2026. Hyperscalers carry the assets, the depreciation and a growing share of debt funding. Whether they earn their cost of capital depends on AI revenue growth and on how long GPUs stay useful.

This guide follows $100 of capex through the stack and computes the revenue a hyperscaler needs at a target return. It reconciles conflicting capex figures and shows how to model all of it in a three-statement model.

Where Does a Dollar of AI Capex Go?

A dollar of AI capex splits between accelerator systems and physical infrastructure such as power, cooling, buildings and networking. A widely used rule of thumb puts accelerators at about half of data center cost. Because margins differ sharply by layer, gross profit concentrates far more than spending does.

The stack has five layers. Chip designers sell accelerators. Foundries and memory suppliers manufacture them. Infrastructure suppliers deliver power, cooling and networking. Hyperscalers and neoclouds own and operate the capacity. Application companies sell AI products on top of it.

Worked example: $100 of capex, followed through the stack

The split below uses the method from Sequoia Capital’s 2024 analysis, which treats GPUs as roughly half of a data center’s total cost of ownership. The 75% margin is Nvidia’s reported gross margin for the quarter ended July 26, 2026, according to its SEC filing. The 25% margin for the infrastructure layer is an illustrative assumption, not a reported figure.

Layer (illustrative)	Share of $100 capex	Gross margin	Gross profit
Accelerator systems	$50.0	75%	$37.5
Power, cooling, buildings, networking	$50.0	25% (assumption)	$12.5
Total	$100.0	50%	$50.0

Accelerator systems receive 50% of the spending but 75% of the gross profit in this example. For your analysis, separate revenue exposure from margin capture. A supplier can receive a large share of capex and still keep little of it as profit.

Gross Margin Is Not Operating Margin
Nvidia reported a 75.0% gross margin and $63.7 billion of operating income on $96.2 billion of revenue for the quarter ended July 26, 2026. That is an operating margin of about 66.2%. Some market commentary quotes 75% as Nvidia’s operating margin, which mixes two different metrics.

Who Captures the Economics of AI Capex Today?

Chip designers capture the highest margins today, infrastructure suppliers capture volume, and hyperscalers and neoclouds carry the assets. Neoclouds carry the sharpest risk because they fund capex largely with debt against GPUs that lose value with each chip generation. The application layer is a profit pool nobody can size yet.

Accelerators: the clearest winner

Nvidia’s data center revenue reached $89.0 billion in the quarter ended July 26, 2026, up 117% from a year earlier. The company guided to about $108.0 billion of revenue for the following quarter at a 74.0% gross margin, according to its quarterly release. A slightly lower margin on much higher revenue still means more gross profit in dollars.

Hyperscalers: spending ahead of revenue

As of July 30, 2026, Amazon guided to about $220 billion of 2026 capex, Alphabet to $195 billion to $205 billion, Microsoft to about $175 billion and Meta to $130 billion to $145 billion, according to a tracker of company guidance. These figures are not on one definition (see the section on capex numbers below).

The funding mix is shifting. Incremental debt rose from 9% of hyperscaler capex in fiscal 2024 to 32% on a trailing basis by mid-2026, according to FactSet.

93%
Estimated share of hyperscalers’ 2026 operating cash flow absorbed by AI capex, up from 33% in 2023.
Source: J.P. Morgan Asset Management → https://am.jpmorgan.com/dk/en/asset-management/institutional/insights/market-insights/investment-outlook/technology-and-ai/

Neoclouds: the levered version of the trade

Neoclouds are GPU-focused cloud providers without a large legacy business to fund their build. CoreWeave reported $2.57 billion of revenue and a $626 million net loss for the second quarter of 2026, with a $104 billion revenue backlog. For full-year 2026, it guided to $12.4 billion to $13.2 billion of revenue and $35 billion to $39 billion of capex, according to Constellation Research. At the midpoints, capex is nearly three times revenue.

Power, cooling and applications

Electrical equipment, cooling and construction firms are capacity-constrained, with data center backlogs extending into 2027 and beyond, according to Ferguson Wellman. The application layer is where long-term value may accrue, but its economics are the least defined today.

Layer	Economic role today	Capital intensity	Main risk
Chip designers	Highest margins	Low (outsourced manufacturing)	Custom chips, slower capex
Power, cooling, construction	Volume and backlog	Moderate	Lumpy orders, overbuilding
Hyperscalers	Own and monetize capacity	Very high	Revenue lags depreciation
Neoclouds	Rent out GPU capacity	Extreme, debt-funded	Refinancing, GPU obsolescence
Applications	Monetize end users	Low to moderate	Unclear willingness to pay

How Much Revenue Does AI Capex Need? (And the Common Mistake in the Math)

A hyperscaler needs enough revenue to cover cash operating costs, depreciation and taxes, and still earn its target return on invested capital. At a 12% after-tax ROIC, 35% cash costs and a five-year asset life, $100 billion of capex needs about $54 billion of annual revenue.

The ROIC method

Return on invested capital (ROIC) is net operating profit after tax (NOPAT) divided by invested capital. The test is whether ROIC exceeds the weighted average cost of capital (WACC).

Illustrative assumptions:

• Invested capital: $100 billion (held at cost for a year-one view)
• Target after-tax ROIC: 12%
• Tax rate: 21%
• Cash operating costs (power, staff, maintenance): 35% of revenue
• Straight-line depreciation, no residual value

The calculation runs in three steps:

1. Required NOPAT: $100 billion × 12% = $12.0 billion
2. Required EBIT: $12.0 billion ÷ (1 − 21%) = $15.19 billion
3. Required revenue: (EBIT + depreciation) ÷ (1 − 35%)

Useful life	Annual depreciation	Required revenue	Revenue per $1 of capex
6 years	$16.67 billion	$49.0 billion	$0.49
5 years	$20.00 billion	$54.1 billion	$0.54
3 years	$33.33 billion	$74.7 billion	$0.75

In Excel: =(Capital*Target_ROIC/(1-Tax_Rate)+Capital/Useful_Life)/(1-Cash_Cost_Pct)

The result matches a published rule of thumb. For a 15% unlevered ROIC, each $1 trillion of ramped capex needs about $500 billion of annual revenue at roughly 30% margins, according to Bessemer Venture Partners.

The capital-recovery method and the double-count

A second method converts capex into an annual payment that returns the capital plus a return on it. In Excel, =PMT(10%,5,-100) gives $26.38 billion per year for $100 billion over five years at 10%.

This payment already includes the return of capital, which is what depreciation represents. You must therefore divide it by a margin before depreciation. At a 65% cash margin, the required revenue is $26.38 billion ÷ 65% = $40.6 billion.

Don’t Count Depreciation Twice
Dividing the capital-recovery payment by an operating margin after depreciation counts depreciation twice. In the example, the operating margin at $40.6 billion of revenue is about 15.7%. Dividing $26.38 billion by 15.7% gives $167.8 billion, roughly four times the correct hurdle.

The two methods give different results because they use different hurdles. The ROIC method targets 12% after tax on capital held at cost. The capital-recovery method targets 10% pre-tax on a declining balance. Pick one method and state it.

Why ROIC matters more than revenue growth

Expected growth in operating income equals the reinvestment rate times the return on capital, according to NYU Stern professor Aswath Damodaran. If ROIC stays below the cost of capital, faster reinvestment destroys value even while revenue grows. For hyperscalers, the question is not whether AI revenue grows but whether it grows faster than the capital base.

Why Do Hyperscaler Capex Numbers Differ Between Sources?

Published AI capex figures differ because they cover different companies, different definitions and different dates. Some count only the four largest hyperscalers, others add Oracle or neoclouds. Some include finance leases, others report cash purchases of property and equipment. Before you compare companies, normalize every figure.

Figure for 2026	Source and date	Companies covered	Definition
$527 billion	Goldman Sachs, Dec 18, 2025	Large public hyperscalers	Consensus estimate, total capex
$650 billion	Forbes / Trefis, Feb 10, 2026	Four largest	Company plans, total capex
$697 billion	J.P. Morgan Asset Management, Jun 15, 2026	Five US companies	Sell-side estimates
About $775 billion	American Century, Jun 18, 2026	Seven, incl. two neoclouds	AI capex estimate
Close to $800 billion	FactSet, Jul 2026	Five, incl. Oracle	Incl. finance leases and customer prepayments

Example: Microsoft’s $15 billion “cut”

Microsoft lowered its calendar-2026 capex outlook from about $190 billion to about $175 billion in July 2026 without cutting planned spending. It extended the estimated useful life of data centers and office buildings from 15 to 25 years. As a result, more future leases qualify as operating leases, which Microsoft excludes from capex, according to CFO Dive.

In the quarter ended June 30, 2026, Microsoft reported roughly $41 billion of capex. That included $5.6 billion of finance leases and $35.8 billion of cash additions to property and equipment, according to a summary of the earnings call. Under ASC 842, finance leases sit on the balance sheet much like purchased assets.

How to Normalize Capex Figures

• Use the same set of companies
• Separate AI capex from total capex
• Include or exclude finance leases for every company
• Match calendar year and fiscal year
• Record the date of each guidance figure
• Note accounting changes that move items between lines

How Does GPU Useful Life Change the Economics?

Useful life decides how fast capex turns into expense. Under ASC 250, a change in useful life is a change in estimate applied prospectively to the remaining book value, with no restatement. In 2025, Meta lengthened server lives and cut depreciation by $2.92 billion, while Amazon shortened some to five years and added $1.4 billion.

Same year, opposite moves

Company	Change effective January 1, 2025	Effect in 2025	Source
Meta	Most servers and network assets extended to 5.5 years	Depreciation −$2.92 billion, net income +$2.59 billion	Form 10-K
Amazon	Subset of servers and networking shortened from 6 to 5 years	D&A +$1.4 billion, net income −$1.0 billion	Form 10-K

Amazon cited the faster pace of technology development in artificial intelligence and machine learning as the reason for its shorter lives. The two companies made opposite judgments about similar equipment in the same year.

Worked example: changing a server cohort’s life

Assume an illustrative $60 billion server cohort with a five-year life and no residual value. Annual depreciation is =SLN(60,0,5) = $12 billion. After two years, net book value is $36 billion.

Scenario after year 2	Remaining life	New annual depreciation	Change vs. $12 billion
Keep 5-year total life	3 years	$12.0 billion	—
Extend to 6-year total life	4 years	$9.0 billion	−$3.0 billion
Shorten to 4-year total life	2 years	$18.0 billion	+$6.0 billion

The formula for the new charge is net book value divided by remaining life. A change in estimate is applied prospectively, as described in PwC’s guide to property, plant and equipment. Past depreciation is not restated.

US GAAP and IFRS Agree Here
Under both ASC 250 and IAS 8, a revised useful life is a change in accounting estimate. It affects depreciation in the current and future periods only. Neither framework restates prior periods for it.

Accounting life vs. economic life

Three lives matter for a GPU. Physical life is how long it runs. Accounting life is the estimate behind depreciation. Economic life is how long it earns a competitive return before newer chips make it uneconomic.

A five- or six-year accounting life is a disclosed estimate, not an error. It becomes a problem only if economic life turns out shorter. In that case, impairments, faster replacement capex or lower returns follow. In your model, treat useful life as an assumption and test it in a sensitivity.

How to Model AI Capex in a Three-Statement Model

Model AI capex in a separate schedule. Forecast capex by driver, depreciate each year’s spending as its own vintage, and keep finance leases visible. Then link net PP&E to the balance sheet, D&A to the income statement and capex to cash flow from investing. Calculate ROIC on the result.

1. Choose the capex driver. Use company guidance for the first year, then a percentage of revenue or capacity in megawatts.
2. Split asset classes. Servers and networking get five to six years, buildings much longer.
3. Build a vintage depreciation waterfall. Depreciate each year’s capex separately, because new spending can follow a different profile than the existing base, as Ed Bodmer explains. A row per vintage and =SUMPRODUCT() or a simple column sum give total depreciation.
4. Treat finance leases consistently. Add finance-lease additions to capex if the company reports them that way.
5. Roll forward PP&E. Ending gross PP&E equals beginning gross PP&E plus capex minus disposals. Net PP&E can never exceed gross PP&E.
6. Link the three statements. D&A reduces EBIT, capex reduces cash flow from investing, and net PP&E feeds the balance sheet.
7. Calculate FCF and ROIC, then run a sensitivity. Flex useful life between three and six years and show the effect on EBIT, ROIC and required revenue.

Common Modeling Errors
Forecasting depreciation as a fixed percentage of capex while capex doubles distorts the timing. Ignoring finance leases understates capex. Applying a useful-life change to past periods restates history, which ASC 250 does not allow.

If you want to build this schedule from a blank Excel sheet with a practitioner, see the one-on-one modeling packages for individuals. For a structured check before you send a model, use this model review process that holds up.

How to Answer an AI Capex Question in an Interview or Stock Pitch

A strong answer separates who earns profit today from who carries the risk. It names one number with its source and date, and states the assumption that would change the conclusion, such as GPU useful life or AI revenue growth. Interviewers test the reasoning, not whether you call a bubble.

A four-part answer structure

1. Layer: State who captures margin today and who carries the assets.
2. Number: Give one figure with its date, for example Nvidia’s 75.0% gross margin in the quarter ended July 26, 2026.
3. Assumption: Name the variable that decides the outcome, such as revenue per dollar of capex or useful life.
4. Trigger: Say what would change your view, such as a shorter useful life or a capex cut.

A sample answer in 60 seconds: “Today the suppliers capture the clearest economics. Nvidia runs a 75% gross margin. Hyperscalers carry the depreciation, and at a 12% ROIC they need roughly 50 cents of annual revenue per dollar of capex. My view depends on useful life. If GPUs last three years instead of five, the hurdle rises to about 75 cents.”

The technical modeling side of interviews is covered in this guide on how investment banking modeling tests are structured.

When this topic matters less

The depth in this guide matters most for TMT coverage, equity research and investors in digital infrastructure. For corporate FP&A outside technology or for restructuring roles, a clear one-minute view is usually enough. Credit roles focus more on how neoclouds are financed, a topic covered in how direct lenders model credit risk.

FAQ

Is AI capex a bubble?
Not necessarily. The measurable test is whether AI revenue grows fast enough to cover depreciation, operating costs and the cost of capital. Headline spending alone does not answer that question.

How much are hyperscalers spending on AI capex in 2026?
Several hundred billion dollars. As of July 30, 2026, guidance ranged from about $130 billion (Meta, low end) to about $220 billion (Amazon). Totals differ by source because of company scope and definitions.

Does extending server useful life inflate earnings?
It lowers annual depreciation and raises reported operating income. Companies disclose it as a change in estimate. It is legitimate if it reflects how long the equipment stays in productive use.

Is capex the same as depreciation?
No. Capex is the cash outflow when an asset is bought. Depreciation spreads that cost over the asset’s useful life on the income statement.

Why does Nvidia’s gross margin matter for hyperscalers?
Because supplier margin is part of the hyperscaler’s cost. Every dollar of that margin raises the capital base the hyperscaler must earn a return on.

Do power and cooling companies benefit from AI capex?
Yes. They receive a large share of data center spending through orders and backlogs. Their margins are typically lower than those of leading chip designers.

The Bottom Line – Conclusion

The suppliers capture the clearest economics of AI capex today, while hyperscalers carry the assets and the risk. Whether that bet pays off is a modeling question with two decisive assumptions: AI revenue per dollar of capex and useful life.

Prioritize four steps:

1. Normalize capex figures before you compare companies.
2. Build a vintage depreciation schedule.
3. Test useful life in a sensitivity.
4. Practice your view as a four-part answer.

Sources

1. NVIDIA, Q2 fiscal 2027 results, Form 8-K, August 26, 2026: https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm
2. NVIDIA, Q2 fiscal 2027 press release: https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/
3. Meta Platforms, Form 10-K for fiscal 2025: https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm
4. Amazon, Form 10-K for fiscal 2025: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm
5. PwC Viewpoint, Property, plant, equipment and other assets, chapter 4: https://viewpoint.pwc.com/dt/us/en/pwc/accounting_guides/property_plant_equip/property_plant_equip_US/chapter_4_depreciati_US/32_determining_the_u_US.html
6. FactSet, Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow, July 2026: https://insight.factset.com/hyperscalers-tap-external-financing-as-ai-capex-outruns-cash-flow
7. J.P. Morgan Asset Management, The AI investment boom is reshaping stock market winners, June 15, 2026: https://am.jpmorgan.com/dk/en/asset-management/institutional/insights/market-insights/investment-outlook/technology-and-ai/
8. American Century Investments, From the Cloud to Chips, June 18, 2026: https://www.americancentury.com/insights/hyperscaler-ai-capex-spending-cycle/
9. Goldman Sachs, Why AI Companies May Invest More than $500 Billion in 2026, December 18, 2025: https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026
10. Forbes / Trefis, $650 Billion AI Capex, February 10, 2026: https://www.forbes.com/sites/greatspeculations/2026/02/10/650-billion-ai-capex-how-marvell-nvidia-micron-benefit/
11. Ferguson Wellman, The Magnificent CapEx, May 8, 2026: https://www.fergusonwellman.com/resources/the-magnificent-capex-ai-infrastructure-spending-and-who-actually-benefits/
12. Analysis Atlas, Hyperscaler AI Capex 2026: https://analysis-atlas.com/research/hyperscaler-ai-capex-2026/
13. CFO Dive, Microsoft holds line on AI spending plans: https://www.cfodive.com/news/microsoft-holds-line-ai-spending-plans/826648/
14. Let’s Data Science, Microsoft keeps AI infrastructure plan: https://letsdatascience.com/news/microsoft-keeps-ai-infrastructure-plan-as-reported-2026-cape-1ce1ad02
15. Constellation Research, CoreWeave Q2 stronger than expected: https://www.constellationr.com/insights/news/coreweave-q2-stronger-expected-touts-backlog
16. Bessemer Venture Partners, Inside the biggest bet in corporate history: https://www.bvp.com/atlas/inside-the-biggest-bet-in-corporate-history
17. Sequoia Capital, AI’s $600B Question, 2024: https://www.sequoiacap.com/article/ais-600b-question/
18. Aswath Damodaran, NYU Stern, Valuation session slides: https://pages.stern.nyu.edu/~adamodar/podcasts/valspr23/session10slides.pdf
19. Ed Bodmer, Depreciation in Corporate Models: https://edbodmer.com/edbodmer-wikispaces-com-depreciation/
20. Featured image: AI-generated illustration created for this article. It does not show real people or a real company model.