Is the AI Capex Supercycle Earning Its Cost of Capital?
Is the AI Capex Supercycle Earning Its Cost of Capital?
Amazon, Microsoft, and Alphabet took combined annual capital expenditures from roughly $120 billion to roughly $500 billion in four years. A ROIC/WACC/EVA read on AWS, Azure, and Google Cloud finds all three still clear their cost of capital — but the spread that made them extraordinary compounders has narrowed sharply at Microsoft and Alphabet, and all three now converge toward a similar, materially lower level of return on new capital.
Read this first. This report is general financial and educational commentary of general and impersonal application, prepared for informational purposes only and not for any particular reader's individual circumstances. It is not investment, tax, or legal advice; it is not a recommendation to buy, sell, or hold any security; and it does not offer an opinion, express or implied, on whether any security discussed is fairly valued, overvalued, or undervalued. All figures are drawn from the author's own financial models for Amazon, Microsoft, and Alphabet, built from company filings, earnings releases, and management guidance through Q2 2026 / FY2026 results, refreshed August 2026. The author does not publish rated equity research — including BUY/HOLD ratings and price targets — on these three companies, and this report contains no rating, price target, or valuation opinion on any of them. The author holds personal long positions in Amazon (AMZN), Microsoft (MSFT), and Alphabet (GOOGL) — the three companies analyzed in this report. Readers should weigh that interest when evaluating the analysis and framing throughout. This report reflects information available as of its stated dates and is not updated on any schedule; see the full Disclosure Statement at the end of this report for complete limitations.
Between 2022 and the guidance issued alongside Q2 2026 earnings, Amazon, Microsoft, and Alphabet took their combined annual capital expenditures from roughly $120 billion to roughly $500 billion — a run rate now closing in on triple that figure once 2027 guidance is included. This report does not ask whether that spending is real, or whether AI demand is real; both are well documented elsewhere in each company’s own filings and disclosures. It asks a narrower, more mechanical question, using the same ROIC/WACC/EVA framework this series has already applied to gaming equities and Las Vegas multifamily real estate: is the capital actually earning more than it costs — and is that gap widening or narrowing as the buildout continues?
Key Findings
- Combined AMZN + MSFT + GOOGL capex has grown roughly roughly 4.5x from 2022 (≈$120B) to 2026E (≈$536B), Microsoft has guided a further step‑up in FY2027 (roughly $175 billion including finance leases); Amazon and Alphabet have signalled continued increases without naming 2027 figures, so this report's 2027E estimates for those two are model projections rather than guidance.
- Run through this report's ROIC/WACC test, all three companies clear their own cost of capital by a wide margin in every year of this study, 2022A through 2030E — this is not, in the base case, a "the spending doesn't pay off" finding.
- ROIC compresses sharply at Microsoft and Alphabet and only modestly at Amazon, with all three converging toward the mid‑20s by 2030E: Microsoft falls roughly 19 points from its 2025 level and Alphabet roughly 21, while Amazon declines just 1.4 points (26.3% to 24.9%, troughing at 22.6% in 2028E before recovering). Because WACC is held constant, the EVA spread narrows by those same amounts — leaving 2030E spreads of +14.7 points (AMZN, against a 10.17% WACC), +15.9 (MSFT, 9.87%), and +14.4 (GOOGL, 10.17%). Every year of the study still clears the bar.
- A second, separate test — not a spread, and unrelated to cost of capital — asks how much annual after-tax operating profit a company added over a period, divided by the cumulative capex it spent during that period. Microsoft's FY2022–FY2025 figure of 28.3%, for example, means $161B of capex coincided with annual NOPAT ending $45.6B higher than it started: roughly 28 cents of added yearly profit per dollar spent. On that measure all three converge into a narrow 11.6%–13.0% band for 2026E–2030E, from a wider and generally higher historical range of 13.2%–28.3%. Microsoft and Alphabet roughly halve; Amazon, already at the low end historically, declines only modestly. This is a proxy, not textbook return on incremental invested capital — the denominator is gross capex, not the change in invested capital. Amazon's shallower decline should not be read as capital-efficiency stability: on the component that does not depend on a margin forecast — revenue generated per dollar of capex — all three deteriorate materially (Amazon −27%, Alphabet −41%, Microsoft −51%), and Amazon's ratio is held up by an assumed margin expansion that, if it does not materialize, would put its figure near 9.7% rather than 11.6%.
- AWS shows the most resilient trajectory of the three cloud segments, consistent with the record 39.4% quarterly operating margin Amazon reported for AWS in Q2 2026. Microsoft’s and Alphabet’s compression is sharper — directionally consistent with the bear case in this author’s own MSFT model ("capex does not earn adequate returns, compressing FCF and ROIC") and with Alphabet's first negative free‑cash‑flow quarter since its 2004 IPO.
- Wall Street has already begun repricing on this exact question. Alphabet's stock fell as much as 6–8% after its Q2 2026 report despite beating on cloud growth by more than either peer, because it raised capex guidance without matching margin or backlog evidence — then recovered above its pre-earnings level within two weeks once Amazon's and Microsoft's results supplied that evidence. Amazon and Microsoft both posted their largest post-earnings stock moves in years (+15.3% and +8–9%, respectively), driven specifically by AWS's record segment margin and Azure's backlog and growth acceleration. Sell-side price targets moved accordingly at all three companies (Street averages near $409 for GOOGL, $322 for AMZN, and $558 for MSFT, alongside a rare batch of target cuts at Alphabet paired with maintained bullish ratings). See the Scenarios and Wall Street Context section below for the full discussion, including why this is evidence worth weighing rather than a confirmed inflection.
- This is a company‑wide test, not a segment test. None of the three companies disclose capital expenditures by segment, so this report cannot isolate AWS's, Azure's, or Google Cloud's capital efficiency from the rest of each company's business. That is a structural data limitation, not a finding — see Limitations.
- Open question this report does not resolve: WACC is held constant at each company's current model‑derived value for the full 2022–2030 window. If financing costs or risk premia move materially over that period — plausible, given the debt‑funded nature of the 2026 capex step‑up at all three companies — every spread in this report moves with it.
- Current Wall Street capex consensus for 2027E, translated to this report's three‑company scope, runs meaningfully above this report's own base‑case model — meaning the ROIC compression shown in the ROIC/WACC test below may understate, not overstate, the actual trajectory. Separately, a live 2026 accounting debate over whether GPU and server useful‑life assumptions are too generous argues the same direction: if depreciation is understated across the sector, reported ROIC in this report's models is running above economic reality. See the Scenarios and Wall Street Context section below.
- This report's figures are not the only published attempt at this question, but they are not directly comparable to the most-cited one. Morgan Stanley's July 2026 frameworks estimate roughly 25%–50% incremental ROIC on AI-specific unit economics — a return on invested capital: annual profit divided by a net capital stock. This report's 11.6%–13.0% is annual profit added divided by cumulative capex spent — a flow-over-flow productivity ratio, not a return on capital. The two use different numerators, different denominators, and different scope (AI-specific workloads vs. whole-company capex including legacy business); none of those differences can be adjusted away to make the figures reconcile, and a materially higher number under Morgan Stanley's approach is fully consistent with, not contradictory to, this report's figure. See the comparison table in the Scenarios and Wall Street Context section for the full breakdown of what each range does and does not measure.
- A case can be made that this is a multi-period inflection rather than one strong quarter, and this report sets it out at full strength as a steelman against its own base case rather than burying it in caveats: sequential acceleration across four to five consecutive quarters at all three companies; contracted backlog growing faster than revenue ($678B, $514B, and $496B respectively); segment margins expanding during the heaviest capex phase rather than compressing; and supply — not demand — cited as the binding constraint by all three managements independently. This report nonetheless retains the more conservative compression case as its base case, principally because the unresolved depreciation question would, if it breaks the wrong way, undercut the margin evidence the inflection case rests on.
The findings are summarized above; what follows is the full framework, the data behind it, and where it does and doesn't hold up.
In This Report
- Translating the Framework: ROIC, WACC, and EVA for Corporate Capex
- The Capex Build-Out, 2022–2030E
- AI/Cloud Revenue Growth, Same Window
- The Core Test: ROIC vs. WACC, Company by Company
- The Sharper Test: Productivity of New Capex
- Where the Three Companies Diverge
- Why This Is a Company-Wide Test, Not a Segment Test
- Applications: Family Offices, Professional Investors, and Corporate Finance Practitioners
- Market Evidence and the Inflection Question
- Scenario Range and Wall Street Context
- Limitations
- What This Means
- Coming Next
- Sources & Citations
- Glossary of Key Terms
Translating the Framework: ROIC, WACC, and EVA for Corporate Capex
This series has applied one consistent lens across very different asset classes — does the capital earn more than it costs? For a gaming equity, that meant comparing operating return on invested capital to a CAPM‑derived cost of capital. For a piece of Las Vegas real estate, a cap rate stood in for ROIC and a blended debt‑and‑equity rate stood in for WACC. For a hyperscaler's AI infrastructure buildout, no substitution is needed: this is corporate ROIC and corporate WACC in their original form, applied to the single largest capital allocation decision three of the world's largest companies have made in decades.
In plain English: ROIC (return on invested capital) asks how much after‑tax operating profit — NOPAT, net operating profit after tax — a company generates each year relative to the capital tied up in the business: its debt and equity, net of cash sitting on the balance sheet earning a market return rather than funding operations. WACC (weighted average cost of capital) is what that capital actually costs, blended across debt and equity in the proportions the company actually uses. EVA (economic value added) is ROIC minus WACC: a positive spread means the business is creating value beyond what its capital costs; a negative spread means it isn't, regardless of whether accounting profit is growing.
Exactly What This Report Calculates — and What It Doesn't
Every figure in this report reduces to one of two measurements, and they answer different questions. Confusing them is the single easiest way to misread this report's findings, so both are stated here explicitly, in full, before either appears again.
1. Company‑wide ROIC (used in the Core Test section) — the return earned on the company's entire capital base, legacy business included:
- NOPAT = operating income (EBIT) × (1 − tax rate). Tax rate treatment differs by company. GOOGL's own model assumes a flat 18% tax rate in every year; this report instead applies a normalized 15% to GOOGL's EBIT for cross-company comparability. That means GOOGL's published NOPAT will not tie to that model file's own DCF Inputs NOPAT row, which reflects the model's native 18% assumption — a deliberate choice, not an error, but one worth knowing before reconciling this report against the raw model. AMZN's own model already uses a comparable two-part convention internally — Amazon's actual historical effective tax rates for 2021A–2025A (ranging roughly 13%–54%, reflecting real one-time items) and a normalized 15% for 2026E–2030E — and this report uses that model's NOPAT figures directly, so AMZN's published figures do tie to its own model file. For MSFT, the model applies Microsoft's own effective rates throughout — roughly 13% to 18% across the historical years (FY2022–FY2025) and settling near 19% across the projection period — rather than a single normalized figure, and also ties directly to its own model. This means all three companies' NOPAT is computed on a different tax basis from one another, and cross-company ROIC comparisons carry that difference; see Limitations.
- Invested capital = total debt (current portion + long‑term) + total shareholders' equity − cash, cash equivalents, and short‑term/marketable investments. This is the standard "excess cash" convention; it does not capitalize operating leases as debt. This choice materially affects the result for cash-rich companies and should be understood before reading any ROIC figure below. Netting the entire cash-and-investments balance against equity produces a small denominator when a company holds cash approaching its equity base — which is why Microsoft's FY2022 figure exceeds 100%. A stricter convention that nets only excess cash (cash above estimated operating needs, often approximated at 1–2% of revenue) would leave a much larger invested-capital base and produce a Microsoft FY2022 ROIC in the 40s–50s rather than above 100%. Neither treatment is wrong; they answer slightly different questions. The direction and shape of the compression documented in this report are similar under either convention, but the absolute levels — particularly the early-period peaks — are not, and cross-company comparisons to figures computed on a different basis will not tie.
- ROIC(t) = NOPAT in year t ÷ invested capital at the start of year t (i.e., invested capital as of the end of year t−1) — a beginning‑of‑period convention, so the return reflects capital already in place, not capital raised mid‑year.
- EVA / spread = ROIC(t) − WACC, where WACC is each company's CAPM-derived figure, held constant across the projection window (a disclosed simplification — see Limitations). Market-wide CAPM inputs are standardized across all three companies at a 4.3% risk-free rate (10-year UST) and a 5.0% equity risk premium, so that differences in cost of capital reflect only company-specific factors — levered beta, pre-tax cost of debt, effective tax rate, and capital-structure weights — rather than differing assumptions about the market. Capital-structure weights are derived from market value of equity against reported total debt (excluding leases, consistent with the invested-capital definition above), not assumed: equity weights of 97.9% (AMZN), 97.9% (GOOGL), and 98.8% (MSFT). Beta reflects each company’s current market-observed level (AMZN 1.20, GOOGL 1.20, MSFT 1.13, per public data as of early August 2026); market values of equity are as of the same date. This produces WACCs of 10.17% (AMZN), 10.17% (GOOGL), and 9.87% (MSFT).
2. Cumulative capex productivity (used in the Sharper Test section) — a narrower, capex‑specific proxy, not a textbook ROIIC calculation:
- Formula = [NOPAT in the final year of the window − NOPAT in the year immediately preceding the first capex year] ÷ [cumulative capex over the window]. The base year sits one year before the window so that the full NOPAT change attributable to that capex is captured; for AMZN's 2022–2025A window, that is (2025A NOPAT − 2021A NOPAT) ÷ (capex 2022+2023+2024+2025). The same convention is applied to both the historical and projected windows for all three companies, so the two are directly comparable.
- What makes this a proxy, not a true ROIIC: a textbook incremental‑ROIC calculation divides ΔNOPAT by the change in invested capital over the same window (net of depreciation, financing mix, and working‑capital movements). This report instead divides by cumulative capex — a simpler, more transparent number pulled directly from each company's cash‑flow statement, but one that is not identical to ΔInvested Capital, since capex additions are partially offset by depreciation and by non‑capex changes in the capital base. The two will move together directionally but will not match numerically.
- No time‑lag. This report compares same‑window capex to same‑window NOPAT change, with no allowance for the one‑ to three‑year lag between a dollar of capex and the revenue it eventually supports — see Limitations for what a lagged version would likely show.
Company‑wide ROIC (test 1) and cumulative capex productivity (test 2) are presented separately throughout this report and should not be added, averaged, or substituted for one another. They answer different questions — "is the whole company earning its cost of capital" versus "is the marginal capex dollar still as productive as it used to be" — and, as the Wall Street Context section below shows, other published estimates of AI‑specific returns use still further conventions (segment‑level invested capital, gross‑of‑depreciation asset bases, or bottom‑up unit economics), none of which are directly comparable to either test in this report without adjustment.
The reason this matters more now than at any point in these three companies' histories as public companies is straightforward: all three have historically been extraordinarily capital‑light relative to their profitability — Microsoft's ROIC peaked above 100% in FY2022 on this report's own corrected model, an extraordinary figure for a company of its size. A capex program that takes combined spending up 4.5x in four years is, mechanically, an experiment in whether that capital‑light advantage can survive a multi‑hundred‑billion‑dollar infrastructure build without the return on capital collapsing toward a normal industrial company's economics. This report is a first attempt to measure that experiment's progress with real numbers rather than management's own framing of it.
The Capex Build-Out, 2022–2030E
The starting fact is not in dispute — each company has said it plainly, repeatedly, and with rising numbers each quarter. Amazon raised FY2026 cash capex guidance to roughly $220 billion (from $200 billion) on higher memory costs and AI demand. Alphabet raised its FY2026 guidance twice over the course of 2026, most recently to $195–205 billion, and posted its first negative free‑cash‑flow quarter since its 2004 IPO along the way. Microsoft guided FY2027 capital expenditures and finance leases to roughly $175 billion, up from $115.9 billion of actual FY2026 cash additions to property and equipment — the single largest guidance revision of the three companies in this reporting cycle.
Two things are worth separating here, because they get conflated constantly in commentary on this spending. The first is whether the underlying AI demand is real — this report takes no position on that question and isn't built to answer it; each company's own backlog disclosures (Alphabet's $514 billion at Q2 2026, up from $106 billion a year earlier; Amazon's $496 billion, up roughly 154% year‑over‑year) are the relevant evidence, and they are substantial. The second, entirely separate question is whether the capital being deployed to meet that demand is earning an adequate return once it's actually in the ground — servers, data centers, and power infrastructure that take years to build and depreciate over long, sometimes recently‑extended, useful lives. That second question is what the rest of this report is built to test.
AI/Cloud Revenue Growth, Same Window
The revenue side of the story is, on its face, straightforward: all three cloud segments have grown substantially, and growth has accelerated in 2026 at all three, not decelerated — a genuinely unusual pattern for businesses already operating at this scale. AWS grew 37% year‑over‑year in Q2 2026, the fastest pace in 18 quarters. Azure grew 43% in Microsoft's fiscal Q4 2026, the fourth consecutive quarter of sequential acceleration, with management stating demand still exceeds available capacity. Google Cloud grew 82% year‑over‑year in Q2 2026, up from 63% the quarter before.
Two asymmetries are worth flagging before moving to the capital‑efficiency test, because they carry through everything that follows. First, only AWS discloses a clean, standalone segment operating margin in its own right — Amazon reports AWS revenue and AWS operating income as a distinct segment. Azure's revenue is disclosed as a product line inside the broader Intelligent Cloud segment, which also includes Server Products and Enterprise & Partner Services, and Microsoft does not disclose Azure's standalone operating margin. Google Cloud's segment operating income is disclosed, and its margin trajectory — a loss as recently as 2022, 23.7% by 2025A — is one of the more striking figures in this data set. Second, and more consequentially for what follows: none of the three companies allocate capital expenditure by segment. Every capital‑efficiency figure in the rest of this report is therefore a company‑wide measurement — see the segment-test limitation below for the full treatment of what that does and doesn't let this report say.
The Core Test: ROIC vs. WACC, Company by Company
Invested capital in this report is calculated as total debt plus total equity, less cash and near‑cash balances — the standard "excess cash" convention, which avoids penalizing a company for holding cash that isn't actually funding operations. ROIC in a given year is NOPAT for that year divided by invested capital at the start of that year, so the test asks what return the capital already in place generated, not a same‑year ratio that partly reflects capital raised mid‑year. WACC for each company is CAPM‑derived using inputs standardized across all three models for comparability rather than taken as-is from the individual model files — 10.17% for Amazon, 10.17% for Alphabet, and 9.87% for Microsoft — and held constant across the full projection window. Both the standardization and the constant-WACC simplification are addressed directly in Limitations.
| Company | 2025A ROIC | 2026E ROIC | 2028E ROIC | 2030E ROIC | WACC | 2030E Spread |
|---|---|---|---|---|---|---|
| AMZN | 26.3% | 25.5% | 22.6% | 24.9% | 10.17% | +14.7 pts |
| MSFT | 44.5% | 42.8% | 31.2% | 25.8% | 9.87% | +15.9 pts |
| GOOGL | 45.7% | 41.4% | 25.0% | 24.6% | 10.17% | +14.4 pts |
Author's calculation from company financial statements (SEC filings, earnings releases) and author's own three‑statement models, refreshed August 2026 for Q2 2026/FY2026 actuals. MSFT figures are fiscal year; AMZN and GOOGL are calendar year.
Read across the row for any of the three companies and the headline is the same: every single year in this study clears WACC, comfortably, in the base case. This is not, on these numbers, a story about capital destruction. But read down each column instead, and the second story — the one this report is actually built to surface — is unmistakable: the spread that made these three companies such extraordinary compounders of capital narrows materially at Microsoft in every year of its projection, and at Alphabet in nearly every year — its ROIC ticks up fractionally in the final year, from 24.4% to 24.6%, a much smaller echo of the same late recovery Amazon shows more dramatically — using separately‑built models with different underlying assumptions. Amazon is the exception in scale: its ROIC declines only to a 22.6% trough in 2028E before recovering to 24.9% by 2030E, a net move of 1.4 points from 2025. Microsoft's decline is the steepest measured from its own peak — from a pre‑buildout ROIC above 100% in FY2022 toward the mid‑20s by 2030E, roughly 78 points, though measured from 2025 alone Alphabet's fall is the larger of the two (about 21 points versus 19) — still roughly 16 points above its own cost of capital, but a fraction of its historical premium. Alphabet's compression is sharpest in its timing, falling from 41% to 30% between 2026E and 2027E alone as this report's modeled 2027E capex of roughly $255 billion — Alphabet's own guidance points to a further increase from 2026 levels without naming a figure — outruns near‑term NOPAT growth — the model‑level echo of the free‑cash‑flow trough Alphabet’s own Q2 2026 results already showed. Amazon's decline is the shallowest of the three, consistent with AWS's segment margin resilience discussed in the divergence discussion below.
The Sharper Test: Productivity of New Capex
Company‑wide ROIC blends a rapidly growing AI‑era capital base with each company's much older, largely depreciated legacy asset base — which means the ROIC figures above, and their compression, partly reflect denominator effects that have little to do with whether the new capex specifically is a good investment. A narrower, more capex‑specific test asks a more direct question: for every incremental dollar of capital spent in a given window, how much incremental after‑tax operating profit showed up by the end of it? This report computes that as change in NOPAT over a window, divided by cumulative capex over that same window — a rough proxy for what corporate finance calls return on incremental invested capital, not a precise like‑for‑like measurement, since it doesn't lag capex to account for construction and ramp time, an important limitation flagged directly below.
| Company | Historical window | Historical ratio | 2026E–2030E ratio | Change |
|---|---|---|---|---|
| AMZN | 2022–2025A | 13.2% | 11.6% | −12% |
| MSFT | FY2022–FY2025A | 28.3% | 13.0% | −54% |
| GOOGL | 2024–2025A | 26.4% | 11.8% | −55% |
Because this ratio is easy to confuse with a cost-of-capital spread — it is not one, and WACC plays no part in it — the full arithmetic is set out below. All figures in $ millions.
| Company | Window | NOPAT, final year | NOPAT, base year | ΔNOPAT | Cumulative capex | Ratio |
|---|---|---|---|---|---|---|
| AMZN | 2022–2025A | 65,427 | 21,757 | 43,670 | 331,192 | 13.2% |
| AMZN | 2026E–2030E | 185,995 | 65,427 | 120,568 | 1,035,000 | 11.6% |
| MSFT | FY2022–FY2025A | 105,907 | 60,268 | 45,639 | 161,115 | 28.3% |
| MSFT | FY2026E–FY2030E | 201,042 | 105,907 | 95,135 | 733,948 | 13.0% |
| GOOGL | 2024–2025A | 109,683 | 71,649 | 38,034 | 143,982 | 26.4% |
| GOOGL | 2026E–2030E | 247,089 | 109,683 | 137,406 | 1,160,000 | 11.8% |
Why Amazon's decline is so much smaller
Amazon falls only 1.6 points on this measure while Microsoft and Alphabet roughly halve, and the reason is not that Amazon's capital is more productive in the ordinary sense. The ratio decomposes into two independent components — revenue generated per dollar of capex, multiplied by the share of that incremental revenue that converts to after-tax operating profit — and the two tell opposite stories.
| Company | Revenue added per $1 of capex | × Incremental NOPAT margin | = Ratio | Blended NOPAT margin, final year |
|---|---|---|---|---|
| AMZN | $0.75 | 17.7% | 13.2% | 9.1% |
| MSFT | $0.71 | 40.2% | 28.3% | 37.6% |
| GOOGL | $0.66 | 39.9% | 26.4% | 27.2% |
Historical windows as defined above. Revenue and NOPAT changes measured from the same base year used in the ratio. Author's calculation.
Amazon is the most efficient of the three at turning capital expenditure into revenue — 75 cents of added revenue per dollar spent, against Microsoft's 71 and Alphabet's 66. What separates it is what happens to that revenue afterward: Amazon converts 17.7% of incremental revenue into after-tax operating profit, while Microsoft converts 40.2% and Alphabet 39.9%. Amazon's blended NOPAT margin of 9.1% against Microsoft's 37.6% reflects a revenue base still dominated by low-margin retail, not capital deployed less effectively.
That reframes the convergence in a way worth stating plainly. Amazon's historical figure was low because of business mix, and it sat close to where all three are now heading — so it has little distance to fall. Microsoft and Alphabet are not becoming less competent allocators of capital; they are becoming more capital-intensive businesses, and their capex-productivity ratios are moving toward the level that a business carrying heavy physical infrastructure has always produced. The convergence documented here is better read as two capital-light software businesses acquiring the economics of an infrastructure operator than as a uniform deterioration across all three.
A caution that materially qualifies Amazon's shallow decline. Running the same decomposition on the projected window shows that revenue generated per dollar of capex deteriorates sharply at all three companies — including Amazon. What separates Amazon is that its incremental margin is projected to rise, offsetting most of that deterioration. At the other two it falls or holds.
| Company | Revenue per $1 capex (hist → proj) | Change (relative %) | Incremental NOPAT margin (hist → proj) | Change (relative %) |
|---|---|---|---|---|
| AMZN | $0.75 → $0.55 | −27% | 17.7% → 21.4% | +21% |
| MSFT | $0.71 → $0.35 | −51% | 40.2% → 37.4% | −7% |
| GOOGL | $0.66 → $0.39 | −41% | 39.9% → 30.1% | −25% |
Projected windows as defined above. Alphabet's 2030E revenue is derived from modeled EBIT at the modeled operating margin. Both "Change" columns are relative percentage change (projected ÷ historical − 1), not percentage-point differences — elsewhere in this report "pts" denotes a point difference; the unlabeled "%" changes in this table are always relative. Revenue per $1 capex = ΔRevenue ÷ cumulative capex for that window. Incremental NOPAT margin = ΔNOPAT ÷ ΔRevenue for that window — for example, Amazon's projected incremental margin of 21.4% is ($185,995M − $65,427M) ÷ ($1,281,353M − $716,900M) = $120,568M ÷ $564,453M, and its Change column of +21% is 21.4% ÷ 17.7% − 1. Author's calculation.
The consequence is direct and worth stating plainly: Amazon's projected 11.6% depends on an assumed margin expansion that has not yet occurred. Hold Amazon's incremental NOPAT margin flat at its historical 17.7% rather than allowing it to rise to 21.4%, and the projected ratio falls to roughly 9.7% — below the 11.6%–13.0% convergence band, and below both Microsoft and Alphabet rather than alongside them. The margin assumption is not arbitrary; it follows from AWS growing faster than retail and this report's modeled AWS segment margin reaching 40.5% by 2030E. But it is an assumption, and it is doing most of the work holding Amazon's figure up.
Read together, this means Amazon's flat-looking trajectory should not be interpreted as capital-efficiency stability. On the one measure that does not depend on a margin forecast — revenue generated per dollar of capex — Amazon deteriorates by 27%, and every company in this study deteriorates materially. Amazon's apparent resilience is a mix-shift story still in progress, and it is more sensitive to AWS's share of the revenue base than either peer's figure is to any single segment.
Read the Microsoft historical row as: across FY2022–FY2025 Microsoft spent $161.1 billion on capital expenditure, and its annual after-tax operating profit ended the period $45.6 billion higher than it began — roughly 28 cents of added yearly profit per dollar of capex. The base-year NOPAT sits one year before the first capex year so the full profit change attributable to that spending is captured. Microsoft's FY2021 base is derived from reported FY2021 operating income at Microsoft's own FY2021 effective tax rate; Alphabet's 2023 base uses the 15% normalized rate applied throughout its model. Author's calculation.
Author's calculation from company financial statements and author's own models. GOOGL's historical window is shorter (two capex years vs. four) because this report's Alphabet model begins at 2023A, which is the earliest available base year under this convention. Amazon's historical figure warrants a caveat beyond the convention: Amazon's 2022 NOPAT was unusually depressed (roughly $5.6B against $21.8B in 2021, reflecting the Rivian mark-to-market loss and margin compression), so any window touching 2022 is sensitive to base-year choice. Using 2022 rather than 2021 as the base would raise Amazon's historical figure to 18.1% and produce an apparent halving; this report uses 2021 because it is the consistent convention, but readers should treat Amazon's historical figure as the least stable of the three.
This is, in this author's view, among the most useful numbers in this report, precisely because it's independent of the ROIC compression above — a different metric, built from different inputs, arrived at through three separately‑built models — and the three converge on a strikingly similar endpoint from very different starting levels: Microsoft and Alphabet see this ratio roughly halve — 28.3% to 13.0% and 26.4% to 11.8% — while Amazon, which already sat near the bottom of the historical range at 13.2%, declines only modestly to 11.6%. The convergence into an 11.6%–13.0% band, not a uniform rate of decline, is the finding. That's not a claim that the spending is unprofitable — an 11.6%–13.0% projected ratio is still a positive return on each marginal capex dollar, and it sits above the 9.87%–10.17% cost-of-capital range these models use, though that comparison should be read loosely rather than as a formal EVA test: this metric's denominator is cumulative capex, not invested capital, so it is not the like-for-like input WACC is designed to be measured against. It's a claim that the wide dispersion in capital productivity that separated these companies before the buildout has collapsed, and that the level all three are converging toward is well below where Microsoft and Alphabet operated as recently as 2025.
Where the Three Companies Diverge
The three companies are not compressing at the same rate, and the divergence tracks reasonably well with the segment‑level evidence already public in each company's own disclosures.
AWS is the most resilient of the three cloud segments on the evidence available. Its Q2 2026 operating margin reached a record 39.4%, up 650 basis points year‑over‑year (520bps excluding a favorable derivative‑accounting item), even as AWS revenue accelerated to 37% growth — margin expansion during an acceleration, not after one, which is the more difficult trick. That shows up directly in this report's company‑wide ROIC test: Amazon's decline from 26.3% to a 2028E trough of 22.6% is the shallowest of the three, and its 2030E figure recovers slightly to 24.9% as AWS's segment margin continues improving toward a modeled 40.5% by 2030E.
The backlog data adds a demand-side confirmation of the same point, and it is a sharper comparison than the margin figures alone. AWS's remaining performance obligations rose $132 billion in a single quarter — from $364 billion to $496 billion — while Google Cloud's rose $54 billion over the identical quarter, from $460 billion to $514 billion. AWS added roughly 2.4 times as much contracted future revenue in the same three months — despite already having the larger current-revenue base ($42.2 billion quarterly for AWS versus $24.8 billion for Google Cloud) and the slower headline growth rate (37% versus 82%). Rate and scale are telling different stories here: Google Cloud's faster percentage growth is partly a function of its smaller base, but AWS is being handed more new contracted demand in absolute dollar terms — the kind of gap that is easy to miss when the conversation centers on the more dramatic percentage.
This matters for the ROIC trajectory specifically because backlog converting to revenue is what ultimately determines whether AWS's margin resilience continues or fades once the current capex wave hits full depreciation. A segment adding $132 billion of contracted demand in one quarter has more revenue already locked in to absorb that depreciation than one adding $54 billion. It is one more reason — alongside the margin evidence already cited — that Amazon's shallow ROIC decline in this report's base case looks more like a genuine structural difference than a modeling artifact, even as the capex-productivity decomposition earlier in this report cautions that part of Amazon's resilience still depends on an assumed margin mix shift rather than an observed one.
Microsoft's compression is the largest of the three measured from its own historical peak — unsurprising given it started from the highest base. A company with a historical ROIC that peaked above 100% in FY2022 has the furthest to fall purely on capital‑base‑growth mechanics, independent of anything going wrong operationally. (Measured from 2025 rather than from that peak, Alphabet's decline is actually the larger of the two — see the Core Test section above.) Consistent with that, Azure's growth has if anything accelerated (43% in fiscal Q4 2026, guided to roughly 45% for fiscal Q1 2027), which is consistent with an infrastructure buildout still ramping faster than the revenue it supports converts to margin, rather than a demand problem. The bear case in this author’s own MSFT model, built before this report, names the exact risk this report is measuring: "heavy capex... does not earn adequate returns, compressing FCF and ROIC."
Alphabet's compression is the sharpest in its timing — a single‑year drop from 41.4% (2026E) to 30.0% (2027E) — which lines up with the fact that Alphabet's FY2026 capex guidance was raised twice during the year and management has signaled a further significant increase in 2027 — which this report models at roughly $255 billion — even as Q2 2026 free cash flow went negative for the first time since the company's 2004 IPO. Google Cloud's own segment margin story is, on paper, the most dramatic of the three — from an 11.3% operating loss in 2022 to 23.7% operating income by 2025A — but Alphabet is also carrying the AI buildout across a business (Search, YouTube) that itself now shows rising R&D and depreciation growing faster than consolidated revenue, per Alphabet’s own reported results.
Why This Is a Company-Wide Test, Not a Segment Test — and What That Hides
This is worth stating as plainly as possible, because it's the most consequential limitation in the entire report and the easiest one for a reader to skate past. None of the three companies disclose capital expenditure by segment. AWS's reported operating income already nets out its own segment's cost base, including depreciation on AWS‑specific infrastructure — so the AWS operating‑margin figures cited throughout this report are, in fact, segment‑specific. But the company‑wide ROIC test and the cumulative‑capex‑productivity test above are not: they use each company's total reported capex, which mixes AI/cloud infrastructure spend with retail fulfillment centers (Amazon), Windows/Xbox/Surface infrastructure (Microsoft), and Search/YouTube/Waymo infrastructure (Alphabet).
What that means in practice: if AI/cloud‑attributable capex specifically is earning a higher return than the legacy‑business capex it's commingled with — plausible, given AWS's and Google Cloud's margin trajectories — then the true AI‑specific ROIC could be running above what this report's company‑wide test shows, and the compression could be partly a legacy‑business drag rather than an AI‑capex problem. The reverse is equally possible: if AI infrastructure is disproportionately funded by debt and carries higher depreciation intensity than the legacy business (both plausible, given the accelerated‑depreciation useful‑life extensions and off‑balance‑sheet lease structures several of these companies have adopted), the AI‑specific return could be running below the company‑wide figure. This report cannot distinguish between those two stories with the data currently disclosed, and any reader treating the ROIC and incremental‑efficiency figures above as an AI‑specific measurement, rather than a company‑wide one, is reading more precision into them than the underlying disclosure supports.
Applications: Family Offices, Professional Investors, and Corporate Finance Practitioners
For family offices and institutional allocators
The value of running one consistent ROIC/WACC/EVA test across asset classes — gaming equities, Las Vegas multifamily real estate, and now megacap technology capex — isn't confined to any single analysis. For a family office or institutional allocator holding positions across public equities, direct real estate, and increasingly AI‑infrastructure‑adjacent private credit and data‑center funds, that consistency is the actual point: it gives an allocator one comparable, dollar‑denominated question to ask of every position in a portfolio — is this specific capital allocation earning more than it costs, or less — rather than evaluating a hyperscaler on P/E multiples, a real estate deal on cap rate, and a data‑center credit fund on yield, using three frameworks that don't speak to each other.
For allocators specifically exposed to the AI infrastructure buildout through direct or fund‑level positions in data‑center real estate, power infrastructure, or specialty credit — rather than through the hyperscalers' own public equity — this report's cumulative‑capex‑productivity test is arguably more directly relevant than the company‑wide ROIC test: it's a closer proxy for the return profile of the underlying infrastructure itself, independent of each hyperscaler's legacy business.
For professional investors and equity research practitioners
This report is not a substitute for equity research on AMZN, MSFT, or GOOGL — it names no rating, no price target, and no position, and it shouldn't be read as one. What it offers instead is a framework and a set of trackable inputs that can sit alongside whatever valuation work a professional investor is already doing, whether that's multiple-based, DCF-based, or sum-of-the-parts.
Three specific applications for that workflow:
A multiple-justification cross-check. If a hyperscaler trades at a premium multiple partly on the strength of its historical capital efficiency, the ROIC/WACC spread trajectory shown earlier in this report is a direct, quantifiable test of whether that historical efficiency is still the right base rate to extrapolate — or whether the multiple is pricing in a persistence of returns the company's own capex trajectory is working against. A widening gap between a stock's implied multiple and its modeled forward spread is a question worth asking, not an answer in itself.
A quarterly tracking checklist. The cumulative-capex-productivity metric introduced earlier in this report is simple enough to update every quarter from each company's own segment and capex disclosures: AWS's and Google Cloud's segment operating margins (both independently disclosed), Azure's product-line revenue growth (disclosed, though its margin is not), and each company's trailing-twelve-month capex against trailing NOPAT growth. Tracked quarter over quarter, a sustained divergence from this report's base-case trajectory — in either direction — is an earlier signal than waiting for full-year results to confirm a trend.
A benchmark against Street consensus, not a replacement for it. The Wall Street consensus comparison later in this report is a starting point for professional investors already tracking sell-side capex numbers, not a substitute for them — it exists to show where an independent, bottom-up model and top-down consensus diverge, and by how much, so that divergence can be investigated rather than assumed away.
For corporate finance practitioners and CFOs
The most directly actionable audience for this report may not be anyone valuing AMZN, MSFT, or GOOGL at all — it's a CFO or finance team running a capex program at a company where none of the three trade. The mechanics this report applies to hyperscaler capex are the same mechanics that apply to any material capital allocation decision, and several of this report's own findings translate into concrete practice:
Track ROIC by initiative, not just company-wide. The segment-test limitation identifies the biggest limitation in this report as its inability to isolate AI/cloud-specific capital efficiency from each hyperscaler's legacy business, because none of the three disclose capex by segment externally. A CFO evaluating a major internal capital program does not have that excuse — initiative-level capex and initiative-level return tracking is achievable internally even when it isn't externally disclosed, and this report's central limitation is a useful illustration of what happens to capital-allocation visibility when that discipline is absent.
Build an incremental-return trigger, not just a payback-period estimate. The cumulative-capex-productivity metric described earlier in this report — incremental profit per dollar of incremental capital deployed, tracked over rolling windows — is a more sensitive early-warning signal than a static payback-period or IRR-at-approval calculation, because it keeps testing a program's efficiency as it scales rather than only at the initial approval decision. A capital committee that reviews this metric quarterly for major programs will see efficiency decay well before it shows up in consolidated margins.
Stress-test depreciation policy before the board does. The GPU/server useful-life debate discussed later in this report, in the Scenarios and Wall Street Context section, is a reminder that depreciation assumptions are a judgment call with a direct, mechanical effect on reported returns on any capital-intensive program — not just hyperscaler AI infrastructure. Presenting capital committee or board materials with a range of useful-life assumptions, rather than a single house view, is cheap insurance against exactly the credibility question now being asked publicly of much larger companies.
Treat WACC as a variable in capital-approval materials, not a constant. This report holds each company's WACC fixed across a nine-year window as an explicit, disclosed simplification — a modeling convenience for a public research report. A CFO building an internal capital-approval package for a multi-year program has less excuse for the same simplification: financing costs move, and a program approved at one WACC can look materially different at a WACC 100–200 basis points higher, particularly for debt-funded capital plans.
Market Evidence and the Inflection Question
Everything in the ROIC, capex-productivity, and divergence sections above is this report's own base case. Before treating that base case as the likely path, it is worth asking what the market itself concluded when these three companies reported the very quarters the base case is built from — and whether that evidence is strong enough to constitute a turning point. This section takes up those questions; the section that follows widens the lens to scenario ranges, sell-side consensus, and other published return estimates.
The market's real-time verdict: how Wall Street read the Q2/Q4 FY2026 prints
This report's base case already incorporates Alphabet's, Microsoft's, and Amazon's Q2/Q4 FY2026 results — the same quarters discussed throughout the cloud-growth, ROIC, and divergence sections above. What those results did to each stock, and to Wall Street's own estimates, is worth examining directly, because it is the most current test available of whether the market is starting to price a different path than this report's base case. The three reactions happened within nine days of each other and, taken together, tell a more precise story than any one of them alone.
Alphabet reported first, after market close on July 22, and its stock fell as much as 6–8% the following session — from a $341.91 close to as low as $316.68 — even though the quarter beat on nearly every headline metric: revenue of $119.8 billion (up 24%) against a roughly $117 billion estimate, and Google Cloud revenue up 82% to $24.8 billion, ahead of the roughly $24.6 billion Street estimate, with Cloud's operating margin reaching 35.6% and its backlog jumping more than $50 billion in the quarter to $514 billion. What moved the stock was guidance: CFO Anat Ashkenazi raised full-year 2026 capex guidance to $195–205 billion, up from $180–190 billion set just one quarter earlier, and second-quarter free cash flow turned negative for the first time since Alphabet's 2004 IPO. Sell-side reaction was a rare case of price-target cuts alongside maintained bullish ratings: Morgan Stanley to $400 from $415, Oppenheimer to $400 from $445, JPMorgan to $420 from $460, Piper Sandler to $395 from $445, UBS to $379 from $400 — while Wedbush held its $445 target and added Alphabet to its Best Ideas List, and Barclays raised its target to $425 from $405. JPMorgan's own framing captured the split: the firm called the selloff "a buying opportunity," citing the same 82% Cloud growth that triggered it — a characterization of JPMorgan's own view, cited here as evidence of the analyst split, not adopted or endorsed by this report. Notably, Alphabet's stock did not stay down — by early August it had recovered to roughly $356, above its pre-earnings level, once Amazon's and Microsoft's results the following week reinforced the industry-wide margin and backlog story.
Amazon closed July 31 up 15.32% to $271.58, its largest single-day gain since 2012, after AWS revenue grew 37% to $42.2 billion — beating the roughly 31% Wall Street had modeled and marking AWS's fastest growth in 18 quarters. AWS operating income rose to $16.6 billion from $10.2 billion a year earlier, a segment margin near 39%, and AWS's backlog reached $496 billion, up 154% year-over-year. Separately, CEO Andy Jassy disclosed that two distinct AWS businesses — its AI services business and its custom-silicon business, the latter spanning Trainium, Graviton, and Nitro combined — had each surpassed a $25 billion annualized revenue run rate, both growing at triple-digit year-over-year rates, and said AWS could become "a trillion-dollar annual revenue business... in time." Sell-side price targets moved accordingly: Goldman Sachs to $375 from $335, JPMorgan to $365 from $330, Benchmark to $400 from $370, Truist to $350 from $320, Morgan Stanley to $335 from $330, with the Street average landing near $322 — roughly 18% above the post-earnings price even after the surge. Not every revision was positive; Cantor Fitzgerald trimmed its target to $320 on valuation-methodology grounds, a useful reminder that the re-rating was not unanimous.
Microsoft rose 8.13% after hours to $422.30 from a $390.54 close, then continued to roughly $427 the next session and to $464.72 within days — adding, by one estimate, some $260 billion in market value in a single evening — after Azure revenue accelerated to 43% (from 40% the prior quarter), beating a Street estimate near 40%, and crossed $100 billion in annual revenue for the first time. Commercial remaining performance obligations — Microsoft's disclosed backlog — jumped $51 billion in the quarter alone, to $678 billion, up 84% year-over-year, and paid Copilot seats crossed 30 million, up from roughly 20 million in April. Operating margin ticked up to 45% even as gross margin absorbed heavier AI infrastructure costs. Sell-side targets moved sharply: Goldman Sachs to $640 from $610, Bernstein to $641, Wedbush to $625, Morgan Stanley to $600, Evercore ISI to $528, with the Street average near $558 — Goldman's own note framed the move as roughly 64% upside from the pre-earnings close. Morgan Stanley's own language is worth quoting directly, because it states plainly, in that firm's own words and as that firm's own view, what this report has been testing quantitatively with its own independent methodology: management's results, the firm wrote, "move the key elements of our investment thesis from expectation to evidence."
Two things about this set of reactions are directly relevant to this report's framework, and one qualification matters before either is taken too far. First, the market's response was not a uniform reward for capex or cloud growth — Alphabet's own initial reaction is the clearest evidence: it beat on cloud growth by more than either Amazon or Microsoft and its stock still fell, because it raised capex guidance without the same margin and backlog evidence the other two supplied. What appears to have driven Amazon's and Microsoft's re-rating specifically — and, on a one-week lag, Alphabet's own recovery — was evidence of conversion: AWS's record segment margin, Microsoft's jump in backlog and Copilot seats, and, once the broader picture came into view, Google Cloud's own 35.6% margin and $514 billion backlog reread more favorably in that context. That is, in miniature, the exact distinction this report's ROIC and cumulative-capex-productivity tests are built to draw, and the market drawing the same distinction in real time, on the same week's data, is a meaningful piece of corroborating evidence for the framework itself, independent of what it implies about either stock.
Second, on valuation: where this report has comparable data, neither Amazon's nor Microsoft's current multiple obviously extends beyond its own recent range. Amazon's forward P/E near 29x sits well below its own trailing three-year average near 47x; Microsoft's forward P/E in the low-to-mid 20s sits below its own three- and five-year averages near 32x. Whether that means either stock is undervalued is not a question this report answers — that is a security-specific judgment this report explicitly does not make, and multiple compression from a company's own historical average can reflect a genuine business slowdown rather than a bargain. What this report can say is narrower: the post-earnings re-rating at Amazon and Microsoft was not, on these figures, a move to an unprecedented multiple: it was a move back toward each company's own longer-run valuation range, funded by upward revisions to the same margin and backlog figures this report's ROIC framework already uses as inputs. Alphabet's own round trip — down sharply, then recovered — is consistent with the same read: the market's initial move priced the capex increase without the offsetting evidence, and the recovery followed once that evidence arrived from elsewhere in the sector. None of the foregoing is an opinion on whether any of the three stocks is currently fairly priced; it is a description of how each company's multiple sits relative to its own history, offered only to show that the re-rating was not, on these particular figures, a move into unfamiliar territory.
The qualification: part of all three headline results was non-operational, and in Alphabet's case decisively so. Roughly two-thirds of Microsoft's EPS beat over consensus, and the large majority of Amazon's GAAP EPS figure, reflected one-time gains from marking Anthropic and OpenAI investment stakes to market — not incremental cloud profit. Alphabet's own headline EPS of $9.11 against a roughly $2.88–2.95 estimate looks like an even larger beat, but once its own Anthropic-stake markup is stripped out, adjusted EPS actually missed consensus by about four cents — a useful reminder that the market's initial capex-driven selloff, not the headline profit number, was the more informative reaction that day. Amazon's free cash flow swung to a $7.6 billion outflow from an $18.2 billion inflow a year earlier, and its net cash position flipped to net debt as borrowing funded the capex step-up; Alphabet's own quarter was free-cash-flow negative for the same reason. Microsoft's reported capex figure for the quarter, which the market read as restraint relative to Alphabet's raised guidance, was itself partly the product of an accounting classification change taking effect in FY2027, not solely a demand-driven spending decision. None of this negates the margin and backlog evidence — Azure's 43% growth and $678 billion backlog, AWS's record segment margin, and Google Cloud's own 35.6% margin and $514 billion backlog are all demand-driven, cash-generating figures no accounting choice can manufacture — but it argues for treating one very strong quarter, reported alongside unusually large one-time items, as a data point rather than a confirmed trend.
What this means for the trajectory shown earlier in this report: this report's base case does not currently assume acceleration — it assumes continued, moderating growth roughly in line with what AWS, Azure, and Google Cloud had already been showing before these three prints. If the margin and backlog evidence behind this re-rating persists through Q3 2026 for Amazon, Q1 FY2027 for Microsoft, and Alphabet's next report — where Microsoft has already guided Azure growth toward roughly 45%, above the roughly 41% Street had modeled — the more resilient end of this report's own scenario range in the next section, not the base case's steeper decline, becomes the more likely path. One week of evidence, partly flattered by one-time investment gains at all three companies, is not enough to revise a base case built on a multi-year trajectory. Whether it is the start of a genuine inflection or the high-water mark of an unusually strong stretch of quarters is a question the next two to three reporting periods — not this report — will answer.
Update, August 3, 2026. The move described above has continued rather than faded. Amazon crossed $3 trillion in market capitalization for the first time this morning — becoming only the fifth company to reach that threshold — with shares up as much as 5.5% intraday on top of the prior week's gains; Alphabet and Microsoft were also higher on the day. Some of this is broad market strength rather than an AI-specific signal: small-cap and industrial indices rallied by similar or larger margins the same morning, so not all of today's move should be read as incremental conviction in the cloud-capex thesis specifically. Taken together with the past week and a half, this extends the observation window this report is tracking but does not change the underlying conclusion above — one additional day of a continuing move is still data to weigh, not a basis to revise the base case.
The case that this is a multi-period inflection, not a single strong quarter
This report has so far treated the evidence conservatively — one strong reporting cycle, partly flattered by one-time gains, insufficient to revise a multi-year base case. That remains this report's position, and nothing in this section changes it. What follows is a deliberate steelman: the strongest version of the argument against this report's own base case, set out at full strength so a reader can weigh it rather than encounter it only as a hedge. Presenting an opposing case fairly is not the same as endorsing it, and the closing subsection explains why this report does not adopt it.
First: the acceleration is sequential, not a single-quarter spike. A one-quarter beat can reflect timing, a favorable comparison, or a pulled-forward contract. What the current data shows instead is consecutive sequential acceleration at all three companies. Azure went from 39% (fiscal Q3 2026) to 40% to 43%, described by Microsoft as its fourth consecutive quarter of sequential acceleration, with management guiding toward roughly 45% for fiscal Q1 2027 — a fifth. Google Cloud went from 63% (Q1 2026) to 82% (Q2 2026). AWS accelerated for a fifth consecutive quarter to its fastest growth in 18 quarters. Businesses at this revenue scale do not typically accelerate at all, let alone for four or five consecutive periods across three separate companies simultaneously. That pattern is structurally harder to attribute to timing noise than a single beat would be.
Second: backlog is a forward-looking, contractually-committed measure, and it grew faster than revenue. This is the most important argument, because backlog is precisely the variable that distinguishes a durable inflection from a good quarter. Microsoft's commercial remaining performance obligations rose $51 billion in a single quarter to $678 billion (+84% YoY); Alphabet's cloud backlog rose more than $50 billion in the quarter to $514 billion, from $106 billion a year earlier; Amazon's reached $496 billion, up roughly 154% YoY. RPO represents contracted revenue not yet recognized — customer commitments already signed. Backlog growing substantially faster than current revenue at all three companies means the revenue acceleration already visible is, to a meaningful degree, already contracted rather than merely hoped for. A demand story that were about to reverse would show backlog decelerating before revenue did; the disclosed data shows the opposite.
Third: margin expansion is occurring simultaneously with the capex peak, which is not the normal sequence. The conventional pattern in an infrastructure buildout is margin compression during the build followed by expansion after it — depreciation and operating costs hit the P&L before the revenue does. AWS instead posted a record 39.4% segment operating margin, up 650 basis points year-over-year, during the heaviest capex quarter in its history and while simultaneously accelerating revenue growth. Google Cloud's margin more than tripled year-over-year to 35.6%. Microsoft's consolidated operating margin ticked up to 45% even as gross margin absorbed AI infrastructure costs. If the AI-era capital base were structurally less productive than the legacy base, the segment margins should be compressing during exactly this phase. They are expanding.
Fourth: supply, not demand, is the stated binding constraint — at all three companies independently. Amazon's CEO stated the company will not have enough capacity to meet 2026 demand and expects the same in 2027. Microsoft stated commercial demand continues to exceed available capacity. Alphabet's management cited compute constraints in explaining its raised capex guidance. When capacity rather than demand is the limiting factor, incremental capex converts to revenue with unusually high confidence and unusually short lag — which is the specific condition under which this report's unlagged incremental-efficiency metric would understate the true return, since the metric charges each period's full capex against that same period's NOPAT change while the revenue conversion is still in front of it.
Fifth: the market's differentiated response suggests it is pricing conversion evidence, not sentiment. Had the July 2026 reaction been an indiscriminate AI-sentiment rally, all three stocks would have moved together on their common cloud beats. Instead Alphabet fell on capex guidance without matching margin evidence, then recovered once peer results supplied that evidence; Meta fell on raised spending without a clear demand narrative; Amazon and Microsoft rose on margin and backlog specifically. That is a market discriminating between capex-with-conversion-evidence and capex-without-it — the same distinction this report's framework is built to measure. A market applying that test consistently is more likely to be responding to a genuine change in the underlying evidence than to a mood.
Sixth: independent third-party analysis points the same direction on AI-specific returns. Morgan Stanley's three bottom-up frameworks estimate roughly 25%–50% incremental ROIC on AI-specific unit economics, and Convequity's independent analysis lands near 29% on gross incremental AI capex. Both are well above this report's 11.6%–13.0% company-wide figure. As the comparison table below explains in full, the gap reflects two compounding differences, not scope alone: Morgan Stanley's and Convequity's figures are returns on a capital stock (net invested capital), while this report's is profit added per dollar of capital spent (cumulative gross capex) — a structurally different, and mechanically lower-reading, construction even before scope is considered. Scope then adds a second, additive gap on top: AI-specific workload economics versus whole-company capex including legacy business. If the AI-specific estimates are directionally right, this report's blended, company-wide figure is understating AI-capex productivity specifically, and the compression this report documents may be substantially a denominator-construction and legacy-business effect rather than a statement about the new capital itself.
What would have to be true for this to be a genuine multi-period inflection: AI/cloud-attributable capex would need to be earning materially above each company's blended company-wide return (supported by segment margin data and by Morgan Stanley's and Convequity's independent estimates); the backlog would need to convert to recognized revenue at roughly disclosed terms (supported by supply-constrained conditions, unproven as to timing); and depreciation schedules would need to be approximately right rather than materially too generous (genuinely contested — this is the strongest argument on the other side). The first two conditions have real supporting evidence today. The third does not, and it is the reason this report does not adopt the inflection case as its base case.
Why this report nonetheless stops short. Three considerations keep the conservative reading as this report's base case. The GPU/server useful-life debate discussed below is unresolved and cuts directly against the margin evidence — if depreciation is understated, reported segment margins and reported NOPAT are both overstated, and the margin expansion cited above is partly an accounting artifact. Backlog conversion is contractual but not certain as to timing, and Azure's backlog specifically carries disclosed counterparty concentration. And a single reporting cycle, however strong across three companies, is a short observation window against a multi-year capital cycle; the 2022–2025 period contains its own examples of quarters that looked like inflections and did not persist. The honest formulation is that the near-term evidence for the inflection case is stronger than it was two months ago, that it remains materially less established than the multi-year evidence underlying this report's base case, and that the next two to three reporting periods — specifically, whether backlog continues outgrowing revenue and whether segment margins hold as depreciation from the 2026 capex wave begins hitting the P&L in earnest — should resolve it in one direction or the other.
Scenario Range and Wall Street Context
The preceding section examined what the market concluded from a single reporting cycle. This section widens the lens: what each company's own bull and bear scenarios imply for the spread, how this report's capex path compares to sell-side consensus, how its return figures compare to other published estimates, and which variables would move the spread from here.
The scenario range, in each company's own words
All three companies' underlying models already contain bull/base/bear scenario tabs, built independently of this report and predating this specific ROIC/WACC question. Two points of method before reading them. First, because the Amazon and Microsoft scenario tabs were built before the post‑Q2‑2026 model refresh, their base cases had drifted from the refreshed projections used elsewhere in this report; the figures below therefore re‑anchor each base case to the refreshed model and preserve each tab's original bull and bear spreads around that new midpoint, so the range reflects current results and guidance rather than a pre‑earnings starting point. Second, these are 2030E operating margins, not ROIC directly — a proxy for the direction and rough magnitude of spread compression, not a restatement of the ROIC series shown earlier in this report.
| Company | Bear | Base | Bull |
|---|---|---|---|
| AMZN | 12.4% | 17.1% | 18.4% |
| MSFT | 41.3% | 46.3% | 52.8% |
| GOOGL | 30.0% | 33.8% | 36.0% |
"Base" is each company's refreshed post‑Q2‑2026 model projection, consistent with the ROIC series shown earlier in this report. Bear and bull preserve the spreads from each model's original Scenario tab, re‑centred on the refreshed base. Author's compilation from AMZN, MSFT, and GOOGL models and Scenario tabs.
The bear case at all three companies names the same failure mode independently — capex outrunning the revenue and margin needed to earn it back — in language written into each model before this report's ROIC/WACC question was asked of it. Microsoft's own bear‑case rationale states plainly that "heavy capex... does not earn adequate returns, compressing FCF and ROIC." Alphabet's bear case cites "capex overbuilt, depreciation compresses margins." Amazon's cites AI demand disappointment and an explicit capex pullback "to protect FCF." None of the three bear cases, on their own terms, imply a spread that goes negative by 2030E — margin compression of this magnitude would narrow the ROIC/WACC spread well below this report's base case, plausibly into the high single digits to low teens rather than the mid‑20s, but not through zero in any of the three models as currently built.
Using the individual-company figures directly — no scope adjustment needed — this report's combined 2027E capex assumption ($630B) sits $132 billion, or about 21%, below the sum of the individual-analyst estimates ($762B). If those estimates prove closer to right than this report's own refreshed model, the ROIC compression shown earlier in this report is more likely to understate the actual 2027E–2028E trajectory than to overstate it. That conclusion follows directly from the gap: because heavier capex spending relative to profit growth is exactly what drives the ROIC compression this report documents, a lower capex input mechanically produces a milder compression result. So if the individual-analyst estimates turn out to be closer to right, the true compression these three companies experience is more likely to come in worse than this report shows than better — this report's base case understates how much capex is coming, not overstates it.
One further, independently‑sourced data point worth citing directly: Goldman Sachs Research estimates sector‑wide depreciation and amortization will rise from roughly 7% of revenue in 2022 to approximately 12% by 2027 — an independent, third‑party estimate of very close to the same mechanism this report's own ROIC compression is built to capture, arrived at through entirely different means.
How this report's figures compare to other published ROIC/ROIIC work
This report's company‑wide ROIC test and its cumulative‑capex‑productivity proxy are not the only published attempts to answer some version of "is hyperscaler AI capex earning its keep." At least three other sources, using three different methodologies, have published their own figures within weeks of this report — comparing them directly is useful for calibration, and the size of the gap between them is itself informative.
| Source | Scope | Denominator | Estimated return | Comparable to this report's Sharper Test? |
|---|---|---|---|---|
| This report — Core Test | Company‑wide, all invested capital | Net invested capital (stock, t−1) | ~23–43% across 2026E–2030E, converging toward ~25% by 2030E | Reference point (this is this report's own ROIC) |
| This report — Sharper Test | Company‑wide, capex‑specific proxy | Cumulative gross capex (flow, unlagged) | 11.6%–13.0% (2026E–2030E projected) | — |
| Morgan Stanley (Jul. 27, 2026) | GenAI‑specific unit economics, three bottom‑up frameworks | Net invested capital (stock) | ~25%–50% headline range; point estimates reported at ~31% / ~46% / ~25% | No — stock vs. flow denominator, plus scope |
| LPL Research (Jul. 2026) | Cloud segment specifically, estimated | Estimated net "deployed infrastructure asset base" (stock) | Scenario framework, not a single headline figure | No — stock vs. flow denominator |
| Convequity (independent, Nov. 2025) | AI‑specific incremental capex, ecosystem‑wide | Gross incremental capex (flow, pre‑depreciation) | ~29% (~$104B ÷ ~$364B) | Partially — same flow-over-flow structure, different window (2yr vs. 4yr) and scope |
This report's figures per the Core Test and Sharper Test tables above. Morgan Stanley figures per BigGo Finance's summary of the firm's July 27, 2026 research note. LPL figures per LPL Research / Advisor Perspectives, "Can Hyperscalers Earn Their AI Ambitions?" (Jul. 2026). Convequity figure computed from the two inputs the source discloses; not the source's own stated headline number. See Sources & Citations for full references.
The gap between 11.6–13.0% and Morgan Stanley's ~25–50% should not be read as this report finding a lower return than Morgan Stanley did — the two figures do not measure the same thing, and treating one as a discount to the other misstates what each represents.
Morgan Stanley's ranges are returns on a capital stock: annual profit divided by net invested capital, the standard construction of ROIC. This report's Sharper Test is profit added divided by capital spent — a flow-over-flow productivity ratio, structurally closer to a payback measure than a return measure. That difference alone is large. Cumulative gross capex over a multi-year build is always bigger than the resulting net invested capital, because depreciation continuously reduces the asset base as spending continues; on assets with 5–6 year useful lives, illustratively grossing this report's 11.6–13.0% up for that effect alone — without changing anything else — could plausibly put it in the high‑10s to mid‑20s range, before the scope difference is even considered. This report has not performed that recalculation with company-specific depreciation schedules, so the adjusted figure is illustrative of direction and rough magnitude only, not a corrected estimate.
Scope is the second, separate difference, and it points the same direction. Morgan Stanley's frameworks isolate AI-specific workload economics — GPU leasing margins, API-service margins. This report's test divides whole-company NOPAT growth by whole-company capex, mixing AI-specific spend with legacy-business capex (retail fulfillment centers, Windows/Xbox infrastructure, Search infrastructure) that this report has already flagged, in the segment-test limitation above, as impossible to separate from disclosed data.
Because both differences run the same direction — a flow-over-stock denominator plus a broader, more diluted scope both push this report's figure below a segment-level, stock-based return — a materially higher Morgan Stanley range is exactly what this report's own construction would predict, not evidence of disagreement between the two. If AI-specific workloads really do earn something closer to Morgan Stanley's range, that is consistent with this report's lower company-wide figure and with a legacy business diluting the blended number — a decomposition this report cannot perform with public disclosures as they currently stand. The honest conclusion is narrower than a side-by-side number comparison suggests: these two ranges corroborate each other's direction without being reconcilable to a common number, and neither should be used to sanity-check the other's magnitude.
LPL Research's approach is the closest published attempt at the segment‑level invested‑capital estimate this report's own segment‑test limitation says is missing — it builds a "deployed infrastructure asset base" specific to the cloud segment using estimated, not disclosed, net PP&E and lease assets. That is a genuine methodological advance over this report's company‑wide approach, purchased at the cost of relying on allocation estimates rather than each company's own reported figures; this report's own Limitations section flags exactly that same trade‑off (estimation versus disclosure) wherever this report itself relies on approximation. Convequity's independent framework is the closest published analog to this report's own incremental‑efficiency test specifically — both divide a NOPAT change by a capex‑based denominator over a multi‑year window — but the two differ in invested‑capital convention (Convequity uses a gross, pre‑depreciation capex base; this report uses cumulative capex without netting depreciation either way, a related but not identical convention) and in window length (two years there, four years here), which is enough to make the ~29% and 11.6–13.0% figures not directly comparable despite the methodological similarity.
How this report's capex path compares to Wall Street's
The most commonly published Street capex figures are reported at the group level — four or five hyperscalers combined, typically adding Meta and sometimes Oracle or SpaceX — which doesn't match this report's three-company scope and would require approximating around the difference. A cleaner test compares this report's own per-company figures directly against named analysts' per-company estimates for the same year: no scope adjustment, no backing anything out. For 2027E specifically, that comparison is available for all three companies:
| Company | This report's model | Named-analyst estimate | Source |
|---|---|---|---|
| AMZN | $225B | ~$250B | Evercore ISI (Mark Mahaney) |
| MSFT | $150B (FY2027E) | ~$262B (FY2027) | BNP Paribas (Stefan Slowinski) |
| GOOGL | $255B | ~$250B | Morgan Stanley (Brian Nowak) |
| Combined | $630B | ~$762B |
Each figure is one named analyst's individual estimate, not a polled consensus. MSFT's estimate is fiscal-year, matching this report's own convention; AMZN's and GOOGL's are calendar-year. On the MSFT scope difference specifically, see the divergence discussion above and the paragraph below. See Sources & Citations.
The gap is not evenly distributed across the three companies. Alphabet's individual estimate ($250B) sits almost exactly on this report's own modeled figure ($255B) — essentially no gap. Amazon's gap is modest (~$25B, or 11%). Microsoft's is the largest of the three in both dollar and percentage terms (~$112B, or 75%) — though part of that gap is scope, not a pure spending disagreement: Microsoft's own FY2027 guidance already points to roughly $175B in cash capex plus finance leases, about $25B above this report's $150B cash-only figure. Even measured against that broader $175B guidance figure, the $262B analyst estimate is still $87B higher, so scope alone doesn't close the gap. As with the other two figures in this table, Microsoft's estimate is one named analyst's individual estimate rather than a polled consensus average, and readers should weigh it with that in mind.
The major factors that would move this spread from here
GPU and server depreciation policy — the single largest lever on the reported number, not just the economic one. A live accounting and disclosure debate centers on whether hyperscalers' 5–6 year useful‑life assumptions for GPU and server hardware are too generous given Nvidia's roughly 2–3 year product‑refresh cycle — a question now drawing attention in accounting and securities‑law commentary because useful‑life estimates are management judgments that flow directly into reported operating income. Investor Michael Burry brought the issue to broad public attention in November 2025 and published his own estimate of the cumulative sector‑wide depreciation understatement for 2026–2028 at roughly $176 billion; that figure is his own opinion, is not audited or independently verified, and is noted here only to indicate the order of magnitude some critics attach to the question, not as a number this report adopts or relies on. If directionally correct, reported NOPAT, and therefore reported ROIC, across all three companies in this report is running above the economic reality that a shorter, more conservative useful‑life assumption would produce — meaning this report's already‑compressing ROIC series could be overstating, not understating, the companies' true capital returns. This is the single factor this report would prioritize modeling explicitly in a future revision.
Financing costs and WACC drift. Big Tech bond issuance in the first quarter of 2026 alone exceeded all of 2025's issuance, and credit spreads on hyperscaler debt have widened sector‑wide through 2026. The important company‑specific nuance: rating agencies continue to place Amazon, Microsoft, and Alphabet well inside their respective downgrade thresholds — net leverage under roughly 1x for what one rating‑agency‑adjacent source called "the four larger names" — with Oracle, not any of this report's three companies, identified as the stressed credit in the sector. WACC drift risk here is real but likely second‑order for this specific trio relative to the sector as a whole, showing up more plausibly as modestly higher marginal cost of debt on new issuance than as a rating‑driven repricing of existing WACC.
Backlog conversion pace and customer concentration. A slower‑than‑guided pace of converting backlog to recognized revenue delays the NOPAT side of the ratio without slowing the capex side of it — mechanically the least favorable combination for near‑term ROIC. Azure's backlog carries the most disclosed counterparty concentration risk of the three, given its OpenAI‑related commitments; AWS's and Google Cloud's backlogs are, on current disclosure, more diversified across customers.
Power and physical supply constraints. These cut in both directions rather than one. A widening US data‑center power shortfall — one widely‑cited estimate points toward roughly 49 gigawatts by 2028 — could delay revenue recognition on capital already deployed, which would weigh on near‑term ROIC. The same constraint could equally act as an external brake on further capex growth once demand catches up to available power, which would show up as improving capital efficiency for reasons unrelated to any deliberate capital‑discipline decision by management.
A shift in capex mix toward leasing. Sell‑side commentary has flagged a strategic move by several hyperscalers toward leasing data‑center capacity rather than owning it outright. This would reduce both reported capex and reported invested capital going forward, which would mechanically support this report's ROIC metric even if the underlying unit economics of the AI buildout were unchanged — a reason cross‑company and cross‑period ROIC comparisons in this series could become progressively less apples‑to‑apples over the 2026–2030 window.
Competitive and pricing dynamics; the macro rate path. Token‑price deflation, the ongoing shift toward custom silicon (AWS Trainium, Google TPU, Azure Maia), and price competition from specialized "neocloud" providers all affect the revenue side of this equation and are not modeled explicitly in this report. On the WACC side, a sustained hawkish Federal Reserve stance and elevated long‑term Treasury yields — the same dynamic this author's real estate report in this series documented in detail — would raise the equity risk premium embedded in all three companies' cost of capital, narrowing the spread from the WACC side rather than the ROIC side.
Limitations
WACC is held constant across the full 2022–2030 projection window — a deliberate choice, with a quantified sensitivity below. Each company's WACC is a single CAPM output, not re-estimated year by year. The argument for varying it is real: all three companies shifted meaningfully toward debt funding during 2026 (Amazon flipped from roughly $58 billion of net cash to net debt in six months; Alphabet's long-term debt rose from $46.5 billion to $98.2 billion between year-end 2025 and Q2 2026), and a strict reading of the framework would re-weight the capital structure each year to reflect that.
Three considerations argue for the constant treatment here. First, the debt weights are small enough that leverage changes are second-order: at market values, debt represents 1.15% to 2.21% of total capital across the three companies, so even doubling a debt weight moves WACC by roughly 10–15 basis points against spreads measured in whole percentage points. Second, year-by-year market-value weights would require forecasting each company's market capitalization to 2030 — which is downstream of the very returns this report is measuring, embedding a valuation inside a valuation input. Third, the variable most likely to move cost of capital materially is not leverage but equity beta: if the AI buildout genuinely converts these from capital-light software businesses into capital-intensive infrastructure operators, beta should rise, and with roughly 98% of the weight sitting in equity, that channel dominates. It is also the input least amenable to credible forecasting, so this report does not attempt it.
The practical consequence is that readers should treat the WACC line as a fixed reference bar rather than a second moving variable. Because it is constant, the EVA spread narrows by exactly the amount ROIC narrows — spread compression in this report is not an independent finding from ROIC compression, it is the same result expressed against a benchmark. It also means invested capital and WACC are treated slightly differently: the denominator reflects each year's actual capital structure while the discount rate does not. The table below sizes what a higher cost of capital would do to the 2030E spreads:
| Company | WACC (base) | 2030E ROIC | Base spread | +100bp | +200bp | +300bp |
|---|---|---|---|---|---|---|
| AMZN | 10.17% | 24.9% | +14.7 | +13.7 | +12.7 | +11.7 |
| MSFT | 9.87% | 25.8% | +15.9 | +14.9 | +13.9 | +12.9 |
| GOOGL | 10.17% | 24.6% | +14.4 | +13.4 | +12.4 | +11.4 |
Spreads in percentage points. The parallel shift above is a simplification, and it understates dispersion. A uniform shift assumes all three costs of capital move together, but the most plausible driver of a genuine repricing — a rising equity risk premium tied to the buildout itself — would flow through each company's beta and equity weight rather than hitting all three equally. A +100bp rise in the equity risk premium specifically would raise Amazon's and Alphabet's WACC by about 117bp each (both beta 1.20) and Microsoft's by about 112bp (beta 1.13), taking 2030E spreads to roughly +13.6 (AMZN), +13.3 (GOOGL), and +14.8 (MSFT) — versus a flat one-point reduction for all three under the parallel-shift assumption. The three betas are close enough now that a uniform ERP shock does not reorder the companies, though Microsoft's modestly lower beta still leaves it with the widest spread under either treatment. Author's calculation. For context on magnitude: eliminating the 2030E spread entirely would require WACC to rise by 14.7 points at Amazon, 15.9 at Microsoft, and 14.4 at Alphabet — to cost-of-capital levels of roughly 25% at each. The constant-WACC simplification is therefore not what determines whether these companies clear their cost of capital in the base case; it affects the precision of the spread, not the direction of the finding.
This is a company‑wide test, not a segment‑isolated test — see the segment-test limitation discussed earlier in this report for the full discussion. This is, by a meaningful margin, the single most consequential limitation in this report.
The cumulative‑capex‑productivity test described earlier in this report does not lag capex for construction and ramp time. Data centers typically take one to three years from capital deployment to full revenue contribution. A more precise version of this test would compare NOPAT growth against capex deployed one to three years earlier, not against same‑window cumulative capex. This report uses the simpler, unlagged version for transparency and ease of replication; a lagged version would likely show a different — this report cannot say with confidence whether higher or lower — efficiency figure for the most recent years, since much of 2025–2026's capex has not yet had time to convert to revenue.
Tax‑rate treatment differs by company, and for GOOGL specifically, differs from that company's own model file. GOOGL's underlying model assumes a flat 18% tax rate throughout; this report instead applies a normalized 15% to GOOGL's EBIT for cross-company comparability, so a reader comparing this report's GOOGL NOPAT directly against that model's own DCF Inputs NOPAT row will find they do not tie — a deliberate choice, not an error. AMZN's own model already blends Amazon's actual historical effective tax rates (2021A–2025A, roughly 13%–54%, reflecting real one-time items) with a normalized 15% for 2026E–2030E, and this report uses that model's NOPAT figures directly, so AMZN's published figures do tie to its own model file. MSFT uses its own actual/guided effective rate throughout (roughly 13%–18% historically, ~19% projected), also tying directly to its own model. This is a modeling simplification stated explicitly rather than left implicit, and it affects the absolute ROIC levels shown, though it should not materially affect the direction of the compression each company shows independently.
Model coverage windows differ by company. The Amazon and Microsoft models used in this report cover 2021 (or FY2021) forward; the Alphabet model covers 2023A forward, with 2022 figures sourced externally from Alphabet's own SEC filings rather than from this author's model directly. Google Cloud's segment operating‑margin figures for 2021–2022 reflect Alphabet's currently‑reported segment definitions, which were revised in 2023 to reclassify certain shared AI research and development costs — meaning the 2021–2022 figures cited are restated, comparable figures as currently reported, not the figures as originally reported in real time.
Microsoft's historical balance sheet (FY2021A–FY2026A) is sourced directly from Microsoft's own 10-K, 8-K, and FY2026 earnings-release figures, cross-checked against multiple independent data aggregators. Every year ties exactly (total assets = total liabilities + equity). A handful of minor FY2021 liability sub-line items (the split between "other current liabilities" and "other long-term liabilities" specifically) are estimated rather than independently sourced, clearly noted as such in the underlying model; they do not affect any total used in this report's ROIC or WACC calculations. FY2027E–FY2030E invested capital is rebased proportionally onto the FY2026A actual, preserving the model's projected capital-structure growth rate rather than rebuilding the projection from first principles.
The core ROIC/WACC test earlier in this report is base‑case‑only; the scenario range in the Wall Street Context section is a rougher, margin‑based proxy, not a full re‑run of the ROIC methodology. That scenario range translates each company's own bull/base/bear operating‑margin assumptions into a directional read on where the spread could move, but it does not rebuild invested capital, NOPAT, or ROIC year‑by‑year under each scenario the way the earlier ROIC test does for the base case. A full treatment would do that; this version flags it as the natural next step rather than attempting it.
The bull and bear scenario ranges for Amazon and Microsoft are re-anchored, not independently rebuilt. Each base case is taken from that company's refreshed post‑Q2‑2026 model, but the bull and bear spreads around it are carried over unchanged from Scenario tabs built before those earnings. The width of each range therefore reflects the author's pre‑earnings judgment about how much upside and downside to allow, applied to a post‑earnings midpoint. A full treatment would re-derive each scenario's underlying assumptions — revenue CAGR, capex intensity, and segment growth — from current guidance rather than preserving the prior spread; this version does not, and the ranges should be read as indicative rather than as independently re-underwritten. Other scenario-tab outputs not used in this report (2030E revenue, capex, and EPS dollar figures) remain on their pre-refresh basis.
Cost-of-capital inputs were standardized for this report and differ from the individual model files. The three underlying company models were built separately and used slightly different market-wide CAPM assumptions — risk-free rates of 4.3% to 4.5% and equity risk premiums of 4.5% to 5.5%. The three models also set capital-structure weights inconsistently — Alphabet's derived equity and debt weights from market values, while Amazon's and Microsoft's used round-number placeholders (95/5 and 94/6) that materially overstated debt relative to each company's actual market-value capital structure. Because this report compares EVA spreads across the three companies side by side, both the market-wide inputs (standardized here at a 4.3% risk-free rate and 5.0% equity risk premium) and the capital-structure weights (now derived for all three from market value of equity against reported total debt, excluding leases for consistency with the invested-capital definition) have been put on a common basis, so that cost-of-capital differences reflect company-specific risk rather than inconsistent assumptions. Correcting the weights raised Amazon's WACC by roughly 18 basis points and Microsoft's by roughly 27. The WACCs used in this report therefore differ from those in the source model files, and any DCF valuation output in those files reflects the original, non-standardized rates.
Beta and market capitalization are refreshed to early-August 2026 levels rather than the values in the original model files, and this materially changed one company's result. Microsoft's model file used a beta of 0.9; current public data (as of early August 2026) puts Microsoft's beta at approximately 1.13, reflecting the market pricing more risk into Microsoft specifically as capex-cycle concerns intensified through 2026. Amazon's beta of 1.2 and Alphabet's market capitalization were updated to current levels as well but did not change materially. The net effect: Microsoft's WACC rose from 8.74% (using the stale beta) to 9.87%, narrowing its 2030E spread from +17.1 to +15.9 points — still the widest of the three, but the gap to Amazon and Alphabet is smaller than an earlier version of this analysis showed. Alphabet's WACC rose from 9.18% to 10.17% on updated market capitalization, converging with Amazon's figure. Because betas and market values move continuously, readers checking these figures against a live data source at a later date should expect some drift from the values used here; the specific levels are dated to early August 2026 and are not held current on any ongoing basis.
Amazon's projected capex-productivity figure depends materially on an unrealized margin assumption. The decomposition in the capex-productivity section shows that Amazon's projected 11.6% rests on incremental NOPAT margin rising from a historical 17.7% to 21.4%, which follows from AWS growing faster than retail and this report's modeled AWS segment margin reaching 40.5% by 2030E. Holding that margin flat at the historical rate instead produces roughly 9.7% — below the convergence band and below both peers. Readers should treat Amazon's relative resilience on this measure as contingent on the mix shift continuing, not as an observed characteristic of its capital deployment. The same decomposition shows revenue generated per dollar of capex falling 27% at Amazon, 33% at Alphabet, and 51% at Microsoft, so the underlying deterioration in converting capital into revenue is common to all three regardless of what margins do.
Third-party research figures are cited from secondary reporting, not from the underlying notes. Institutional research from Morgan Stanley, Goldman Sachs, CreditSights, LPL, and Wells Fargo referenced in this report is licensed, subscription-only product that the author has not reviewed directly. Every such figure is taken from press and trade-media accounts of those notes, and in several cases different outlets report slightly different framings of the same research (Morgan Stanley's GenAI return work, for example, is reported both as a headline 25%–50% range and as three point estimates near 31%, 46%, and 25%). These figures should be treated as directionally indicative of what those firms published rather than as verified quotations of their work, and readers with access to the primary notes should rely on those instead.
Analyst price targets and consensus averages are aggregator-computed and vary by provider. The average and high price targets cited in this report come from public aggregation services (TipRanks, Benzinga, and similar), which differ in which analysts they include, how far back they look, and how quickly they reflect revisions. A consensus figure drawn from a subscription terminal such as FactSet or Bloomberg may differ. Price targets also change continuously and are stated here only as of the dates noted.
The Wall Street consensus figures cited earlier in this report cover a different company set than this report — typically four or five hyperscalers including Meta and sometimes Oracle or SpaceX, against this report's three. The comparison is included to show direction and scale, not as a precise apples‑to‑apples benchmark; see the Scenario Range and Wall Street Context section for the adjustment attempted and its limits.
What This Means
Run through a disciplined ROIC/WACC test, the AI capex buildout at Amazon, Microsoft, and Alphabet is not, on the evidence in these three companies' own base‑case models, destroying value — every year of this study clears each company's own cost of capital, often by a wide margin. But "still profitable" and "as profitable as it used to be" are different findings, and this report's second and, in this author's view, more important result is that they've come apart: cumulative capex productivity converges into a narrow 11.6%–13.0% band across all three companies by 2026E–2030E — down from a historical range of 13.2% to 28.3% — with three separately‑built models arriving at strikingly similar endpoints. Microsoft and Alphabet roughly halve; Amazon, which already sat at the low end of that historical range, declines only modestly. That compression is the more useful thing to watch going forward than the capex headline number itself — and it's a number this report can actually track, quarter by quarter, as each company reports.
The most recent evidence available as this report goes to publication sits at the less severe end of that compression range, though it covers a single reporting cycle. All three companies' Q2/Q4 FY2026 results — the same results this report's base case is built from — moved their stocks by unusually large amounts: Amazon's and Microsoft's largest earnings-day gains in years, and, in Alphabet's case, a sharp initial decline followed by a full recovery once the same margin and backlog evidence arrived from its peers. All three moves were driven by the same underlying question this report's framework treats as decisive — is the capital converting to margin and backlog, not just revenue — and Wall Street's own price-target revisions moved by a similar magnitude at all three companies.
As the inflection discussion above sets out, that evidence is stronger than a single good quarter: the acceleration is sequential across four to five consecutive periods, contracted backlog is outgrowing revenue at all three companies, segment margins are expanding rather than compressing during the heaviest capex phase, and all three managements independently describe supply rather than demand as the binding constraint. Any one of those could be noise. Together, at three separate companies, in the same reporting cycle, they warrant more than dismissal. How much weight they deserve against a multi-year capital cycle is a judgment on which reasonable analysts will differ, and this report's own weighting is set out below.
This report nonetheless keeps the compression case as its base case, for one principal reason: the unresolved question of whether GPU and server depreciation schedules are too generous cuts directly at the margin evidence the inflection case depends on. If depreciation is materially understated, then reported segment margins and reported NOPAT are both overstated, and the most persuasive pillar of the inflection argument weakens considerably. That question is live, contested, and not resolvable from current disclosure. What will resolve it is specific and observable: whether backlog keeps outgrowing revenue over the next two to three reporting periods, and whether segment margins hold as depreciation from the 2026 capex wave begins hitting the P&L in earnest. Those two indicators, more than any capex headline, are what this framework will be tracking.
Coming Next
A natural next installment would extend this report's cumulative‑capex‑productivity test with proper time‑lagging (construction‑to‑revenue ramp, discussed in Limitations), and would fully rebuild the ROIC series from earlier in this report under each company's own bull and bear scenarios rather than the operating‑margin proxy used in the Scenarios and Wall Street Context section. A second, arguably more consequential extension: re‑running this report's NOPAT and ROIC figures under a shorter, Burry‑style GPU/server useful‑life assumption in place of each company's own disclosed schedule, to size directly how much of this report's ROIC compression is economic versus an artifact of current depreciation policy. A further extension worth flagging: applying this same framework to Meta and Oracle, the two other companies most frequently named alongside AWS, Azure, and Google Cloud in current AI‑infrastructure capex commentary, neither of which is covered in this report.
Sources & Citations
All figures in this report trace to the sources below or to the author's own financial models, which are themselves built from company SEC filings and earnings releases. Chart‑specific and table‑specific notes appear as captions throughout this report; this section consolidates the full source list.
Amazon.com, Inc. (AMZN): Form 10‑K filings (SEC EDGAR); Q2 2026 earnings release and supplemental financial information (July 30, 2026); author's AMZN financial model, refreshed August 2, 2026.
Microsoft Corporation (MSFT): Form 10‑K filings (SEC EDGAR), including FY2021 Form 8-K comparative financial statements; FY2026 Q4 earnings release (July 29, 2026); author's MSFT financial model, refreshed August 1, 2026, with balance sheet figures (FY2021A–FY2026A) cross-checked against StockAnalysis.com/Fiscal.ai standardized financials and CliffsNotes financial-statement compilation (accessed Aug. 2026).
Alphabet Inc. (GOOG/GOOGL): Form 10‑K filings (SEC EDGAR), including segment revenue and operating income disclosures for Google Cloud, Google Services, and Other Bets; Q1 2026 and Q2 2026 earnings releases; author's Alphabet financial model, refreshed August 2, 2026.
Prior reports in this series: Cap Rates vs. Cost of Capital, for the ROIC/WACC/EVA framework applied here.
Individual-company 2027E capex estimates (Scenarios and Wall Street Context section): Evercore ISI research (Mark Mahaney), as reported via MEXC/crypto.news (Mar. 20, 2026); Yahoo Finance, "Alphabet Just Tied Amazon's $200 Billion Capex Guidance" (Jul. 2026), citing BNP Paribas (Stefan Slowinski) on Microsoft FY2027 and Morgan Stanley (Brian Nowak, via CNBC "Power Lunch") on Alphabet 2027E.
Comparison to other published ROIC/ROIIC estimates (Scenarios and Wall Street Context section): Morgan Stanley research by Brian Nowak, Stephen C. Byrd, and Adam Wood (Jul. 27, 2026), as reported by Seeking Alpha (Jul. 28, 2026), Yahoo Finance/Stocktwits, and BigGo Finance; the author has not reviewed the underlying Morgan Stanley note, which is a licensed institutional product; LPL Research, "Can Hyperscalers Earn Their AI Ambitions?," syndicated via Advisor Perspectives and AdvisorAnalyst.com (Jul. 2026); Convequity, "Notes: AI Bubble? Refining the Forward ROIC" (Nov. 21, 2025, independent analysis, not affiliated with the author).
Wall Street capex consensus (Scenarios and Wall Street Context section): Goldman Sachs Research, "Why AI Companies May Invest More than $500 Billion in 2026" (Dec. 18, 2025) and subsequent 2027E commentary (June 2026); CreditSights, "Technology: Hyperscaler Capex 2026 Estimates" (Nov. 2025); Morgan Stanley hyperscaler capex model revisions (2027E ~$1.2T, 2028E ~$1.4T), as reported by Seeking Alpha (Jul. 28, 2026) and BigGo Finance (Jul. 2026); Wells Fargo hyperscaler compute‑capacity research, as reported via Yahoo Finance (Feb. 2026); Moody's hyperscaler capex and credit commentary, as reported via Data Center Dynamics (Mar. 2026) and Fortune (Jul. 2026).
GPU/server depreciation debate (Scenarios and Wall Street Context section): National Law Review, "Artificial Intelligence GPU Depreciation Debate and Earnings Risk," for the accounting and disclosure framing; Michael Burry public commentary (X/Twitter, Nov. 2025), as reported via CNBC (Nov. 11, 2025). Burry's cumulative sector estimate is his own published opinion, is not audited or independently verified, and is cited here as one contested input to a live accounting debate rather than as a finding this report adopts.
Hyperscaler credit and financing (Scenarios and Wall Street Context section): CreditSights hyperscaler leverage data, as reported via MUFG Americas; Fitch Ratings commentary on 2026 corporate bond issuance (Jul. 2026).
August 3, 2026 dateline update (Scenarios and Wall Street Context section): Bloomberg, "Amazon Joins Elite List of Stocks to Top $3 Trillion in Value" (Aug. 3, 2026); Reuters, via Yahoo Finance, "Amazon enters $3 trillion club as AI, cloud growth power rally" (Aug. 3, 2026).
Market reaction and analyst price targets, Alphabet (Scenarios and Wall Street Context section): Investing.com, "Alphabet Q2 2026 slides" and earnings call transcript (Jul. 22, 2026); Yahoo Finance / IndMoney, "Why Did Google Stock Fall After Q2 Earnings" (Jul. 2026); Yahoo Finance / TradingView, "GOOGL Stock Sinks 8% A Day After Strong Q2 Earnings" (Jul. 23, 2026); Benzinga analyst price-target aggregation (accessed Aug. 2026); TipRanks, "GOOGL Stock Price Forecast 2026" (Jul. 2026); CNN Markets and Yahoo Finance historical price data for GOOGL (accessed Aug. 1–2, 2026).
Market reaction and analyst price targets (Scenarios and Wall Street Context section; also cited in Key Findings): CNBC, "Amazon (AMZN) Q2 earnings report 2026" and "Analysts react to Microsoft's earnings results" (Jul. 30, 2026); Investing.com earnings call transcripts and analyst-ratings coverage (Amazon, Microsoft, Jul.–Aug. 2026); TipRanks, "Goldman Sachs Raises Microsoft Stock Price Target to $640" (Jul. 2026); Yahoo Finance / 24/7 Wall St., "Microsoft Surges 9% as Azure Tops $100B" (Jul. 30, 2026); TheStreet and AOL/Yahoo syndication of Morgan Stanley and JPMorgan Amazon research notes (Jul. 31–Aug. 1, 2026); StockAnalysis.com valuation and statistics pages for AMZN and MSFT (accessed Aug. 2, 2026); GuruFocus MSFT valuation page (accessed Aug. 1, 2026).
Glossary of Key Terms
- ROIC (Return on Invested Capital)
- NOPAT divided by invested capital. How much after‑tax operating profit a company generates each year relative to the capital tied up in the business.
- NOPAT (Net Operating Profit After Tax)
- Operating income (EBIT), tax‑effected, before financing costs — the numerator in this report's ROIC calculation.
- WACC (Weighted Average Cost of Capital)
- The blended cost of all capital financing a business — the interest rate on debt, weighted by the debt share of the capital structure, plus the return equity investors require, weighted by the equity share.
- EVA (Economic Value Added)
- ROIC minus WACC. A positive spread means the business earns more than its capital costs; a negative spread means it earns less, regardless of whether accounting profit is positive or growing.
- Invested Capital
- Total debt plus total equity, less cash and near‑cash balances not funding operations — the "excess cash" convention used throughout this report.
- Cumulative Capex Productivity (a gross-capex yield, not ROIIC)
- Change in NOPAT over a period, divided by cumulative capital expenditure over that same period — this report's rough proxy for return on incremental invested capital, not a precise, properly time‑lagged ROIIC calculation. See Limitations.
- Segment Operating Margin
- Operating income divided by revenue for a specific reported business segment (e.g., AWS, Google Cloud), as disclosed in a company's own segment reporting footnote.
- Backlog / Remaining Performance Obligations (RPO)
- Contracted future revenue not yet recognized — a forward‑looking demand indicator distinct from, and not directly comparable to, the capital‑efficiency metrics in this report.
Full Disclosure Statement
Not Investment, Financial, Tax, or Legal Advice
This report is published for informational and educational purposes only. Nothing in this report constitutes investment advice, financial advice, a securities recommendation, tax advice, legal advice, or an offer or solicitation to buy, sell, or hold any security or financial instrument of any kind. No advisor‑client, fiduciary, or professional‑services relationship of any kind is created between the author and any reader by virtue of publishing, reading, or relying on this report. Readers should not rely on this report as the basis for any investment decision regarding Amazon, Microsoft, Alphabet, or any other security. Any such decision should be made solely in consultation with qualified, licensed financial, tax, and legal advisors, based on the reader's own individual circumstances.
General and Impersonal Commentary
This report is written and published for general circulation to readers of gregg-carlson.com and is not prepared for, directed to, or tailored to the financial situation, objectives, or risk tolerance of any specific individual or entity. It does not recommend any specific transaction in any security and is not furnished in connection with, or as consideration for, any securities transaction by the author or any other party. The author receives no compensation from any party tied to any transaction in AMZN, MSFT, or GOOGL in connection with this report, and this report is not published in exchange for, or as promotion of, any securities transaction. This report is offered as bona fide financial and economic commentary of the kind regularly published by the author, not as individualized advisory services. The author is not disinterested in the subject matter: personal long positions in all three companies discussed are disclosed below and should be read together with this section.
No Rated Equity Research on the Companies Discussed
The author does not publish rated equity research — including BUY/HOLD/SELL ratings, price targets, or investment theses — on Amazon (AMZN), Microsoft (MSFT), or Alphabet (GOOGL). This report contains no rating, no price target, and no valuation opinion on any of these three companies, and none should be inferred from any figure, comparison, or observation in it. Where this report cites price targets or ratings, they are those of third-party sell-side firms, attributed as such and reproduced to document the range of professional opinion at a point in time — not adopted, endorsed, or independently verified by the author. This report belongs to the author's CFO Insights series of general financial and educational commentary and should be read only in that capacity.
Third-Party Commentary and Quotations
This report quotes and summarizes published commentary from sell-side analysts, financial journalists, and other third parties, including statements characterizing specific stock reactions (for example, one firm's description of a post-earnings decline as a "buying opportunity"). Every such statement reflects the view of the third party attributed and is reproduced solely to document the range of professional opinion at a point in time. Quotation or summary of a third party's statement in this report is not an adoption, endorsement, verification, or restatement of that view by the author, and should not be read as this report's own opinion or recommendation.
Personal Investment Position — Material Conflict of Interest
The author holds personal long positions in Amazon (AMZN), Microsoft (MSFT), and Alphabet (GOOGL) as of the publication date of this report. These are the three companies analyzed throughout this report. This is a direct financial interest in the subject matter and readers should treat it as a material conflict of interest when evaluating every figure, comparison, framing choice, and conclusion presented here — including the selection of which metrics to emphasize, which scenarios to model, and how the evidence is characterized. The author has made no commitment to hold, add to, or dispose of any of these positions, may transact in any of them at any time without notice, and undertakes no obligation to update this report or this disclosure to reflect any such change. The author also holds a personal long position in Red Rock Resorts, Inc. (NASDAQ: RRR), a company unrelated to and not discussed in this report, disclosed here as a standing general practice across this author's published work.
Business-Relationship Disclosure
The author's independent advisory practice serves clients across technology, gaming and hospitality, real estate, cannabis, and family‑office contexts, among other industries. This report is published on the author's own website, which also describes that practice, and the author therefore has a general interest in the report being read and well regarded. That interest is disclosed here so readers can weigh it. It does not extend to any transaction in any security discussed: the author is not compensated by any party in connection with this report, does not solicit securities business, and is not registered as an investment adviser or broker-dealer.
Model-Based Analysis — Not Independently Audited
All ROIC, WACC, EVA, and cumulative‑capex‑productivity figures in this report are the author's own calculations, derived from the author's own three‑statement financial models for AMZN, MSFT, and GOOGL, which are themselves built from public company filings and management guidance. These models have not been independently audited or verified by any third party, involve numerous modeling judgments and simplifications disclosed in the Limitations section above, and should not be treated as equivalent to each company's own reported financial statements.
Professional Credentials — Scope Limitation
The author holds a Certified Public Accountant (CPA) credential that is currently inactive in the State of Nevada. This report does not constitute the practice of public accounting, auditing, assurance, or any regulated professional service. The author is not registered with the SEC or any state securities regulator as an investment adviser or broker‑dealer.
Use of AI Tools in Research, Analysis, and Report Preparation
The author used Claude (Anthropic) throughout the preparation of this report — to extract and cross‑reference data from the underlying financial models and company filings referenced above, to construct the ROIC/WACC/EVA calculations, charts, and tables, and to draft and revise the written analysis, with the resulting research, calculations, and drafting reviewed by the author at each stage. This is a third‑party tool; the author has no control over, and makes no representation regarding, its underlying accuracy, reliability, or fitness for any purpose, and AI‑generated or AI‑assisted output can contain errors that are not always apparent on review. Use of this tool does not constitute a warranty of accuracy for any figure, calculation, or conclusion in this report. Regardless of the tools used in its preparation, the author retains final responsibility for all research, analysis, calculations, and conclusions presented in this report.
Forward-Looking Statements and Speculation
Statements in this report regarding future capital expenditures, revenue growth, margins, and returns on capital are inherently forward‑looking and speculative, reflect the author's models and analytical judgment at the time of writing, and are subject to change without notice. No representation is made that any forecast, estimate, or projection in this report will prove accurate.
Information as of Stated Dates; No Obligation to Update
Stock prices, analyst price targets, ratings, and similar market data in this report are stated as of the specific dates noted in the text and change continuously; a figure that was current when written may be stale by the time it is read. Occasional dated updates (such as the entry in this report noting conditions as of August 3, 2026) reflect information available to the author as of that specific date only and do not represent a commitment to provide ongoing or periodic updates. The author undertakes no obligation to update, correct, or supplement this report to reflect subsequent events, new information, or changes in circumstances, except as the author may elect in its sole discretion.
Past Performance
Historical financial data presented in this report reflects past performance. Past performance is not indicative of future results.
No Liability; No Warranty
This report and all data, analysis, commentary, and opinions contained in it are provided "as is" and "as available," without warranty of any kind, express or implied. To the fullest extent permitted by applicable law, the author disclaims all liability for any direct, indirect, incidental, consequential, special, or punitive loss, damage, or expense arising directly or indirectly from any use of or reliance on the information, analysis, data, or opinions contained in this report.
Jurisdictional Restriction
This report is published from, and primarily intended for readers in, the United States and is not directed to any person in any jurisdiction where its publication or availability would be contrary to local law, regulation, or registration requirements. Readers outside the United States are responsible for informing themselves of, and complying with, any such restrictions that apply to them.
Copyright and Attribution
This report is the original work of Gregg Carlson and is published at gregg-carlson.com. All rights reserved. Factual data underlying this report's figures and charts is sourced from the third‑party filings named in Sources & Citations above and is not subject to the author's copyright. Chart design, analytical framework, written analysis, and all original content are the proprietary work of the author.
This Disclosure Statement was last updated August 2026. Readers accessing archived or redistributed versions of this report should confirm the current disclosure status at gregg-carlson.com.
I am a fractional CFO, Controller, and CPA (inactive, NV) with 25+ years of senior finance experience in gaming, cannabis, family offices, real estate, and institutional investment, and $700M+ in closed transactions. I provide institutional‑quality financial modeling — ROIC, WACC, EVA, and DCF analysis — to real estate, family office, and institutional clients. Learn more at gregg-carlson.com.
