Is AWS's $220B AI Buildout Earning Its Cost of Capital?

CFO Insights · Technology & Capital Markets

Is AWS's $220B AI Buildout Earning Its Cost of Capital?

A segment-level ROIC / WACC / EVA read on the AWS capex cycle, FY2026E–FY2030E — built from AWS's own 10-K disclosures, a live Bull / Base / Bear scenario model, and a reality check against Morgan Stanley's and Goldman Sachs' most recent estimates.

By Gregg Carlson — Fractional CFO & Controller · CPA (inactive, NV) · August 2026

$496B

AWS backlog,
Jun-26

2.9x

Backlog / ann.
revenue run-rate

10.4%

Estimated
WACC

3 of 3

Independent methods
clear the WACC hurdle

1.00x

Capital-discipline throttle
(never engages, Base Case)

Disclosure. This report is general financial and educational commentary, not investment advice, a rating, or a price target. Author holds a personal long position in AMZN, the company analyzed. All figures are drawn from a segment-level financial model built from Amazon's SEC filings (10-Ks through FY2025, 10-Q/8-K through Q2 2026) and company earnings-call commentary, current as of August 31, 2026. No BUY/HOLD/SELL rating or price target is expressed or implied.

AI disclosure. This report was developed through an iterative research collaboration with Claude (Anthropic). I originated the analytical questions addressed here — including framing AWS's capex cycle through a segment-level ROIC-versus-WACC lens, introducing the utilization- and yield-driven approach to testing whether incremental AI capacity is being absorbed at positive economics, and the concept of using backlog growth to validate, rather than assume, that capacity utilization is not speculative. Claude built the underlying models, formulas, and much of the drafting based on that direction. I reviewed, validated, and take responsibility for the data, assumptions, and conclusions presented here.

A note on how I use these tools. I am a professional user of AI in my analytical consulting practice, and this report reflects that broader pattern, not an isolated experiment: I have built and maintain multiple internally developed use cases beyond this report, and applying AI through disciplined, purpose-built workflows has materially advanced both my productivity and the depth of analysis I can deliver as a sole practitioner. That gain has not reduced how much AI capacity the work consumes — my own token usage has risen as these tools became more central to my practice, even as I've optimized their use. I mention this because it's the honest context for how a report like this gets built, not because I think it should change how you weigh the analysis itself.

ROIC · WACC · EVA

In This Report

1. 01The Bottom Line
2. 02The Backlog Signal
3. 03AWS Segment Economics, FY2018A–FY2025A
4. 04Three Paths for FY2026E–FY2030E
5. 05Does the Capex Clear the Cost of Capital?
6. 06Economic Value Added
7. 07The Tripwire Framework
8. 08Probability-Weighted Outcome
9. 09Reality-Checking Against the Sell-Side
10. 10AI vs. Non-AI Cloud
11. 11Capacity, Utilization & Yield
12. 12What Else Could Move This Analysis
13. 13The Verdict: Bubble, or Not?
14. 14Methodology & Sources
15. 15Footnotes
16. 16Glossary

The Bottom Line

AWS's FY2026E–FY2030E capex cycle looks demand-backed, not speculative — and the segment's own numbers, not just management's framing, say so.

Amazon has raised full-year 2026 capex guidance to $220B¹, with “nearly all” of the increase AWS-directed, against an AWS backlog (remaining performance obligations) that has gone from $156.6B in June 2024 to $496B in June 2026 — roughly 2.9x AWS's own annualized revenue run-rate.² That backlog is contracted revenue, not a demand forecast, which is why this report treats it as the load-bearing fact in the bull case.

The question this report answers is narrower and more mechanical than “is the AI buildout justified”: does AWS's own segment-level ROIC — computed from its disclosed operating income, PP&E, and depreciation — clear a reasonable cost of capital through FY2030E, under a range of growth, margin, and capex assumptions? The answer, worked through below, is yes in the Bull and Base cases, and only in the Bear case (a combination of AI pricing compression and a moderate overbuild) does the spread over WACC turn negative, and not until FY2027E.

The Backlog Signal

AWS backlog has grown faster than AWS revenue in five of the last six quarters — the opposite of what a business building ahead of demand would show.

AWS remaining performance obligations (RPO), selected periods, $B. Source: Amazon 10-Q/8-K filings²; Reuters, “Amazon lifts investment plans after strong cloud sales,” 7/30/26⁴.

At AWS's disclosed weighted-average contract life of roughly 3.9 years¹, the $496B backlog implies about $127B/year of contractually-locked revenue recognition — already close to all of FY2025A AWS revenue ($128.7B) — before a single new booking. That reframes the headline that backlog grew 36% quarter-over-quarter while revenue grew 37% year-over-year: new bookings are simply keeping pace with a rapidly expanding recognition base, not racing ahead of it.

AWS Segment Economics, FY2018A–FY2025A

The segment's own 10-K disclosures show a business whose capital base grew 72% in FY2025A while operating income grew 15% — the mechanical signature of a capex-first, revenue-catches-up-later cycle, not a margin problem.

AWS segment net sales ($B, bars) and operating margin (%, line), FY2018A–FY2025A. Source: Amazon 10-K Segment Information notes, FY2019–FY2025 filings.¹

AWS property & equipment, net (period-end) vs. net additions to PP&E, $B. Source: Amazon 10-K Segment Information notes.

AWS ROIC (NOPAT ÷ average PP&E, net; 24% assumed cash tax rate) vs. an estimated 10.4% WACC. WACC: CAPM build, 4.7% risk-free rate, 1.20 beta, 5.0% ERP, 5.0% pre-tax cost of debt — see Methodology.

Key read: ROIC on average PP&E compressed from ~33% (FY2024) to ~23% (FY2025) even as AWS operating income grew 15%, because AWS's net additions to PP&E jumped 81% to $96.5B. That is exactly the pattern Amazon's own shareholder letters describe for the first AWS buildout cycle (2018–2020): capex outruns revenue for a period, then capacity comes online and the ratio inverts. The FY2026E–FY2030E scenarios below test whether that inversion actually happens this time.

Three Paths for FY2026E–FY2030E — Plus What Wall Street Models

Bull, Base, and Bear scenarios carry independent revenue growth, margin, and capex paths; a fourth, non-weighted Street Case translates Morgan Stanley's and Goldman Sachs' own disclosed estimates into the same framework.

BULL — capacity constraints persist, pricing holds BASE — consensus-consistent growth, margin dips then recovers BEAR — AI pricing compression, capex pulled back FY2028E+ STREET — built from Morgan Stanley & Goldman Sachs estimates, not the author's own view

AWS net sales by scenario, $B. Base Case FY2026E of $168.0B is consensus-consistent with the ~$169B annualized run-rate AWS exited Q2 2026 at. The Street Case tracks closer to the Bull Case than the Base Case for most of the forecast window. Bull/Base/Bear assumptions are the author's own; Street Case is derived from sell-side notes — see table and Sheet 4 of the companion workbook.

Scenario	FY2026E rev. growth	FY2030E rev. growth	FY2030E op. margin	FY2026E–30E capex CAGR
Bull	35.0%	20.0%	43.0%	+13.1%
Base	30.5%	15.0%	39.0%	+8.1%
Bear	26.0%	10.0%	26.0%	−1.1%
Street	33.0%	20.0%	43.0%	+14.5%

Street Case revenue growth blends Goldman Sachs⁸ (33% FY26E, 35% FY27E, 29% four-year CAGR FY25E–FY29E) and Morgan Stanley⁷ (41% FY27E); capex blends Morgan Stanley's explicit FY27E/FY28E figures ($325B/$350B) with Goldman's implied figures (~$292B/$315B, back-solved from its $827B FY26E–FY28E cumulative estimate). Margin path is the author's own estimate — neither desk discloses a specific AWS margin forecast.

Does the Capex Clear the Cost of Capital?

AWS PP&E, net is rolled forward year-by-year (beginning balance + capex − D&A) to compute NOPAT ÷ average invested capital in each scenario, then compared against an estimated 10.4% WACC.

AWS ROIC by scenario vs. the estimated WACC hurdle, FY2026E–FY2030E. D&A is projected as 14–17% of beginning PP&E (scenario-dependent), calibrated to AWS's actual FY2025A D&A/average-PP&E ratio of 14.3%. Invested capital = AWS PP&E, net only (excludes AR and operating-lease ROU assets); NOPAT assumes a 24% cash tax rate.

The Bull and Base cases clear WACC in every modeled year, though the spread compresses steadily as the capital base keeps compounding — from an 8.7-point spread in the Bull Case's FY2026E to a 3.8-point spread by FY2030E. The Base Case spread narrows from 7.2 points to just 1.2 points by FY2030E: still positive, but a much less comfortable margin of safety. The Bear Case crosses below WACC in FY2027E and the gap widens to roughly 2.9 points of value destruction by FY2030E. (Modeled — PP&E roll-forward and ROIC by scenario, Sheet 5 of the companion workbook.)

The Street Case — higher capex than even the Bull Case, but also higher revenue growth — lands at 13.1% FY2030E ROIC, between the Base and Bull cases. The two effects roughly offset: taking Wall Street's own numbers at face value does not produce a worse ROIC outcome than this report's own framework already contemplated, provided the revenue call proves out alongside the capex.

Economic Value Added, FY2026E–FY2030E

EVA = (ROIC − WACC) × average invested capital — the dollar-denominated version of the same story.

Annual AWS EVA by scenario, $B. Cumulative FY2026E–FY2030E EVA: Bull $144.0B, Base $65.0B, Bear −$31.6B, Street $115.3B. EVA measures dollar value created (or destroyed) above the cost of capital charge on invested capital — not accounting profit. Modeled — Sheet 5 of the companion workbook.

Even the Bear Case does not show AWS becoming unprofitable at the segment level — NOPAT still grows from $40.7B to $54.5B FY2026E–FY2030E. What it shows is the spread over WACC going negative once the capital base outgrows earnings by enough: EVA turns negative in FY2028E and stays there, consistent with Risk #4 in the source framing for this thesis — AI compute pricing compression hitting NOPAT directly, independent of any change in demand volume. The Street Case shows no such air pocket — EVA stays positive and growing in every year, tracking closer to the Bull Case than the Base Case.

Real-Time Monitoring: The Tripwire Framework

The bearish case requires five conditions to hold together, not any one alone. As of the Q2 2026 print, four of five point the other way.

Condition	Threshold	Current (Q2'26)	Status
AWS revenue growth, YoY	< 25%	36.7%	OK
AWS operating margin	< 30%	39.4%	OK
Backlog growth, QoQ	< 15%	36.3%	OK
Total co. capex growth, YoY	> 30%	71.5%	TRIGGERED
Mgmt says capacity no longer constrained?	Yes	No	OK

All five conditions met (bearish signal)? Not triggered. Rapid capex growth alone is not a red flag — it is exactly what a demand-backed buildout looks like. It only becomes concerning alongside decelerating revenue, compressing margin, slowing backlog growth, and management signaling capacity has caught up with demand. None of the other four conditions are currently close to triggering.

Probability-Weighted Outcome

Weighting the three scenarios 60% Bull / 25% Base / 15% Bear produces a single blended answer to the question this report set out to test.

FY2030E ROIC by scenario and probability-weighted blend, vs. the estimated 10.4% WACC hurdle. The hatched bar is the Street Case, shown for reference only — it is not assigned a probability or included in the blend. Probability weights are the author's own judgment, not derived from market pricing or analyst consensus.

A probability-weighted FY2030E AWS ROIC of 12.5%, roughly 2.1 points above an estimated 10.4% WACC, is the single most decision-useful number in this analysis: it says the $220B FY2026E buildout is, on balance of evidence available today, being funded into a genuinely high-return asset base rather than a speculative one — while leaving real, quantified downside if AI compute pricing compresses faster than backlog conversion can offset it. The weighting itself is a judgment call, not a market-implied probability, and is worth revisiting every quarter against the tripwire dashboard above. The Street Case's 13.1% sits just above that probability-weighted figure — an independent, sell-side-derived data point that corroborates the blend rather than contradicting it. (Modeled — Sheet 6 of the companion workbook.)

Reality-Checking Against the Sell-Side

Morgan Stanley (Brian Nowak) and Goldman Sachs (Eric Sheridan) are, as of their most recent notes, modeling more AWS/total-company capex for FY2027E–FY2028E than this report's own Bull Case — and Goldman's revenue growth call sits almost exactly on this report's Bull Case, not its Base Case.

Total company capex, $B. Morgan Stanley (7/27/26 note)⁷: explicit FY2027E $325B, FY2028E $350B. Goldman Sachs (7/9/26 note)⁸: implied from a disclosed $827B FY2026E–FY2028E cumulative estimate, apportioned using Morgan Stanley's year-over-year shape as a guide. All figures are total-company capex, the basis both banks report on.

Both desks' capex estimates run above this report's Bull Case for FY2027E and FY2028E — Morgan Stanley by 15–23%, Goldman by a similar margin. Neither bank publishes an explicit AWS-segment WACC or a formal ROIC-vs-cost-of-capital spread; Morgan Stanley's closest analogue is a bottom-up, AI-specific unit-economics framework — three separate models estimating 25–50% incremental ROIC⁷ on hyperscaler GPU rental (~31%), proprietary-infrastructure model APIs (~46%), and third-party-infrastructure model APIs (~25%).

Why those numbers look so different from this report's own 11–19% range: they are answering a different question, not disagreeing with this one. Morgan Stanley's figures isolate the marginal return on new AI-specific capacity alone; this report's ROIC blends that new capacity with AWS's entire existing base of storage, databases, and legacy compute, which structurally pulls the blended figure below the return on the newest, highest-margin slice — the same reason a company's overall ROIC typically runs below its ROIIC. Timing compounds the gap: this report's "incremental ROIC" memo line (Sheet 5 of the companion workbook) divides one year's NOPAT growth by the full prior year's capex, effectively a one-year payback test, while AWS itself discloses a 6–24 month lag between capex and first billing and Jassy's own shareholder letter³ describes returns as "cumulatively quite attractive a couple years after being in service." Both figures can be correct at once.

Translating both desks' numbers directly into this report's framework (the Street Case, built and charted above) is the cleanest way to reconcile them: it shows that Wall Street's more aggressive capex path, combined with Wall Street's own more aggressive revenue call, does not produce a worse ROIC outcome than this report's Bull Case — it produces one nearly identical to it. Both Morgan Stanley (Overweight, $335 PT, bear case $215)⁷ and Goldman Sachs (Buy, $375 PT, raised from $335 on 8/10/26 after AWS's fastest growth in 18 quarters)⁸ remain bullish-rated, and Goldman's own language now explicitly ties its higher price target to "visible demand, bigger backlog, and better economics over time" — the same demand-backed-versus-speculative distinction this report's tripwire framework is built to test.

A different Goldman voice asks a different question

Sheridan's price target and capex estimates are not the only view inside Goldman Sachs worth weighing. Jim Covello, the firm's Head of Global Equity Research, has been one of Wall Street's most consistent AI skeptics since co-authoring "Gen AI: Too Much Spend, Too Little Benefit?" in June 2024, and published an update in mid-2026 maintaining the same position.²⁸ Covello's question is deliberately not company-specific: not "will AWS out-execute its hyperscaler peers," but "do enterprises make or save enough money implementing AI to justify the industry's aggregate spend at all." That is a different, broader question than anything else in this report addresses — this report's entire framework tests whether AWS earns an adequate return on its own capex assuming AI demand and pricing behave as modeled; Covello's framework questions whether the demand and pricing assumptions themselves are durable across the industry. Worth holding both: Goldman's own equity research (Sheridan) has gotten more bullish on AMZN specifically as 2026 progressed, while Goldman's own head of research remains skeptical of the AI capex cycle broadly — the same firm, reaching different conclusions depending on whether the question is "which company wins" or "does the spending pay off industry-wide."

AI vs. Non-AI Cloud: Where Does the Return Actually Sit?

Decomposing the Base Case into an AI pool and a Non-AI/legacy-cloud pool reverses the popular framing of where this cycle's risk lives.

Heavy caveat, stated up front: AWS does not disclose an AI/non-AI split of revenue, capex, or PP&E — and much of AWS's infrastructure (power, networking, some storage/compute) is genuinely shared between AI and non-AI workloads, unlike a single-purpose GPU cloud. Every figure in this section is the author's own estimate, triangulated from partial disclosures. Treat it as a stress-test of a tension in the bull thesis, not a forecast.

Amazon has disclosed AI-specific revenue run-rate figures: AI services crossed a $15B annualized run-rate in Q1 2026 (~10% of AWS revenue), and AI and custom silicon (Trainium/Graviton) each independently crossed $25B annualized by Q2 2026 (~15% of AWS revenue)⁹ — a share that roughly doubled in a single quarter. Extrapolating that trajectory, and applying Morgan Stanley's own 60–70% incremental-EBIT-margin framework for hyperscaler GPU rental economics to the AI-attributed revenue⁷, produces two very different pools:

AWS capex split into an estimated AI pool and Non-AI/legacy-cloud pool, Base Case, $B. AI capex share of AWS total: 55% (FY26E) rising to 70% (FY30E), below CreditSights'¹⁰ ~75% cross-industry 2026 estimate to reflect AWS's larger shared-infrastructure base.

Non-AI cloud capex holds roughly flat at $79–87B/year — consistent with steady-state maintenance and modest organic growth capex for a mature business — while AI capex nearly doubles, from $103B to $185B. That split alone is unremarkable. What it implies for returns is not:

AI pool and Non-AI/legacy-cloud pool ROIC, Base Case, vs. the estimated 10.4% WACC hurdle. Both reconcile exactly to the blended AWS figures (AI revenue + Non-AI revenue = total; AI op. income + Non-AI op. income = total, by construction). Non-AI operating income is the residual of total AWS operating income less AI operating income — it is not independently assumed.

AI pool ROIC rises from 12.0% (FY2026E) to 15.2% (FY2030E), comfortably above WACC throughout and actually improving with scale — the only sub-component of this entire model that gets better over the forecast rather than compressing. Non-AI/legacy-cloud pool ROIC falls sharply, from 22.3% to 7.0%, crossing below the 10.4% WACC hurdle around FY2029E–FY2030E. The blended AWS ROIC shown throughout the rest of this report (17.5%→11.6%, Base Case) is a weighted average of these two increasingly divergent trajectories — and it is the Non-AI line doing almost all of the compression work, not the AI line. (Modeled — Sheet 7 of the companion workbook.)

This is partly mechanical: fixing AI's incremental margin at 65% while total AWS margin only grows modestly forces the residual (non-AI) margin to absorb the difference, compressing from 30.9% to 17.7% by FY2030E. If AI's true incremental margin runs below Morgan Stanley's 60–70% range, or if total AWS margin expands faster than this report's Base Case assumes, the non-AI compression shown here would be less severe. But the mechanism itself is real and worth watching independent of the exact numbers: as AI captures a growing share of AWS's revenue and a disproportionate share of its margin dollars, the legacy cloud business is left carrying a shrinking share of profit against capex that, while flat in dollar terms, is not shrinking. The practical implication: the AI capex/ROIC debate that dominates the headlines may not be where this cycle's real balance-sheet risk sits. AWS's non-AI growth rate and non-AI margin trajectory — neither of which Amazon discloses separately today — may be worth watching at least as closely. One more risk this split doesn't capture: the 20% AI-pool D&A rate assumes AI hardware holds its assumed 5-year useful life, but faster hardware refresh cycles or more compute-efficient model architectures could shorten real-world useful life further, compressing AI-pool ROIC below what's shown above even if revenue and pricing hold up exactly as modeled.

Amazon's own framing of this risk, direct from the source: CFO Brian Olsavsky on the Q2 2026 earnings call (7/30/26) split the investment into two separate capital cycles. Data centers carry "30-plus-year useful lives,"³⁶ across which AWS expects "at least five to six generations of server economics" — with each successive generation's economics improving further, Olsavsky explained, since the upfront data-center investment doesn't need to be repeated. Servers and networking equipment break even in "a little less than three years," with meaningful free cash flow in the two to three years after that. That's a real, disclosed mechanism supporting this section's D&A assumptions³⁶ — and a genuine tension worth naming: reporting a few weeks after that call, some coverage highlighted³⁷ that a subset of Amazon's Trainium 2 AI chips were already being replaced after roughly 20 months in service, well inside the 5-year useful life this report (and Amazon's own stated policy) assumes for AI-specific hardware. Both things can be true: the 30-plus-year data-center shell is a real, durable asset, while the AI-specific silicon inside it may refresh faster than the blended assumption captures — which is exactly the AI-pool D&A risk named above, now with a concrete example attached to it. A third, independent data point leans the same direction as the Trainium 2 evidence, and it comes from how capital is actually priced rather than from commentary: Goldman Sachs describes the private credit market's own adaptation to GPU-backed lending as treating GPUs as "short-duration, rapidly depreciating assets" facing obsolescence risk that traditional collateral frameworks weren't built to underwrite — in practice, GPU-fleet financing now uses shorter loan tenors and embedded refresh provisions rather than treating GPUs as long-lived collateral.⁵⁰ That's a market actually pricing risk with real capital at stake, not a forecast, and it's a data point on the faster-depreciation side of this section's tension — set against Pilling's rising-old-chip-price argument on the other side. Both are real; this report doesn't resolve which dominates.

Capacity, Utilization & Yield: Is the Capex Actually Speculative?

A bottom-up cross-check, built from AWS's own disclosed gigawatt plan rather than a revenue growth-rate assumption — and an explicit, formula-driven test of when Amazon's own capital discipline would actually curtail spending.

The source analysis's central claim is that AWS's capex is demand-backed, not speculative, because backlog already covers years of contracted revenue. This section tests that claim mechanically rather than narratively: it builds AWS's disclosed GW capacity plan, applies evidence-based utilization and revenue-per-GW ("yield") assumptions, and defines a capital-discipline rule as a formula — not an assurance — for when Amazon's own stated behavior (curtail capex once utilization or yield turn speculative) would actually bind.

AWS power capacity, GW (year-end). Base Case splits Amazon's own "double total capacity by end of 2027" target evenly (8→16 GW); Street Case uses Morgan Stanley's (Nowak)¹⁴ explicit +6 GW (FY26E) / +8 GW (FY27E onward) estimates — which run ahead of Amazon's own guided pace, the same pattern found in the capex comparison earlier in this report. YE2025A anchor (~8 GW) reconciles Jassy's "3.9 GW added in 2025" and "double again by 2027" disclosures¹³ against public cross-industry capacity baselines; not an Amazon-confirmed total.

The demand side of this plan got a concrete, dated data point after this report's initial draft: Nvidia's Q2 FY2027 results (reported 8/26/26)³⁴ showed Data Center revenue of $89.0B, up 117% year-over-year, with management guiding to ~70% revenue growth for fiscal 2028 — a figure Nvidia explicitly frames as a supply-constrained outlook against demand growth "closer to 100%." Alongside those results, Nvidia and AWS announced AWS will deploy an additional 2 million Nvidia GPUs across 2027–2028 (Blackwell Ultra, Rubin, and Rubin Ultra)³⁵, on top of the 1 million GPUs already announced in March 2026 — a direct, chip-level corroboration of continued capacity expansion from AWS's largest compute supplier, independent of anything Amazon itself has disclosed.

Utilization: validated against backlog coverage, not assumed

FY2026E–FY2027E utilization is held near-full (95%) because Amazon states demand exceeds capacity through 2027 and already describes 2028 demand as "striking." Independent estimates of the broader industry-wide buildout disagree with each other on magnitude, which is itself informative — genuine uncertainty at the input level, not false precision. An earlier (June 2026) SemiAnalysis count put new data-center capacity energized at 20 GW in 2026, with ~30 GW more expected in 2027¹¹; the same analyst, Dylan Patel, gave notably higher figures directly in an August 2026 interview — incremental global compute additions of roughly 30 GW in 2026, 50 GW in 2027, 70 GW in 2028, and 90–100 GW in 2029⁴⁶, an accelerating flow, not a plateauing one. Goldman Sachs's own economists, in an April 2026 research note, describe global data-center capacity (a stock, not an annual flow) growing from 30 GW in 2019 to 57 GW in 2024, with a further 65 GW projected online by 2030⁴⁷ — implying roughly 122 GW of total global capacity by 2030, a meaningfully smaller number than Patel's flow figures would imply if summed over the same years. Some of the gap may be definitional (global compute vs. data-center capacity specifically; different base years and methodologies), but not obviously all of it, and this report cannot fully reconcile two credible, differently-sourced estimates that disagree this much — nor can it fully reconcile one analyst's own estimate roughly doubling within two months. The honest takeaway is that the pace of the buildout is a live, disputed question even among specialists, not a settled fact this report can cite with confidence in either direction. What sources agree on, and what actually matters more for the utilization assumption above, is pricing: rental prices for old (Hopper-generation) GPUs are rising rather than falling as newer chips arrive — the signature of persistent scarcity, not a glut — and Patel separately describes base compute pricing (roughly $10–15 per megawatt-year, in millions) as having already begun to "inflect up" toward levels he expects frontier labs to be paying by end-2027 ($50M+/megawatt-year, equivalent to $50B+/GW-year). FY2028E–FY2030E tapers modestly to 88%, on the assumption that industry-wide supply gradually narrows the gap — a reasonable middle-ground assumption given the buildout-pace uncertainty above, though the pricing evidence on its rising rather than stabilizing.

Backlog-coverage check: at 2.9x backlog-to-run-rate coverage (Jun-26) against a 1.5x minimum threshold for a "non-speculative" utilization assumption, this section's 95% FY2026E–FY2027E utilization is backlog-validated, not merely asserted — nearly three years of AWS's current revenue is already contracted, before counting any new bookings.

Is the shortage real, and how long might it last?

The utilization case above rests on AWS-specific and SemiAnalysis evidence. It's worth checking against the wider compute market, because "shortage" is really three separate bottlenecks with three different clocks, not one story with one end date.¹⁵ At the chip layer, TSMC's advanced-node (N3) capacity has run at or above 100% utilization since late 2025, with leadership reportedly describing demand as roughly 3x what the company can produce¹⁶; new fab capacity in Arizona and Kumamoto doesn't reach volume production until 2027–2028, and TSMC's own executives have said shortages will likely persist "until 2027 and potentially beyond."¹⁷ High-bandwidth memory is a tighter chokepoint still — all three major suppliers were reportedly sold out for 2026, with Synopsys's CEO putting the "crunch" at continuing through at least 2027¹⁸, since new memory capacity takes a minimum of two years to come online.

Power is the longest-duration constraint, and the one this report's GW framework is built around. J.P. Morgan Asset Management's¹⁹ read is the most precise available: a data center itself can be permitted and built in 2–3 years, but the power to run it can take 5–7 years for natural gas, 10+ years for nuclear, and grid interconnection queues alone are running 5+ years in many U.S. markets — a structurally longer bottleneck than the chip shortage this section's FY2028E–FY2030E utilization taper assumes will start closing the gap. Goldman Sachs's own research (Global Investment Research/SUSTAIN, February 2026) puts this even more starkly for the markets that matter most: grid interconnection queues in the highest-priority data-center markets stretch 8–12 years — long enough to span two to three full GPU hardware generations before a single connection clears.⁴⁸ Those queues aren't purely physical: permitting timelines and local opposition to data-center power and water usage have already delayed or blocked projects in several U.S. markets, adding regulatory friction on top of the engineering constraint. The same J.P. Morgan source reports that 92% of data-center capacity currently under construction industry-wide is already pre-leased — an industry-wide analog to AWS's own 2.9x backlog coverage figure above. A more recent (August 2026) argument from investor Gavin Baker²⁰ frames the shortage as persisting through 2028 specifically, citing sub-one-year data-center paybacks (Nebius has disclosed 9–10 month paybacks) and low current AI penetration (fewer than 10 million heavy AI users against 1.5 billion knowledge workers who haven't adopted these tools yet) — a bull argument, not a neutral one, but from a named source with checkable claims rather than sentiment.

Net effect on this section's own assumptions: if power genuinely is the binding constraint and runs on a 5–10 year timeline, the FY2028E–FY2030E utilization taper (95%→88%) and the yield-growth reversal (flat, then −1% to −2%/year) may both be closing the supply/demand gap earlier than the physical evidence supports — which would mean this section, like the top-down Base Case elsewhere in this report, is understating rather than overstating the outcome.

But this cuts both ways — a risk the utilization case doesn't address. A persistent shortage is bullish for the returns on whatever capacity gets built, but it says nothing about whether Amazon can build the planned capacity on schedule. This report's GW paths (Base Case: +4–5 GW/year; Street Case: +6–8 GW/year) treat capacity additions as a controllable input — Amazon choosing how fast to build. If power generation and grid interconnection are genuinely multi-year physical processes outside Amazon's control, "double capacity by 2027" is racing a constraint Amazon doesn't fully own, and realized GW growth could fall short of the modeled path for reasons that have nothing to do with demand or capital discipline. That would show up as lower absolute revenue than modeled, not as impaired ROIC on the capacity actually delivered — a build-execution risk, distinct from every risk this report's tripwire dashboard is built to catch, and worth tracking alongside it.

Yield: does incremental capacity clear its own cost of capital?

Assumed yield ($B revenue per utilized GW) vs. the breakeven yield required to clear the estimated 10.4% WACC on a $38B/GW capacity investment (Epoch AI's 2026 cost-per-GW estimate¹²) at a 37% reference margin. FY2026E yield anchored to AWS's actual Q2 2026 annualized revenue run-rate ($169B) over an estimated ~10 GW of capacity at that point in the year.

Assumed yield clears its breakeven by a 20–24% cushion in every year through FY2030E — incremental AWS capacity does not need heroic pricing power to earn its cost of capital under this section's assumptions. This is the sharpest single test of whether the buildout is speculative: a cushion below 1.0x would mean even fully-utilized capacity destroys value; a cushion of 1.2–1.24x means there is real margin for error.

The yield-decline assumption for FY2028E–FY2030E deserves harder scrutiny than a single sentence, because there is now peer-reviewed evidence that AI inference pricing is falling faster than headline numbers suggest. A 2026 quality-adjusted price index³¹ built from 21,024 posted-price observations across 3,208 models and 86 providers finds that inference prices, adjusted for the capability customers actually receive, have fallen at roughly 0.6–0.7 log points per year — the study estimates 87% of that decline is invisible to the naive, matched-model methods that statistical agencies apply to software pricing generally. Read narrowly, that supports this section's already-conservative yield-decline assumption. Read more broadly, the picture is not a simple one-way collapse: separate market tracking shows AI pricing has genuinely split into two tiers — commodity/mid-tier model pricing continuing to fall³² (~36% year-over-year in mid-2026), while frontier-tier pricing has actually risen roughly 100% since January 2026 as newer, more capable generations command premium pricing. And Gartner's own 2026 forecast³³, while projecting per-token costs to keep falling sharply through 2030, explicitly warns enterprises not to "confuse the deflation of commodity tokens with the democratization of frontier reasoning" — falling unit prices have driven higher, not lower, total inference spend industry-wide, since cheaper tokens unlock more agentic, higher-volume workloads (a Jevons Paradox dynamic). For AWS specifically, this cuts both ways: if AWS's AI revenue mix skews toward frontier and enterprise workloads rather than commodity inference, realized yield could hold up better than this section's flat assumption; if AWS increasingly competes on commodity capacity, the decline could be sharper than modeled.

Fresh, granular pricing evidence from SemiAnalysis's Dylan Patel points toward the optimistic side of that range, at least for frontier-lab demand. He puts current base compute pricing at roughly $10–15 million per megawatt-year ($10–15B/GW-year in this section's units — consistent with, not far from, this section's own $16.9–17.4B/GW yield assumption) and describes that price as having already begun to "inflect up," with a specific expectation that frontier-lab compute could reach $50 million+ per megawatt-year ($50B+/GW-year) by the end of 2027 — roughly triple this section's current assumption.⁴⁶ Two caveats matter before reading this as a reason to revise the model upward: first, this is one analyst's forward projection, not a disclosed transaction price, and Patel himself flags some of his adjacent estimates (interest rates, specifically) as "vibing a number" rather than a rigorously derived one; second, and more importantly, this pricing is specific to frontier AI-lab demand (OpenAI, Anthropic), not AWS's blended yield across its entire customer base (enterprise, non-AI workloads, smaller AI customers included) — the two are related but not identical, and this section's yield figure is necessarily a blend, not a frontier-lab price. Directionally, though, it reinforces the same conclusion as the GW-buildout revision above: this section's flat-to-declining yield path may be the conservative assumption, not the aggressive one.

The chip-generation curve: what actually drives the yield assumption

The yield-decline assumption above has a specific, nameable mechanism behind it, not just "competitive pressure" in the abstract. Nvidia's own disclosures quantify successive GPU generations' cost-per-token improvement directly: Blackwell Ultra delivers 50x higher throughput and 35x lower token cost versus Hopper; the newer Rubin platform delivers a further 10x reduction in cost per token versus Blackwell.⁴² AWS's own G7 (Blackwell-based) instances deliver 4.6x the inference performance of the prior G6 generation.⁴³ These are not small numbers, and AWS is deploying this exact roadmap: on top of the 1 million Nvidia GPUs (Blackwell/Rubin) already committed in March 2026, AWS and Nvidia announced in August 2026 a further 2 million GPUs (Blackwell Ultra, Rubin, Rubin Ultra) landing in 2027–2028.⁴⁴ SemiAnalysis's Dylan Patel independently corroborates the general magnitude, from the compute-supply side rather than Nvidia's own marketing materials: newer accelerators (GB300s, TPUv7s, Trainium3s) are running "3–5x more performance per watt than the prior-generation chips" in his account — a smaller multiple than Nvidia's own 10–50x cost-per-token figures (different metric, and Nvidia's own numbers naturally sit at the favorable end of the range), but an independent, non-vendor source confirming the direction and rough scale of the effect.⁴⁶

This cuts in two directions, and it matters which one dominates. If AWS captures the generational cost decline as margin — holding pricing roughly firm while its own cost per unit of compute falls — that is directly ROIC-positive and would make this section's AI-pool margin assumption (65% incremental margin, Sheet 7) conservative. But AWS does not have exclusive or even early access to this roadmap: Nvidia's own Rubin launch materials name Google Cloud, Microsoft, and OCI as deploying the identical platform on essentially the same timeline⁴², alongside neoclouds (CoreWeave, Lambda, Nebius, Nscale). A cost curve every major provider rides simultaneously is not a source of AWS-specific competitive advantage — it is an industry-wide shift that competition has historically passed through to customers as lower prices, which is precisely the mechanism behind the inference-pricing deflation documented above. Put plainly: the chip roadmap explains why yield decline is a reasonable assumption to model, not a reason to expect AWS's margin to structurally outperform what this section already assumes.

One nuance this section's framework doesn't capture and is worth naming: because later years' incremental GW capacity is built with progressively more efficient hardware (Blackwell → Blackwell Ultra → Rubin → Rubin Ultra), a gigawatt commissioned in FY2028 should generate more useful compute output than a gigawatt commissioned in FY2026, all else equal. This section's yield-per-GW assumption treats a GW as a homogeneous unit across the full forecast window; in reality, later-vintage capacity is likely more productive than earlier-vintage capacity purely from the hardware curve. That is a reason the bottom-up method's already-above-Base-Case revenue path (chart above) may have more legitimate upside room than modeled, not a reason to revise the yield assumption itself — this report does not have the fleet-mix data (what share of AWS capacity runs each chip generation, by year) needed to quantify the effect responsibly, so it is flagged as a directional consideration rather than built into any figure.

Corroborating evidence from the buy side, not just chip specs: a September 2026 interview with Daniel Pilling (co-PM, Sands Capital Global Growth Fund)⁴⁵ offers an independent read consistent with this section's shortage and yield-cushion evidence: GPU/server payback periods have compressed to roughly 12 months (from 12–18), and — the more unusual signal — rental prices have risen even for already-depreciated, older-generation chips, which Pilling argues is evidence against the market's fear that AI hardware depreciates faster than assumed, not for it. That sits in tension with the Trainium 2 replacement-cycle evidence earlier in this section; this report presents both rather than resolving the question, since the two data points concern different things (Amazon's own custom-silicon refresh decisions versus market-clearing prices for GPUs generally) and a single data point on either side is not conclusive.

A third, independent revenue cross-check

Bottom-up revenue (utilized GW × yield) vs. this report's own top-down scenarios, $B. Bottom-up method uses only the Base Case GW path; utilization and yield are this section's own assumptions.

The capacity-utilization-yield method runs above this report's top-down Base Case throughout — by 15% in FY2026E, widening to 30–31% by FY2028E–FY2030E — and above even the Bull Case by FY2029E–FY2030E. That is a third independent line of evidence, alongside the sell-side comparison earlier in this report, pointing the same direction: if Amazon executes its disclosed GW plan and utilization/pricing hold up anywhere near current levels, this report's own Base Case is more likely to prove conservative than aggressive. It is not proof — utilization and yield remain estimates — but three genuinely different methodologies converging on the same conclusion is meaningfully stronger evidence than any one of them alone.

The capital-discipline rule: when would Amazon actually curtail?

Amazon's own framing is that capex is customer-committed and predictable, not a blind bet — and that the company has "been through this cycle" before and "liked the results." The implicit promise is that Amazon would slow spending if utilization or yield stopped clearing a reasonable return. This report turns that into an explicit rule: full planned capex growth is allowed at 85%+ utilization; below 70%, the rule allows no further capex growth until utilization recovers; between the two, allowed growth tapers linearly.

Does the rule ever bind? Not under this report's own assumptions — the throttle factor holds at 1.00x (full growth allowed) in every year through FY2030E, because assumed utilization (88–95%) never approaches the 85% threshold where tapering would begin. Utilization would need to fall roughly 10 points from this report's own FY2026E assumption before the rule engaged at all — a large, sudden shock relative to 2.9x backlog coverage and SemiAnalysis's industry-wide evidence of demand still outrunning supply. The rule is not vacuous (it is calibrated well below the assumed utilization range specifically so it is a genuine test), but nothing in evidence today indicates it would bind.

Translating the bottom-up method into this report's own ROIC framework (same invested-capital base as the Base Case, so only the revenue/NOPAT side changes) produces a FY2030E ROIC of 15.1% — above the top-down Base Case's 11.6% and closer to the Bull Case's 14.2%, consistent with the higher bottom-up revenue. Three independently-built methods in this report now converge: the top-down scenarios, a direct translation of sell-side estimates, and this capacity/utilization/yield build all clear the estimated WACC hurdle, and the latter two both run at or above this report's own Base Case. (Modeled — Sheet 8 of the companion workbook.)

What Else Could Move This Analysis

Three risks this report hasn't yet quantified — each capable of changing the verdict on its own — plus two more worth a paragraph each.

1. Financing and leverage: is the balance sheet keeping pace with the buildout?

Amazon's debt has grown faster than almost any other figure in this report. Total face value of long-term debt went from $68.8B (Dec-25) to $122.6B (Mar-26) to $133.0B (Jun-26) — essentially doubling in two quarters — and on June 8, 2026 Amazon added a three-year, $17.5B senior unsecured Delayed Draw Term Loan facility on top of that. Fitch assigned the new debt an AA- rating in July 2026²¹, and Moody's affirmed Amazon's A1 senior unsecured rating with a stable outlook in February 2026²², despite the capex plan rising more than 50% — both solidly investment-grade and not deteriorating on Amazon specifically.

Metric	Dec-25	Mar-26	Jun-26
Total face value of long-term debt	$68.8B	$122.6B	$133.0B
Credit rating (Moody's / Fitch)	A1 stable (Moody's, 2/26) · AA- (Fitch, 7/26)

Per Amazon's Q1/Q2 2026 10-Qs, Note 5 — Debt: total face value of long-term debt (all unsecured notes plus other long-term debt) of $68,836M (Dec-25), $122,632M (Mar-26), and $132,995M (Jun-26). This is very slightly broader than the "unsecured senior notes outstanding" figure Amazon cites in its own prose ($68.0B/$121.8B/$132.1B for the same dates) — the gap is under $1B at any point and doesn't change the trajectory, but this report uses the fully-inclusive figure consistently across all three dates.

But the Amazon-specific picture isn't the whole picture. Moody's flagged the sector on 7/24/26²³: combined capex across the six hyperscalers/AI infrastructure names it tracks (Microsoft, Amazon, Alphabet, Meta, Oracle, CoreWeave) is projected to hit $785B in 2026 and roughly $1 trillion in 2027, forcing even the most cash-rich companies to lean on debt, off-balance-sheet leases, and equity issuance in a way that "threatens credit quality" across the group. The off-balance-sheet piece is the one this report's WACC calculation doesn't capture: Moody's puts combined lease commitments across the six names at $1.2 trillion, with more than $820B of that from leases that haven't even started yet. This report's WACC uses Amazon's $133.0B of on-balance-sheet debt as its leverage input; if long-term data-center lease commitments were added, Amazon's true fixed-obligation load — and potentially its cost of debt — would be higher than modeled here.

A broader, more recent tally sharpens this further: a Wall Street Journal analysis published 8/17/26²⁹ found nine major technology companies (adding chipmakers and SpaceX to the hyperscaler group above) carrying roughly $3 trillion in combined off-balance-sheet commitments — about $1.2T in signed-but-not-yet-started leases and $1.9T in purchase commitments for chips, memory, and networking gear — against roughly $600B of actual trailing-twelve-month capex across the same group, meaning the contracted future obligation is roughly five times the current run-rate of recognized spending. Separately, Morgan Stanley's own research (distinct from the WSJ tally, using a narrower seven-company hyperscaler-plus-chipmaker sample) put combined purchase and lease commitments at roughly $2.8T as of late August 2026³⁰, and its "Mapping AI's Circularity" research separately projects industry-wide AI spending approaching $3T by 2028. These are related but not identical figures from different analyses; what they agree on is the direction — reported capex meaningfully understates the total forward financial commitment being taken on across the industry, Amazon included. A third, still larger and longer-dated estimate comes from SemiAnalysis's own modeling: Dylan Patel puts total global AI infrastructure capex at roughly $11 trillion cumulative from 2024 through 2029, of which about $6T is expected to be funded from cash flow and roughly $5 trillion from debt issuance across the ecosystem.⁴⁶ That figure spans a longer window and a broader definition of "AI infrastructure capex" than either the WSJ or Morgan Stanley tallies above, so it isn't directly additive to them — but it is a third, independently-derived estimate landing in the same range and pointing the same direction: whatever the precise number, the industry-wide financing need this decade is measured in trillions, not hundreds of billions, and a meaningful share of it is expected to be debt-funded rather than self-funded from operating cash flow.

That debt has to go somewhere, and where it goes is itself a risk this report hasn't priced. Goldman Sachs's own capital-markets research puts AI-related debt issuance from hyperscalers and AI infrastructure companies at roughly 6.6% of the entire $1.83 trillion U.S. investment-grade bond market in 2025, projected to rise toward 20% in 2026 as unsecured and project-finance issuance combine.⁴⁹ That is a distinct mechanism from either the Fed-policy rate channel or Patel's broader credit-crowding-out argument above: institutional fixed-income portfolios carry issuer- and sector-concentration limits that were never calibrated for one thematic category approaching a fifth of the entire index. As those limits bind, the practical effect is to push incremental hyperscaler issuance — Amazon's included — into non-USD currencies and private credit markets, potentially at wider spreads than the benchmark IG market would otherwise demand, regardless of Amazon's own standalone credit quality. This doesn't change this report's WACC estimate, which already uses Amazon's actual disclosed debt terms, but it is a specific, checkable reason the cost of Amazon's future debt issuance could run higher than its cost of debt today, independent of the rate and credit-spread channels already discussed.

The free cash flow side of this, stated plainly: Fitch's own estimate²¹ puts Amazon's free cash flow at roughly negative $40B annually in both 2026 and 2027. ROIC measures whether capital earns an adequate return over its life — it says nothing about whether the company is comfortable, cash-wise, in the years before that return shows up. Both things can be true at once: attractive multi-year ROIC, and real near-term cash strain funded by debt markets that remain open and investment-grade today, but aren't guaranteed to stay that way if the pace of issuance continues.

Amazon is not the most exposed hyperscaler on this dimension, and the contrast is worth naming rather than leaving implicit. Oracle³⁸ — the hyperscaler most frequently cited as vulnerable — carries roughly $125B in debt against a BBB credit rating (well below Amazon's AA-/A1) and a debt-to-equity ratio near 500%, funding an AI buildout roughly a quarter of Amazon's scale (~$55.7B FY2026 capex, rising to ~$70B FY2027)⁴⁰ almost entirely through debt rather than internally generated cash flow; one Wall Street estimate has flagged³⁹ Oracle could face a financing gap as soon as November 2026 absent additional capital raises. Microsoft, by contrast, generated ~$68B of free cash flow in FY2026 despite ~$116B of capex⁴¹ — comfortably positive, unlike Amazon's negative estimate above. Even Google, historically the most cash-comfortable of the group, posted its first-ever negative quarterly free cash flow (-$5.9B) in Q2 2026 after raising its own FY2026 capex guidance to $195–205B, up from $180–190B just one quarter earlier⁵² — a reminder that this isn't an Amazon-specific or even an Amazon-and-Oracle-specific pattern; it is reaching companies previously assumed to be immune to it. Amazon sits between Oracle and Microsoft: not cash-flow-comfortable like Microsoft, but nowhere near as leveraged as Oracle. That relative positioning matters for how much weight the sector-wide Moody's warning above should carry for Amazon specifically — the risk is real, but it is not evenly distributed across the hyperscaler group, and Amazon's own scale and cash-generation capacity (retail, advertising, and AWS combined) give it more room to absorb a disappointing outcome than smaller, more leveraged peers.

2. Backlog concentration: how much rests on a handful of customers?

This report has treated the $496B backlog as a single, load-bearing fact. It isn't evenly distributed. Two named AI-lab customers alone account for well over a quarter of it at face value:

Customer	Announced	Total contract value	Term	Capacity secured
OpenAI24	11/3/25	$38B	7 years	Hundreds of thousands of GPUs, scaling toward tens of millions of CPUs
Anthropic25	4/20/26	$100B+	10 years	Up to 5 GW of Trainium/Graviton, current and future generations
Combined		$138B+

These are announced total contract values over multi-year terms, not confirmed dollar-for-dollar components of the $496B RPO figure — Amazon doesn't disclose how backlog maps to individual customers, so this is directional, not precise. But directionally, it's hard to read $138B+ from two customers against a $496B total (27.8% of it) and conclude concentration risk is immaterial. (Modeled — Sheet 9 of the companion workbook.) Worth noting too: Amazon's own equity stake in Anthropic has grown to roughly $33B cumulative (an $8B initial stake plus a new $25B commitment, $5B of it immediate) alongside Anthropic's $100B+ compute commitment — a structure where Amazon is simultaneously the capital provider and the compute vendor to the same counterparty. That isn't evidence of anything improper, and Anthropic's own disclosed revenue run-rate reportedly tripled²⁶ to roughly $30B over the past year, suggesting real, fast-growing revenue behind the commitment. But it is the kind of circular-financing structure that has drawn scrutiny industry-wide, and it means backlog quality, not just backlog size, is worth tracking — a single large customer's distress would show up as a backlog impairment this report's framework has no mechanism to flag.

3. WACC sensitivity: how much room is there before the verdict flips?

Every scenario in this report holds WACC fixed at the estimated 10.4%. Given the Verdict below finds the spread over WACC compressing toward zero in every scenario, the natural next question is what happens if the hurdle itself moves — from rate changes, credit-spread widening (see above), or a re-rating of AI-linked risk generally.

Scenario	WACC 8%	WACC 9%	WACC 10.4% (base)	WACC 11%	WACC 12%	WACC 13%
Bull	+6.2%	+5.2%	+3.8%	+3.2%	+2.2%	+1.2%
Street	+5.1%	+4.1%	+2.7%	+2.1%	+1.1%	+0.1%
Base	+3.6%	+2.6%	+1.2%	+0.6%	−0.4%	−1.4%
Bear	−0.5%	−1.5%	−2.9%	−3.5%	−4.5%	−5.5%

The Base Case has less room than it looks like. FY2030E ROIC by scenario — Bull 14.2%, Street 13.1%, Base 11.6%, Bear 7.5% — is also each scenario's breakeven WACC. The Base Case needs WACC to rise only 1.2 percentage points from today's 10.4% estimate to flip its FY2030E conclusion from value-creating to value-destructive. Given this report's own WACC rests on a placeholder beta and a risk-free rate that moves with Fed policy and Treasury issuance, a 1.2-point move is not a remote scenario — it is well within the range these inputs have moved over shorter periods historically. The Bull and Street cases have considerably more room (3.8 and 2.8 points respectively); the Bear Case is already past its breakeven at today's WACC. (Every cell in the table above is a live formula linked to this report's actual WACC and scenario ROIC cells — Sheet 9 of the companion workbook.)

The rate-path discussion above treats moves in the 10-year Treasury as the primary channel — but SemiAnalysis's Dylan Patel raises a distinct, credit-market-specific mechanism worth naming alongside it: if AI-linked capex issuance absorbs enough of the credit market's capacity (his own estimate, above, puts industry-wide AI debt issuance at roughly $5 trillion through 2029), the resulting crowding-out could push corporate borrowing costs up independent of Fed policy or Treasury yields — he specifically floats Amazon paying something like 8% on new debt versus the 5–6% peers have recently priced.⁴⁶ Mechanically, this section has already shown that a change in Amazon's own cost of debt moves WACC only weakly (Amazon's capital structure is roughly 95.5% equity-weighted, so even a 100bp move in cost of debt shifts WACC by only a few basis points) — so a credit-spread-driven channel, on its own, would need to be far larger than the rate-driven channel to matter as much. Where this argument has more force is as a second, correlated driver of the same risk already identified above: if AI-linked credit demand pushes benchmark rates up broadly (not just Amazon's own spread), that operates through the equity-side risk-free rate channel already shown to move WACC at close to full strength — making this less a new, independent risk than a plausible amplifier of the one already quantified.

Three more worth a paragraph each. This report analyzes AWS in isolation, without reference to Azure, Google Cloud, or the neoclouds (CoreWeave and similar) — Moody's $785B 2026 / ~$1T 2027 industry-wide capex figure above spans six companies, of which AWS is one. "Demand exceeds AWS's capacity" is consistent with total AI compute demand growing, but it doesn't rule out AWS losing relative share to competitors building or pricing more aggressively — and that isn't a hypothetical gap in this report's coverage, it's visible in the most recent quarter. Google Cloud revenue grew 82% year-over-year to $24.8B in Q2 2026 (versus AWS's 37% and Azure's 43%), and Google Cloud's own backlog grew from $462B to $514B over the same quarter — now larger than AWS's own $496B backlog, from a smaller revenue base, growing faster.⁵² None of this weakens this report's core finding: AWS clearing its own cost of capital and AWS losing relative share to Google Cloud are logically independent questions, and this report only answers the first one. But "demand-backed" should not be read as "AWS-specific moat" — the same demand growth this report treats as validating AWS's capex could, in the scenario where Google (or another competitor) captures a disproportionate share of it, still leave AWS's absolute growth rate lower than modeled here without AWS's ROIC math being wrong in any way. This report has no relative-positioning view, and that gap matters more given this data point than it did in the abstract. Separately, this report tests whether AWS's capex clears AWS's cost of capital — it does not test whether the customers paying AWS have business models that can sustain those payments. Anthropic's revenue growth (above) is a reassuring data point for one large customer; a rough back-of-envelope check offered independently by Sands Capital's Daniel Pilling⁴⁵ (backing out revenue-per-gigawatt from Anthropic's public disclosures at roughly $20–30B/GW, against its disclosed forward gigawatt commitments) implies a revenue base broadly consistent with, not obviously unable to support, its $100B+ AWS commitment — corroborating, not proving, the scale is realistic. OpenAI is a starker case, with total infrastructure commitments across all its vendors reported near $1.4 trillion against a revenue run-rate closer to $10–13B annually.²⁷ A customer's inability to pay would show up in this report's framework as a backlog or revenue miss, with no earlier warning — the tripwire dashboard catches AWS-side deterioration, not customer-side distress. It's also worth naming the counter-argument directly, since it complicates rather than resolves this concern: Pilling⁴⁵ frames enterprise AI adoption as closer to early-1900s electrification than to a normal ROI-justified purchase decision — a competitive-necessity dynamic where businesses adopt because falling behind peers who do is the larger risk, independent of whether any individual customer can cleanly prove a positive return. If that framing is right, durable demand may not require the customer-level economics this paragraph worries about to work out cleanly for any single AI lab — though it would still leave the concentration risk above intact if a specific large customer stumbles regardless of the category's overall durability.

The version of this risk worth naming as its own point, not just an extension of the Anthropic and OpenAI discussion above: the customer-credit analysis this report has done so far is idiosyncratic, name-by-name. The more consequential version is systemic. AI-lab funding has so far come disproportionately from a wave of large, correlated capital events — mega-rounds, rumored IPOs, vendor-financed compute deals — rather than from broad-based, diversified revenue. If AI-lab funding conditions cooled broadly (venture and private-equity AI allocations slowing at the same time, a high-profile IPO disappointing or being pulled), multiple large AWS customers could face funding pressure simultaneously and in a correlated way, not one at a time. Neither this report's tripwire dashboard nor its backlog-concentration table is built to catch that pattern — both are constructed to flag AWS-side deterioration or a single named customer's distress, not a correlated, ecosystem-wide funding contraction that could hit several large customers' ability to convert backlog into paid revenue in the same window. This is a real gap in coverage, not a prediction that it will happen.

The Verdict: Bubble, or Not?

Not a bubble in the sense of capex chasing demand that isn't there — but "not a bubble" is not the same as "riskless," and the real risk sits somewhere more specific than the headline debate.

Direct answer: across four scenarios and three independently-built methodologies, AWS's ROIC clears its estimated cost of capital in every case except one narrowly-defined downside scenario — and the mechanism that would actually signal a bubble forming (backlog decoupling from capacity, utilization collapsing, capex outrunning contracted demand) is not showing up in the data as of the most recent quarter.

Why the evidence points away from "bubble"

• The backlog math is the load-bearing fact, and it holds up. $496B in contracted revenue (Jun-26), up from $156.6B two years earlier, growing faster than revenue in five of the last six quarters — the opposite signature of speculative building.
• AWS has never failed to clear its cost of capital since 2019 — even as ROIC has compressed from the 33–37% range to ~23% in FY2025A, it remains more than double the ~10.4% WACC estimate.
• Three genuinely different methods converge on the same answer: the top-down scenarios (Base 11.6%, Bull 14.2%, Street 13.1% all clear WACC at FY2030E), a direct translation of sell-side estimates that run more aggressive than even this report's Bull Case on both capex and revenue, and a bottom-up capacity/utilization/yield build grounded in AWS's disclosed GW plan and SemiAnalysis's supply/demand evidence — all three land at or above this report's own Base Case.
• The capital-discipline check never binds. Formalizing "Amazon would pull back if utilization or yield turned speculative" into an actual formula, current evidence sits well clear of the threshold where that rule would even start tapering capex.

One systemic data point worth holding alongside the AWS-specific evidence above, from a source with no AMZN-specific position to defend: Goldman Sachs's own economists estimate current U.S. AI investment at roughly 1.2% of annual GDP — and even under the most optimistic projections for continued growth, expect it to remain well below the 3–4.5% of GDP the railroad buildouts of the 1800s reached at their peak.⁵¹ That doesn't settle the AWS-specific ROIC question this report exists to answer, and a systemic comparison says nothing about whether any single company's capital is well spent — but it is a useful sanity check against the framing that this buildout is unprecedented in scale relative to the economy as a whole. By this one historical yardstick, it isn't, at least not yet.

Worth noting as one more corroborating line, not a fourth methodology this report built itself: an independent capex/backlog analysis prepared outside this report's own framework reaches a similar probabilistic conclusion by a different route — assigning roughly an 80% probability that Amazon's AI/AWS investment ultimately earns above its cost of capital, versus roughly 10–15% for material value destruction. That is directionally consistent with, though more optimistic than, this report's own probability-weighted FY2030E outcome (12.5% ROIC against an estimated 10.4% WACC, roughly a 2.1-point spread) — two independently-constructed probability assessments landing on the same side of the question, even if not the same magnitude of confidence.

Where the real risk actually sits

ROIC minus WACC, by scenario, FY2026E–FY2030E. The spread compresses in every single scenario, every year, with no exception — only the rate of compression differs. Bear Case crosses into value-destructive territory (spread < 0) starting FY2027E.

Several findings complicate a simple "no bubble" answer and deserve equal weight against the reassuring evidence above. First, the spread over WACC is compressing in every single scenario, every year, with no exception. Base Case spread narrows from 7.2 points (FY2026E) to just 1.2 points by FY2030E. Bull compresses from 8.7 to 3.8; even Street, despite higher revenue, compresses from 8.5 to 2.8. Nothing in this model shows the spread stabilizing or reversing — only the rate of compression slowing, which is a different thing from the risk going away.

Second, the Bear Case is a specific, trackable scenario, not a tail risk included for completeness. It crosses below WACC in FY2027E and destroys a cumulative $31.6B of EVA by FY2030E. The mechanism isn't "demand disappears" — it's AI compute pricing compression (more industry-wide GPU capacity eventually competing on price) combined with a modest overbuild. Worth being precise about what that mechanism is not: none of this report's four scenarios models a broad macroeconomic recession hitting enterprise cloud demand generally, as opposed to AI-specific competitive dynamics specifically. A recession scenario would be a different shape of risk than the Bear Case already built here — it would likely hit revenue growth assumptions across both the AI and non-AI pools simultaneously, while having an ambiguous, not obviously favorable, effect on WACC (a lower risk-free rate in a flight to quality, plausibly offset by a wider equity risk premium and wider credit spreads in a broad selloff). That is a genuinely different scenario than a more severe version of the Bear Case already modeled, and this report does not construct it.

Third, that 1.2-point cushion is thinner in practice than it looks on paper. Because Amazon's capital structure is roughly 95.5% equity-weighted, a change in the 10-year Treasury yield passes through to WACC at nearly full strength (~95% of the move, holding beta and the equity risk premium constant) — debt and credit-spread effects are a rounding error by comparison. The 10-year has already moved 53 basis points over the trailing twelve months and was trading above this report's own 4.7% assumption as this report was finalized. This is not a remote tail scenario the Base Case needs to worry about; it is within the range 10-year yields move routinely. Separately, the backlog this report treats as its load-bearing fact is more concentrated than the headline number suggests (two named customers account for well over a quarter of it at face value), and Amazon's own balance sheet has taken on debt at a pace Moody's has flagged as a sector-wide credit-quality concern, even though Amazon-specific ratings remain strong today. None of these change the conclusion on their own, but together they are a reminder that the WACC side of this report's central test is not a fixed reference point — it is a live variable with real, current movement of its own.

Most importantly, the AI-vs-non-AI decomposition suggests the aggregate "ROIC > WACC" answer may be masking a problem in a specific sub-segment. Under that framework, AI-specific ROIC actually improves over the forecast (12.0%→15.2%) — the AI capex itself does not look like the risky part. It is the non-AI/legacy cloud business whose implied ROIC falls from 22.3% to 7.0%, crossing below WACC by FY2029E–FY2030E, as AI captures a growing share of both revenue and margin dollars. If that holds, the question worth tracking isn't "is AI overbuilt" — it's whether AWS's core, non-AI cloud pricing and margin hold up while capital gets reallocated toward AI.

Bottom line

The evidence in this report supports "demand-backed with a thinning margin of safety and a specific, identifiable pressure point," not "speculative overbuild." A genuine bubble read would require backlog decoupling from capacity, utilization falling meaningfully, or the discipline rule actually engaging — none of which the data shows as of Q2 2026. The more useful question going forward isn't "is this a bubble" but three narrower, trackable ones: does the spread over WACC keep compressing toward zero on AWS's own operating performance; does that compression show up first in the AI pool or the legacy cloud pool; and does WACC itself move before either of those does, given how directly it now sits exposed to the rate environment. That is exactly what the tripwire dashboard, the AI/non-AI split, and the WACC-sensitivity work in this report are built to catch in real time.

This conclusion is analytical commentary on publicly disclosed information, not investment advice or a rating — see full disclosure at the top of this report.

Methodology & Sources

Two kinds of citation in this report, deliberately kept distinct. Numbered footnotes (see Footnotes, next section) mark facts and figures drawn from an external, independently-checkable source — a filing, a news article, an analyst note. A separate (Modeled — Sheet N) marker, in italics, appears instead wherever a figure is this report's own calculated output rather than a sourced fact — an AWS-segment ROIC, an EVA dollar figure, a WACC-sensitivity cell. Neither marker implies the other's reliability standard: a footnote means "you can verify this independently"; a Sheet marker means "here is exactly how this was calculated, and you can trace or challenge every input." Any calculated figure in this report without either marker draws on the same companion workbook and methodology described below.

• Segment basis. All revenue, operating income, PP&E, and D&A figures are AWS-segment (not consolidated Amazon), per the Segment Information note in Amazon's 10-Ks. This is a real methodological dependency worth naming directly, not just a citation: segment reporting reflects "how management views the business," not a fixed, externally-audited allocation rule. Amazon has real discretion over how shared costs — corporate overhead, shared R&D, infrastructure that serves multiple segments — get allocated between AWS and the rest of the company, and that discretion isn't disclosed at the level of granularity this report would need to detect a change in allocation methodology versus a genuine change in underlying economics. If Amazon were to shift how it allocates shared AI-related costs between AWS and other segments, this report's entire ROIC time series could show a discontinuity that reflects an accounting choice, not a change in AWS's actual capital productivity — and this report has no independent way to distinguish the two. This is a foundational dependency of the whole framework, not a peripheral caveat.
• Invested capital. Average AWS PP&E, net (beginning + ending balance ÷ 2). Excludes accounts receivable and operating-lease right-of-use assets, which a broader (more conservative) memo measure on the companion workbook adds back.
• NOPAT. AWS operating income × (1 − 24% assumed cash tax rate).
• D&A. Projected as a percentage of the beginning-of-period PP&E balance (14–17%, scenario-dependent) to avoid circularity in the roll-forward; calibrated to AWS's actual FY2025A ratio of 14.3%, consistent with Amazon's disclosed 5–6 year blended useful-life policy for servers and networking equipment.
• Street Case. Not the author's own view — a direct translation of Morgan Stanley's (Brian Nowak) and Goldman Sachs' (Eric Sheridan) most recently disclosed capex and AWS revenue-growth estimates into this report's ROIC framework, shown for comparison only and excluded from the probability-weighted blend.
• AI vs. Non-AI split. Built for the Base Case only. AI revenue share anchored to disclosed run-rate figures (~10% of AWS exiting Q1 2026, ~15% exiting Q2 2026); AI capex share estimated below CreditSights' ~75% cross-industry 2026 AI-capex-intensity figure; AI incremental margin set at 65% (midpoint of Morgan Stanley's disclosed 60–70% range); Non-AI operating income and PP&E are residuals, not independent estimates. None of this is company-disclosed — treat FY2029E–FY2030E figures especially as illustrative.
• Capacity-utilization-yield model. GW capacity and additions are company-disclosed (Base Case) or named-analyst (Street Case, Morgan Stanley/Nowak) figures. Utilization, yield, yield growth, and the capital-discipline thresholds are the author's own estimates informed by SemiAnalysis's published compute supply/demand tracking (Dylan Patel) and are not company-disclosed. Cost-per-GW ($38B, Epoch AI) is used only as a cross-check, not to derive AWS's modeled capex. The bottom-up ROIC reuses the Base Case's invested-capital base; only revenue/NOPAT differ.
• WACC. CAPM build: 4.7% risk-free rate (10-yr UST, 8/31/26)⁵, 1.20 levered beta (placeholder — verify against a live data source), 5.0% equity risk premium, 5.0% pre-tax cost of debt, 24% tax rate, ~$2.8T market capitalization⁶, $133.0B face value of long-term debt (6/30/26)² → 10.4% WACC.
• Primary sources. Amazon 10-K filings (FY2019–FY2025), 10-Q and 8-K filings for Q2 2026, Amazon CEO Andy Jassy's 2025 Letter to Shareholders; Reuters (7/30/26); Trading Economics (10-yr UST yield); stockanalysis.com (market capitalization).
• Sell-side sources. Morgan Stanley (Brian Nowak): notes of 7/27/26, 7/30/26, and 8/16/26, via TheStreet, Seeking Alpha, and BigGo Finance. Goldman Sachs (Eric Sheridan): notes of 4/30/26, 7/9/26, and 8/10/26, via TipRanks, Investing.com, and Prismnews. Neither bank's full research report was directly available; figures are as reported by financial media citing each note.
• AI-revenue sources. Amazon CEO Andy Jassy, April 2026 shareholder letter and Q1/Q2 2026 earnings-call commentary (AI services and custom silicon run-rate disclosures); CreditSights, "Technology: Hyperscaler Capex 2026 Estimates" (cross-industry AI capex-intensity figure).
• Capacity/compute sources. SemiAnalysis (Dylan Patel): earlier (June 2026) Datacenter Industry Model and public commentary on 2026–2027 GW buildout and GPU rental pricing, presented alongside his notably higher August 2026 figures (Dwarkesh Podcast interview, below) as a disclosed, unreconciled range rather than a supersession. Goldman Sachs Global Institute, "Tracking Trillions" (April 2026, global data-center capacity trajectory, a smaller stock figure than Patel's flow estimates would imply). Epoch AI (2026), "Total cost of ownership of a one-gigawatt AI data center." Amazon Q3 2025/Q1 2026/Q2 2026 earnings calls and shareholder letters (GW capacity disclosures). Morgan Stanley (Nowak), 8/16/26 note (GW build-pace estimates).
• Compute-shortage sources. Apollo Global Management, "The Growing Compute Shortage" (6/15/26, TSMC/HBM capacity); TweakTown citing TSMC leadership (4/21/26); CNAS, "American AI Companies Can't Get Enough Chips" (citing SemiAnalysis's "The Great AI Silicon Shortage," 3/12/26, and Epoch AI); CNBC citing Synopsys CEO Sassine Ghazi (1/26/26, memory shortage); J.P. Morgan Asset Management, "Is AI Running Out of Compute?" (4/17/26, power/grid timelines, 92% pre-lease figure); BigGo Finance summarizing investor Gavin Baker on the a16z Podcast (8/26, shortage-duration argument).
• Financing and leverage sources. Investing.com citing Fitch Ratings (7/8/26, AA- rating, debt/liquidity/FCF figures, $17.5B term loan); Yahoo Finance citing Moody's (2/20/26, A1 stable outlook); CNBC citing Moody's sector research (7/24/26, $785B/~$1T industry capex, $1.2T lease commitments).
• Backlog concentration sources. AboutAmazon.com and CNBC (11/3/25, OpenAI $38B AWS deal); Anthropic.com, GeekWire, and The New Stack (4/20–21/26, Anthropic $100B+ AWS deal, 5GW Trainium/Graviton, $25B incremental Amazon investment); Reuters (Hu and Seetharaman, Anthropic revenue run-rate and 2026 targets); Medium/Tech Buster and CNBC (OpenAI total infrastructure commitments and revenue run-rate context).
• Sell-side and industry-skeptic sources. Goldman Sachs Research, "Gen AI: Too Much Spend, Too Little Benefit?" (Jim Covello, originally June 2024, update via Fortune, 6/5/26 and 5/6/26). Wall Street Journal, "Why Big Tech's AI Spending Is $3 Trillion Higher Than It Seems" (Rudegeair and Santilli, 8/17/26), via TipRanks and Axios (8/27/26) for the related, narrower Morgan Stanley off-balance-sheet estimate.
• Inference-pricing sources. "The Price of Intelligence: A Quality-Adjusted Price Index for AI Services" (arXiv:2608.29843, Aug. 2026); Axis Intelligence Research, LLMflation Index (7/29/26); Gartner AI inference cost forecast (3/26, via NeuralWired, 6/20/26).
• Nvidia/AWS demand sources. NVIDIA Corp. Form 8-K, Q2 FY2027 results (8/26/26, SEC EDGAR); CNBC and 24/7 Wall St. (8/26/26, Data Center revenue, AWS 2M-GPU commitment for 2027–2028).
• Server/data-center economics sources. Amazon Q2 2026 earnings call transcript (Brian Olsavsky, CFO, 7/30/26, via Roic.ai and Yahoo Finance) for the 30-plus-year data-center life and multi-generation server economics quotes; Moneywise/Yahoo Finance (8/26, Trainium 2 replacement-cycle reporting).
• Hyperscaler leverage-comparison sources. Yahoo Finance/24/7 Wall St., "AI CapEx Risk: Amazon And Oracle Are The Most Vulnerable Hyperscalers" (8/18–26/26); Tomasz Tunguz, "Is Your AI Funded by Junk Bonds?"; EnkiAI (6/17/26, Oracle capex and debt trajectory); The Motley Fool (8/3–20/26, hyperscaler capex/FCF comparison).
• Chip-generation and buy-side sources. NVIDIA Corp. Form 10-K and DEF 14A, FY2026 (Rubin/Blackwell Ultra cost-per-token figures); NVIDIA Newsroom, "AWS and NVIDIA to Deliver 2 Million Additional GPUs..." (8/26/26); TechCrunch and Yahoo Finance (8/26/26, deal context, custom-silicon commitments); "Pitch The PM" (YouTube, @PitchThePM, hosted by Doug Garber), interview with Daniel Pilling (Sands Capital Global Growth Fund), September 2026 (GPU payback periods, old-chip pricing, Anthropic revenue-per-GW estimate, enterprise-adoption framing).
• Companion deliverable. A fully-formulaed Excel workbook (nine tabs: Assumptions, AWS Historicals, Backlog & Capex Bridge, Projections, ROIC-WACC-EVA Model, Sensitivity & Tripwires, AI vs. Non-AI ROIIC, Capacity-Utilization ROIC, Additional Risk Factors) underlies every figure in this report, including the financing/leverage, backlog-concentration, and WACC-sensitivity tables above, and is available on request.

Footnotes

Every numbered citation in this report links back here — each entry traces the specific claim it supports to its source.

1. Amazon Q2 2026 Form 8-K / earnings release (7/30/26) — FY2026 capital investment guidance ("approximately $220 billion," AWS-directed). ↑
2. Amazon Form 10-Q, Q2 2026 (period ended 6/30/26), SEC EDGAR — remaining performance obligations ($496.0B) and Note 5, Debt (face value of long-term debt). ↑
3. Amazon CEO Andy Jassy, 2025 Letter to Shareholders (April 2026), aboutamazon.com. ↑
4. Reuters, "Amazon lifts investment plans after strong cloud sales," 7/30/26. ↑
5. Trading Economics, U.S. 10-year Treasury yield, retrieved 8/31/26 and 9/1/26. ↑
6. stockanalysis.com, AMZN market capitalization, retrieved August 2026. ↑
7. Morgan Stanley (analyst Brian Nowak), research notes of 7/27/26, 7/30/26, and 8/16/26, as reported via TheStreet, Seeking Alpha, and BigGo Finance; full report not directly available. ↑
8. Goldman Sachs (analyst Eric Sheridan), research notes of 4/30/26, 7/9/26, and 8/10/26, as reported via TipRanks, Investing.com, and Prismnews; full report not directly available. ↑
9. Amazon CEO Andy Jassy, April 2026 shareholder letter and Q1/Q2 2026 earnings-call commentary — AI services and custom-silicon annualized run-rate disclosures. ↑
10. CreditSights, "Technology: Hyperscaler Capex 2026 Estimates" — cross-industry AI capex-intensity figure. ↑
11. SemiAnalysis (Dylan Patel), Datacenter Industry Model and public commentary on 2026–2027 GW buildout and GPU rental pricing. ↑
12. Epoch AI (2026), "Total cost of ownership of a one-gigawatt AI data center." ↑
13. Amazon Q3 2025, Q1 2026, and Q2 2026 earnings calls and shareholder letters — GW capacity disclosures. ↑
14. Morgan Stanley (Nowak), 8/16/26 note — GW build-pace estimates, via BigGo Finance. ↑
15. Apollo Global Management, "The Growing Compute Shortage," 6/15/26. ↑
16. TweakTown, citing TSMC leadership commentary, 4/21/26. ↑
17. Center for a New American Security (CNAS), "American AI Companies Can't Get Enough Chips," citing SemiAnalysis's "The Great AI Silicon Shortage" (3/12/26) and Epoch AI. ↑
18. CNBC, citing Synopsys CEO Sassine Ghazi, 1/26/26. ↑
19. J.P. Morgan Asset Management, "Is AI Running Out of Compute?" 4/17/26. ↑
20. BigGo Finance, summarizing investor Gavin Baker on the a16z Podcast, August 2026. ↑
21. Investing.com, citing Fitch Ratings, 7/8/26 — AA- rating, debt/liquidity/FCF figures, $17.5B term loan facility. ↑
22. Yahoo Finance, citing Moody's, 2/20/26 — A1 senior unsecured rating, stable outlook. ↑
23. CNBC, citing Moody's sector research, 7/24/26 — $785B/~$1T industry capex, $1.2T lease commitments. ↑
24. AboutAmazon.com and CNBC, 11/3/25 — OpenAI $38B, 7-year AWS compute deal. ↑
25. Anthropic.com, GeekWire, and The New Stack, 4/20–21/26 — Anthropic $100B+, 10-year AWS deal; up to 5GW Trainium/Graviton; $25B incremental Amazon investment. ↑
26. Reuters (Krystal Hu and Deepa Seetharaman), "Exclusive-Anthropic aims to nearly triple annualized revenue in 2026, sources say" — Anthropic revenue run-rate disclosures and 2026 targets, on-the-record company sourcing. ↑
27. Medium/Tech Buster and CNBC — OpenAI total infrastructure commitments and revenue run-rate context. ↑
28. Goldman Sachs Research (Jim Covello), "Gen AI: Too Much Spend, Too Little Benefit?" (goldmansachs.com, June 2024, freely accessible); update reported via Fortune, 6/5/26 and 5/6/26. ↑
29. Wall Street Journal (Rudegeair and Santilli), "Why Big Tech's AI Spending Is $3 Trillion Higher Than It Seems," 8/17/26. ↑
30. TipRanks and Axios, 8/27/26 — Morgan Stanley's narrower, seven-company off-balance-sheet commitment estimate and "Mapping AI's Circularity" research. ↑
31. "The Price of Intelligence: A Quality-Adjusted Price Index for AI Services," arXiv:2608.29843, August 2026. ↑
32. Axis Intelligence Research, LLMflation Index, 7/29/26. ↑
33. Gartner, Inc., "Gartner Predicts That by 2030, Performing Inference on an LLM With 1 Trillion Parameters Will Cost GenAI Providers Over 90% Less Than in 2025" (press release, gartner.com, 3/25/26); secondary analysis via NeuralWired, 6/20/26. ↑
34. NVIDIA Corp. Form 8-K, Q2 FY2027 results, 8/26/26, SEC EDGAR. ↑
35. NVIDIA Newsroom, "AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI," 8/26/26; also reported by CNBC and 24/7 Wall St. ↑
36. Amazon Q2 2026 earnings call transcript (Brian Olsavsky, CFO), 7/30/26, via Roic.ai and Yahoo Finance — data-center useful life and multi-generation server economics. ↑
37. Moneywise/Yahoo Finance, August 2026 — Trainium 2 replacement-cycle reporting. ↑
38. Yahoo Finance / 24/7 Wall St., "AI CapEx Risk: Amazon And Oracle Are The Most Vulnerable Hyperscalers," 8/18–26/26. ↑
39. Tomasz Tunguz, "Is Your AI Funded by Junk Bonds?" ↑
40. EnkiAI, 6/17/26 — Oracle capex and debt trajectory. ↑
41. The Motley Fool, 8/3–20/26 — hyperscaler capex/free-cash-flow comparison. ↑
42. NVIDIA Corp. Form 10-K and DEF 14A, FY2026, SEC EDGAR — Rubin and Blackwell Ultra cost-per-token and throughput figures. ↑
43. NVIDIA Newsroom / AWS — G7 (Blackwell) vs. G6 instance performance comparison. ↑
44. TechCrunch and Yahoo Finance, 8/26/26 — AWS/Nvidia deal context and Amazon custom-silicon commitment figures. ↑
45. "Pitch The PM" (@PitchThePM), hosted by Doug Garber — YouTube interview with Daniel Pilling (co-PM, Sands Capital Global Growth Fund), youtu.be/Ym7ExbGUP-U, September 2026 — GPU/server payback periods, older-chip pricing, Anthropic revenue-per-GW estimate, enterprise-adoption framing. ↑
46. Dwarkesh Podcast, hosted by Dwarkesh Patel — interview with Dylan Patel (founder, SemiAnalysis), "Dylan Patel — 'Two Labs Will Soon Control Most of the World's Workforce,'" youtu.be/aV26V1UvkJw, published approximately August 2026 — incremental global compute buildout (30/50/70/90–100 GW, 2026–2029E), chip-generation efficiency (3–5x performance/watt), compute pricing per megawatt, industry-wide CapEx and debt-funding estimates ($11T CapEx / $5T debt, 2024–2029E), and a credit-market channel for rising interest rates. ↑
47. Goldman Sachs Investment Banking, "Harnessing AI for the Real Economy" (2026), citing its own 2025 "Powering the AI Era" report — global data-center capacity growth (30 GW in 2019 → 57 GW in 2024, +65 GW projected online by 2030). ↑
48. Goldman Sachs Investment Banking, "Harnessing AI for the Real Economy" (2026), citing Goldman Sachs Global Investment Research/SUSTAIN (2/5/26) — grid interconnection queues of 8–12 years in the highest-priority U.S. data-center markets. ↑
49. Goldman Sachs Investment Banking, "Harnessing AI for the Real Economy" (2026), citing Bloomberg/Dealogic/Goldman Sachs estimates — AI-related hyperscaler and infrastructure debt issuance at ~6.6% of the $1.83T U.S. investment-grade bond market in 2025, projected toward ~20% in 2026. ↑
50. Goldman Sachs Investment Banking, "Harnessing AI for the Real Economy" (2026), citing Applied Digital's June 2026 senior secured notes offering and broader private credit market commentary — GPU-fleet-specific financing structures (shorter tenors, embedded refresh provisions). ↑
51. Goldman Sachs Investment Banking, "Harnessing AI for the Real Economy" (2026), citing Goldman Sachs Global Investment Research, "The AI Spending Boom Is Not Too Big" (Datastream/Factset, October 2025) — U.S. AI investment at ~1.2% of GDP versus 3–4.5% for the 1800s railroad buildouts. ↑
52. CNBC, "Alphabet earnings takeaways" (7/22/26); Alphabet/blog.google, Sundar Pichai Q2 2026 earnings call remarks — Google Cloud revenue $24.8B (+82% YoY), Cloud backlog $514B (up from $462B in Q1 2026), Alphabet's first-ever negative quarterly free cash flow (-$5.9B), and raised FY2026 capex guidance ($195–205B, up from $180–190B). ↑

Glossary

Reference material, not required reading — standard finance terms used throughout, plus a handful of terms coined specifically for this report's framework.

Standard Financial Terms

Backlog / Remaining Performance Obligations (RPO)

Contracted future revenue a company has already secured from customers but not yet recognized as revenue. Unlike a demand forecast, backlog reflects revenue customers have already agreed to pay — the central fact this report treats as evidence the AWS buildout is demand-backed.

CAPM (Capital Asset Pricing Model)

A standard formula for estimating a company's cost of equity: Cost of Equity = Risk-Free Rate + (Beta × Equity Risk Premium). Used in this report to build the WACC estimate (Cover & Assumptions tab of the companion workbook).

D&A (Depreciation & Amortization)

The non-cash expense that spreads the cost of long-lived assets — data centers, servers, networking equipment — over their useful life. Modeled here as a percentage of beginning-period PP&E, calibrated to AWS's actual disclosed FY2025A ratio.

EVA (Economic Value Added)

The dollar amount of value created, or destroyed, above the cost of capital: EVA = (ROIC − WACC) × Average Invested Capital. Positive EVA means a business earns more than its capital costs; negative EVA means it destroys value even while remaining profitable on a NOPAT or GAAP basis.

Free Cash Flow (FCF)

Cash generated by the business after capital expenditures — distinct from ROIC or NOPAT, which measure whether capital earns an adequate return over its life, not whether the company is cash-comfortable in the near term. A company can show attractive multi-year ROIC while FCF is deeply negative during a heavy build-out phase, as this report's Financing & Leverage discussion lays out for Amazon specifically.

Incremental margin

The operating margin earned on the newest slice of revenue growth, as distinct from the blended margin across an entire business. Used here specifically in Morgan Stanley's AI-economics framework (assumed at 60–70%).

Invested Capital

The capital base a business needs to generate its operating income. Defined in this report as average AWS property & equipment, net (beginning + ending balance ÷ 2) — the denominator in the ROIC calculation.

NOPAT (Net Operating Profit After Tax)

Operating income adjusted for the cash taxes a business would pay if it carried no debt: NOPAT = Operating Income × (1 − tax rate). The numerator in the ROIC calculation.

PP&E, net (Property, Plant & Equipment, net)

The book value of a company's physical, long-lived assets — data centers, servers, networking equipment, land, buildings — after subtracting accumulated depreciation.

ROIC (Return on Invested Capital)

The core profitability metric in this report: ROIC = NOPAT ÷ Average Invested Capital. Measures how efficiently a business converts its capital base into after-tax operating profit.

ROIIC (Return on Incremental Invested Capital)

A stricter, marginal version of ROIC measuring the return on only the newest capital added: ROIIC = ΔNOPAT ÷ prior-year capex. Because it isolates only the newest investment, ROIIC runs more volatile than blended ROIC and is especially sensitive to the lag between when capital is spent and when it starts generating revenue.

Spread

Shorthand for ROIC minus WACC. A positive spread means a business earns more than its cost of capital; a negative spread means the opposite, even if the business remains profitable in accounting terms.

WACC (Weighted Average Cost of Capital)

The blended, minimum required rate of return a company must earn on its invested capital to satisfy both equity and debt investors. Built here via CAPM for cost of equity, a market-based cost of debt, and weights based on market capitalization and total debt.

This Report's Framework Terms

Terms tagged This report's term below do not have a standard external definition — they were coined for this report's specific framework and are defined here, not elsewhere in the finance literature.

Bull / Base / Bear CaseThis report's term

Three scenarios reflecting the author's own judgment about AWS revenue growth, margin, and capex, ranging from most optimistic (Bull) to most conservative (Bear).

Street CaseThis report's term

A fourth scenario built not from the author's own judgment but as a direct translation of Morgan Stanley's and Goldman Sachs' most recently disclosed capex and revenue estimates. Shown for comparison only; excluded from the probability-weighted blend.

Probability-weighted outcomeThis report's term

A single blended figure produced by weighting the Bull/Base/Bear scenarios by the author's own judgment-based probabilities (60%/25%/15%), not a market-implied or consensus weighting.

Tripwire dashboardThis report's term

Five specific, checkable conditions — growth, margin, backlog growth, capex growth, and management commentary — that, if ALL triggered together, would signal the buildout has shifted from demand-backed to speculative. Built for quarterly monitoring, not a one-time assessment.

GW (Gigawatt)

A unit of electrical power capacity, and the standard measure used across the AI/hyperscaler industry (including by SemiAnalysis) to describe data-center scale — because power availability, not physical space, is the binding constraint on how much compute capacity can be built.

UtilizationThis report's term

The percentage of built data-center/compute capacity actively generating revenue at a given time, as opposed to sitting idle or still ramping after construction.

Yield (per GW)This report's term

This report's shorthand for revenue generated per unit of utilized capacity ($ of AWS revenue per gigawatt) — used as a bottom-up cross-check on whether capacity is being monetized at healthy rates.

Breakeven yieldThis report's term

The minimum yield a unit of capacity must generate to earn enough NOPAT to clear its own cost of capital (WACC), given its cost to build and the business's operating margin. A cushion above breakeven means capacity has room for error before turning value-destructive.

Yield cushionThis report's term

Actual yield divided by breakeven yield. A cushion above 1.0x means capacity earns more than its minimum required return; below 1.0x would mean capacity destroys value even at full utilization.

Capital-discipline throttleThis report's term

A formula-based rule — not a narrative assumption — built to test when Amazon's own stated capital discipline would actually curtail capex growth, based on whether utilization falls below defined thresholds.

AI Pool / Non-AI PoolThis report's term

The author's own estimated split of AWS revenue, capex, and PP&E into an AI-attributed portion and a non-AI/legacy-cloud portion. AWS does not disclose this split; Non-AI figures are calculated as the residual of AI figures from AWS's actual disclosed totals.

Forward-Looking Statements

Figures for FY2026E through FY2030E throughout this report are estimates and projections, not historical fact. They rest on assumptions — disclosed and flagged throughout this report and the companion workbook — that involve significant uncertainty. Actual results may differ materially from any figure presented here, and no representation is made that any estimate will be achieved.

No Compensation or Affiliation

The author has not received, and does not expect to receive, any compensation, consideration, or benefit from Amazon.com, Inc., Morgan Stanley, Goldman Sachs, SemiAnalysis, Epoch AI, CreditSights, or any of their affiliates in connection with this report. References to these companies and their research are for identification and analytical purposes only and do not imply endorsement, affiliation, or sponsorship.

As of Date; No Obligation to Update

This report reflects information available as of August 31, 2026. The author undertakes no obligation to update or revise any statement, figure, or conclusion to reflect subsequent events, new information, or changes in circumstances.

Not a Substitute for Professional Advice

Nothing in this report should be construed as legal, tax, accounting, or personalized investment advice. Readers should consult their own qualified advisors before making any financial decision.

Gregg Carlson · Fractional CFO & Controller · CPA (inactive, NV) · CFA Institute Member
gregg-carlson.com · CFO Insights: Technology & Capital Markets
© 2026 Gregg Carlson. General financial and educational commentary only — not investment advice.