Phase Space AI

02 Financial Model Notes

Amazon [AMZN]

Amazon.com, Inc. [AMZN] — Financial Model Notes

Task 2 · AMZN_Underwriting_Model_2026-07-27.xlsx · 6 tabs, live formulas


1. Structure

Tab Purpose
Assumptions Every driver, blue = input / black = formula. Segment growth, EBITDA margin, capex, asset-class shares, useful lives, placed-in-service lags, tax, SBC, WACC, terminal growth. Base-rate check block at the foot.
Depreciation Engine The analytical core. Vintage straight-line schedule: each forecast year's capex is split into servers/network, buildings/shell and other equipment, each depreciated over its own life after its own CIP lag. Includes a full parallel calculation at a 4-year server life.
Income Statement Segment revenue → consolidated → EBITDA → less modelled D&A → EBIT → pre-tax → EPS. Plus an earnings-quality overlay: EPS excluding non-operating income, and EPS at a 4-year server life.
Cash Flow & BS OCF, capex, economic FCF (SBC treated as a real cost), cash roll-forward including debt issuance, net debt, lease liabilities, off-balance-sheet memo.
Scenarios Bull / Base / Bear with terminal revenue linked live to the Income Statement, exit EV/NOPAT, discounting to today, probability weighting, E[R] and the cash-hurdle comparison. Base-rate check per scenario.
DCF & Valuation Unlevered FCF, a five-year fade period in which capex reverts to maintenance, Gordon terminal value, sensitivity grid, comps, football field.

2. Verification — performed as the skill requires

All workbooks were closed first (an open workbook serves a stale in-memory copy), the file was opened in Excel via AppleScript, and computed cells were read back. openpyxl writes formulas as strings without evaluating them, so a formula that looks right can be silently wrong; only a readback proves otherwise.

Readback results (2026-07-27):

Revenue  : $716,924  $818,001  $915,415  $1,014,763  $1,114,876  $1,213,848
EBIT     : $79,994   $94,530   $103,297  $108,924    $112,043    $117,120
D&A      : $41,860   $56,800   $75,209   $96,058     $119,851    $140,215
D&A@4yr  : $53,358   $72,213   $95,388   $121,670    $151,955    $177,635
EPS      : $7.24     $8.26     $7.76     $8.23       $8.50       $8.93
EPSexNonOp:$5.95     $7.07     $7.76     $8.23       $8.50       $8.93
EPS@4yr  : $6.39     $7.11     $6.25     $6.29       $6.07       $6.08
EconFCF  : (19,044)  (50,686)  (59,576)  (47,729)    (27,940)    (7,019)
DCF/sh = $14.20   TV% = 189%   Scenario PW = $116.81   E[R] = -49.5%

FY2025A ties to the 10-K: revenue $716,924m ✓ (exact), segment D&A $41,860m ✓ (exact), operating income $79,994m vs $79,975m reported (0.02% residual from the margin input precision — disclosed, not hidden).

Three model errors were found and fixed during verification, each of which would have failed silently: 1. The D&A fade formula referenced capex as a signed (negative) figure, driving terminal D&A negative and the DCF to −$168/share. 2. The cash roll-forward omitted debt issuance, so cash went deeply negative and inflated scenario net debt. 3. The original five-year DCF terminated while capex was still 1.8x D&A — not a steady state. A five-year fade period was added so the terminal value is reached from a normalised position.

3. The Depreciation Engine — calibration

The engine is fitted to reported depreciation, then extrapolated. Fitting first is what makes the extrapolation credible.

Parameter Fitted value Method
Servers / network share of capex 52% From FY2025 gross PP&E movement by class (R58)
Buildings / shell share 30% same
Other equipment share 18% residual
Server useful life 5.5 years 10-K discloses "Five to six years"
Buildings life 30 years 10-K: "lesser of forty years or remaining life"; 30 used as an effective average
Other equipment life 9 years 10-K: "Three to ten years"
Placed-in-service lag, servers 0.5 years Fitted
Placed-in-service lag, buildings 1.0 years Fitted — construction in progress was $71.7bn at Dec-25, depreciating nothing
Calibration RMSE, FY2022–FY2025 $2.9bn on a $25–42bn series Grid search over lags and scale

The CIP lag is not a modelling convenience — it is the mechanism. $71.7bn of Amazon's $534.1bn gross PP&E was construction in progress at Dec-2025 and was depreciating nothing at all. That is 13% of the asset base sitting outside the income statement, waiting.

4. Base-rate checks — every scenario's growth assumption

Reproduced on the Assumptions and Scenarios tabs.

AWS revenue growth. AWS grew 28% in Q1-2026, its fastest in 15 quarters, off a $128.7bn base. The reference class says cloud-infrastructure growth at >$100bn scale decelerates: AWS itself went from 37% (2021) to 13% (2023) before re-accelerating. The model assumes 27% → 14% over five years — deceleration, i.e. WITH the base rate, and on the conservative side of the current run-rate. No override is being claimed.

North America retail operating margin. FY2025 was 6.95%, up from 3.86% in 2022. ~$0.9bn of the FY2025 improvement is the heavy-equipment useful-life extension (10→13 years), not operating leverage — 16% of the year-on-year segment gain. The model expands retail margin only ~150bp over five years, deliberately modest, because part of the observed trend is an accounting estimate that does not repeat.

Capex. FY2025 PP&E additions of $142.4bn were +66% year on year. No company in the reference class has sustained >50% capex growth for four consecutive years. The model assumes deceleration to +1.6% by FY2030 ($200bn → $247bn). This is an override of the recent trend in the conservative direction and is named as such.

Consolidated operating margin. The model's FY2030 operating margin of 9.6% is below FY2025's 11.2%. That is not a bearish assumption imposed on the business — it is the arithmetic consequence of the depreciation engine running against a rising EBITDA margin. Stated explicitly so it is not mistaken for a demand call.

5. Duration-matched, two-tranche treatment

Per references/regime-change-test.md. Amazon discloses no contracted revenue tranche, so duration matching can only be applied to the cost side — and it is:

Tranche Contracted duration Model treatment
Leases not yet commenced ($106.3bn) Disclosed ladder: $5.1bn 2026, $9.6bn 2027, $7.2bn 2028, $7.3bn 2029, $7.3bn 2030, $69.8bn thereafter Modelled at the disclosed ladder within the horizon
Unconditional purchase obligations ($103.8bn, incl. energy) $19.1bn 2026, $15.8bn 2027, then declining Modelled at the ladder
Long-term debt ($122.6bn face, to 2076) Full ladder disclosed Interest modelled through the forecast
AWS revenue None disclosed Fully merchant; mean-reverting growth throughout. No contracted tranche exists to model at contracted terms.

Contracted share of revenue by forecast year: 0% in every year. That is the honest expression of the uncertainty — and it is why the regime-change test fails for Amazon.

6. Known limitations, stated rather than buried

  1. Segment forecasts are not built bottom-up from units. Amazon does not disclose units, capacity, or AI-specific revenue, and §3 of the research document records the corpora attempted.
  2. Non-operating income beyond FY2026 is set to zero. The Anthropic mark is unforecastable by construction — it depends on a private financing round. Setting it to zero is a choice, disclosed, and it is why modelled EPS falls from $8.26 (FY2026E) to $7.76 (FY2027E).
  3. The DCF has terminal value at 189% of enterprise value and is labelled as having low information content rather than being presented as a target. The reverse DCF is the informative version.
  4. Comps use trailing reported figures, not NTM consensus, because Alpha Vantage was exhausted. Fiscal-year ends differ and are labelled. Alphabet's FY2025 was not retrievable under the standard XBRL tag; FY2024 is shown and flagged.
  5. Capacity in GW is derived, not disclosed. Amazon has used the word "gigawatt" once in eighteen quarters of earnings releases.