Phase Space AI

01 Company Research

Amazon [AMZN]

Amazon.com, Inc. [AMZN] — Company Research

Task 1 · investment-memo skill v1.4.0 · as of 2026-07-27 · spot $231.39

---

Updated 2026-07-29 — Criteria + two-horizon valuation. The numbered Gates are retired and replaced by named Criteria with types (BINDING / MEASURED) returning PASS / FAIL / INDETERMINATE. This memo no longer issues a position verdict — no Long, Short, Watchlist or Avoid. The expected-return-versus-cash-hurdle test is replaced by a reverse-DCF implied-path test plus a 12-month target, with sensitivity run over the exit multiple rather than over scenario probabilities. Momentum is entry timing only and vetoes nothing. Sections below that predate this update are retained as the historical record and are annotated where superseded.

Reading the retired "Gate N" numbering. Any Gate N below is historical. The mapping is: Gate 1 → Quality Criteria (BINDING) · Gate 1b → Short Mechanism Criteria (MEASURED) · Gate 2 / 2A / 2B / Path B → dissolved, absorbed by the Valuation Criteria · Gate 2C → Peer Spread Criteria (MEASURED) · Gate 3 → Catalyst Criteria (MEASURED) · Gate 4 / 4a → Valuation Criteria (BINDING) · Gate 5 → Liquidity Criteria (BINDING) · Gate 6 → Momentum Criteria (MEASURED, entry timing only) · (new) Downside Criteria (MEASURED). A Criteria returns PASS / FAIL / INDETERMINATE; a missing input is INDETERMINATE, never FAIL. Historical calibration items keep the old numbering on purpose — the record is the record.

0. Stage C.6 variant-likelihood pre-flight — RUN FIRST, BEFORE ANY VIEW WAS FORMED

The skill requires this for a heavily-covered mega-cap. It was run before the model was built and before any directional view existed. Verdict recorded here unedited.

Candidate modelling questions identified in the pre-flight (fifteen-minute honest effort):

1. What is in "consensus EPS"? Amazon books its Anthropic stake through non-operating income. Q1-2026 carried $16.8bn of pre-tax gains from that stake. Some analysts include it, some strip it. This is almost certainly what produces the ±15% EPS dispersion against ±3% revenue dispersion.
2. The two-directional useful-life change. Amazon shortened AI server lives and lengthened fulfilment heavy-equipment lives in the same month. The segment-level effect runs in opposite directions.
3. Amazon Leo capitalisation timing. Amazon currently expenses satellite network costs and will capitalise them "once the service achieves commercial viability, including sales to customers."

PRE-FLIGHT VERDICT (recorded before the work): variant UNLIKELY. All three questions are about how reported earnings are constructed, not about what the business will earn. None of them changes my view of Amazon's revenue or operating income versus the Street's. The correct prior is that Gate 2 fails.

DID THE PRE-FLIGHT PROVE RIGHT? Yes, on the estimate variant — and it under-rated the earnings-quality findings. Sections 3–5 below turned up materially more than expected on financing structure and accounting, all of it quantified from primary filings. But none of it converted into a differentiated number for FY2026 or FY2027 revenue or operating income. Gate 2 fails, as predicted. Sections 3–5 are reported because they are genuinely informative about the quality and durability of reported earnings, not because they rescue a thesis.

---

1. Screen provenance — AMZN did NOT clear the 2026-07-27 screen

Verified against reports/screens/Screen_2026-07-27.json. Amazon does not appear on either the long or the short shortlist, and did not reach Stage C.

The screen's Stage B took the top and bottom ~8% of a 4,302-name universe by 12-1 momentum. The lowest-ranked name on the long shortlist carried 12-1 momentum of roughly +300%. Amazon's 12-1 momentum is +0.5%. It was eliminated at Stage B, along with essentially every mega-cap.

What this means, stated plainly:

• Amazon carries no base-rate support from the empirical-signal cluster this system is built to exploit. The declared edge hypothesis is selective breadth — screening the whole market and spending underwriting effort only where signals cluster. Amazon is the opposite case: a user-directed name with no signal.
• It is not evidence Amazon is a bad investment. The screen is a momentum-and-quality funnel, and a $2.5 trillion company mechanically cannot produce 300% twelve-month momentum. The screen is silent on mega-caps by construction, not by judgment.
• The honest consequence is that the bar for this memo to produce an actionable position is higher, not lower, than for a screened name — because the base rate that would otherwise support it is absent.

---

2. Transcript mention-frequency — RUN FIRST, GENERATIVELY, BEFORE ANY VIEW

2.1 Corpus and method

Alpha Vantage EARNINGS_CALL_TRANSCRIPT was exhausted for the day, so a first-party corpus was used instead.

• Source (pinned, single, named): SEC EDGAR 8-K Exhibit 99.1 earnings releases, amzn-YYYYMMDDxex991.htm.
• Window: 18 consecutive quarters, 2021Q4 through 2026Q1. Every quarter present; no gaps; no source mixing.
• 100% company-authored prepared text. No Q&A, therefore no analyst-prompted contamination — every mention is unprompted, which the reference file identifies as the materially stronger signal.
• Normalised per 10,000 words. This is not optional here: the corpus length halves mid-series, from ~8,600 words (2024Q1) to ~5,300 (2024Q2), and stays there. Raw counts across that break are meaningless. See chart 13_corpus_word_counts.png.
• Word-boundary matching. Homonyms checked by eye — see the backlog finding in §2.4.

2.2 Step 2 — every emerging and decaying term listed as an OPEN QUESTION, no interpretation attached

This list was written before any investigation. Interpretations were added only in §2.3, after checking independent corpora.

2.3 Step 3 — each question investigated against an INDEPENDENT corpus

2.4 Required output — the mention-frequency table

Per 10,000 words. Source: SEC EDGAR 8-K Ex-99.1, 2021Q4–2026Q1 (18 quarters). Chart: 11_, 12_.

2.5 Hypothesis provenance — generated here versus carried in as a prior

Required by the reference file. Recorded honestly.

Two of five hypotheses that carry weight in this memo were generated by the mention-frequency pass rather than carried in. The Nova/Anthropic crossover in particular is the kind of finding the method exists to produce.

---

3. Unpublished scoping — what is NOT in the filings, and what was actually attempted

Per references/unpublished-scoping.md: a "not disclosed" finding is a research task, not a limitation. Every item below states the corpora attempted and what they returned.

3.1 R-file check — performed FIRST, per the MU lesson

The MU memo declared two disclosure gaps that turned out to be wrong, because XBRL company-facts strips dimensional members. The R-files were checked before anything was declared undisclosed here. They contained substantially more than company-facts did:

Three of the six items would have been wrongly reported as "not disclosed" had only company-facts been used.

3.2 Power and permitting — the corpus the assignment identified as richest

Attempted: FERC Section 206 show-cause dockets (six issued 2026-06-18 to all RTOs), PJM and ERCOT interconnection-queue reporting, MISO Long-Range Transmission Plan Tranche 2.1, state utility filings.

What it returned: - PJM's interconnection queue stands at ~275 GW of generation projects; >55 GW has cleared; ~30 GW of transition queue remains for 2026 processing. PJM expects to host as much as 70% of US data centres and projects 30 GW of new demand by 2030. - National filings for >150 GW of new data-centre power capacity as of January 2026. - FERC issued six show-cause orders in June 2026 forcing every RTO to revise or justify large-load interconnection rules — an active, dated regulatory process.

What it did NOT return, and this is the honest limitation: interconnection queues are filed by project entity, and hyperscalers routinely file through LLCs and third-party developers whose beneficial ownership is not disclosed in the queue. I could not attribute a specific GW figure to Amazon from the queue data. That is a genuine determinability limit, not a filing gap — the information exists but is deliberately obscured at source.

So the capacity number had to be inferred from Amazon's own capital spending instead:

Flagged as a derived estimate throughout. Amazon's refusal to publish GW (§2, Q7) is what forces this.

3.3 Items genuinely NOT determinable — with corpora attempted

3.4 Job postings and product changelogs

Attempted: AWS release-notes cadence as referenced in the releases; product-launch descriptions in Ex-99.1.

Returned: a coherent, dated agent product family — Kiro agent, AWS Security Agent, AWS DevOps Agent, described as "frontier agents… to solve tasks more autonomously and over longer durations." This independently corroborates the agent mention spike (Q1). It does not size the opportunity.

---

4. The Anthropic loop — the single most important finding on this name

Generated by the mention-frequency pass (Q2/Q3), then quantified from primary filings.

4.1 The numbers, all from filings

4.2 Why this matters — the loop, drawn explicitly

Every leg below is separately disclosed. The loop is not commonly drawn:

1. Amazon invests equity and convertible notes in Anthropic, and as of April 2026 also extends it credit.
2. Anthropic spends on AWS compute (Project Rainier).
3. AWS revenue accelerates — 28% in Q1-2026, Amazon's own words, "our fastest growth in 15 quarters."
4. Anthropic raises at a higher valuation; the observable price change marks Amazon's holding up.
5. Amazon books the mark as $16.8bn of non-operating income, 43% of quarterly pre-tax profit.

This is not an allegation of impropriety. The accounting is correct under ASC 321 measurement-alternative rules and is fully disclosed. The analytical point is narrower and harder to dispute:

A material and growing share of Amazon's reported earnings is an unrealised, non-cash mark on a private company whose principal expense is buying Amazon's own product — and Amazon is now also that company's lender.

4.3 Financing structure — Amazon's answer is debt, on balance sheet

The assignment flagged financing structure as under-modelled. For Amazon it is on balance sheet and it moved violently in Q1-2026:

$53.8bn of debt issued in a single quarter, some of it 50-year paper. This happened after the FY2025 10-K, which is where most sell-side models were last fully rebuilt.

Contrast with Meta: Amazon has no unconsolidated data-centre VIE. Search of the FY2025 10-K for "variable interest entity" returns no match. Amazon is funding this on its own balance sheet, in public debt markets, with full disclosure. On the specific axis the assignment asked about, Amazon's structure is the cleaner of the two — see the Meta memo §4 for the contrast.

---

5. The depreciation useful-life check — Amazon went the OPPOSITE way to the hypothesis

The assignment flagged the NBIS precedent (a 4→5-year extension flattering near-term margins) and asked whether either company had done the same. For Amazon the answer is no — it did the reverse, and disclosed why.

5.1 The full history, from the filings

Amazon's stated reason for shortening AI server lives, verbatim from the 10-K:

"The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning."

5.2 The finding that is actually under-modelled

The headline "Amazon shortened server lives" is a quality signal. But the two changes landed in opposite segments, and that is not widely modelled:

• AWS took the hit: −$1.4bn of additional FY2025 D&A, explicitly allocated primarily to AWS.
• Retail took the benefit: +$0.9bn of FY2025 operating income, explicitly allocated to Fulfilment, i.e. to North America and International.

FY2025 North America + International operating income was $34,369m, up $5,610m year on year. Approximately $0.9bn — 16% of the entire year-on-year improvement in retail segment operating income — is an accounting-estimate change, not operating leverage.

Anyone extrapolating "Amazon's retail margin is inflecting" from the FY2025 segment print is extrapolating a one-time, non-recurring accounting benefit as if it were a trend. This is the sharpest genuinely under-modelled item on Amazon. It is also, honestly, too small to move a $2.5 trillion stock — which is exactly why it does not rescue Gate 2.

5.3 Where Amazon stands versus peers on server life

Amazon and Meta moved in opposite directions in the same month on substantially the same asset class. That single fact is the connective tissue between these two memos, and it is developed further in the Meta memo §5.

---

6. Bottom-up TAM — the incremental AI opportunity, not the total

Per references/tam-sizing.md. For a mega-cap the live question is the incremental AI opportunity against an already-enormous base. No company TAM framing is used as an input — the NET lesson (a company expanding its own denominator by press release) applies.

6.1 Units build

6.2 Penetration path and time-to-revenue

Time-to-revenue: short, and this is Amazon's real structural advantage over pre-revenue AI names. Capacity placed in service bills within one to two quarters. There is no clinical trial, no regulatory gate, no qualification cycle. The FY2025 capex is already earning in FY2026.

6.3 REQUIRED implied-penetration statement, in physical units

Amazon's $2,488bn market capitalisation, capitalised at 25x sustainable earnings, requires roughly $100bn of annual net income. FY2025 net income was $77.7bn, of which $15.2bn was an unrealised mark on Anthropic — underlying earnings were therefore about $62.5bn. Closing the ~$37bn gap from AI compute, at AWS's 35.4% operating margin and a 20% net margin, requires roughly $185bn of incremental annual AI revenue. At $20m of revenue per MW-year, that is approximately 9 GW of net new, fully-utilised, fully-monetised AI capacity — against the ~3 GW Amazon added in FY2025 at a record $96.5bn of AWS capital spending, and against a US electricity system that added roughly 50 GW of total new generation capacity in 2025.

Honest reading of that sentence: it is demanding, but it is achievable. Three years of building at the current record pace, fully monetised, gets there. That is a materially less extreme requirement than the SMR case (18.5 GW implied against 0.0 GW of binding orders) or the MU case ($83bn of sustainable earnings against a 16-year average of $3.4bn). This is precisely why Amazon is not a short. The implied penetration is a stretch, not a fantasy.

6.4 Does consensus already embed it? — Gate 2B leg 4

Yes, and more. Street consensus price target is $312.19 across 36 analysts (high $370, low $175) against a $231.39 spot — the Street is 35% above the market. Consensus already models AWS at ~$40.5bn for Q2-2026 (+28% year on year) and carries positive revision drift. There is no gap between the Street's numbers and this TAM build. The TAM work here is context, not edge, and this memo says so.

6.5 EV/TAM secondary check, with its denominator's elasticity stated

EV ≈ $2,569bn (including lease liabilities). Against a bottom-up FY2030 AI-compute revenue pool of ~$250bn plus Amazon's existing ~$717bn revenue base, EV/forward-revenue is ~2.1x. Stated as a sanity check only. The denominator is elastic — Amazon could redefine "AI revenue" to include Alexa+, Rufus, advertising ranking, and robotics at will. Never used as the headline, never compared across sectors.

---

7. Regime-change test — tiered evidence, duration matched

Per references/regime-change-test.md. The claim under test: AI has structurally changed AWS's growth and margin trajectory.

7.1 Duration matching — and why it cannot be applied here

The reference file's core discipline is to model the contracted tranche at contracted terms for exactly as long as the contracts run, then revert. That discipline cannot be applied to AWS revenue, because Amazon discloses no contracted revenue tranche at all. There is nothing to duration-match.

It can be applied to the cost side, and the model does so:

7.2 Principle 5 signature test — does the evidence raise the floor or the ceiling?

A genuine regime change truncates the left tail. Amazon's Tier 1 evidence is all obligations and no entitlements, which does the opposite: it raises the floor on costs while leaving revenue merchant.

Conclusion: Amazon fails the regime-change test. The AI story is a Tier 3/4 argument about demand, wearing an evidence costume made of Amazon's own spending commitments. The model therefore mean-reverts AWS growth from 27% to 14% over five years, which is with the historical base rate, not against it.

---

8. Business, management, competition, ownership

8.1 What the company is, in numbers (FY2025, from the R-files)

AWS is 18% of revenue, 57% of operating income, and 68% of capital spending. Amazon is a cloud-infrastructure company with a very large retail business attached, and the capital allocation says so unambiguously.

By product line: Online stores $269.3bn; 3P seller services $172.2bn; AWS $128.7bn; Advertising $68.6bn; Subscription $49.6bn; Physical stores $22.6bn; Other $5.9bn.

By geography: United States $489.7bn (68%); Germany $45.9bn; UK $43.2bn; Japan $30.7bn; Rest of world $107.5bn.

8.2 Management and capital allocation

Andy Jassy (CEO since 2021, previously AWS CEO). The capital-allocation record in the AI era is the thing to judge, and it is unusually aggressive: PP&E additions went $48.3bn (2023) → $85.8bn (2024) → $142.4bn (2025), funded increasingly with debt ($53.8bn issued in Q1-2026 alone, some maturing 2076). Trailing-twelve-month free cash flow has collapsed from $25.9bn to $1.2bn.

Two capital-allocation decisions in April 2026 deserve note: the $10.9bn Globalstar acquisition (satellite, with accompanying Apple agreements) and the credit line to Anthropic. Both are disclosed as subsequent events in the Q1-2026 10-Q and post-date most published models.

8.3 Competition

AWS versus Microsoft Azure and Google Cloud in infrastructure; versus Anthropic/OpenAI at the model layer, where Amazon has visibly chosen partnership over competition (§4, the Nova→0 finding). In retail, Walmart, Temu/Shein at the low end, and a rising share of commerce search moving to AI assistants. In advertising, a $68.6bn business growing ~22% and taking share from Google and Meta at the bottom of the funnel.

8.4 Ownership and positioning

Positioning conclusion: the Street is 35% above the market and has been raising targets into a flat tape. That is a bearish configuration in the abstract — estimates that have not yet been cut. But it is not a variant, because I have no differentiated number to put next to theirs.

---

9. Risks

To any long: the depreciation wall (D&A modelled to rise from $41.9bn to $140.2bn by FY2030, absorbing all of the EBITDA margin expansion); economic free cash flow negative through FY2030 on the base case; $122.6bn of debt and rising; earnings quality contaminated by Anthropic marks; power availability as a hard physical constraint; antitrust exposure in retail and advertising.

To any short: Amazon is a genuine compounder with a 35.4%-margin cloud business re-accelerating; the Street is 35% above spot with positive revision drift; momentum is neutral not negative; the useful-life shortening is a sign of conservative accounting; there is no leverage crisis and no covenant risk; and — decisively — there is no negative variant versus consensus. Shorting a mega-cap on valuation alone is the exact error this skill exists to prevent.

---

10. What would change this memo

1. AWS growth decelerating below ~20% while capex guidance stays above $200bn — the first sign the spend is not earning.
2. Any Anthropic down-round, or a reversal of the accumulated ~$32bn of marks. This would hit reported EPS by several dollars with no change in the business.
3. Retail segment operating margin failing to expand in FY2026, revealing the FY2025 improvement as the accounting change it partly was.
4. A further useful-life extension on AI servers — which would reverse the quality signal in §5.
5. Amazon beginning to disclose AWS RPO or contracted backlog, which would for the first time make the regime-change test applicable.