Amazon (AWS)

Refresh delta (since 2026-07-22)

What changed — six material moves

1 · The capex guide was RAISED to $220B, the basis was stated for the first time — "cash CapEx" — and the driver named was memory. Jassy, verbatim: "We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from our prior estimate of about $200 billion. Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027, too.". In the same passage he retroactively re-characterised the February guide the same way: "Earlier this year, we said we plan to invest approximately $200 billion in cash CapEx in 2026." The prior dossier's single largest open question — what basis is the $200 billion, when the four defensible bases differ by ~$46B annualised — is now settled from management's own mouth. It is the net cash line, Amazon's own MD&A convention and the denominator of its FCF definition.

2 · AWS accelerated again, to +37%, and the five-quarter TTM margin decline broke. AWS Q2 net sales $42,232M, +37% y/y (identical ex-FX), operating income $16,621M, margin 39.4%. The prior dossier flagged TTM AWS operating margin falling five consecutive quarters, 37.5% → 35.2%. It is now 36.85% `` — up 165bp. Both of the prior dossier's headline AWS conditions cleared.

3 · TTM free cash flow went negative — the single condition the destructive reading needed, and it is the one it got. Filed, non-GAAP reconciliation: TTM FCFFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. $(7,604)M versus $18,184M a year earlier. The prior dossier's mark was "it is $1.2B; a Q2 with capex > OCF flips it." Q2 OCF was $45,387M against $53,076M of net cash capex. It flipped.

4 · The backlog went $364B → $496B and the weighted-average life 5.5 → 6.4 years — but the driver is once again ONE counterparty. Verbatim: "those commitments not yet recognized were approximately $496 billion as of June 30, 2026. The weighted-average remaining life of our long-term contracts is 6.4 years.". Attribution, same note: "In Q2 2026, AWS and Anthropic announced an expansion of the strategic collaboration and existing multi-year commitment by more than $100.0 billion over 10.0 years, which includes contractual obligations related to the performance of AWS chips." The prior dossier's condition was "backlog grows again and the growth is not dominated by a single counterparty." It grew; the breadth half failed for the second consecutive quarter, with a different single name.

5 · Amazon's balance sheet became, in one quarter, a leveraged holder of two private AI labs — and 81% of the quarter's net income is that mark. Other income (expense), net was $53,415M in Q2, of which $50,486M is "upward adjustments relating to equity investments in private companies… primarily from our nonvoting preferred stock in Anthropic". Net income $62,647M; $50,486 ÷ $62,647 = 80.6% . Carrying values at 6/30/26: Anthropic nonvoting preferred **$92.5B** and Anthropic convertible notes at fair value **$97.9B** (a combined **~$190.4B** against $18.0B of cash actually invested ``); OpenAI Series C $28.7B, with the remaining $21.3B of the commitment invested after quarter-end. Total assets crossed $1,095,689M — Amazon is now a trillion-dollar-balance-sheet company, and roughly $122.3B of that is Level 3 private marks.

6 · New: a $20.0B vendor-financing facility for Anthropic that unlocks against AWS compute delivery. Verbatim: "we entered into a financing arrangement to make available to Anthropic an aggregate facility not to exceed $20.0 billion… At inception, there is no amount available to be drawn against and as we reach certain delivery milestones of compute capacity under the amended commercial arrangement, amounts under this facility are made available for Anthropic to draw upon at its discretion." Reduced to $15.0B after Amazon put $5.0B into Series H under it. This is a materially different object from the equity stakes the prior dossier described. It is a supplier extending credit to a customer, released by the supplier's own delivery. Say what it is; do not editorialise past it.

What held

The structural thesis stands, and it strengthened. Three engines; AWS now 21% of revenue (from 18%) and 61% of consolidated operating income ``; advertising $19,809M, +26%; the widest moat stack in mega-cap tech. Nothing cracked.

The desk's memory-premium argument survived and got a number — but it landed somewhere the prior dossier did not predict. The prior dossier's destructive condition read "memory costs pass through to AWS gross margin rather than being absorbed." Memory costs did not hit the AWS margin, which rose. They hit the capital programme: +$20B, +10% on a $200B base, named by the CEO. That is the sharpest mechanism this refresh found, and it generalises: at a hyperscaler, memory-price inflation is a capex event first and a depreciation event later — not a gross-margin event. It gets capitalised into a 5–6-year asset and released into the P&L over the following half-decade rather than compressing this quarter's margin. Anyone watching cloud gross margin for evidence of the memory squeeze is watching the wrong line.

Management's defence of 5–6 year server lives hardened, and it added a number the desk did not have. Jassy: "For servers and networking equipment, on average, it takes a little less than three years to break even on that investment. The servers currently have a useful life of at least five to six years, and most of our AI capacity these days is being contracted for at least five-year terms.". Contract duration matched to asset life is the strongest form of the argument available, and it is new disclosure.

Scoring the prior dossier's falsifiable conditions

The prior pass wrote two lists precisely so this one could grade them. It graded 4½ / 5 one way and 1½ / 5 the other.

The honest read of that scorecard: the print resolved decisively toward the productive reading on every operating metric, and the two things that got worse are both disclosure problems rather than performance problems — counterparty concentration and silicon opacity. That is a specific, testable shape, and it is not the same as "the bear case is dead."

---

Refreshed lenses

Segments (the quarterly series, filed)

$M, all for the 2026-Q2 column; earlier columns.

(Q3/Q4 2025 and Q1 2026 consolidated rows are segment sums ``; the Q2 2026 column is filed directly.)

Three things to read off this table, in order of how easy they are to get wrong.

One — the consolidated +20% is flattered by a calendar shift and the Q3 guide is depressed by the same shift. Prime Day moved from July 2025 to June 2026. The Q2 guide said so explicitly, and the Q3 guidance says "Excluding the impact of Prime Day in both 2025 and 2026, third quarter 2026 year-over-year growth would be nearly 400 basis points higher". So Q2's +20% and Q3's guided +9–12% are the same underlying business seen from two sides of a shifted event; H1's +18% is the honest number. AWS is untouched by this — Prime Day does not move cloud.

Two — the AWS margin print contains a non-operating gain, and management said so. The 10-QForm 10-QThe quarterly version of the annual report. Lighter, and not audited. discloses "net unrealized gains of $551 million in Q2 2026… recorded within 'Technology and infrastructure' and primarily impacting our AWS segment" from energy-contract derivative remeasurement. Strip it: 38.05% ``, +515bp y/y. Olsavsky gave ~520bp of underlying improvement on the call, which reconciles. Amazon's Q3 guidance explicitly "assumes… no impact from energy derivative contract remeasurements" — so this item does not recur by assumption, and a Q3 margin comparison against the reported 39.4% will look worse than the business is.

Three — consolidated operating income also carries ~$1.2B of one-off benefit. $640M of IEEPA tariff refunds recorded as a reduction of cost of sales, "primarily impacted our North America segment", plus the $551M derivative gain — $1,191M combined, which the call summarised as "a $1.2 billion benefit from tariff refunds and energy derivative accounting gains." The 10-Q adds that the tariff refund "represents the significant majority of refunds we expect to receive" — so it is close to non-repeating. Operating income ex both items ≈ $26,270M ``, still above the top of the $20.0–24.0B guide.

---

Capex — the definition question is answered, and the number went up

The basis, settled

The prior dossier laid out four defensible "capex" numbers differing by ~$46B annualised and said a CFO reconciliation "or its continued absence" would answer it either way. Management answered it directly: cash capex, twice, in one paragraph. That is the net line — "Purchases of property and equipment, net of proceeds from sales and incentives" — Amazon's own MD&A term and its FCF denominator.

The full ladder, refreshed

$M. Cash-flow lines and the supplemental cash-flow note from the Q2 2026 10-Q.

(Every input row is a filed line item; the subtotals are arithmetic. The ladder's accrual subtotal runs ~$450M above the filed "net additions" line, which is net of disposals — use the filed figure where one exists.)

Two things the prior dossier called correctly and one it under-weighted. It said finance leases were not the material swing — $4,048M on a $169,007M TTM base is +2.4%, so that holds. It said the genuinely material definitional swing was the payables accrual — and that has now grown to +$29,267M TTM (+17.3%), up from $16,967M a quarter earlier. Amazon is running a $29B-a-year supplier float on its capital programme, and it is growing faster than the programme. What it under-weighted: on an accrual basis the TTM figure is already $202.3B `` — i.e. Amazon has already built $200B+ of property and equipment in the last twelve months, before the raised guide takes effect.

The $220B pace mark, for the Q3 print

The single most useful Q3 number: Amazon's capex is Q4-weighted, so a straight-line split overstates what Q3 must carry. A Q3 cash-capex print in the $56–62B range is consistent with the guide; below ~$54B and the raise is not being executed on the cash line, whatever the guide says. That is a clean, falsifiable mark.

How the capex is being funded — the debt page turned

Operating cash flow TTM $161,403M against TTM cash capex $169,007M → TTM FCF $(7,604)M. The gap is entirely debt-funded.

• Face value of long-term debt: $68,836M (12/31/25) → $132,995M (6/30/26) — +93.2% in six months ``
• The 10-Q discloses issuances the prior dossier did not have: March 2026 US$37.0B (2028–2076, stated rates to 6.05%), March 2026 €14.5B ($16,550M carrying), May 2026 CHF2.8B ($3,487M), June 2026 C$14.0B ($9,853M). Four currencies in four months.
• Subsequent to quarter-end, the US$25.0B July deal the prior dossier covered (to 6.25%, 2066) → pro-forma face ≈ $158.0B ``
• New and previously unseen: a $17.5B unsecured delayed-draw term loan entered in June 2026, SOFR + 0.625–0.875%, three-year maturity, drawable in a single draw through 2026-09-30, undrawn at quarter-end. That is $17.5B of committed-but-invisible capacity with a hard expiry inside Q3 — a specific thing to watch on the Q3 balance sheet.
• Interest expense Q2 $1,314M vs $516M (+154.7%); H1 $2,114M vs $1,057M
• Estimated fair value of the Notes $123.8B against $132,995M of face — the debt trades ~6.9% below par ``, consistent with a long book (combined weighted-average remaining life 14.2 years) issued into a higher-rate market
• Total contractual commitments $650,034M at 6/30/26, from $569,280M at 3/31/26 (+14.2% ``) — including leases not yet commenced $137,214M (from $106,347M) and unconditional purchase obligations $130,065M (from $103,768M)
• No buybacks in H1 2025 or H1 2026; $6.1B remains authorised
• Commercial paper: zero outstanding; revolvers $20.0B undrawn

Two plain observations. First, Amazon has issued in USD, EUR, CHF and CAD inside five months and taken a $17.5B term loan on top — that is a funding programme being deliberately diversified across every available pool, which is what a borrower does when it expects to keep coming back. Second, $20.7B of the Euro- and Canadian-dollar notes are designated as net investment hedges, so the currency diversification is doing balance-sheet work as well as funding work.

---

AWS growth and the performance-obligation backlog

Growth

• AWS Q2 net sales $42,232M, +37% y/y — described on the call as an annualised run rate of $169 billion and the fastest growth in 18 quarters
• Sequential add +$4,645M vs Q1 2026 `` — versus +$2.0B in Q1, itself called the largest Q4-to-Q1 add ever
• Five consecutive quarters of acceleration: 17% → 20% → 24% → 28% → 37%
• AWS AI business >$25B annual revenue run rate, triple-digit % growth — up from >$15B disclosed in Q1
• AWS TTM revenue $148,404M ``, +27.6% on the prior TTM

The backlog

Derived ``:

• Q2 change: +$132B (+36.3%) and +0.9 years of duration
• Two-quarter change: +$252B (+103.3%) — which is where management's "growing at a triple-digit percentage rate" comes from
• Backlog ÷ AWS TTM revenue = 496,000 ÷ 148,404 = 3.34× (was 2.66× at Q1)
• Naive annualised release = 496,000 ÷ 6.4 = $77,500M/yr, i.e. the backlog underwrites 52.2% of the current AWS run-rate (was 48%)

The recognition disclaimer travels with the number, unchanged and verbatim: "The amount and timing of revenue recognition will be driven by customer usage and our performance in accordance with contractual obligations, which can extend beyond the original contractual duration and commitment.". $496B ÷ 6.4 years is not a revenue annuity; Amazon says so in the same paragraph that gives the figures.

And the concentration caveat has now compounded rather than resolved. Two consecutive quarters of ~$100B+ backlog additions, each attributed by the filing to a single named counterparty — OpenAI ($100.0B / 8.0 years, Q1) then Anthropic (>$100.0B / 10.0 years, Q2). Amazon holds $28.7B of OpenAI equity (plus $21.3B invested after quarter-end, $50.0B total) and ~$190.4B of Anthropic carried value against $18.0B invested, and has extended Anthropic a $15.0B undrawn facility that unlocks as AWS delivers compute. Every one of those is a filed fact. What the pattern means is a judgement this pass does not make; see Lens 13.

---

Earnings calls — the sentiment arc, now three deep

Shelf went 2 → 3 transcripts. The arc across Q4 2025 → Q1 2026 → Q2 2026 is unusually legible.

Q4 2025 (2026-02-05) — assertion. "We expect to invest about $200 billion in capital expenditures… because we have very high demand." Olsavsky, pressed by Evercore on ROIC duration and an FCF floor, gave neither.

Q1 2026 (2026-04-29) — justification. Jassy pre-emptively laid out the 6–24-month monetisation lag and the 30+/5–6-year asset lives before anyone asked, and first named memory: "the cost of components, particularly memory, has skyrocketed."

Q2 2026 (2026-07-30) — quantification and commitment. The tone moved a third step: from explaining why the FCF line looks bad to naming the payback period, the contract duration, and the abandonment option, and to raising the number rather than defending it. The load-bearing new passage:

"There are two major parts of the investment, the data centers and the servers and networking equipment that go into them. These have different capital cycles. Data center capital is spent starting two years before we can put servers into them to start monetizing… Servers and networking equipment operate on a shorter cycle. We typically purchase these a few months before putting them into service, so we have strong visibility into customer demand before we trigger the spend. If the demand isn't there, we won't spend the capital. For servers and networking equipment, on average, it takes a little less than three years to break even on that investment. The servers currently have a useful life of at least five to six years, and most of our AI capacity these days is being contracted for at least five-year terms."

Why that paragraph matters more than anything else on the call. The prior dossier modelled a single blended 6–24-month lag. Management has now split it into two cycles with very different risk profiles: a two-year, largely irreversible shell/land/power commitment, and a few-months-ahead, cancellable server commitment. If that decomposition is accurate, the abandonment risk in the $220B programme is concentrated in the data-centre leg, not the silicon leg — which is the opposite of how the capex-overbuild argument is usually framed. It is management's claim, not a verified fact, and "if the demand isn't there, we won't spend the capital" is exactly the sentence a Q3 or Q4 capex miss would be explained with.

Ambition re-based, twice over. "We long believed AWS could become a few hundred billion-dollar revenue business and now believe it'll be at least double that, and very possibly be a trillion-dollar annual revenue business for us in time." And on capacity: "Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027… In fact, the demand we already have for 2028 is striking."

New risk language. Olsavsky: "we face heightened transportation costs driven by fuel inflation from the conflict in the Middle East and higher line haul rates from driver capacity limitations." Shipping costs $27,913M, +19% y/y. Memory chips remain in the formal safe-harbour list, as they were in Q1.

A strategy disclosure worth its own line. Jassy confirmed Amazon is building its own frontier model: "we are pursuing our own frontier model… it just gives us additional control over cost… My view of it is that within the next few years, you're going to have at least a half dozen models that are comparably good to each other. They'll all be in Bedrock, one of them will be ours." Amazon is simultaneously the largest external investor in two frontier labs, their compute supplier, their distribution channel via Bedrock, and now a declared competitor. That is four positions in one value chain.

---

Custom silicon — the numbers moved, the opacity did not

Hard — filed or explicitly quantified

The scope caveat did not just survive — it got stronger against the desk's interest. The $25B run-rate is still Graviton + Trainium + Nitro combined, and in the same quarter Amazon put Graviton5 into general availability and disclosed Graviton revenue commitments up nearly 3× quarter-over-quarter. Mechanically, that makes the Graviton share of the $25B larger, not smaller. Any model treating >$25B as an AI-accelerator number is now overstating Trainium by more than it was a quarter ago. Trainium-only revenue: n/a, and the disclosure moved away from sourceability this quarter, not toward it.

The prior dossier's second open question also went unanswered. It asked whether the >$225B of Trainium revenue commitments sits inside or outside the $364B backlog. The Q2 10-Q gives a $496B backlog and two filed contract expansions that "include contractual obligations related to the performance of AWS chips" — but no reconciliation between chip commitments and the performance-obligation total. Still n/a.

The margin claim, unchanged and still forward-looking

Jassy's Q1 claim — "at scale, we expect Trainium will save us tens of billions of dollars of CapEx each year and provide several hundred basis points of operating margin advantage versus relying on other chips for inference" — was not repeated with a realised number on the Q2 call. It remains an expectation. Against a $220B programme, "tens of billions" is a ~10–20% capex avoidance ``.

And the counter-fact hardened into a guidance revision. Amazon designs its own accelerator and still buys HBM/DRAM in a market it does not control — and this quarter that fact cost it $20 billion, by the CEO's own attribution. Custom silicon removes the merchant-accelerator margin; it does not remove the memory premium. The desk's argument is now sourced twice from the buyer, and the second time it came with a price tag.

Merchant Trainium — still reported intent, no signed transaction

External Trainium sales remain talks. The strongest management framing available is "a real chance we'll do that in the future", against Peter DeSantis's June/July confirmation of early-stage negotiations. No signed, disclosed transaction. Treat as intent.

---

Price catalysts

Window summary: $247.55 (2026-07-21 close, prior dossier) → $279.30 (2026-08-10, 11:43 ET) = +12.8% . Market cap **$3.01T** — Amazon crossed $3 trillion in this window. 52-week range 196.00–287.20, so the stock sits ~2.8% off its high .

The catalyst pattern held exactly as the prior dossier described it. AWS growth rate moved the stock; the negative FCF print — which was the most alarming single number in the release — did not. A 15% single-day move on an acceleration from 28% to 37% while TTM FCF went from +$1.2B to −$7.6B is a clean, dated demonstration that this market is currently pricing AWS revenue growth over cash conversion. That is a fact about the market's present preference, not a claim that the preference is correct.

Analyst reaction: more than a dozen banks raised targets within hours — Benchmark to $400 (from $370), JPMorgan to $365 (from $330), Citi to $350 (from $325), Morgan Stanley to $335 (from $330). Consensus average target moved $313.13 → $324.94 (60 analysts, Strong Buy; low $230, high $400).

---

Management (correcting the prior dossier)

The prior dossier said "No management change disclosed." That was wrong at the time of writing.

Dave Brown, SVP and an S-team member with ~19 years at Amazon, was announced as departing on 2026-07-15 — seven days before the prior dossier's boundary — with his last day at end-July. Dave Treadwell, previously leading the e-commerce foundation unit and also an S-team member, took over effective 2026-08-01. Brown ran the compute organisation that owns EC2 and the custom-silicon-adjacent surface, so this is not a peripheral seat.

Handling note: the announcement predates the boundary (a prior-dossier miss, corrected here); the effective transition is post-boundary. Both are stated so the record is honest about which is which.

Other: no change at CEO or CFO. Insider selling of ~$46.6M over the trailing three months; Bezos's November 2025 10b5-1Rule 10b5-1 planA schedule an insider sets in advance saying when their shares will be sold, so the sales cannot be timed on private news. A sale made OUTSIDE such a plan is a live decision, which is why it carries more information. plan (up to 15M shares through February 2027) continues. No buybacks in H1 2026.

---

Depreciation policy and the conveyor

Policy — no revision, and a strengthened defence

No useful-life revision was disclosed in the Q2 10-Q. Note 1 carries the boilerplate only: "We review the useful lives of equipment on an ongoing basis." The full history stands as the prior dossier corrected it — 4→5 years (Jan 2022), 5→6 (Jan 2024, worth +$3.1B of 2024 operating income), and a subset of servers and networking equipment 6→5 effective 2025-01-01 (+$1.4B D&A, −$1.0B net income, −$0.10/sh, already inside the FY2025 base).

What is new is the defence, quoted in full in Lens 6: break-even "a little less than three years", useful life "at least five to six years", AI capacity "contracted for at least five-year terms", and an explicit claim of "a strong track record of… finding ways to extend the useful life of this equipment." Read plainly: management is signalling the next revision is more likely to be a lengthening than a shortening. The February 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes. remains the likelier venue for any change.

Sizing the conveyor

$M, ``.

The fact that does not fit the simple bear model. AWS depreciation grew 66.7% y/y while AWS revenue grew 37% — and the AWS margin still rose 515bp on an underlying basis. Depreciation outran revenue by ~30 points and the margin expanded anyway. Whatever "efficiency gains and server capacity optimization" means operationally, the arithmetic says the operating leverage on the rest of the AWS cost base absorbed a two-thirds increase in the depreciation line. That is the single strongest data point available for the productive reading of this build, and it is filed.

The fact that does not fit the simple bull model. AWS net additions to P&E were $48,604M in one quarter against $42,232M of quarterly revenue — AWS is adding property and equipment faster than it is billing. Some large share of that is construction in progress that has not started depreciating; the 10-Q does not disclose interim CIP, so the size of the not-yet-depreciating pool at 6/30/26 is n/a. It was $71.7B at FY2025 year-end and rising 53.8% y/y. The conveyor's next twelve months are already bought and paid for; they are simply not in the D&A line yet.

Sensitivity for the 3-year-life debate: unchanged and still blocked. Amazon does not disclose the compute-only share of net P&E, so a precise sensitivity is n/a. Write nothing plausible in its place.

---

Earnings quality — the widest GAAP/operating gap this name has printed

The quarter, two ways

Method A — strip all non-operating income at the disclosed effective rate ``: pre-tax $80,857M − other income (expense) net $53,415M = $27,442M; H1 effective rate 23.0% ($27,759 ÷ $120,691) → tax $6,312M → operating-quality net income ≈ $21,130M; ÷ 10,903M diluted → operating-quality Q2 EPS ≈ $1.94.

Method B — strip only the private-company mark and its own disclosed tax : pre-tax $80,857M − $50,486M = $30,371M; the 10-Q discloses **$15.9B of net discrete tax expense "primarily attributable to the upward adjustments to our investments in Anthropic"** for H1, of which the Q2 share pro-rated on the mark is ~**$12,780M** → Q2 tax ex-mark ≈ $5,419M → net ≈ $24,952M → EPS ≈ $2.29.

Both methods beat the $1.82 consensus. Reported GAAP EPS was $5.75. So the operating beat was real but modest — roughly 7% to 26% ahead depending on method — and the headline was a 216% beat. The prior dossier predicted this shape precisely: "the street is not modelling another large investment mark. If one lands, the GAAP print beats by a wide margin on non-cash income, exactly as in Q1." It landed.

The trailing multiple is now the most misleading number on the tape

At $279.30 (2026-08-10):

Read the top and bottom rows together. Since the prior dossier the stock is up 12.8% and the trailing P/E has fallen from 29.6× to 22.5× — because a $50B non-cash private mark entered the denominator. On operating-quality earnings the multiple went the other way, 38.5× → 41.2×. Amazon looks meaningfully cheaper on the screen and is meaningfully more expensive on cash earnings. That is a screening artefact of some size, and it is the kind of thing worth naming in public.

A consensus-line warning that travels with it. FY2026 consensus EPS of $12.30 embeds H1's actual GAAP $8.53, most of which is the mark. FY2027 at ~$9.46 does not assume a repeat. So the consensus series shows an apparent −23% EPS decline in 2027 that is nothing of the kind — it is the mark washing out. Anyone comparing FY2026 and FY2027 P/Es on this name without that adjustment is reading a fake earnings collapse.

Equity quality, same point from the other side. Total stockholders' equity $551,620M includes AOCI of $66,287M, of which $69,647M is unrealized gains on available-for-sale debt securities — the Anthropic convertible notes. A material slice of the six-month equity build is an unrealized Level 3 private mark, not retained operating earnings.

Regulatory (not re-run this pass). regulatory/regulatory-findings.md still carries last_fetched: 2026-07-22 — 0 SEC Litigation Releases, 0 AAERs over 2021-07-22 → 2026-07-22. Not refreshed here; treat the SEC read as 19 days stale. Two items from outside EDGAR that the prior dossier did not carry:

• Xockets, Inc. v. Amazon/AWS/Annapurna/NVIDIA/Microsoft — an ITC complaint filed May 2026 over EC2 P6e-GB200 UltraServers, DGX Cloud on AWS, SageMaker HyperPod and EKS, seeking injunctive relief; the ITC instituted an investigation in June 2026. Separate district-court complaints from June 2025 target the AWS Nitro System. An ITC exclusion order is a different risk shape from damages litigation.
• EU cloud-competition probe into Microsoft Azure and AWS under the digital-competition rulebook. Pre-boundary, newly surfaced.
• FTC Prime settlement ($2.5B, Sep 2025) remains closed and booked.

---

Forward projection

Model workbook — integrity flags, stated before any number

our model was re-run after the CSVs were populated. It produced a workbook with no usable outputs, and this dossier does not cite one. The flags, verbatim from the tooling:

• quarters: 3 (2025-Q2 → 2026-Q2) — "only 3 quarter(s) of financials — LTM figures are partial"
• "fewer than 8 quarters — revenue growth seeded at 8%" (an arbitrary default, not a research input)
• opening BS: partly sourced (9/10 lines) — "opening balance sheet is partly conventional, not sourced… Every valuation output rests on it."
• our model: "No computed values. The workbook has 51 formulas and no cached results… Nothing below is a real number yet."
• Share price on the Assumptions sheet: blank

So every figure below is hand arithmetic, labelled ``, with the working shown — not a model output. The workbook is a scaffold for a future pass, and the honest statement of its current status is that it computes nothing.

The base the projection starts from

FY2025 actuals ``: revenue $716,924M (cross-checks exactly against the $716.92B on ), operating income $79,975M.

FY2026 build ``: H1 operating income $51,313M (filed) + Q3 guide midpoint $24,500M + Q4 at +30% on Q4 2025's $24,977M ≈ $32,500M → FY2026 operating income ≈ $108,300M, +35.4% on FY2025.

Operating-quality FY2026 EPS ``: $108,300M, net interest roughly neutral (interest income ~$5.0B against interest expense running toward ~$6B annualised post-issuance), taxed at 23% → ~$83,400M ÷ ~10,950M diluted → ≈ $7.61.

Cross-check that this is not fantasy: H1 operating-quality EPS is $3.65 ``, implying ~$3.96 in H2. Consensus FY2026 GAAP EPS of $12.30 less H1 actual $8.53 implies $3.77 for H2 — i.e. the street is modelling essentially no further marks and operating earnings a touch below this estimate. The two reconcile to within 5%.

Three-year forward operating-quality EPS path

These are mark-free operating EPS. They are NOT comparable to GAAP consensus, which embeds private-company marks. Do not put them side by side without the adjustment.

All ``; arithmetic is $108.3B × (1+g)³, less net interest, at a 23% tax rate, over a diluted count grown ~1%/yr for SBC net of the (currently dormant) buyback.

What each case requires, stated so it can be falsified:

• Bear (10%): AWS growth decelerates back toward the high teens as the 2026–27 capacity lands, the depreciation conveyor from a $220B programme catches the margin, memory pass-through finally reaches the P&L, and interest expense compounds on a $158B+ debt book. Roughly consistent with the AWS TTM margin resuming its decline.
• Base (22%): AWS holds the high-20s/low-30s for another year then decelerates; advertising compounds mid-20s; the operating leverage that absorbed 66% AWS depreciation growth this quarter keeps working; retail margin grinds up. This is close to the FY2027 consensus check below.
• Bull (30%): the $496B backlog converts at or above its naive schedule, AWS approaches the "at least double a few hundred billion" trajectory management now claims, Trainium's stated inference margin advantage becomes realised rather than expected, and the external-Trainium option converts.

FY2027 sanity check ``: base gives $108.3B × 1.22 = $132.1B operating income → ~$8.79 operating-quality EPS; bull gives ~$9.39. The market's FY2027 implied consensus of ~$9.46 therefore sits at or slightly above this bull case on an operating basis — meaning consensus is either assuming continuing non-operating income, a lower tax rate, or faster operating growth than modelled here. That gap is stated, not resolved. It is the most useful single disagreement this pass surfaced.

Peer comps (Lens 7 comparables): not re-run this pass. Forward multiples and 5-year average ROE for MSFT/GOOGL/ORCL remain n/a, carried from the previous dossier.

No Brier forecast logged — dispatched, unattended run; no committed base case (see the binding rules).

---

Bull vs bear, with a pre-mortem

Bull, in one paragraph

AWS is accelerating (17% → 37% over five quarters) with a $496B / 6.4-year contracted backlog behind it, and the margin is expanding while depreciation grows 67% — which is the empirical answer to the "AI capex destroys cloud economics" argument. Management has now put a payback number on the asset (<3 years to break-even against 5–6-year lives and ≥5-year contracts), named the abandonment option on the server leg, and raised the capital programme 10% because demand exceeds a $200B build. The retail engine is quietly excellent — advertising +26%, NA margin 7.9%, worldwide paid units +17%. And the balance sheet holds ~$190B of Anthropic against $18B invested plus $50B of OpenAI, with an Anthropic IPO reportedly targeted at ≥$1T. On operating-quality earnings the stock is ~41× — expensive, but against a business compounding operating income >30%.

Bear, in one paragraph

TTM free cash flow is −$7.6B and heading further negative as a $220B programme runs against $161B of operating cash flow; the funding gap is being closed with debt that nearly doubled in six months to $133B face (pro-forma ~$158B, plus $17.5B of undrawn term loan), at coupons up to 6.25%, on assets management itself dates at 5–6 years — long money against short assets. Eighty-one percent of the quarter's net income is a Level 3 mark on a private company, and $92B of stockholders' equity is an unrealized gain on convertible notes in that same company; both reverse if the AI funding market re-rates. Two consecutive quarters of $100B+ backlog additions came from two counterparties Amazon simultaneously funds — with a $15B facility that unlocks as AWS delivers compute, which is a supplier financing its customer's purchases. Meanwhile the thing that is supposed to be the cost advantage — custom silicon — has produced no disclosed Trainium-only revenue in five quarters of asking, while the combined chips number is increasingly Graviton (commitments up 3× q/q).

Pre-mortem — it is 2028 and this was a mistake. What happened?

1. The backlog was circular and it unwound together. One or both of the two counterparties re-rated in a private-market correction; Amazon wrote down $100B+ of marks, the backlog additions attributed to them were renegotiated rather than paid, and the vendor facility converted from an asset into a loss. The tell would be a downward adjustment appearing in "Other income (expense), net" — the same line, in reverse.
2. The two-year data-centre leg was the irreversible one and demand arrived late. Management's own decomposition says the shells are committed two years ahead. If AI demand growth slows in 2027, the server leg can be cancelled but ~$137B of leases-not-yet-commenced and $130B of unconditional purchase obligations cannot.
3. Memory kept going. The +$20B raise was the first pass-through, not the last. A second and third raise turns a 10% capex overrun into a structural margin problem once the assets enter service on 5–6-year lives.
4. The useful life was too long after all. Management is signalling lengthening, not shortening. If a competitive reset forces 3–4-year compute lives across the industry, the FY2025 base is understated and every forward margin in this dossier is too high.
5. Interest compounded faster than AWS did. Interest expense grew 155% y/y. On a $158B+ book at 5–6%, that is $8–9B/yr of fixed cost against a business whose cash generation is currently negative.

The falsifier for each is a specific line in a specific filing — the marks line, the commitments table, the capex guide, Note 1's useful-life language, and interest expense. That is deliberate.

---

Devil's advocate

Against the bull, the sharpest available objection is not the FCF number — it is the composition of the growth. AWS accelerated from 28% to 37% in the quarter in which two counterparties Amazon funds signed $200B+ of combined commitments. The 10-Q does not disclose AWS revenue by customer, so the share of the acceleration attributable to Anthropic and OpenAI usage is n/a. It could be small. It could be most of it. Nobody outside Amazon can tell from the filings, and the absence of that disclosure — in a quarter when the company voluntarily disclosed Graviton commitment growth, AI run-rates, and satellite counts — is itself information about what is comfortable to publish.

Against the bear, the sharpest available objection is that the margin arithmetic refuses to cooperate. The bear case requires the depreciation conveyor to compress AWS margins. This quarter AWS depreciation grew 66.7% against revenue growth of 37% — the conveyor arrived, at nearly twice the rate of revenue — and the underlying margin still expanded 515bp. That is not a forecast; it is a filed outcome, and it is the hardest single fact for the overbuild thesis to metabolise.

Against both, and against this dossier: the prior pass made the point that survives everything above — Jassy's disclosed lag means Q2's capex tells you about 2027 revenue, not 2026. That is now more true, because management split the lag into a two-year data-centre cycle and a few-months server cycle. Any judgement rendered on 2026 numbers about whether a 2026 capex programme "worked" is a category error on management's own stated timeline. The desk should keep saying so.

Three things this pass could not resolve, listed so they are not quietly filled in:

1. Trainium-only revenue — n/a, five quarters running, and the disclosure moved further away this quarter.
2. Whether the >$225B of Trainium commitments sits inside or outside the $496B backlog — n/a.
3. Interim construction in progress at 6/30/26 — not disclosed in a 10-Q; blocks a precise estimate of the not-yet-depreciating pool.

---

Carried by reference (unchanged since the prior dossier)

One item the prior two dossiers missed entirely, surfaced here for the record (pre-boundary, not a delta): on 2026-04-13 Amazon signed a definitive merger agreement to acquire Globalstar, Inc. for cash and stock — $90.00/share or 0.3210 AMZN shares, ~$10.9B including debt, expected to close in 2027, with side agreements with Apple (Globalstar's largest customer). It sits alongside Project Kuiper / Amazon Leo at ~400 satellites in orbit, sufficient to begin initial service later in 2026. Neither is load-bearing for the AI thesis; both belong on the record.

---

The five questions the Q3 2026 print can answer

1. Does cash capex step up to the $56–62B/quarter the raised guide requires? H1 ran $96.3B of a $220B guide. A print below ~$54B means the raise is not being executed on the cash line.
2. Does the backlog grow on breadth this time? $496B / 6.4 years is the mark. Two consecutive quarters of single-counterparty additions makes the third the tell.
3. Does the AWS TTM margin hold above 36% without a derivative gain? Q3 guidance explicitly assumes no energy-derivative remeasurement, so Q3 is the clean read the Q2 print was not.
4. Is there finally a Trainium-only number, and does the $17.5B term loan get drawn before it expires on 2026-09-30? One is a disclosure question, one is a funding question, and both are answered by the same filing.
5. Does another private-company mark land — and if the AI funding market cools, does one land in the other direction? The line to watch is "Other income (expense), net," which is now capable of moving GAAP EPS by $4+ in a quarter in either direction.

Secondary watch items: interest expense annualising off $1,314M · whether Note 1 changes any useful-life language · memory pass-through to AWS cost of revenue rather than to capex · the Xockets ITC investigation calendar · Anthropic IPO timing and what it does to the Level 3 marks · external Trainium moving from "a real chance" to a signed transaction · Globalstar closing conditions · Amazon Leo initial service launch.

---