Amazon.com, Inc. — Q2 2026 earnings review

Verdict: ACCELERATING, decisively — and the acceleration is happening in the highest-margin segment. Revenue growth has climbed for five consecutive quarters from a +8.6% trough to +19.6%, with two consecutive ~300 bps steps. AWS accelerated for a fifth straight quarter to +36.8%, the fastest in 18 quarters at a revenue base more than twice the size it was the last time it grew this fast, and AWS operating margin flipped from four straight quarters of year-over-year compression to +645 bps expansion. That flip is the single most important second-derivative datapoint in the print: the capacity being added is landing at high incremental margin, not at a discount. This is a market leader accelerating off a trough with its best business doing the accelerating, which is precisely the set-up the framework rewards.

The headline numbers are both flattered, and you should hold the clean versions. GAAP diluted EPS of $5.75 (+242% YoY) contains $53.4B of non-operating pre-tax other income, primarily the Anthropic investment mark. Adjusted for it, EPS is ~$1.95, a +7.1% beat rather than +216%. Operating income of $27.461B (+43.2%) contains $1.2B of one-time expense relief the CFO disclosed unprompted — ~$600M of tariff refunds in North America and ~$600M from an energy-derivative fair-value remeasurement, primarily in AWS. Clean, operating income is $26.3B, +37.0%. And roughly 430 bps of the revenue acceleration is Prime Day landing in Q2 this year versus entirely in Q3 last year, putting clean revenue growth near +15.3%. Every one of those adjustments still leaves a very strong quarter — but the difference between +43% and +37%, and between +20% and +15%, is the difference between a narrative and an underwriting.

The beat is an AWS beat, full stop. Strip the two one-offs and the entire upside sits in AWS revenue acceleration and AWS margin. Retail did not contribute: North America operating margin is −18 bps YoY excluding the tariff refund, and International margin is flat at 4.1%. Advertising is the second accelerant — $19.8B, +26.2%, breaking out of four straight quarters parked at ~+22%, driven by agentic ad surfaces where a sponsored-prompt click converts 48% more often and spends 21% more.

Guidance reads as a deceleration and is not one. Q3 net sales of $197.0–202.0B print as +9% to +12%, an optical 800–1,100 bps step down from +20%. Management itself quantified ~400 bps of Prime Day timing and ~80 bps of FX inside that; underlying guided growth is ~15.5%, essentially flat against Q2's ~16% underlying. Operating income guided to $22.5–26.5B versus $17.4B last year is where the real conservatism sits — against the clean $21.7B prior-year base it implies only +24 bps of margin expansion after delivering +166 bps clean this quarter, and it explicitly assumes no derivative benefit, no settlements, no restructurings. Nine months of FY2026 operating income at the guide LOW is $73.8B, already $7.9B above the $65.9B full-year street EBIT consensus in the file, with the seasonally largest profit quarter still to come. That number has to move.

Tone moved from asking for patience to publishing unit economics. In Q1 the word "confident" appeared twice and management asked investors to trust the capex. In Q2 "confident" appears zero times and is replaced by an unprompted, structured ROIC disclosure in the opening remarks: servers break even in "a little less than three years," useful life "at least five to six years," most AI capacity contracted for "at least five-year terms," data centres delivering "five to six generations of server economics" over a 30-plus-year life. "Excited" went 1 to 6, "enthusiastic" 0 to 3. Management stopped asking to be believed and started publishing the math.

The constraint reversed from demand to supply, and that is the whole story. Q1: "we've done a good job in making sure that we're not capacity constrained." Q2: "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 fact, the demand we already have for 2028 is striking." Backlog (RPO) went $364B to $496B in one quarter, growing triple digits year-over-year, roughly 2.5x the Q3 2025 level while AWS revenue over the same span grew ~28% — backlog is compounding about three times faster than revenue, and it is contracted, not modelled.

Contradictions: 10 found, 4 high-severity — and all four sit under the same beam. The capex/ROIC defence. Nova was "our own frontier model" in 2025 but Amazon is now only "pursuing" one; memory cost was "in the right spot" in April and drove a +$20B capex raise in July; Amazon cut server useful life from six years to five because of AI in 2024 and the Q2 2026 ROIC bridge now assumes "at least five to six years" plus further extension; and "we purchase servers a few months before service, if the demand isn't there we won't spend the capital" contradicts the prior quarter's "6 to 24 months" and a Trainium4 book that was reserved 18 months before broad availability. None of these touch the AWS operating facts — +36.8%, five straight accelerations, $496B backlog, 39.4% segment margin are facts, not narrative. They touch the capital math around it, and they say to underwrite the returns bridge on Amazon's older, more conservative disclosures.

The single deteriorating line is cash. TTM free cash flow is −$7.6B, from +$18.2B a year ago and +$53.0B two years ago, on capex at 27.0% of revenue and FY2026 cash capex guided up to ~$220B from ~$200B. Under the framework this is the textbook "increasing revenue with declining FCF" red flag. The offsets are real — TTM operating cash flow grew +33% to $161.4B, and a $496B contracted backlog gives visible return on the spend — so it reads as a capital-allocation bet rather than operating deterioration. But management was asked directly how it funds this and answered "nothing to share today." That is a genuine disclosure gap, and it is the reason this is a BUY on operations with a capped score on financial trends, not an unqualified one.

Consolidated, 10 quarters

| Quarter | Revenue $M | YoY | Gross margin | GM YoY bps | EBITDA $M | EBITDA margin | Op income $M | Op margin | GAAP EPS | |---|---:|---:|---:|---:|---:|---:|---:|---:|---:| | Q2'26 | 200,606 | +19.6% | 52.3% | +44 | 47,449 | 23.7% | 27,461 | 13.7% | $5.75 | | Q1'26 | 181,519 | +16.6% | 51.8% | +127 | 42,797 | 23.6% | 23,852 | 13.1% | $2.78 | | Q4'25 | 213,386 | +13.6% | 48.5% | +113 | 44,448 | 20.8% | 24,977 | 11.7% | $1.95 | | Q3'25 | 180,169 | +13.4% | 50.8% | +175 | 34,218 | 19.0% | 17,422 | 9.7% | $1.95 | | Q2'25 | 167,702 | +13.3% | 51.8% | +168 | 34,398 | 20.5% | 19,171 | 11.4% | $1.68 | | Q1'25 | 155,667 | +8.6% | 50.6% | +123 | 32,667 | 21.0% | 18,405 | 11.8% | $1.59 | | Q4'24 | 187,792 | +10.5% | 47.3% | +179 | 36,834 | 19.6% | 21,203 | 11.3% | $1.86 | | Q3'24 | 158,877 | +11.0% | 49.0% | +146 | 30,853 | 19.4% | 17,411 | 11.0% | $1.43 | | Q2'24 | 147,977 | +10.1% | 50.1% | +176 | 26,710 | 18.1% | 14,672 | 9.9% | $1.26 | | Q1'24 | 143,313 | +12.5% | 49.3% | +255 | 26,991 | 18.8% | 15,307 | 10.7% | $0.98 |

Three things jump off this table. (1) Revenue growth has accelerated five straight quarters with no down quarter — +8.6% to +19.6%, roughly 1,100 bps of improvement on a $200B-per-quarter base. (2) Operating margin expanded 226 bps to 13.7% while growth accelerated. Growth accelerating with margins expanding is the strongest possible combination and the opposite of the "buy growth with margin" pattern that normally accompanies a capex cycle this size. (3) The one line that is not cooperating is gross margin: at a record 52.3% it expanded only +44 bps against a +113 to +175 bps run rate, because D&A moved to 10.0% of revenue from 9.1% as the AI data-centre build depreciates into cost of sales. That is the first place a capex-digestion problem would show up — watch it.

Revenue, AWS and EBITDA, 10 quarters

Year-over-year growth — the acceleration is broad

Revenue drivers

| Driver ($M) | Q2'25 | Q3'25 | Q4'25 | Q1'26 | Q2'26 | YoY | |---|---:|---:|---:|---:|---:|---:| | AWS net sales | 30,873 | 33,006 | 35,579 | 37,587 | 42,232 | +36.8% | |   AWS operating margin | 32.9% | 34.6% | 35.0% | 37.7% | 39.4% | +645 bps | |   AWS backlog / RPO ($B) | n/d | 200 | 244 | 364 | 496 | triple digits | | Advertising services | 15,694 | 17,703 | 21,317 | 17,243 | 19,809 | +26.2% | | Online stores | 61,485 | 67,407 | 82,988 | 64,254 | 70,432 | +14.6% | | Third-party seller services | 40,348 | 42,486 | 52,816 | 41,578 | 46,780 | +15.9% | | Subscription services | 12,208 | 12,574 | 13,122 | 13,427 | 13,730 | +12.5% | | Worldwide paid units YoY | +12% | +11% | +12% | +15% | +17% | volume-led | | Worldwide shipping costs YoY | +6% | +8% | +10% | +14% | +19% | fuel, line-haul |

Amazon has no same-store-sales metric, so the volume/price decomposition is units against GMV-proxy revenue. Worldwide paid units accelerated from +8% at the Q1 2025 trough to +17% while online-stores revenue accelerated from +5.0% to +14.6%. Units are outrunning revenue dollars, meaning the growth is genuinely volume-led rather than price or mix. That is the healthier version of the same number. The offsetting cost line is shipping, growing +19% against +17% units — management attributes the gap to Middle East fuel inflation and higher line-haul rates from driver-capacity limits, partly recovered by the April FBA fuel and logistics surcharge.

Segments

| Segment | Q2'25 | Q3'25 | Q4'25 | Q1'26 | Q2'26 | YoY | |---|---:|---:|---:|---:|---:|---:| | North America revenue ($M) | 100,068 | 106,267 | 127,083 | 104,143 | 116,177 | +16.1% | |   NA operating margin | 7.5% | 4.5% | 9.0% | 7.9% | 7.9% | +34 bps | |   NA margin ex-tariff-refund | 7.5% | — | — | — | ~7.3% | −18 bps | | International revenue ($M) | 36,761 | 40,896 | 50,724 | 39,789 | 42,197 | +14.8% | |   International operating margin | 4.1% | 2.9% | 2.1% | 3.6% | 4.1% | flat | | AWS revenue ($M) | 30,873 | 33,006 | 35,579 | 37,587 | 42,232 | +36.8% | | Revenue mix NA / Intl / AWS | 60/22/18 | 59/23/18 | 60/24/17 | 57/22/21 | 58/21/21 | — |

Mix shift is doing quiet, compounding work. AWS went from 15–18% of revenue through FY2024 to 21% in Q2 2026, and it carries a 39.4% operating margin against North America's 7.9%. That mix shift alone is worth roughly 100 bps a year of consolidated margin at the current trajectory, before any segment-level improvement. It also means the consolidated margin story is increasingly an AWS story and increasingly not a retail story — which is exactly what the ex-one-timer segment margins show.

The capital cost of the acceleration

Five-year annual trend

| Year | Revenue $M | YoY | AWS $M | AWS YoY | Gross margin | EBITDA margin | Op margin | Diluted EPS | Capex % of rev | FCF $M | |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:| | FY2021 | 469,822 | — | 62,202 | — | 42.0% | 12.6% | 5.3% | $3.24 | 13.0% | (14,726) | | FY2022 | 513,983 | +9.4% | 80,096 | +28.8% | 43.8% | 10.5% | 2.4% | $(0.27) | 12.4% | (16,893) | | FY2023 | 574,785 | +11.8% | 90,757 | +13.3% | 47.0% | 14.9% | 6.4% | $2.90 | 9.2% | 32,217 | | FY2024 | 637,959 | +11.0% | 107,556 | +18.5% | 48.9% | 19.0% | 10.8% | $5.53 | 13.0% | 32,878 | | FY2025 | 716,924 | +12.4% | 128,725 | +19.7% | 50.3% | 20.3% | 11.2% | $7.17 | 18.4% | 7,695 | | TTM through Q2'26 | 775,680 | — | 148,404 | — | 50.8% | 21.8% | 12.1% | — | 22.3% | (7,604) |

The five-year shape. Revenue compounded ~11.2% a year with the YoY rate accelerating every year since FY2022, and the current quarter runs at +19.6%. Gross margin expanded 830 bps over five years, EBITDA margin 770 bps, operating margin 590 bps. Every profitability line is up and to the right. The single degrading line is free cash flow — $32.9B in FY2024, $7.7B in FY2025, and −$7.6B on a TTM basis — entirely a function of capex going from 9.2% of revenue in FY2023 to 27.0% this quarter.

This quarter versus every available bar

| Metric | Bar | Actual | Variance | Result | |---|---|---:|---:|---| | Net sales vs street | $197.03B (31 analysts) | $200.606B | +$3.57B / +1.81% | Beat | | Net sales vs company guide | $194.0–199.0B | $200.606B | +$1.61B above the high end | Beat | | Operating income vs company guide | $20.0–24.0B | $27.461B | +$3.46B above the high end | Beat | | Operating income, ex-$1.2B one-offs | $20.0–24.0B | $26.261B | +$2.26B above the high end | Beat | | Diluted EPS, GAAP | $1.82 | $5.75 | +215.9% | Not meaningful | | Diluted EPS, ex-Anthropic mark | $1.82 | ~$1.95 | +7.1% | Beat | | Diluted EPS, fully clean | $1.82 | ~$1.86 | +2.4% | Narrow beat | | AWS revenue | ~$40.54B (press-coverage bar) | $42.232B | +4.17% | Beat |

Adjusted-EPS derivation, auditable. Reported pre-tax income $80,857M less "Other income (expense), net" of $53,415M gives adjusted pre-tax income of $27,442M. Taxed at the quarter's reported effective rate of 22.51% ($18,199M on $80,857M) and less $11M of equity-method activity gives adjusted net income of $21,254M, on 10,903M diluted shares = $1.95. Removing a further $1,200M of one-time operating benefits gives $20,325M = $1.86. The method deliberately keeps interest income and expense (both recurring) and strips only the investment-mark line. The figure is sensitive to the tax assumption and ranges $1.95 to $2.47 across reasonable treatments — every point in that range is a beat, so the verdict is robust even though the magnitude is not precise.

Beat rates and the magnitude question

| Window | Revenue beat rate | EPS beat rate | Combined | |---|---:|---:|---:| | Last 12 quarters (Q3'23 – Q2'26) | 11 / 12 = 91.7% | 11 / 12 = 91.7% | 22 / 24 = 91.7% | | Last 4 quarters (Q3'25 – Q2'26) | 4 / 4 = 100% | 3 / 4 = 75% | 7 / 8 = 87.5% |

Pattern: consistent beater. Two misses in 24 observations over three years, both rounding-level — a $0.02 EPS miss in Q4 2025 and a −0.46% revenue miss in Q2 2024. There is no chronic-misser risk here.

Magnitude cuts two ways, and this is the finding that matters.

| Trailing 4-quarter block | Revenue surprise | Reported EPS surprise | Clean EPS surprise | |---|---:|---:|---:| | Q3'23 – Q2'24 | +2.39% | +31.9% | +31.9% | | Q3'24 – Q2'25 | +1.31% | +23.6% | +23.6% | | Q3'25 – Q2'26 | +1.60% | +77.4% (artifact) | +8.5% |

Reported EPS surprise is unusable — both of the last two quarters are dominated by non-operating investment marks ($15.6B in Q1'26, $53.4B in Q2'26). On clean numbers the street's cushion has compressed by roughly two-thirds, from +31.9% to +8.5%. Read alone, that is a deteriorating-magnitude red flag.

But the beat versus management's own bar is at a series high and widening:

| Guided on | Quarter | Revenue guide | vs high end | Op income guide | vs high end | |---|---|---|---:|---|---:| | 2025-02-06 | Q1 2025 | $151.0–155.5B | +$0.17B | $14.0–18.0B | +$0.41B | | 2025-05-01 | Q2 2025 | $159.0–164.0B | +$3.70B | $13.0–17.0B | +$2.17B | | 2025-07-31 | Q3 2025 | $174.0–179.5B | +$0.67B | $15.5–20.5B | +$1.2B clean | | 2026-02-05 | Q1 2026 | $173.5–178.5B | +$3.02B | $16.5–21.5B | +$2.35B | | 2026-04-29 | Q2 2026 | $194.0–199.0B | +$1.61B | $20.0–24.0B | +$3.46B |

These are not in conflict. The street now models Amazon's operating trajectory better than it did in 2023-24, which is why the clean EPS surprise has shrunk. Management's own guidance philosophy has stayed conservative, and outperformance against that guide is widening. With five straight quarters of AWS acceleration and +520 bps of clean AWS margin expansion, the guidance-relative trend is the higher-signal read: management is under-promising into a re-accelerating business and delivering by more each quarter. The Q3 2025 apparent shortfall was ~$4.3B of disclosed special charges, not an operational miss.

What management said drove it

| Driver | Assessment | |---|---| | AWS acceleration | "Up 36.7% year-over-year, driven by both core and AI services"; "$169 billion annualized revenue run rate"; "we see a strong linkage between AI spend and core growth." Real and recurring — the dominant driver. | | AWS margin | "+650 basis points year-over-year, 520 basis points if you exclude the derivative accounting gain… disciplined efficiency gains, capacity optimization, which we benefited quite a bit from in Q2." +520 bps of clean expansion is the largest single contributor to the operating-income beat. | | Tariff refunds, $600M, North America | "Represents the significant majority of refunds we expect to receive." Non-recurring. Do not extrapolate. | | Energy-derivative fair value, $600M, AWS | "These adjustments have not been significant in prior quarters." Non-recurring and non-cash, and explicitly excluded from the Q3 guide. | | Prime Day timing | Prime Day landed in Q2 for most large countries; in 2025 it was entirely in Q3. Optical revenue tailwind worth ~400 bps, reversing in Q3. | | Retail cost-to-serve | "Shipping costs grew more slowly than worldwide unit growth" only after excluding fuel and line-haul inflation. Mildly negative — retail was not the source of the beat. |

Credit where due: the CFO volunteered the $1.2B bridge unprompted in prepared remarks rather than letting the street find it. Under the guidance-accuracy lens that is a positive.

What was issued on 2026-07-30

| Item | Guidance | |---|---| | Q3 2026 net sales | $197.0B – $202.0B, "to grow between 9% and 12%" versus Q3 2025 | | Q3 2026 operating income | $22.5B – $26.5B, versus $17.4B in Q3 2025 | | Prime Day timing | Excluding Prime Day in both years, YoY growth would be ~400 bps higher | | FX | ~80 bps unfavourable at current rates | | Energy derivatives | Guide assumes no remeasurement impact (Q2 carried a +$600M benefit) | | M&A, restructuring, legal | Guide assumes none concluded | | FY2026 cash capex | ~$220B, raised from ~$200B — "the higher cost of memory pushing this number up" | | EPS, gross margin, EBITDA, segment revenue, AWS | Not guided. Amazon guides two lines, one quarter forward. |

Why "9 to 12 percent" understates the business

Reported guided growth of +10.7% at the midpoint contains ~480 bps of pure optics. Underlying guided growth of ~15.5% is essentially unchanged from the ~16% underlying rate delivered in Q2. Underlying growth is guided flat, not decelerating — and Amazon has beaten the guide midpoint by an average 2.75% in each of the last four quarters, which on the base rate puts Q3 underlying near +18%, still accelerating.

The operating income guide is where the conservatism is

The headline "+41% operating income growth" is flattered by prior-year charges. On a clean basis the guide is +12.9%, and the implied margin is 12.28% versus a clean 12.04% prior-year base — just +24 bps of expansion, from a company that has just delivered +166 bps clean and is riding a mix shift into a 39.4%-margin AWS. The guide also explicitly assumes no energy-derivative benefit, no settlements and no restructurings; three of the last four quarters contained at least one of those.

FY2026, and the number that has to move

| Metric | Guide low | Guide mid | Consensus | Read | |---|---:|---:|---:|---| | FY2026 revenue, 9M locked plus Q3 guide ($B) | 579.1 | 581.6 | 825.6 (full year) | Implies Q4 of $244.0B, +14.3% YoY — achievable, not a gimme | | FY2026 operating income, 9M locked plus Q3 guide ($B) | 73.8 | 75.8 | 65.9 (full year) | 9M at the guide LOW already exceeds the full-year street EBIT by $7.9B | | FY2026 cash capex ($B) | — | ~220 | — | Raised from ~$200B; ~26.7% of revenue versus ~18.4% in FY2025 |

This is mechanical, not a judgement call. H1 2026 net sales of $382,125M and H1 2026 operating income of $51,313M come straight off the consolidated statements in the release. Add the Q3 guide and nine months of FY2026 operating income lands at $73.8B to $77.8B — against a full-year street EBIT consensus of $65.9B, with the seasonally largest profit quarter ($25.0B last year) still to come. The FY2026 EBIT consensus needs to move up by at least $33B — and closer to $37B at the top of the guide. Even assuming Q4 merely repeats last year's $25.0B with zero growth, 9M at the guide low plus a flat Q4 is $98.8B against $65.9B. Revenue consensus, by contrast, is close to fair.

The capex raise

The $20B raise is cost inflation, not incremental capacity — Jassy attributed it entirely to memory pricing. This is the only negative revision in the quarter, and it is to a cost input, not to demand or to the revenue and margin outlook. It is also the one that compresses free cash flow and re-prices nothing in the existing $496B backlog.

Guidance integrity — do they hit what they guide to?

| Quarter | Revenue guide mid | Actual | vs mid | Op income guide mid | Actual (clean) | Clean vs mid | |---|---:|---:|---:|---:|---:|---:| | Q2 2025 | 161.5 | 167.7 | +3.8% | 15.25 | 19.17 | +25.7% | | Q3 2025 | 176.75 | 180.2 | +1.9% | 18.00 | 21.70 | +20.6% | | Q1 2026 | 176.00 | 181.5 | +3.1% | 19.00 | 23.85 | +25.5% | | Q2 2026 | 196.50 | 200.6 | +2.1% | 22.00 | 26.26 | +19.4% | | Average | | | +2.75% | | | +22.8% |

Four for four above the high end on both lines once disclosed charges are normalised. Applying the four-quarter average beat to the Q3 2026 midpoint gets to roughly $205B of revenue and ~$30B of operating income. That is the base rate, not a forecast — but it is the right prior when reading a guide that optically decelerates to 9–12%. The Q4 2025 guidance range is missing from the truncated transcript in the run folder and has been left blank rather than estimated.

Tone: Q1 2026 versus Q2 2026

| Term | Q1 2026 | Q2 2026 | Direction | |---|---:|---:|---| | "excited" | 1 | 6 | up | | "enthusiastic" | 0 | 3 | up | | "strong" | 13 | 19 | up | | "demand" | 16 | 19 | up | | "tariff" | 1 | 7 | up (now a P&L benefit, not a threat) | | "confident" / "confidence" | 2 | 0 | down | | "constrained" / "capacity constrained" | 3 | 0 | down | | "cautious" / "uncertain" | 0 | 0 | flat |

Transcripts are comparable in length (49.1k characters each), so the counts are meaningful. The highest-quality shift is "confident" going to zero and being replaced by arithmetic. The Q1 call defended the investment; the Q2 call underwrote it with cycle economics, contract durations, break-even periods and a 2028 demand reference. Hedges are up in number but narrower in scope — energy-derivative noise, transport inflation, component pricing on new contracts. None of them touch the top line.

| Metric | Q3'24 | Q4'24 | Q1'25 | Q2'25 | Q3'25 | Q4'25 | Q1'26 | Q2'26 | |---|---:|---:|---:|---:|---:|---:|---:|---:| | Revenue YoY | 11.0% | 10.5% | 8.6% | 13.3% | 13.4% | 13.6% | 16.6% | 19.6% | | Revenue accel (bps) | — | (55) | (187) | +471 | +7 | +23 | +298 | +301 | | Operating income YoY | 55.6% | 60.5% | 20.2% | 30.7% | 0.1% | 17.8% | 29.6% | 43.2% | | Op income accel (bps) | — | — | — | +1,042 | (3,060) | +1,774 | +1,180 | +1,365 | | Op margin YoY (bps) | +314 | +352 | +114 | +151 | (129) | +42 | +132 | +226 | | GAAP EPS YoY | 52.1% | 86.0% | 62.2% | 33.3% | 36.4% | 4.8% | 74.8% | 242.3% |

The seven inflection points

1. Q1 2025 — the trough at +8.6% revenue. The low point of the whole post-2023 cycle. Everything since has been up and to the right.
2. Q2 2025 — inflection one, +471 bps. The largest single-quarter acceleration in the series. This is where AWS re-acceleration started carrying the consolidated line.
3. Q2 2025 to Q4 2025 — a three-quarter plateau. +13.3% to +13.4% to +13.6%, effectively flat. AWS was already climbing but was not yet large enough relative to Stores to move the total.
4. Q1 2026 to Q2 2026 — inflection two, and the real story. Two consecutive ~300 bps accelerations taking the line from +13.6% to +19.6%. This is AWS reaching escape velocity: $37.6B (+28.4%) then $42.2B (+36.7%), the fastest in 18 quarters and the fifth consecutive quarter of acceleration.
5. Q3 2025 — the EPS pothole. Not an operating event: $2.5B FTC settlement plus $1.8B severance turned +24.8% underlying operating income growth into +0.1% reported.
6. Q4 2025 — the EPS trough at +4.8%. Purely a base effect from lapping +86%. Operating income still grew +17.8%.
7. Q2 2026 — the EPS spike at +242%. Non-operating. Ignore it; use +43.2% operating income growth, or +37.0% excluding the $1.2B of one-timers.

Two honest deductions, and what survives them

First deduction: the Q2 headline is flattered by Prime Day. Management sized the Q3 drag at "nearly 400 basis points," which implies roughly 430 bps of the Q2 acceleration is timing, putting clean Q2 revenue growth near +15.3% — a small step down from Q1's clean +16.6%, not an acceleration. Second deduction: EPS is uninformative in two of eight quarters and outright meaningless in Q2 2026. Track operating income instead.

What survives both deductions is the part that matters. AWS is the entire acceleration story and it carries no timing asterisk: +16.9% to +17.5% to +20.2% to +23.6% to +28.4% to +36.7%, five consecutive quarters, $169B annualised run rate, a $496B backlog growing triple digits, and AWS operating income of $16.6B against $10.2B a year ago. Advertising re-accelerated to +26% after four quarters parked at +22%. Stores did the unglamorous work — cost-to-serve down, robotic-arm fleet set to more than double in 2026 — and held North America margin near 7.9% through fuel and line-haul inflation. A market leader accelerating off a trough with the highest-margin segment doing the accelerating.

| # | Catalyst | Timing | Management said | What consensus expects | Gap | |---|---|---|---|---|---| | 1 | AWS sixth straight acceleration | Q3 print 2026-10-29 | "Accelerating for the fifth straight quarter, our fastest growth in 18 quarters"; added $4.6B QoQ, "about 80% more than our largest increase ever"; "we will still not have enough capacity to meet all the demand we have in 2026… also true in 2027" | Pre-print street had AWS at ~31% for Q2 and got 36.7%. The street has been chronically 500 bps light. | Widest gap in the story. RPO of $496B growing triple digits is contracted, not modelled. 2027 capacity "largely reserved," 2028 demand "striking." The street models deceleration by default. | | 2 | First FY2027 capex guide | Q4 print, ~early Feb 2027 | FY26 raised to ~$220B from ~$200B on memory cost. Unprompted ROIC framework: sub-3-year server break-even, 5–6 year lives, most AI capacity contracted for 5-plus years, data centres 30-plus years and "five to six generations of server economics." | FY27 street (pre-revision): revenue $937.7B, EBIT $74.9B, EPS $10.15. | Management is telling you FY27 capex goes up. The market took the $220B raise well; a $280B-plus FY27 number is the real test. | | 3 | Memory / DRAM cost inflation | Continuous, re-rated each print | The entire $20B raise was pinned on memory. "Inflated prices right now on some of the components like memory and hard drives and SSDs." "Resource and supply volatility, including for memory chips" is a new addition to the safe-harbor risk language. | Third-party research has DRAM up sharply in CY26 with further ASP increases in CY27, and memory rising toward ~36% of hyperscaler capex. (Industry research, not company-guided.) | Asymmetric downside. Signed contracts lock the price, not the cost. Inflation lands on AWS margin, not the customer, for the duration of the backlog. | | 4 | Anthropic IPO | Targeted listing ~October 2026 | Not addressed on the call. Q2 carried a $53.4B non-operating pre-tax gain, primarily Anthropic; Q1 2026 carried $15.6B. | Amazon holds a reported ~15–21% stake. (Press reporting; the stake is not company-disclosed — treat as unverified.) | Signal-quality catalyst, not an earnings catalyst. It converts a mark into an observable price and makes EPS comparability worse. Watch the reverse: a down-round mark is a multi-billion hit to a line the street ignores. | | 5 | OpenAI scale-up | Tranches through 2026-27; capacity 2027-plus | "Multi-year, multi-gigawatt commitments to Trainium" from "the two leading AI labs in the world, Anthropic and OpenAI." | Reported: ~$50B Amazon investment in OpenAI, AWS agreement expanded ~$100B over eight years, ~2GW of Trainium. (Press and company announcements.) | Underappreciated second-order effect: these are capex commitments Amazon must fund, and management named no funding source. Balance-sheet expansion is coming. | | 6 | re:Invent 2026 / Trainium4 | Nov 30 – Dec 4, 2026 | Chips run rate over $25B, triple-digit growth. Graviton in 98% of top-1,000 EC2 customers; Graviton revenue commitments ~3x QoQ; Graviton5 ramping ~2x faster than Graviton4. Nvidia partnership explicitly reaffirmed. | Third-party: Trainium4 targeted late-2026 to early-2027. | The densest catalyst date before the Q4 print. Trainium4 specs plus any merchant-silicon news land there. | | 7 | Merchant Trainium sales | Unscheduled — "in the future" | "An increasing number of customers who are interested in us providing the Trainium chips to them, separate from our cloud… I expect there's a real chance we'll do that in the future." Second consecutive quarter the topic has come up, with firming language. | Not in any published model I can verify. Worth zero in consensus. | The cleanest "management is the contrarian" set-up in the name. Low near-term revenue probability, high re-rating-headline probability. | | 8 | EU DMA cloud gatekeeper decision | Final decision before end-2026, then six months to comply | Not addressed. Amazon's public response calls the preliminary assessment one that "disregards the breadth of cloud services available to European customers." | Preliminary findings issued 2026-06-25; would impose interoperability, portability, anti-self-preferencing and anti-bundling obligations. | Modest P&L risk, real structural risk. Egress-fee and switching-cost economics are what make International AWS sticky, and International margin is already flat at 4.1%. Not priced. | | 9 | FTC v. Amazon monopoly trial | Bench trial 2027-03-29 | Not addressed. Guidance assumes no legal settlements concluded. | Discovery and expert disclosures through H2 2026 generate headline risk. | Far-dated and well-telegraphed; the push from February removed it from the FY2026 window. Low priority. | | 10 | Amazon Leo initial service | "This year" (2026) | "Close to 400 satellites in orbit, enough to begin initial satellite internet service this year. We already have meaningful revenue commitments from enterprises and government customers, and more than 20 partners." | Beta partners reported to include Verizon, AT&T, Vodafone, JetBlue, NASA. An FCC milestone of 1,618 satellites by 2026-07-30 was not met; a two-year extension has been requested. (Press reporting.) | Revenue is immaterial in FY26-27. The catalyst is binary regulatory — denial of the extension is a low-probability, high-severity write-down on a business the street values at roughly zero. |

Secondary catalysts, all disclosed as operating metrics

| Catalyst | Management datapoint | Why it matters | |---|---|---| | Prime Day timing | Moved into Q2 for most large countries; ex-Prime Day, Q3 growth would be ~400 bps higher, plus ~80 bps FX | The Q3 optical deceleration is roughly 480 bps mechanical. Do not read Q3 as a demand break. | | Prime Big Deal Days | Not mentioned; expected ~mid-October 2026 | Falls in Q4, so it does not rescue the Q3 optics | | Same-day perishables | Second-largest US grocer, over $150B gross merchandise sales; monthly active perishables customers +50% since the start of the year; 2,300 US cities; perishables orders carry 3x more units | Jassy's strongest grocery language yet — "changing the trajectory of our everyday essentials business." But see contradiction C-5: the city count has not moved in three quarters. | | Amazon Now, 30-minute delivery | +80% sales and units QoQ, +60% more customers QoQ, 80 new US cities in Q2, live in nine countries and 250-plus cities | The fastest-scaling disclosed metric in Stores | | Alexa+ and Alexa for Shopping | 350M-plus customers used Alexa for Shopping in the last 12 months; active users nearly doubled QoQ; Alexa+ triers sign up for Prime at ~25% higher rates | The monetisation is the ad unit, not the subscription — ties directly to the advertising acceleration | | Advertising | $19.8B, +26%, after four straight quarters at +22%; Thursday Night Football, NBA, WNBA and NASCAR inventory all sold out | The highest-margin revenue line in the company just accelerated 400 bps and is now structurally linked to agentic shopping surfaces | | Amazon Supply Chain Services | Launch customers include P&G, 3M, Lands' End, American Eagle Outfitters | Converts fulfilment fixed cost into third-party revenue. Zero in consensus. | | Robotics and cost to serve | Expect to more than double the fleet of robotic arms (Cardinal, Sparrow) in 2026 | The offset to fuel and line-haul inflation |

Calendar

| Date | Event | |---|---| | Aug–Sep 2026 | Anthropic public S-1 and roadshow (reported) | | ~mid-Oct 2026 | Prime Big Deal Days (dates unconfirmed) | | ~Oct 2026 | Anthropic Nasdaq listing (reported target) | | 2026-10-29 | Q3 2026 earnings — AWS sixth acceleration, Q4 guide, Leo initial service confirmation | | Before end-2026 | EU DMA final gatekeeper decision on AWS; six-month compliance clock starts | | 2026-11-30 to 12-04 | AWS re:Invent 2026 — Trainium4, agentic stack, possible merchant-silicon news | | ~Feb 2027 | Q4 2026 earnings — first FY2027 capex guide, the big one | | 2027-03-29 | FTC v. Amazon bench trial begins (pretrial 03-15, briefs 03-22) | | 2027 | Trainium4 volume delivery; OpenAI ~2GW capacity build |

What would invalidate the top three

• AWS acceleration: RPO growth decelerating below revenue growth; any language shift from "supply constrained" back to "demand pipeline"; AWS sequential dollar adds below the $4.6B set this quarter.
• Capex ROIC: AWS operating margin falling below ~33% (the year-ago base) while capex still grows; server useful-life assumption revised down in the 10-K; break-even guidance extended past three years.
• Memory: a second capex raise attributed to components within two quarters, or AWS gross-margin compression management attributes to input cost rather than mix.

| Metric | Q2 2026 | |---|---| | Analysts who asked | 6 (Anmuth, Post, Nowak, Sebastian, Gawrelski, Sheridan) | | Discrete sub-questions | 10 | | Well answered | 4 of 10 (40%) — AWS margin drivers, frontier model, application layer, grocery | | Deflected or avoided | 6 of 10 (60%) — every one on capex, capacity, financing or returns | | Questions taken by the CFO | 1 of 10 | | Questions on the $53.4B Anthropic gain | 0 |

Call depth is unchanged — six analysts fielded in each of the last three quarters — so the low answer rate is not a shorter call. It is a boundary. Management answered generously on demand and stonewalled on capital. The call functioned as a demand advertisement, not a capital-allocation disclosure.

The six deflections, consolidated

| Analyst | The ask | What management gave | Why it matters | |---|---|---|---| | Justin Post (BofA) | Is the AWS acceleration demand or capacity finally landing? Second-half versus first-half gigawatts; 2027 | Re-stated the previously-issued "double power capacity by end-2027 versus 2025" | Determines whether +36.7% is durable demand or a supply catch-up bulge that decays | | Brian Nowak (Morgan Stanley) | Can long-lived data-centre spend slow at all in 2027? | A demand "barbell" narrative; no capex figure, not even a directional no | The two-year-lead, non-discretionary half of capex — the free-cash-flow swing factor | | Brian Nowak (Morgan Stanley) | ROIC on selling Trainium to third-party data centres | A soft yes on the when; complete silence on ROIC | Merchant silicon carries structurally lower returns than renting the same chips at a 39% segment margin | | Colin Sebastian (Baird) | Sources of capital for the build-out — asked to the CFO by name | Taken by Jassy: "You've seen us issue debt this year… nothing to share today" | TTM FCF −$7.6B, FY26 capex ~$220B, funding plan undisclosed. The CEO intercepting a question addressed to the CFO makes it more notable, not less. | | Ken Gawrelski (Wells Fargo) | What does the $496B RPO imply for 2028-plus capacity? | "We know about that backlog, so that's all taken into account in our CapEx projections" | Logically circular — the capacity plan he referenced was set before the backlog reached $496B. He declined the 2028 question entirely. | | Ken Gawrelski (Wells Fargo) | Do long-term contracts preserve stable return profiles under cost inflation? | Explained price-lock mechanics; never answered "stable" | $496B of backlog was priced before the memory spike that just added $20B to capex, and cannot be re-cut. Arguably worse than silence. |

The four that were answered well

AWS margin drivers (Anmuth to Olsavsky). Quantified, split out the one-time item unprompted, and named three specific drivers — efficiency gains, capacity optimisation, fixed-cost discipline. The sustainability half was answered only with "these margins will fluctuate."

The frontier model (Anmuth to Jassy). The most newsworthy disclosure of the call: "AWS and Amazon can have a wildly successful business without its own frontier model… All that said, we are pursuing our own frontier model," for control over cost, prioritisation and speed. See contradiction C-1 — this is also the cleanest credibility flag in the quarter.

Application layer (Sebastian to Jassy). Genuine strategic answer with named products and named customers, and a coherent "we built it for ourselves first" logic. Zero revenue disclosed for any application-layer service.

Grocery (Sheridan to Jassy). The most quantified answer on the call, with an unusually concrete mechanism — co-locating perishables in existing same-day facilities rather than building new formats. The geography half of the question was ignored entirely.

The dog that did not bark

Not one analyst asked about the $53.4B gain, and management never mentioned it. A word-level scan of the transcript confirms it: "Anthropic" appears three times, all three about Trainium capacity commitments, never the investment gain. The words net income, earnings per share and diluted appear zero times in the entire call. "Free cash flow" appears six times, all forward-looking and qualitative — the actual TTM figure of −$7.6B was never stated by management and no analyst asked for it.

Both readings are informative. Management deserves credit for not leading with a headline EPS number it knew was non-economic; Jassy opened on revenue and operating income, the right metrics. But the sell-side letting a $53.4B swing and a negative free-cash-flow print pass without a single question, in a call where six of ten capital questions were deflected, says the street is currently underwriting the AI narrative and not the balance sheet. That is a sentiment condition, and sentiment conditions reverse.

Carry-forward into 2026-10-29. None of the six questions were retired. Expect capex funding, the 2027 capex trajectory and backlog margin durability at the top of the list — and management now has one quarter's notice to prepare real answers.

Checked and found consistent

Recorded so the reader knows these were tested, not skipped. Segment margins tie to the filing exactly (North America $9,123M on $116,177M; International $1,717M on $42,197M). The AWS margin bridge ties — 39.4% versus 32.9% is +645 bps, and the 520 bps ex-derivative figure implies ~$540M of the $600M energy benefit sat in AWS, consistent with "primarily." Backlog math ties — $496B versus Q4 2025's disclosed $244B is consistent with triple-digit YoY growth and with the analyst's uncontested "two and a half times Q3 2025." International ex-FX ties at +15%. Grocery gross merchandise sales are consistent across calls at over $150B. Prime Day country split is consistent. Chips run-rate progression is consistent at $20B-plus to $25B-plus. Tariff-refund treatment is reconcilable — "we largely absorbed these costs" and the "limited set of circumstances where we can trace" pass-through are subset statements, not conflicting ones.

Analyst conclusion on the contradictions

They cluster, and they cluster in one place: the capex and ROIC defence. C-2, C-3 and C-4 are all load-bearing beams under the argument that $220B of 2026 capex is demand-verified, cancellable, and depreciable over a comfortable asset life. Each is contradicted by Amazon's own statements from either three or eighteen months earlier. That does not break the AWS thesis — +36.7% growth, five consecutive quarters of acceleration, a $496B backlog and a 39.4% segment margin are facts, not narrative. But it means the returns bridge should be underwritten on the company's older, more conservative disclosures (five-year lives, 6-to-24-month cash-out lead times), not the Q2 2026 restatements.

C-1 is the credibility item to carry forward. C-7 and C-8 are the quality-of-earnings items. C-5 and C-6 suggest the everyday-essentials narrative is running ahead of its metrics. Under the "you are not contrarian, only management is" lens, the honest contrarian case here is Jassy's own repeated, street-doubted claim that AWS can be a $1 trillion revenue business — a claim he has now escalated twice. The contradictions above do not touch that claim; they touch the near-term capital math around it. Score them separately.

The macro signal management put in writing

Amazon rewrote its own safe-harbor risk paragraph between Q3 2025 and Q4 2025 and has kept the new version through Q2 2026. That edit is the cleanest read on what this management team actually thinks the risks are.

| Quarter | Risk-factor language | |---|---| | Q2 2025 / Q3 2025 | "…fluctuations in foreign exchange rates, changes in global economic and geopolitical conditions, tariff and trade policies and customer demand and spending…" | | Q4 2025 through Q2 2026 | "…fluctuations in foreign exchange rates and energy prices, changes in global economic and geopolitical conditions, tariff and trade policies, resource and supply volatility, including for memory chips, and customer demand and spending…" |

The worry set rotated from tariffs and the consumer to power prices and memory supply. Two new named line items, both on the cost side of the AI build, neither on the demand side. This is a supply-constrained macro, not a demand-constrained one, and Amazon is saying so in the one place it is legally obliged to be accurate.

The macro reads, ranked

| Theme | Evidence | Read-through | |---|---|---| | Memory and storage inflation | Quarterly capex $43.2B to $53.1B; FY26 guide $200B to $220B, attributed solely to memory. "No secret to any company in the world that there are inflated prices right now on memory and hard drives and SSDs." | A $20B unit-cost pass-through in one guidance revision on flat-to-up physical capacity. A buyer with Amazon's scale and a pre-positioned supply book still absorbed a 10% capex-dollar increase for the same build. Everyone smaller faces worse. | | Pricing power passes through only on new contracts | "The deals that you sign, those will be the prices… New agreements that you sign, you always take into account what your costs are." | The $496B RPO is priced at pre-inflation cost assumptions and cannot be repriced. Mildly negative gross-margin read for every cloud and AI provider carrying a large multi-year backlog signed in 2025. AI unit economics get harder before they get easier. | | Energy procurement at derivative scale | A $600M fair-value benefit from electricity contracts "to secure electricity supply for existing and future operations," previously immaterial. | Hyperscaler power buying is now long-tenor, fixed-price and large enough to be derivative-accounted and P&L-moving. Constructive for merchant and IPP power generators signing hyperscaler PPAs. It also means reported AWS margin now carries a mark-to-market investors should strip. | | Freight and fuel | "Heightened transportation costs driven by fuel inflation from the conflict in the Middle East and higher line haul rates from driver capacity limitations," partly offset by the April FBA surcharge. | Second consecutive quarter of the same headwind, i.e. persistent. Pricing-positive for freight carriers; cost-negative for every shipper without Amazon's surcharge leverage. | | Tariffs flipped from cost to price | $600M of refunds booked, "the significant majority" of what is expected; Amazon is not the importer of record for most units; "we'll utilize refunds to continue to invest in low prices." | Do not model a recurring benefit. Importer-of-record retailers should be receiving proportionally larger refunds. And the refund gets spent on price — disinflationary for goods, competitively negative for any retailer that intended to bank it. | | Consumer: conspicuously no caution | Worldwide units +15% to +17%; North America +12% to +16%; International ex-FX +11% to +15%; Prime membership growth quantified as double-digit. Zero cautionary consumer language anywhere in the call. | The consumer is spending, and spending faster. But it skews to value: Amazon Haul selection expanded ~20x since launch to over 6 million items under $10, and Prime Day was "more than 80% at our lowest price of the year." Positive for volume-driven consumer names and discounters; the print does not validate premium discretionary demand. | | Enterprise IT sentiment | "85% of global IT spend is still on premises. That equation is going to flip." Plus the Q1 mechanism: memory scarcity is "a further impetus pushing companies who have on-premises infrastructure into the cloud" because "suppliers are prioritizing their very largest customers, which cloud providers are." | A rare macro dynamic where an input shortage transfers demand rather than destroying it. Enterprises literally cannot buy the memory to refresh on-prem. | | Rates and capital markets | "You've seen us issue debt this year… nothing to share today," against TTM operating cash flow of $161.4B (+33%) and TTM FCF of −$7.6B. | The largest cash-flow generator in the index is now funding AI capex partly with debt and declining to rule out more. IG issuance supply from mega-cap tech is a live and growing flow. | | FX | Q2 was +$0.1B favourable; Q3 guide assumes ~80 bps unfavourable, a swing from Q1's +180 bps favourable. | Modest mechanical headwind to reported Q3 growth for US multinationals. |

Company-level read-throughs

| Company or cohort | What was said | Direction | |---|---|---| | Nvidia | "Deep partnership with Nvidia… customers who will run on Nvidia for as long as we can foresee… choice is good for customers, competition, and driving the cost of inference down." | Modestly worse. Note the tonal downgrade from Q1's "immense respect… continue to order substantial quantities." Q2 drops the volume commitment, reframes Nvidia as customer preference rather than Amazon's, then pivots to inference-cost reduction. Essential, but no longer described as the growth vector. | | Anthropic | Named one of "the two leading AI labs in the world," a runaway app (Claude Code), and a multi-gigawatt Trainium committer. Also the source of the $53.4B non-operating gain. | Better. Reinforces Anthropic as an AWS-native, Trainium-native franchise. | | OpenAI | Same multi-gigawatt Trainium commitment; GPT models in Bedrock, "Bedrock managed agents powered by OpenAI." | Better for OpenAI's compute economics, worse for Microsoft's exclusivity. Multi-gigawatt volumes on a non-Nvidia, non-Azure stack is the clearest evidence yet of compute diversification. | | Merchant AI accelerators | "Actively having those conversations… a real chance we'll do that in the future" on selling Trainium outside AWS — firmed from Q1's "a good chance over the next couple of years." | The single most consequential quarter-over-quarter change on the call. A $25B-plus chips run rate growing triple digits, converting from captive cost advantage into a merchant competitor. Track this exact sentence next quarter. | | Uber, Pinterest, and ~9 named AI startups | "Larger companies like Uber and Pinterest all adopting Trainium," plus a named-logo startup list that roughly doubled. | Worse for merchant GPU demand at the startup tier — the most price-elastic accelerator buyers are migrating. Better for the adopters' inference COGS. | | Microsoft (Copilot) | "Unlike other offerings in this space, Q also lets you manage across leading SaaS tools like Slack, Salesforce, Jira, Teams, and ServiceNow." | Worse. An unmistakable jab at Copilot's suite-locked scope, backed by 15 named production logos across industrials, pharma, autos, energy, financials and sports. | | Salesforce, Atlassian, ServiceNow, Slack | Same quote. | Mixed and time-dependent. Short term better (usage driven into their systems of record); long term worse — the agent layer sits above the SaaS UI and does the work, which is how seat-based pricing gets disintermediated. | | IT services and systems integrators | "AWS Transform, which automates software migration, growing quickly." Q1: "customers have used Transform to save over 1.56 million hours of manual effort." | Worse. Billable-hours destruction stated as a feature, aimed at core SI revenue pools. | | Application-security vendors | AWS Continuum "discovers, prioritizes, validates, and remediates code vulnerabilities… validates in a sandbox so teams aren't chasing false positives." | Worse for standalone AppSec. Full workflow, bundled into AWS, sold into a captive install base, attacking the loudest complaint about incumbent scanners. | | Contact-centre software | Amazon Connect "used by all five major leading airline providers, as well as many of the leading banks and healthcare companies." | Worse. A near-total vertical sweep, and the natural landing zone for the automated-customer-service use case Jassy names as the top enterprise ROI category. | | Walmart, Target, Costco and big-box generally | Prices "on average 14% less than other retailers, according to Profitero," and refunds will be spent on lower prices. | Worse. A quantified, third-party-verified price gap plus stated intent to widen it. Amazon is choosing price share over margin. | | Kroger, Albertsons, Sprouts, Walmart grocery | Second-largest US grocer, $150B-plus GMS, perishables customers +50% YTD, 2,300 cities, 3x units per order. | Materially worse for conventional grocers. Amazon solved perishables without new store capital by bolting fresh into existing same-day facilities — a scalable model with a national footprint already in place. | | Instacart, DoorDash, Uber Eats | Amazon Now: 30-minute delivery, +80% sales and units QoQ, +60% more customers QoQ, 80 new US cities in Q2. | Worse. Amazon's fulfilment cost base with Prime bundling the delivery fee away, at the fastest QoQ growth rate disclosed on the call. | | Temu, Shein, dollar stores | Amazon Haul selection up ~20x since launch, over 6 million items under $10. | Worse. A structural answer to the cross-border direct-from-China model, and negative for the dollar-store value proposition. | | UPS, FedEx, truckload carriers | Fuel inflation plus "higher line haul rates from driver capacity limitations"; Amazon's own April surcharge. | Mixed. Better for carrier pricing (capacity is tightening); worse for carrier volume (Amazon in-sources and surcharges rather than paying third parties). | | GXO, C.H. Robinson, XPO, DHL Supply Chain | "Amazon Supply Chain Services. Any business can move, store, and deliver everything from raw materials to finished products using the same supply chain that supports Amazon" — launch customers P&G, 3M, Lands' End, American Eagle. | Worse, and underappreciated. This is full contract logistics, not just fulfilment, competing for enterprise-scale 3PL contracts on day one. | | Alphabet, Meta, Criteo, retail media networks | Ads $19.8B, +26% after four quarters at +22%. Sponsored-prompt clicks convert 48% more often and spend 21% more. | Worse for competing ad platforms. The acceleration source is agentic surfaces — a placement type competitors do not have at Amazon's purchase-intent proximity. Budget share is moving to lower-funnel, conversion-verified inventory. | | Disney, Fox, Warner Bros. Discovery, Comcast | Thursday Night Football, NBA, WNBA and NASCAR inventory "all sold out"; 30-plus new NBA advertisers in year one; NBA Game Seven peak of 6.5M US viewers, "outperforming Game Seven on broadcast a year ago." | Worse for linear sports broadcasters. Sold out means no discounting to fill. The broadcast comparison is a deliberate audience-transfer claim. | | NBA, NFL, WNBA, NASCAR (rights holders) | Same. | Better for the leagues — a sold-out, audience-growing streaming buyer strengthens the next rights negotiation. | | Netflix, Roku, The Trade Desk | Q1 context: Amazon Audiences applying Amazon's signals to Netflix inventory; Comcast Advertising partnership; interactive video starting with Samsung TVs. | Mixed. Better near term for Netflix, Comcast and Samsung. Worse structurally for independent DSPs and identity vendors — Amazon is becoming the identity and measurement layer for third-party CTV, which is The Trade Desk's franchise. | | Starlink/SpaceX, EchoStar, Viasat, Globalstar | Amazon Leo "close to 400 satellites in orbit, enough to begin initial satellite internet service this year… meaningful revenue commitments from enterprises and government customers… more than 20 partners." Constellation went from "over 250" in Q1. | Worse for incumbent satellite broadband. ~150 satellites added in one quarter with service starting this year makes Leo a real second entrant in the contested enterprise, government and aviation connectivity pool. | | Warehouse automation suppliers; warehouse labour | "We expect to more than double our fleet of robotic arms, like Cardinal and Sparrow, in 2026," with retrofit and greenfield demand. | Better for automation component suppliers (vision, actuators, motion control, sensors). Worse for warehouse labour demand and staffing exposed to fulfilment. | | Merchant power and IPPs | Electricity contracts large and structured enough to require derivative accounting, producing a $600M fair-value benefit. | Better for power producers. The accounting treatment is the tell: long-dated, fixed-price, physically-settled-shaped contracts at a size that moves a $27.5B operating-income line. | | Dell, HPE, Supermicro, IBM, on-prem storage | "85% of global IT spend is still on premises… that equation is going to flip," plus component allocation favouring hyperscalers. | Two-sided. Worse for the enterprise on-prem segment specifically — enterprises cannot buy the memory to refresh. Offsetting: the same vendors' hyperscaler and ODM-direct businesses benefit from a $220B Amazon build. Net direction depends on each vendor's mix. |

Cross-cutting conclusions

1. The macro that matters here is a supply macro, not a demand macro. Amazon changed its own risk language to name energy and memory, raised capex $20B purely on memory cost, and flagged fuel and driver-capacity inflation for a second straight quarter — while the demand commentary contains no caution at all. Position for cost inflation in physical inputs, not for demand deterioration.
2. The tariff cycle flipped from cost to price, and Amazon has committed to spending the windfall on lower prices rather than banking it. Disinflationary for goods; negative for competitors' retail gross margin.
3. The Trainium merchant-sale language is the most consequential quarter-over-quarter change, against a chips business already over $25B run rate growing triple digits with both Anthropic and OpenAI on multi-gigawatt commitments.
4. Amazon is now a direct competitive threat in five markets it was not credibly in two years ago — contract logistics, fresh grocery, application security, cross-vendor enterprise agents, and satellite broadband. Each is specific, named and quantified, not aspirational.
5. The honest bear case is Jassy's own. He described AI demand as "very barbelled" — labs and a handful of runaway consumer apps at one end, cost-avoidance projects at the other — and volunteered that the broad enterprise middle, which he expects to be "the largest absolute segment," may not ramp on "the same wildly steep trajectory." Backlog of $496B is real, but the reservation base is concentrated in the narrow end of that barbell, and it is priced at pre-inflation cost assumptions that cannot be repriced. That is where the risk sits.

| Watch item | The bar | Why it decides something | |---|---|---| | AWS growth | Above +36.7% for a sixth straight acceleration; sequential dollar adds above $4.6B | The entire thesis. Anything below and the "supply-gated not demand-gated" framing weakens | | AWS backlog / RPO | Above $496B, still growing faster than revenue | If RPO growth falls below revenue growth, the acceleration is being consumed rather than replenished | | AWS margin ex-derivative | Hold near 38% | Where the "AI margins track core, a little ahead" claim resolves, and where memory-cost inflation on a price-locked backlog would first show | | Revenue versus the guide | Above $202B (the high end) | Four straight quarters above the high end; the base rate says ~$205B | | Operating income versus the guide | Above $26.5B | Four straight quarters above the high end on a clean basis; the base rate says ~$30B | | Q4 revenue guide | Above $244B | The FY2026 revenue consensus of $825.6B requires it | | FY2026 capex | Held at ~$220B | A second raise attributed to components within two quarters invalidates the memory framing | | Free cash flow trajectory | TTM outflow no worse than −$7.6B | Jassy pre-committed to FCF pain "until these data centers come online." Hold him to it | | Funding disclosure | Any answer at all | Asked directly and refused this quarter. With ~$62B of quarterly capex in H2, silence a second time is a real negative | | North America margin | Above 7.3% clean | Ex-tariff-refund margin went backwards this quarter. Two in a row makes the cost-to-serve claim unsupportable | | Trainium merchant sales | Any firming beyond "a real chance" | Two quarters of escalating language on a line worth zero in consensus | | Same-day perishables cities | Above 2,300 | Frozen for three quarters while the narrative escalated. Either it moves or the metric was abandoned |