| Jul 30, 2026 | +15.32% | Q2 FY2026 | Read transcript briefingQuarter in one view- Revenue $200.6B, +20% YoY (Olsavsky frames it as ex-FX; includes Prime Day shifted into Q2 for largest geos). OI $27.5B, +43% — but includes ~$1.2B of one-time benefits: ~$600M tariff refunds (NA) and ~$600M energy-derivative fair-value gain (AWS). Ex-items OI ≈$26.3B, still above the $20–24B guide top end.
- AWS $42.2B, +36.7% — fifth straight acceleration, fastest in 18 quarters, +$4.6B QoQ ("~80% more than our largest increase ever"), $169B run rate. Backlog $496B, triple-digit YoY growth (vs. $364B ex-Anthropic last quarter — basis not restated this quarter).
- AI run rate "over $25B" (from >$15B last quarter); chips run rate "over $25B" (from >$20B). Both growing triple digits.
- Segments: NA $116.2B (+16%), OI $9.1B / 7.9% (flat QoQ despite tariff-refund help); International $42.2B (+15% ex-FX), OI $1.7B / 4.1% (from 3.6%). AWS OI $16.6B, ~39.3% margin, +650bps YoY (+520bps ex-derivative gain).
- Cash capex $53.1B in Q2; 2026 capex guidance raised ~$200B → ~$220B, explicitly attributed to memory cost inflation. Jassy conceded capacity shortfalls through 2027 and flagged "striking" 2028 demand.
- Q3 guide: revenue $197–202B (Prime Day shift worth ~400bps; FX ~80bps headwind), OI $22.5–26.5B.
- Jassy delivered his first full ROIC framework for the capex cycle (data centers 30+ yr life; servers ~3-yr breakeven, 5–6-yr life, mostly 5-yr contracts) and acknowledged debt issuance as a funding source — the first capital-structure acknowledgment on these calls.
What management is focused on- The ROIC defense, upgraded from assertion to arithmetic: data-center capital spent ~2 years pre-monetization then monetized for 30+ years across "five to six generations of server economics"; servers bought months before service with demand visibility, breakeven "a little less than three years," then 2–3 years of "significant free cash flow" per server cohort. Near-term FCF headwinds conceded again; "revenue growth outpaces incremental capex growth... will happen at some point" — still undated.
- AWS as a potential "$1 trillion annual revenue business" — new long-range framing, doubling the prior "few hundred billion" language. AI margins "tracking what we saw with Core at the same point of evolution, actually a little ahead."
- The barbell demand narrative: labs + runaway apps (Claude Code, ChatGPT) on one end, enterprise cost-avoidance use cases on the other, and the unconverted middle (existing production workloads) as "the largest absolute segment" — this is the answer to the concentration critique, asserted rather than evidenced.
- Agentic application layer as a real business: Kiro (tripled QoQ, "up to 50% more cost-effective"), Amazon Q (named enterprise list: 3M, Allianz, AstraZeneca, BMW, Exxon, NFL...), Connect (all five major airlines), Transform, and new AWS Continuum (agentic security remediation). Management is now explicitly building the app layer, per Sebastian's question.
- Own frontier model: Jassy now confirms "we are pursuing our own frontier model" — a direct answer to the Nova-prioritization question he deflected last quarter, justified on cost control, prioritization, and speed, while insisting AWS can be "wildly successful" without one.
- Grocery/perishables as the everyday-essentials flywheel: same-day perishables in 2,300 U.S. cities, 9 of top 10 sellers perishable in those geos, monthly perishable customers +50% YTD, Whole Foods Daily Shop format "expanding very rapidly."
- New B2B surface: Amazon Supply Chain Services (P&G, 3M, Lands' End, American Eagle) — selling the logistics network itself.
Key numbers and quarter mechanics- Revenue $200.6B (+20%); OI $27.5B (+43%); the ~$1.2B of benefits (~$600M tariff refunds, "significant majority" of expected refunds; ~$600M energy-derivative gain, AWS) means ~4.3% of OI was non-operational in nature. Derivative gains/losses recur quarterly and "can vary."
- AWS: $42.2B, +36.7%; OI $16.6B; margin ~39.3%, +650bps YoY, +520bps ex-derivative. Olsavsky credited "capacity optimization, which we benefited quite a bit from in Q2" — implying some margin upside was utilization timing, not pure structure.
- NA: $116.2B (+16%), OI $9.1B / 7.9% — margin flat vs. Q1's 7.9% despite the $600M tariff refund landing in NA; underlying NA margin ex-refund ≈7.4%. The ~$1B Leo headwind guided for Q2 was not explicitly reconciled on the call.
- International: $42.2B (+15% ex-FX), OI $1.7B / 4.1% — inflection extended (2.1% → 3.6% → 4.1%).
- Units +17% (from +15%); shipping costs ex-fuel/line-haul grew slower than units "at a pace relatively consistent with last quarter"; fuel inflation from Middle East conflict + driver-capacity line-haul rates partially offset by the April FBA surcharge. No seller-attrition data offered.
- Ads $19.8B, +26% (from +22%) — acceleration; sponsored-prompt clickers convert 48% more often and spend 21% more; Ads Agent users see 8% lower CPM / 6% lower CPA; TNF/NBA/WNBA/NASCAR inventory sold out.
- Capex $53.1B in Q2 (from $43.2B); H1 total ~$96.3B against the new ~$220B full-year plan — implies H2 acceleration to ~$124B. Memory inflation named as the driver of the raise; Gawrelski's pricing question confirmed existing contracts are fixed-price (inflation absorbed in-contract; new deals priced to costs) — a margin risk embedded in the $496B backlog.
- Q3 guide: $197–202B revenue (sequential deceleration is Prime Day mechanics: ~400bps drag; FX ~80bps headwind), OI $22.5–26.5B. Guidance carve-out for acquisitions/restructurings/settlements retained.
- Not disclosed again: FCF, 3P mix, Rainier chip counts (third quarter), Alexa+ subscribers/revenue, Leo revenue, OpenAI deal sizing. Rufus was not mentioned at all — the consumer-AI shopping narrative has been rebranded to "Alexa for Shopping" (350M users, interactions +5x).
Product and launch scorecard- Trainium: Anthropic + OpenAI "multi-year, multi-gigawatt commitments" restated; startup adopters named (Neurorobotics, Odyssey, Twelve Labs, Poolside, etc.) plus Uber and Pinterest — the first breadth evidence beyond labs. Third-party/rack sales: "a real chance we'll do that in the future" — moved from floated idea to active conversations. No T2/T3/T4 subscription-status update this quarter (last quarter's "largely sold out"/"nearly fully subscribed" language not refreshed).
- Graviton: revenue commitments ~3x QoQ; Graviton5 growing ~2x faster than Graviton4; 98% of top-1,000 EC2 customers (held). Price-performance claim moderated to "30–40% better" (from "up to 40%").
- Bedrock: Anmuth cited customers spending more in Q2 than all prior quarters combined (not stated by management in prepared remarks); frontier-model strategy clarified (selection + own model in progress).
- Kiro: usage tripled QoQ, cost-effectiveness claim vs. Claude Code/Codex — first competitive positioning numbers in coding agents.
- Amazon Q: autonomous background agents, activity feed, 16 integrations (Adobe, Moody's, Snowflake); long named-customer list — now a production enterprise product, not a preview.
- AWS Continuum (new): agentic vulnerability discovery/validation/remediation; unsized, but security is the stated #1 enterprise AI objection.
- Alexa for Shopping: 350M users LTM, active users ~2x, interactions +5x YoY; Alexa+ expanded to Germany, Austria, France, Brazil; Alexa+ triers convert to Prime at ~25% higher rates; Alexa shoppers spend 40%+ more per order. Still zero revenue/subscriber disclosure for the paid tier.
- Amazon Now: 30-minute delivery; +80 U.S. cities + Egypt in Q2; 9 countries, 250+ cities; sales/units +80% QoQ, customers +60% QoQ.
- Grocery: perishables same-day in 2,300 cities; perishable MAUs +50% YTD; fresh = 6 of top 20 amazon.com bestsellers; Pharmacy new customers +2x H1, same-day prescriptions ~5x, ~$250M customer savings (+400%).
- Leo: "close to 400 satellites" (from 250+), "enough to begin initial satellite internet service this year," 20+ distribution partners, "meaningful revenue commitments." Note: the committed Q3 commercial launch was not reaffirmed with a date, and the Q4 capitalization trigger went unmentioned — language drifted from "commercial launch Q3" to "initial service this year." Verify in filings whether the schedule slipped.
- Ads: sponsored prompts now have conversion economics (48%/21%); NBA season one added 30+ advertisers; multi-sport advertisers get 2.3x reach, +12% spend, +17% orders.
Sell-side read-through- Anmuth (J.P. Morgan) asked the right margin question: is 39% AWS margin sustainable given AI-mix dilution assumptions? Olsavsky: +650bps YoY, +520bps ex-derivative, "isn't random," but "will fluctuate" — and credited Q2 capacity optimization, hinting at timing benefit. His frontier-model question got the first straight answer in three quarters: yes, building one.
- Post (BofA) tried to quantify capacity adds (H2 vs H1, 2027 gigawatts) — Jassy gave no numbers, only "on pace" to double power capacity by end-2027 vs. 2025. The gigawatt disclosure peers provide remains withheld.
- Nowak (Morgan Stanley) asked whether long-lived data-center spend can slow in 2027 — effectively the capex-peak question, third variant in three quarters. Answer: no ("so much demand... 2027 largely reserved... quite a bit reserved for 2028"). His Trainium rack-sale question got the most concrete answer yet: active conversations, "real chance."
- Sebastian (Baird) got the quarter's capital-structure admission: "You've seen us issue debt this year. We have a lot of options... nothing to share today." First acknowledgment that the build-out needs external funding; no quantum, no framework.
- Gawrelski (Wells Fargo) surfaced RPO at 2.5x the Q3 2025 level and pressed on 2028+ capacity and contract pricing vs. inflation. Jassy confirmed backlog is in capex projections but gave no 2028 capacity framework, and confirmed existing contracts are fixed-price — meaning memory/SSD inflation is absorbed within current contract terms. That is the most underappreciated margin risk disclosed this call.
- Sheridan (Goldman) pivoted to commerce adoption — got grocery/perishables metrics but no country-level divergence data and no unit economics.
- Cross-read: five questions (Helfstein no-showed). Nobody asked about the $1.2B one-time items' quality, FCF (seventh straight call without disclosure), the Leo launch-date drift, Rufus's disappearance, 3P mix, or NA margin ex-tariff-refund. The Q&A remains friendly; the hardest question (capex deceleration) got its third consecutive non-answer, now upgraded with a $220B raise.
Management credibility- Delivered: AWS acceleration continued as the supply-constraint narrative implied (28% → 36.7%); the AI run-rate disclosure given last quarter (>$15B) was updated upward (>$25B) rather than buried — good disclosure hygiene; Q2 revenue and OI both landed above guide top end ($200.6B vs. $199B; $27.5B vs. $24B); International margin inflection extended to 4.1%; the FBA surcharge mechanism described last quarter was confirmed operating; the ROIC framework promised implicitly for quarters was finally delivered with specifics (breakeven periods, useful lives, contract durations).
- Slipped/softened: Leo language moved from "commercial launch Q3" to "initial satellite internet service this year" with no Q3 reaffirmation and no Q4 capitalization mention — a quiet schedule/scope softening inside a quarter that ends in September; Graviton price-performance claim trimmed to "30–40%"; no Trainium generation-by-generation subscription update after last quarter's detailed T2/T3/T4 read; Rufus metrics (MAU +115%, engagement +400%) replaced wholesale by Alexa-for-Shopping metrics — the second consumer-AI metric rotation.
- New credibility positives: the capex raise was attributed to a named cause (memory) rather than opaquely; the $1.2B one-time items were proactively itemized with segment placement and the ex-item margin bridge (+520bps) volunteered; tariff-refund treatment (proactive customer refunds where pass-through is traceable) was explained in unusual detail.
- Consistency issues: backlog $496B "growing triple digits" vs. last quarter's $364B ex-Anthropic — the Anthropic inclusion basis was not restated, so QoQ comparability is unverifiable; "AI margins tracking ahead of Core at the same point" is unverifiable assertion; the $1T AWS framing is aspiration presented alongside contracted-demand evidence; fixed-price contracts vs. rising memory costs received no margin-impact quantification despite a direct question.
- Under pressure: capex-deceleration question → non-answer #3; gigawatt disclosure → declined; capital sources → "nothing to share" after conceding debt issuance; 2028 capacity framework → declined. The pattern holds: management answers strategy questions richly and capital-boundary questions minimally.
What changed versus the prior quarter- AWS: 28% → 36.7% (18-quarter best); run rate $150B → $169B; backlog $364B-ex-Anthropic → $496B (basis unclear); AI run rate >$15B → >$25B; chips >$20B → >$25B; AWS margin ~37.8% → ~39.3% (+520bps ex-derivative).
- Capex: ~$200B → ~$220B for 2026, memory inflation named; quarterly print $43.2B → $53.1B; first-ever ROIC mechanics disclosed; first acknowledgment of debt funding.
- Frontier models: from deflecting the Nova question to confirming an in-house frontier model effort — a real strategy disclosure.
- Trainium: from "sold out/subscribed" status updates to third-party rack-sale conversations; no per-generation supply update (disclosure granularity decreased).
- Consumer AI: Rufus disappeared; Alexa for Shopping (350M users) is now the lead metric; Alexa+ gained Prime-conversion data (25% higher sign-up) but still no revenue.
- Leo: 250+ → ~400 satellites; "Q3 commercial launch" → "initial service this year"; Q4 capitalization unmentioned; 20+ partners new.
- Margin quality: Q1 was apparently clean; Q2 is clean of charges but carries $1.2B of one-time gains — the "clean quarter" test now cuts the other way (verify recurring-derivative treatment in the 10-Q).
- Retail: units 15% → 17%; NA margin flat at 7.9% despite $600M refund; International 3.6% → 4.1%; ads 22% → 26%; Amazon Now 9 countries held, cities to 250+.
- Omissions now standing: FCF (7 calls), 3P mix (2), Rainier counts (3), Rufus revenue (3+), Alexa+ revenue (3+), Leo revenue/pricing, OpenAI deal size.
Bull case- The demand thesis keeps outrunning the build: 36.7% growth on a $169B run rate, $496B backlog growing triple digits, 2027 capacity "largely reserved," 2028 reservations underway, and management still capacity-constrained at $220B capex — the constraint is supply, not demand, for the fifth consecutive quarter of acceleration.
- Margin structure is holding through the AI mix shift: AWS ~39% (+520bps ex-one-time), with management claiming AI margins track ahead of Core's historical curve and Trainium/Graviton providing structural cost advantage — the "AI is dilutive" assumption took a direct hit this quarter.
- The ROIC framework is now falsifiable: ~3-year server breakevens, 5–6-year lives, 5-year contracts, 30-year data centers with 5–6 server generations — if revenue growth outpaces incremental capex "at some point," the FCF inflection is a mechanical outcome of cohort math, and 2028 demand is already partially contracted.
- Breadth evidence improved materially: Trainium adopted beyond the two labs (Uber, Pinterest, named startups), Graviton commitments ~3x QoQ, Q in production at a dozen-plus blue chips, Bedrock spend exceeding all prior quarters combined.
- Consumer flywheel compounding: units +17%, perishables +50% MAUs with 3x basket attachment, Alexa-for-Shopping at 350M users with 5x interaction growth, ads +26% with agentic formats showing 48% higher conversion — retail engagement and monetization are reinforcing.
- International margin inflection is now a trend (2.1% → 3.6% → 4.1%), and new revenue surfaces (Supply Chain Services, Continuum, Leo commitments) keep stacking.
Bear case- One-time items flattered the print: $1.2B (~4.3% of OI) from tariff refunds and derivative gains; NA margin ex-refund ≈7.4%, below the reported 7.9% and well below the 9.0% Q4 peak; AWS margin ex-derivative is +520bps, not +650bps. The beat-over-guide quality is lower than the headline.
- Capex raised to ~$220B with the cause (memory inflation) explicitly unhedged inside fixed-price contracts — Gawrelski's answer confirmed existing backlog reprices only at renewal. Rising input costs against $496B of fixed-price paper is a margin squeeze mechanism nobody quantified.
- Capital structure is now in play: debt issuance acknowledged with "nothing to share" on plans; FCF undisclosed for a seventh call; the capex-deceleration question has now been asked three ways and answered zero — the peak is not in sight by management's own telling ("true in 2027 too").
- Leo schedule drift: "commercial launch Q3" became "initial service this year" with the Q4 capitalization trigger unmentioned — in the quarter the launch was supposed to happen. Combined with still-zero revenue disclosure, the "many billion-dollar business" remains unpriced and undated.
- Metric rotation continues: Rufus vanished after two quarters of non-repetition of the $10B claim; Trainium per-generation subscription detail dropped after one quarter; backlog basis (Anthropic in or out) not restated. Metrics appear when strong and rotate when they stop improving.
- Concentration risk unaddressed: the barbell framing concedes labs and two apps drive current consumption; the "middle of the barbell" (enterprise production inference) is asserted as the largest future segment with no adoption data; fixed-price lab contracts plus memory inflation plus circular OpenAI/Anthropic economics remain unexamined.
- Q3 guide implies deceleration even adjusting for Prime Day (~400bps) and FX (~80bps): $197–202B is +~9–12% on a headline basis against +20% in Q2 — the stores business is lapping into a much tougher H2.
Next-quarter watchlist- 10-Q reconciliation: treatment and recurrence of the $600M energy-derivative gain (Olsavsky flagged it "can vary" — a Q3 reversal would hit AWS margin optically); confirmation no severance/restructuring recurred; TTM FCF print; debt balance and any new financing structures; 3P mix; whether backlog/RPO footnotes clarify the Anthropic basis.
- Leo: did "initial service" actually begin in Q3 as implied? Is the Q4 capitalization trigger still operative (NA-margin optics)? Launch cadence vs. ~400 satellites; first revenue or per-subscriber pricing; Globalstar close status (unmentioned this call).
- AWS: does growth hold above ~30% as compares toughen; backlog progression on a stated basis; depreciation step-up vs. the "capacity optimization" benefit that flattered Q2 margin; any quantification of memory-cost absorption in fixed-price contracts; rack-sale decision ("real chance" → commitment?); gigawatt disclosure (declined twice now).
- Capex/funding: H2 implied ~$124B pace vs. the $220B guide; size/terms of debt issuance; any 2027 capex framework; whether "revenue growth outpaces capex growth at some point" gets a date.
- Margins: NA ex-tariff-refund trajectory vs. the 7.9% reported plateau; International holding ~4%; AWS margin ex-derivative as the clean trend line; fuel/line-haul inflation vs. FBA surcharge offset and any seller attrition data.
- Q3 delivery: revenue $197–202B and OI $22.5–26.5B — where in the range, and whether the ex-Prime-Day ~400bps adjustment holds as described.
- Metric continuity: does Alexa-for-Shopping data repeat (or rotate like Rufus); Trainium per-generation subscription status; Alexa+ first subscriber/revenue number; Rainier chip counts (three quarters overdue); OpenAI deal sizing; Kiro/Q revenue contribution vs. usage stats.
|
| Apr 29, 2026 | +0.77% | Q1 FY2026 | Read transcript briefingQuarter in one view- Revenue $181.5B, +17% reported / +15% ex-FX ($2.9B FX tailwind, +180bps). OI $23.9B, OPM 13.1% — management calls it the highest operating margin ever. No special charges were mentioned on the call; verify cleanness in the 10-Q after two consecutive charged quarters ($4.3B in Q3, $2.4B in Q4).
- AWS accelerated again: +28% (from 24%), fastest in 15 quarters, +$2B QoQ ("largest Q4→Q1 AWS increase ever"), $150B annualized run rate. AI revenue run rate disclosed for the first time at "over $15B" — a long-standing watchlist item is now answered.
- Backlog disclosed at $364B, explicitly excluding the newly announced >$100B Anthropic deal — effectively >$464B of contracted demand vs. $244B last quarter. This is the quarter's central disclosure.
- Segments: NA $104.1B (+12%), OI $8.3B / 7.9%; International $39.8B (+11% ex-FX), OI $1.4B / 3.6% — a real international inflection vs. 2.1% last quarter. Companywide OPM 13.1%.
- Cash capex $43.2B in Q1, "primarily AWS and generative AI." Jassy, asked directly for a capex update: "our plan is largely the same" — i.e., ~$200B for 2026 reiterated, no new guardrails.
- Leo moved from story to schedule: commercial launch now committed to Q3 (was "later this year"), capitalization of production/launch costs starting Q4 (was "later in the year"), 250+ satellites, Globalstar acquisition announced, Apple direct-to-device agreement, Delta/NASA added to the customer list.
- Q2 guide: revenue $194–199B (FX ~10bps headwind — the tailwind flips), OI $20–24B, embedding SBC step-up, ~$1B YoY Leo cost in NA, and fuel-inflation transport costs partially offset by a new FBA fuel/logistics surcharge.
What management is focused on- Contracting demand as the capex defense: the narrative shifted from "monetizing as we install" to showing signed paper — $364B backlog (ex-Anthropic), >$225B of Trainium revenue commitments, >$15B AI run rate, and a named deal list (OpenAI, Anthropic, Meta, NVIDIA, Uber, U.S. Army, Bloomberg, Cerebras, AT&T...). Jassy now also concedes the FCF mechanics: cash laid out 6–24 months before billing, "early years' free cash flow is challenged" — an admission, framed with the 2003-era AWS analogy.
- Elevating the silicon franchise: chips ">$20B run rate," triple-digit YoY, +~40% QoQ; a new counterfactual — "if standalone, $50B run rate" and "top 3 data-center chip business in the world." Trainium economics quantified for the first time: "tens of billions of CapEx saved each year" and "several hundred basis points of operating margin advantage" vs. third-party chips for inference. Rack sales floated as "a good chance over the next couple of years."
- The stateful-agent architecture pitch: AWS's differentiation claim moved from model selection to stateful runtime — Bedrock managed agents co-built with OpenAI (preview), Strands (25M+ downloads, 3x QoQ), AgentCore (a deployment every 10 seconds). This is the moat argument replacing "we have many models."
- Agentic AI as a CPU story: Meta committing "tens of millions of Graviton cores" is being used to argue that inference/agentic workloads pull CPUs, positioning Graviton (98% of top-1,000 EC2 customers) alongside Trainium.
- Leo as a business, not a cost: launch commitment, named enterprise/government commitments (Delta "at least half of fleet from 2028," NASA), Globalstar spectrum acquisition for direct-to-device, Apple relationship, and an explicit Q4 capitalization date that will optically relieve NA margins.
- Internal AI productivity as an operating thesis: the 5-engineers-in-65-days rebuild story (vs. "40–50 people for a year") is management explicitly linking agentic coding to workforce efficiency — my inference: this quietly rationalizes the recurring severance charges.
- Retail speed/spatial expansion: 1-hour and 3-hour delivery tiers (90K+ items), Amazon Now now in 9 countries, Health AI launch, and grocery framed as "#2 grocer in the U.S."
Key numbers and quarter mechanics- Segments: NA $104.1B (+12%), OI $8.3B / 7.9% (vs. 9.0% in the Q4 peak quarter); International $39.8B (+11% ex-FX), OI $1.4B / 3.6%; AWS $37.6B (+28%, +480bps acceleration), OI $14.2B / ~37.8% margin — margin up from ~35% despite the depreciation headwind flagged last quarter.
- Units +15% YoY — "highest since the tail end of COVID lockdowns" — while outbound shipping grew +12% and fulfillment expense +9% (both FX-neutral): continued cost-to-serve leverage.
- Ads $17.2B, +22% YoY (matches last quarter's growth rate on the Q1 seasonally smaller base).
- Capex $43.2B cash in Q1 — annualizes to ~$173B, below the reiterated ~$200B plan; either back-loaded or supply-paced. FCF was not disclosed or discussed — the sixth-decline trend remains unresolved by silence.
- FX mechanics: Q1 +180bps favorable; Q2 guide assumes ~10bps headwind. Prime Day timing shifts comps: Q2 (June) for largest geos including the U.S.; Q3 for Australia, Brazil, India, Japan (2025: all in Q3).
- Q2 guide details: OI $20–24B; includes seasonal SBC step-up; ~$1B YoY NA Leo headwind repeats; fuel-driven transportation costs partly offset by the newly implemented FBA fuel/logistics surcharge (cost pushed to sellers — watch 3P reaction; 3P mix was not updated this call).
- Grocery: $150B+ gross sales in 2025, claimed #2 U.S. grocer; perishables +40x YoY; 9 of top 10 same-day items; perishable same-day shoppers add ~3x items, spend +80%; Whole Foods 550+ stores, 100+ more planned.
- Delivery: 1B+ items same-day/overnight YTD; 1-hour in hundreds of cities, 3-hour in 2,000+ cities, on 90K+ items; average prices on Amazon.com decreased YoY in Q1 (management claim).
- Guidance carve-out retained: excludes future acquisitions/restructurings/settlements — relevant given the announced Globalstar deal and the recurring-severance history.
Product and launch scorecard- Trainium: commitments >$225B (new disclosure granularity). T2 "largely sold out" — note the softening from last quarter's "fully subscribed" language. T3 "nearly fully subscribed" (consistent with the prior "nearly all supply committed by mid-2026"). T4, ~18 months from broad availability, already "much reserved" — earliest-ever reservation signal in the roadmap. Chip-count/Rainier disclosure remains absent for a second straight quarter.
- OpenAI on Bedrock: delivered — GPT-5.4 live yesterday, 5.5 "in the next couple of weeks," plus the stateful managed-agents preview co-built with OpenAI. The November agreement now has product substance; the financial sizing of the deal is still undisclosed.
- Bedrock breadth: 125K+ customers, ~80% of Fortune 100, customer spend +170% QoQ (vs. +60% QoQ in Q4), Q1 tokens exceeded all prior years combined, majority of inference on Trainium. This is the partial answer to the "concentration vs. breadth" flag — breadth is arriving via Bedrock abstraction, not direct chip buyers.
- Graviton: first named mega-commitment (Meta, "tens of millions of cores"); 98% of top-1,000 EC2 customers (from 90%+); "up to 40% better price performance than any other x86." The CPU-for-agentic framing is new and testable.
- Q ("QRO" in transcript): developers >2x QoQ, enterprise usage ~10x; Transform 1.56M hours saved; Quick new customers >4x QoQ with a new desktop app — the agent suite now has usage metrics, not just launch language.
- Rufus: MAU +115% YoY, engagement +400% YoY; sponsored prompts inside Rufus show ~20% conversation continuation. The prior $10B incremental-sales claim remains un-repeated — second watchlist miss.
- Amazon Now: 3 countries → 9 countries; India orders +25% MoM; Prime tripling-frequency claim repeated.
- Health AI (new): 24/7 clinical agent with One Medical backing — booking, prescriptions, treatment facilitation. Early, unsized.
- Leo: Q3 commercial launch committed; 250+ satellites; 20+ launches 2026 / 30+ 2027; 2x downlink / 6x uplink performance claims vs. "existing alternatives"; Globalstar acquisition (spectrum + satellite know-how) and Apple direct-to-device agreement are the quarter's two largest incremental strategic moves. Delta commitment starts 2028 — revenue is still mostly forward.
- Alexa+: geographic expansion (Mexico, U.K., Italy, Spain) and engagement stats repeated; still zero subscriber/revenue disclosure for the $19.99 tier.
- Zoox: public in Las Vegas and SF, testing 8 more cities, ~2M miles / 350K+ riders, Uber-app distribution announced — first third-party distribution channel.
- Ads tooling: Creative Agent expanded to 7 new markets; Netflix inventory addressable via Amazon Audiences; Comcast local-ad partnership; Samsung interactive formats — supply-side expansion continues.
Sell-side read-through- Sheridan (Goldman) asked the capex-return question again (scale of investment vs. backlog). Answer: "no new update on capital... plan is largely the same" plus once-in-a-lifetime framing. The guardrail non-answer from last quarter was simply repeated, one quarter on.
- Nowak (Morgan Stanley) extracted the quarter's key data point: backlog $364B excluding the >$100B Anthropic deal, with breadth asserted ("not just 1 or 2 customers") but not evidenced. His agentic-commerce milestones question got engagement metrics and the repeated critique of third-party horizontal agents (wrong prices, no history) — same answer as last quarter, no partnership or monetization news.
- Post (BofA) got confirmation OpenAI's full suite lands on Bedrock and that rack sales are genuinely contemplated. His Nova-focus sub-question was deflected — Jassy never addressed own-model investment prioritization.
- Sanderson (Loop) pushed Leo sizing: got qualitative TAM ("billions without broadband"), a "very large many billion-dollar revenue business" claim, and the AWS-bundling pitch for governments/enterprises. His orbital-data-center/long-tail question was explicitly skipped ("I'll try and address as many as I can") — a clean partial evasion.
- Khajuria (Wolfe) surfaced the newest disclosed risk: memory/storage inflation. Jassy conceded cost "skyrocketed," claimed supply secured via strategic partners, asserted "not capacity constrained," and spun it as a cloud-migration accelerant. No quantification of capex impact for 2026–27 was given despite being asked directly. Her agentic-ads question produced the first substantive response to the funnel-compression risk Morton raised last quarter: multi-turn conversations create more surfaces, sponsored prompts, and cheaper creative → "advertising will do well in a world of agentic commerce." Asserted, not measured.
- Sebastian (Baird) on demand breadth: Jassy conceded labs are the outsized spenders ("several of those labs") while claiming enterprise production projects are scaling. The internal-AI answer contained the call's most consequential operational claim: 5 engineers rebuilt a service in 65 days vs. a 40–50 person-year norm — read alongside recurring severance, this is a headcount-deflation argument management will not make explicitly (my inference, not their statement).
- Cross-read: six questions again, and again none on FCF (undisclosed this call), the NA margin's 9.0%→7.9% move, the missing 3P mix figure, the prior quarter's garbled guidance resolution, or whether Q1 was actually charge-free. The Q&A remains structurally friendly; the single hardest question (capex guardrails) received its second consecutive non-answer.
Management credibility- Delivered: AWS re-accelerated (24%→28%) as the "supply-constrained" claim implied; the AI absolute revenue number finally appeared (>$15B run rate) — a multi-quarter watchlist item closed; OpenAI models shipped to Bedrock on the timeline implied by the November deal; Trainium3 supply "nearly fully subscribed" consistent with the mid-2026 commitment; Leo's launch window narrowed to Q3 and the capitalization date pinned to Q4 — both vaguer promises converted to dated ones; International margin delivered its inflection (2.1%→3.6%) despite guided investment.
- Slipped/softened: T2 language moved "fully subscribed" → "largely sold out" (minor but real); Rainier chip counts absent for a second quarter; Rufus $10B claim still un-revised; Alexa+ monetization data still withheld after the pricing was announced two quarters ago; FCF simply not disclosed on a call where capex was the first question.
- Consistency issues: "chips >$20B run rate" vs. "$50B if standalone" is a counterfactual the CFO cannot validate; "top 3 data center chip businesses in the world" is self-assessed; backlog "breadth" is asserted one quarter after Trainium commitments were framed around "the 2 leading AI labs"; the "$23.9B actual vs. presumed $16.5–21.5B Q1 guide" gap (pending press-release verification of the garbled low end) is a ~$2.4B-above-top-end beat that raises its own questions about guide conservatism or efficiency timing.
- Under pressure: capex-guardrail question #2 answered with same philosophy; memory-cost quantification declined; orbital-data-center question skipped; Nova prioritization deflected. Jassy's "this isn't some sort of quixotic top line grab" from Q4 has now evolved into showing contracts — materially better evidence, still no returns math (no ROIC percentages, no FCF floor, no cycle duration).
- New positive disclosure discipline: backlog given with explicit Anthropic exclusion (unusual specificity), Trainium commitment dollar total ($225B) disclosed, and Q2 guide itemized (SBC step-up, ~$1B Leo, fuel offset via surcharge).
What changed versus the prior quarter- AWS: 24%→28% (15-quarter best vs. 13-quarter); run rate $142B→$150B; backlog $244B→$364B plus >$100B Anthropic disclosed separately; AI run rate >$15B disclosed for the first time; Q1→Q2 guidance mechanics now include Prime Day timing shifts.
- Chips: run rate >$10B→>$20B; Trainium commitments $225B newly totaled; T2 "fully subscribed"→"largely sold out"; T4 now "reserved" not just "substantial interest"; Graviton gained a named anchor (Meta) and the agentic-CPU thesis emerged.
- AI-revenue transparency: from refusal to disclose (prior quarters) to a voluntary absolute number — the single largest disclosure change of the quarter.
- Margin mix: company OPM 13.1% (record per management); NA down sequentially 9.0%→7.9% with the ~$1B Leo drag repeating into Q2; International up 2.1%→3.6% — the long-deferred inflection printed.
- Charges: none mentioned after $2.4B (Q4) and $4.3B (Q3) — first apparently clean test passed verbally; confirmation required in the 10-Q and in whether severance recurs despite the guidance assumption.
- Capex: from guidance (~$200B) to first actual print ($43.2B); the plan was reaffirmed but the Q1 pace annualizes below it.
- Leo: 180→250+ satellites; "later this year"→Q3 launch; "later in the year"→Q4 capitalization; customer list broadened (Delta, NASA, Vodafone, DP World Tour); Globalstar acquisition and Apple direct-to-device agreement are entirely new.
- Retail: units 12%→15%; Amazon Now 3→9 countries; grocery framed as #2 U.S. player; prices claimed down YoY; FBA fuel surcharge introduced (new inflation pass-through lever).
- Omissions that changed: Rainier chips and the 1M target (second consecutive silence), 3P mix (not disclosed), TTM FCF (not disclosed), Rufus $10B (still absent), Alexa+ revenue (still absent).
- Risk flags: memory inflation elevated from a preamble clause to a named Q&A topic with an asserted-but-unquantified mitigation; FX flipped tailwind→slight headwind.
Bull case- Demand is now contracted, visible, and accelerating: $364B backlog ex-Anthropic, plus >$100B Anthropic, plus >$225B Trainium commitments, against 28% growth on a $150B run rate — the supply-constraint claim is now backed by signed multi-gigawatt paper, not just management assertion.
- The silicon model is demonstrating self-funding economics: tens of billions in annual CapEx avoidance and "several hundred basis points" of margin advantage vs. third-party inference silicon, at a >$20B run rate growing triple digits — if the T3/T4 subscription data holds, AWS's margin structure improves while rivals rent GPUs.
- Breadth is finally showing: Bedrock at 125K customers and ~80% of the Fortune 100 with +170% QoQ spend growth answers the concentration critique better than last quarter's rhetoric; Graviton adds a second, CPU-sided franchise with a named anchor (Meta).
- Margin structure is inflecting broadly: record 13.1% company OPM, AWS ~37.8%, International 3.6% and improving, with fulfillment (+9%) and shipping (+12%) growing below units (+15%).
- Leo crossed from concept to countdown: dated Q3 launch, dated Q4 capitalization (optically relieving NA), real enterprise/government commitments, exclusive Apple distribution, and scarce spectrum via Globalstar — the "many billion-dollar revenue" framing is at least now testable.
- Multiple consumer-AI vectors are compounding with metrics: Rufus engagement +400%, Amazon Now +25% MoM in India with tripling frequency, ads +22% with new Netflix/Samsung supply and sponsored-prompt formats native to agentic interfaces.
Bear case- Capital consumption dominates and remains unbounded: $43.2B in one quarter, ~$200B reiterated, "significant capital over the coming years," FCF undisclosed on the call after six straight TTM declines, and no floor/governor/duration offered for the second consecutive quarter — the ROIC debate is being answered with backlog paper, not returns math.
- The demand paper is concentrated in counterparties trained to spend ahead of revenue: two AI labs anchor Trainium commitments and the largest single deal (>$100B Anthropic) had to be excluded from backlog to make the breadth claim; OpenAI simultaneously collects AWS compute commitments while AWS resells OpenAI models — circularity and counterparty-quality risk are unexamined on the call.
- Self-referential silicon economics: Trainium is "sold out" largely into Bedrock/Amazon's own consumption; the "$50B if standalone" and "top 3 globally" framings are marketing arithmetic with no external validation.
- Memory inflation is a new, acknowledged cost vector with an unquantified capex impact; mitigation ("we're not capacity constrained") is asserted, and the only disclosed offset mechanism (FBA surcharge) transfers cost to sellers — against a 3P mix that already slipped to 61% and wasn't updated.
- NA margin fell 9.0%→7.9% with another ~$1B Leo headwind in Q2 and fuel inflation rising; the Q4 9% peak may mark the ceiling of the old model rather than a base.
- Accumulating silence list: Rainier chip counts (two quarters), Rufus revenue attribution, Alexa+ monetization, TTM FCF, 3P mix, Leo revenue timing — each individually small, collectively a pattern of rotating away from metrics that stop cooperating.
- Internal-AI productivity (5 engineers doing 50-person work) implies structural headcount reduction that management will not state — while severance charges have recurred in two of the last three quarters; if the Q1 clean-quarter claim breaks in the 10-Q, the credibility of both threads degrades together.
Next-quarter watchlist- 10-Q reconciliation: confirm Q1 was truly charge-free; TTM FCF print vs. the $11.2B prior read; 3P mix direction; stock-comp step-up; any new leases/financing structures; whether the press release validated the presumed $16.5B Q1 OI guide low end.
- Q2 delivery: revenue $194–199B with Prime Day shifted into the quarter; OI $20–24B against SBC step-up + ~$1B Leo + fuel — where in the range it lands, and whether the FBA surcharge triggers seller attrition or mix shifts.
- AWS: does 28% hold against tougher compares; backlog progression including vs. excluding Anthropic; first revenue contribution from the OpenAI cohort; Trainium3 ramp vs. "nearly fully subscribed"; any rack-sale decision; depreciation step-up as 2025–26 capacity enters service.
- Capex path: is $43.2B a trough or the run rate vs. ~$200B; memory-cost pass-through quantification (Khajuria's unanswered question, redux); any first acknowledgment of financing/leasing/capital-return framing.
- Leo mechanics: actual Q3 commercial launch; Q4 capitalization trigger date and its NA-margin optics; launch cadence vs. 20+ planned for 2026 (250+ satellites implies the cadence must steepen); Globalstar close conditions and Apple economics; first revenue disclosure and per-subscriber pricing.
- Concentration reconciliation: backlog breadth evidence (customer count, ticket mix) vs. the "not just 1 or 2 customers" assertion; sizing of the OpenAI agreement; sustainability of "several labs" spending.
- Retail margins: NA margin trajectory vs. the 9% Q4 peak; International holding ~3.6% while seller fees fall in Europe/Brazil; units vs. shipping/fulfillment cost spread persistence; Amazon Now unit economics in any new market.
- Returns filings: any ROIC/FROI quantification to back the "compelling" adjective; Trainium margin-advantage claim ("several hundred basis points") showing up in AWS OI trend; Rufus $10B-type attribution returning with methodology or staying buried; first Alexa+ subscriber/revenue datapoint; Rainier chip counts (now two quarters overdue).
|
| Feb 5, 2026 | -5.55% | Q4 FY2025 | Read transcript briefingQuarter in one view- Revenue $213.4B, +12% YoY ex-FX (150bps favorable FX). Reported OI $25.0B includes $2.4B of special charges: $1.1B Italy tax-dispute resolution + lawsuit settlement (International), $730M severance (all segments), $610M asset impairments primarily physical stores (NA). Ex-charges OI ~$27.4B.
- AWS accelerated again to 24% ($35.6B, $142B run rate) — fastest in 13 quarters, +$2.6B QoQ, +~$7B YoY. AWS OI $12.5B, margin ~35% (up 40bps YoY per Olsavsky). Backlog $244B, +40% YoY, +22% QoQ.
- 2026 capex guided to ~$200B, "predominantly in AWS" — a step-change from 2025's ~$125B. TTM FCF fell again to $11.2B (from $14.8B) — sixth consecutive sequential decline. FY2025 operating cash flow $139.5B, +20%.
- Q1 2026 guidance given on the call (filling last quarter's record gap): revenue $173.5–178.5B (~180bps favorable FX); OI stated as "$616.5 billion to $21.5 billion" — an obvious transcript error, almost certainly $16.5B; verify against the press release. Guidance embeds ~$1B YoY NA cost increase from Amazon LEO.
- NA segment: revenue $127.1B (+10%), OI $11.5B / 9.0% margin (vs. 8% in Q4 2024). International: $50.7B (+11% ex-FX), OI $1.0B / 2.1% (ex-charges, margins expanded YoY). Units +12% — highest quarterly growth of 2025. 3P mix 61% (down from 62%).
What management is focused on- The ROIC defense as the central debate: Mahaney opened the Q&A asking directly how investors will see returns, duration of the capex cycle, and minimum FCF floors. Jassy/Olsavsky's answer: capacity is monetized as installed, backlog and commitments evidence demand, AWS margin at 35% despite depreciation headwind, and analogy to early core-AWS investment patterns. No FCF floor, no cycle duration, no numeric guardrails were given — Anmuth's follow-up on "financial guardrails or governors" was answered with "we are going to invest aggressively."
- Custom silicon as a sized business: chips (Trainium + Graviton) now >$10B annualized run rate, growing triple digits. Trainium2: 1.4M chips landed ("fastest ramping chip launch ever"), fully subscribed, multibillion-dollar run rate, 100K+ companies using it, majority underpinning of Bedrock usage. Trainium3 launched/shipping, ~40% better price-performance, "nearly all supply committed by mid-2026." Trainium4 in build for 2027 with "very substantial interest"; Trainium5 conversations already underway. Graviton: multibillion-dollar, +50% YoY, 90%+ of top 1,000 customers.
- The "barbell" AI demand framing: AI labs + runaway apps on one end, enterprise productivity workloads on the other, and the middle — enterprise production workloads — as "the largest and most durable," still to come. This is the answer to the concentration concern.
- Supply as the binding constraint: "every provider... could actually grow faster if we had all the supply." 3.99GW added in 12 months (2x the entire 2022 footprint), 1.2GW added in Q4 alone, doubling again by 2027, "a lot more in '26 and '27 and '28."
- Amazon LEO (rebranded from Kuiper): 180 satellites launched, 20+ launches planned 2026, 30+ in 2027, commercial launch expected 2026 with wider rollout "later this year." Dozens of commercial agreements signed (AT&T, DIRECTV Latin America, JetBlue, Australia NBN). Costs currently expensed (~$1B YoY NA headwind in Q1); satellite manufacturing and launch costs to be capitalized "later in the year" — a material accounting inflection.
- Quick commerce ("Amazon Now"): 30-minute delivery live in India, Mexico, UAE; testing in US/UK. India Prime members triple shopping frequency after adoption. Everyday essentials now 1 of 3 units sold, growing ~2x other categories.
- Grocery scale-up: $150B+ gross sales, 150M+ Americans reached, 100+ new Whole Foods stores planned "over the next few years," perishables same-day in 2,300+ cities with 2x shopping frequency and 3x items per same-day order.
Key numbers and quarter mechanics- Segments: NA $127.1B (+10%), OI $11.5B / 9.0%; International $50.7B (+11% ex-FX), OI $1.0B / 2.1%; AWS $35.6B (+24%), OI $12.5B / ~35%.
- Units +12% (highest of 2025) vs. revenue +12% ex-FX — aligned again. 3P mix 61%, down 100bps sequentially from the 62% record.
- Advertising $21.3B, +22% YoY; $12B+ incremental ad revenue added in 2025. Prime Video ad-supported audience 315M globally (from 200M in early 2024), now in 16 countries.
- Charges: $2.4B total ($1.1B Italy tax/lawsuit, $730M severance, $610M store impairments). Third consecutive quarter with severance-type charges ($1.8B in Q3, $730M in Q4) — the "one-time" framing is weakening.
- Capex: ~$200B planned for 2026 (vs. ~$125B in 2025) — a ~60% increase. FCF $11.2B TTM, sixth straight decline. OCF $139.5B FY2025, +20%.
- Delivery: 8B+ items same/next-day for US Prime in 2025 (+30%); same-day items +~70% YoY; ~100M US customers used same-day; rural same-day monthly customers ~2x YoY. "Add to Delivery" already ~10% of weekly Prime network volume six months post-launch.
- Q1 2026 guide: revenue $173.5–178.5B; OI low end misstated in transcript ("$616.5B" — presumably $16.5B); includes ~$1B LEO cost headwind in NA and stepped-up International investment (Amazon Now, price/seller-fee competitiveness).
Product and launch scorecard- Trainium: strongest evidence yet — 1.4M Trainium2 chips landed, fully subscribed, 100K+ companies, majority of Bedrock on Trainium, Trainium3 shipping with supply nearly committed by mid-2026, roadmap visibility to Trainium5. The concentration caveat from prior quarters was not re-addressed directly, but the "100,000+ companies" claim (if accurate) is a material broadening data point. Rainier: Anthropic training next Claude on Trainium2; the 1M-chip year-end target from Q3 was not confirmed — Jassy said "500,000 chips... you will see that continuing to increase," which reads as a slip or a re-scope. Flag.
- Bedrock: multibillion-dollar run rate, customer spend +60% QoQ — first growth-rate disclosure for the service.
- NovaForge (new): pre-training customization of Nova models with customer data ("novellas") — launched; no adoption data, positioned as differentiated ("nothing else out there like this").
- Frontier Agents (new at re:Invent): Kiro autonomous coding agents, DevOps agents, security agents — "already making a big difference," unquantified. Kiro developers +150% QoQ.
- Rufus: 300M customers in 2025 (up from 250M actives last quarter), 60% higher purchase completion, now with auto-buy price tracking and Buy for Me across "tens of millions" of off-Amazon items. The prior $10B incremental-sales claim was not repeated or updated — flag.
- Alexa+: now available to all US customers — free for Prime, $19.99/month non-Prime. First actual monetization structure after two quarters of engagement-only metrics. No subscriber or revenue figures.
- Amazon LEO: 180 satellites, LEO Ultra terminal (1Gbps down / 400Mbps up), named commercial agreements (AT&T, DIRECTV LatAm, JetBlue, Australia NBN), commercial launch 2026. First customer names and a firm-ish timeline after last quarter's silence.
- Ads: ads agent and creative agent launched (campaign creation "weeks → hours"); TNF averaged 15M+ viewers (+16%, third straight double-digit year); Packers-Bears wild card 31.6M — most-streamed NFL game ever.
- Amazon Now: live in 3 countries, testing US/UK; India tripling of Prime frequency is the standout metric. Early, unsized.
Sell-side read-through- Mahaney (Evercore): asked the question of the call — ROIC visibility, capex-cycle duration, FCF floors. Got philosophy, not numbers: "monetizing as fast as we install," 35% AWS margin cited, no floor, no duration. The non-answer is itself the answer: management will not commit to guardrails.
- Anmuth (JPMorgan): pressed Rainier chip counts (500K vs. the prior 1M commitment) and financial guardrails. Got "continuing to increase" on chips — no reaffirmation of 1M — and a deflection on guardrails. Two partial evasions in one exchange.
- Sandler (Barclays): AI demand concentration and the OpenAI relationship. Jassy introduced the barbell framing and confirmed the November OpenAI agreement is "a big one," while emphasizing breadth ("thousands of companies"). No sizing of OpenAI deal.
- Morton (MoffettNathanson): agentic commerce and ad-funnel compression risk. Jassy argued retailer-native agents win on selection/price/speed/trust, reiterated third-party agent criticisms (wrong prices, no history), and kept partnership optionality open. The funnel-compression risk to sponsored ads was not directly addressed.
- Nowak (Morgan Stanley): retail efficiency vs. investment split. Got regionalization extended (8 → 10 regions), inbound regionalization, units-per-box gains, 1M+ robots — but no quantified cost-savings targets.
- Sheridan (Goldman): backlog ($244B, +40% YoY, +22% QoQ — the quarter's biggest new disclosure), internal vs. external AI demand (internal "a very small fraction"), and supply/demand balance (supply-constrained; 3.99GW added, 1.2GW in Q4).
- Cross-read: six questions, zero on the $2.4B charges, the recurring severance, FCF's sixth decline, the 3P mix downtick, or the Italy settlement. The Q&A remains management-friendly; the hardest exchange (Mahaney/Anmuth on guardrails) produced the call's most important non-answer.
Management credibility- Delivered: last quarter's "monetizing as fast as we add capacity" claim is now backed by a second acceleration (20.2% → 24%) and a $44B sequential backlog jump ($200B → $244B). The October deal-cohort claim from Q3 appears validated by backlog math. Trainium3 shipped on the promised timeline (preview end-2025, volume early-2026). Alexa+ monetization arrived as floated. LEO timeline was re-anchored with specifics after a quarter of silence.
- Slipped/softened: Rainier's 1M-chip year-end target was not reaffirmed ("500,000... continuing to increase"). Rufus's $10B incremental-sales claim was dropped from the narrative. The severance-is-one-time implication from Q3 is now contradicted by a second severance charge ($730M) in Q4.
- Consistency issues: Q1 OI guidance low end garbled in transcript ("$616.5B") — must be verified. "100,000+ companies using Trainium" sits awkwardly with last quarter's "small number of very large customers" — possibly breadth via Bedrock abstraction rather than direct chip buyers; needs clarification.
- Under pressure: the ROIC/guardrails questions — the most important of the call — were answered with conviction language and historical analogy, no numbers. Jassy's "this isn't some sort of quixotic top line grab" is a tell that management knows the market's central doubt.
- Positive discipline: backlog disclosed proactively with growth rates; LEO cost headwind ($1B) and capitalization timing flagged in guidance walk — unusually specific forward disclosure.
What changed versus the prior quarter- AWS accelerated again (20.2% → 24%), the fastest in 13 quarters; backlog jumped $200B → $244B (+22% QoQ) — the demand story strengthened materially.
- Capex guidance stepped from "~$125B in 2025, higher in 2026" to a specific ~$200B for 2026 — the consumption problem is now quantified and much larger. FCF fell a sixth straight quarter ($14.8B → $11.2B TTM).
- Q1 guidance was actually delivered on the call, closing last quarter's record gap (though the OI low end is garbled).
- Charges recurred: $2.4B (Italy tax, severance, impairments) after $4.3B in Q3 — special charges are becoming a structural feature, not an exception.
- NA margin printed 9.0% (vs. 4.5% reported / 6.9% ex-charges in Q3) — a strong peak-quarter recovery. International stayed weak at 2.1% with more investment guided.
- 3P mix ticked down 62% → 61%; units re-accelerated 11% → 12%.
- Kuiper rebranded to Amazon LEO with first commercial agreements, a 2026 commercial-launch commitment, and a disclosed capitalization plan — from silence to specifics.
- Alexa+ moved from engagement metrics to a priced product ($19.99/month non-Prime, free with Prime).
- Trainium narrative broadened: from Anthropic-anchored to "100K+ companies," chips business sized at >$10B run rate, roadmap extended to Trainium5.
- New named AWS customers include OpenAI (November agreement confirmed as "big") — a notable shift given Anthropic exclusivity optics in prior quarters.
Bull case- AWS is compounding acceleration: 24% on a $142B run rate with $2.6B QoQ adds, backlog $244B (+40% YoY), and management claiming supply — not demand — is the constraint, with 1.2GW added in Q4 alone and doubling by 2027.
- The silicon franchise is real and scaling: >$10B run rate, triple-digit growth, Trainium3 nearly sold out through mid-2026, Graviton at 90%+ of top-1,000 customers — a structural cost/margin advantage that also funds the AI capex case.
- NA retail margin hit 9% in peak quarter with units +12%, same-day volume +70%, and "Add to Delivery" at 10% of weekly Prime volume in six months — the fulfillment flywheel is still tightening.
- Ads at $21.3B (+22%) with a 315M Prime Video ad audience and agentic creative tools — full-funnel buildout continues.
- LEO has real customers (AT&T, JetBlue, Australia NBN) and a 2026 commercial launch; capitalization of satellite costs later in 2026 will optically relieve the P&L.
- Alexa+ is finally monetized; Rufus at 300M users with 60% higher completion; Amazon Now showing 3x frequency lifts — multiple consumer-AI vectors with early evidence.
Bear case- ~$200B 2026 capex against $11.2B TTM FCF and falling — management explicitly refused to offer FCF floors, OI governors, or cycle duration when asked twice. The capital-consumption risk is now the entire debate, and the call provided no new tools to underwrite it.
- Special charges are recurring: $4.3B in Q3, $2.4B in Q4, with severance now in consecutive quarters — the "non-recurring" framing is eroding, and Q1 guidance explicitly excludes any new restructurings or settlements (i.e., more would be off-guide).
- International margin stuck at 2.1% with guided incremental investment (Amazon Now, price aggression) — the international profitability inflection keeps deferring.
- Rainier's 1M-chip commitment quietly went unconfirmed; Trainium's "100K+ companies" claim needs reconciliation with admitted large-customer concentration; the OpenAI deal is unsized.
- 3P mix slipped to 61% — a small but notable reversal of a long-running mix tailwind.
- LEO is a ~$1B quarterly NA cost headwind before any revenue, with commercial launch still ahead; the capitalization shift will flatter future OI without changing cash economics.
- The Rufus $10B claim disappeared without update — a reminder that self-reported AI-attribution metrics can vanish as easily as they appear.
Next-quarter watchlist- Q1 OI guide: confirm the low end ($16.5B presumed) from the press release; track whether LEO's ~$1B headwind and International investment land as described.
- Capex/FCF: the actual Q1 capex print against the ~$200B plan; whether FCF troughs; any first discussion of financing, leasing mix, or capital returns. Watch for depreciation step-up as 2025–26 capacity enters service.
- AWS: does ~24% hold; backlog trajectory vs. the $244B print; Trainium3 supply commitment by mid-2026; first sized disclosure of the OpenAI agreement; whether the AI-revenue absolute number finally appears.
- Rainier: chip count progression toward (or away from) the 1M target; Anthropic's next-Claude training milestones on Trainium2.
- Charges: whether Q1 is clean — the first real test of the "one-time" claim after two consecutive charged quarters; any FTC/Italy-style regulatory recurrence.
- LEO: commercial launch timing ("later this year"), the capitalization trigger date and its OI optics, launch cadence vs. 20+ planned, revenue disclosure.
- Alexa+: first subscriber or revenue data points for the $19.99 tier; Prime-member engagement now that it's free and GA.
- Rufus/agentic commerce: whether the $10B incremental-sales claim returns with methodology; any third-party agent partnership announcements; Buy for Me expansion.
- Retail: NA margin ex-charges vs. the 9% peak print; International investment drag magnitude; 3P mix direction; Amazon Now US/UK test expansion and any unit economics.
- Reconciliation items: the "100K+ Trainium companies" claim vs. concentration; the garbled Q1 OI figure; whether severance appears in Q1 results despite guidance language excluding it.
|
| Oct 30, 2025 | +9.58% | Q3 FY2025 | Read transcript briefingQuarter in one view- Revenue $180.2B, +12% YoY ex-FX (90bps favorable FX) — above the $174–179.5B guide. Reported OI $17.4B includes two special charges: $2.5B FTC settlement (NA segment, other operating expense) and $1.8B severance (all three segments). Ex-charges OI $21.7B, $1.2B above the high end of the $15.5–20.5B guide — a seventh consecutive OI beat, confirming last quarter's read that the $5B-wide range was a staged low bar.
- AWS reaccelerated to 20.2% ($33.0B, $132B run rate) — the largest growth rate in 11 quarters, +270bps sequentially, with $2.1B QoQ revenue adds. AWS OI $11.4B, margin ~34.5% (recovering from 32.9%). Backlog $200B, and Jassy says October's unannounced deals alone exceed all of Q3's deal volume.
- TTM FCF fell again to $14.8B (from $18.2B) — fifth consecutive sequential decline. Cash capex $34.2B in Q3, $89.9B YTD; FY2025 guided to ~$125B with an explicit statement that 2026 will be higher.
- Net income $21.2B includes a $9.5B pretax Anthropic investment gain (non-operating) — headline EPS quality needs adjusting.
- Notably, no Q4 guidance appears in this transcript — neither revenue nor OI ranges are stated in the prepared remarks or Q&A. Flag as a record gap; the press release presumably carries it.
What management is focused on- AWS capacity as the monetization story: 3.8GW of power added in 12 months ("more than any other cloud provider"), double 2022 capacity, doubling again by 2027, +1GW+ in Q4 alone. Jassy's new framing: "as fast as we're adding capacity right now, we're monetizing it" — a direct answer to the capex-return question, though unquantified.
- Trainium as a business, not a project: Trainium2 is "fully subscribed," a "multibillion-dollar business" growing 150% QoQ; Project Rainier live with ~500K Trainium2 chips for Anthropic, going to 1M+ by year-end. Trainium3 previews end of 2025, volume early 2026, claimed ~40% better than Trainium2, with "medium-sized" customer interest broadening the base.
- The severance/layoff defense: Jassy insists the $1.8B charge and recent job cuts are "not really financially driven, and it's not even really AI-driven, not right now" — framed as culture/de-layering ("world's largest start-up"). This directly counters the AI-driven-headcount narrative Mahaney raised.
- Grocery as a strategic pillar: $100B+ GMS ex-Whole Foods/Fresh ("top 3 grocery in the U.S."), perishables same-day expanding 1,000 → 2,300 cities by year-end, perishable shoppers returning 2x as often; Whole Foods "daily shop" small-format concept (3 stores) expanding.
- Agentic commerce positioning: Rufus at 250M active customers, +140% monthly users, 60% higher purchase completion, "on track to deliver over $10 billion in incremental annualized sales" — the first dollar-quantified AI-attributed retail metric Amazon has given. Jassy also opened the door to partnering with third-party agents while criticizing their current experience (wrong prices, wrong delivery estimates).
- Advertising full-funnel + DSP: Netflix, Spotify (400M listeners), SiriusXM (160M) integrations layered on Roku; upfront commitments for 2025–26 live sports "exceeded our own expectations."
Key numbers and quarter mechanics- Segments: NA revenue $106.3B (+11%), OI $4.8B / 4.5% margin; ex-FTC charge $7.3B / 6.9% (down from 7.5% in Q2 — severance also embedded). International $40.9B (+10% ex-FX), OI $1.2B / 2.9% (down from 4.1%; ex-severance, margins expanded YoY per Olsavsky). AWS $33.0B (+20.2%), OI $11.4B / ~34.5%.
- Units +11% worldwide (vs. +12% in Q2) — roughly in line with ex-FX revenue growth again; no inversion. 3P mix 62%, now +200bps YoY (vs. +100bps last quarter) — the record level held and widened.
- Advertising: $17.6B (Jassy) / $17.7B (Olsavsky) — a $100M internal discrepancy in the transcript; +22% YoY, "accelerated for the third consecutive quarter" (19% → 22% → 22% per the record; the "third consecutive acceleration" claim doesn't square with flat 22% — flag).
- Capex: $34.2B Q3, $89.9B YTD, FY2025 ~$125B, "will increase in 2026." FCF $14.8B TTM — fifth straight decline, no trough called, no capital returns mention (streak continues).
- Charges: $2.5B FTC settlement (NA), $1.8B severance (all segments, mostly tech/infrastructure, S&M, G&A). Ex-charges OI $21.7B vs. $20.5B guide high end.
- Anthropic: $9.5B pretax gain in non-operating income; net income $21.2B.
- Retail inputs: inbound lead time −4 days YoY; perishables same-day in as little as 5 hours; 3-hour delivery rolling out in select cities; $4B rural network commitment, rural same/next-day communities +60%, ~half of year-end plan.
Product and launch scorecard- Trainium2/Rainier: strongest quarter of evidence yet — fully subscribed, multibillion-dollar revenue, +150% QoQ, Rainier live at ~500K chips scaling to 1M+ by year-end, majority of Bedrock token usage already on Trainium. Still vendor-benchmarked price-performance (30–40%), and customer concentration admitted ("small number of very large customers").
- AgentCore: 1M+ SDK downloads; named production users — Ericsson, Sony, Cohere Health (30–40% medical review time reduction). First named-customer production evidence for the agentic stack.
- Kiro: 100K+ developers in first days, "more than doubled since," trillions of tokens processed, weekly actives "growing fast." Still preview; no GA or monetization.
- Transform: 700K hours of manual effort saved YTD (335 developer-years); Thomson Reuters at 1.5M lines of code/month; ~1B lines of mainframe code analyzed. Quantified but self-reported.
- Quick Suite (new): consumer-AI-style work assistant; claims of 80%+ time savings, 90%+ cost savings — self-validated, no adoption numbers.
- Connect: crested $1B annualized revenue run rate; 12B AI-handled interaction minutes; Capital One, Toyota, American Airlines, Ryanair named. First hard revenue number for any AI-adjacent service.
- Rufus: 250M actives, +140% monthly users, +210% interactions, 60% higher completion, $10B incremental annualized sales run-rate claim — the most concrete retail-AI monetization claim to date; methodology undisclosed.
- Alexa+: engagement metrics only (2x conversations, 4x shopping conversations ending in purchase, 2.5x Fire TV usage) — still no user count update beyond last quarter's "millions," no monetization progress on the ads/subscription optionality floated in Q2.
- Kuiper: 150+ satellites in orbit, >1Gbps demonstrated on enterprise terminal ("first commercial phased array" to clear that). No commercial-service timing update — last quarter's "later this year or early next year" not revisited; no revenue or customer counts.
- Zoox: robotaxis live for riders in Las Vegas; D.C. announced as eighth testing location. First public rider availability — a real milestone, unsized.
- Ads: Netflix DSP access is the headline add; Spotify/SiriusXM broaden audio; agentic creative studio launched. NBA opening night 1.25M average viewers, double-digit increase vs. cable last season.
Sell-side read-through- Post (BofA): capacity-constraint status and Trainium demand beyond anchor customers. Got the "monetizing as fast as we add" line and Trainium3 timing (preview end of 2025, volume early 2026) plus "medium-sized" customer interest — the broadening-base claim.
- Nowak (Morgan Stanley): pressed the Trainium-vs-third-party positioning and adoption hurdles. Jassy committed to continued large NVIDIA purchases ("not constrained in any way in buying NVIDIA") while arguing price-performance will drive Trainium3 adoption; named software ecosystem as the work item.
- Anmuth (JPMorgan): Rainier architecture and replicability — Jassy framed it as Anthropic-specific but said other customers want large Trainium clusters via Trainium3.
- Mahaney (Evercore): grocery strategy (does perishables same-day obviate Fresh stores? — Jassy: physical experiments continue but perishables is "very significant") and the headcount question that produced the "culture, not AI, not financial" defense.
- Sheridan (Goldman): robotics/physical AI — 1M+ robots, more coming, humans-plus-robots framing; no quantified cost impact.
- Blackledge (TD Cowen): agentic commerce — the most forward-looking answer of the call: Rufus/Buy for Me today, third-party agent partnerships expected, but current third-party experience criticized (no personalization, wrong prices/ETAs) and "right exchange of value" required.
- Sebastian (Baird): asked the two disaggregation questions management didn't fully answer — AI vs. core split of the AWS acceleration (Jassy listed growth areas but gave no split) and ads decomposition into stores/DSP/Prime Video (qualitative only; video called "already a very large amount," DSP "growing really quickly").
- Cross-read: nobody challenged the FTC settlement, the severance charge's recurrence risk, FCF's fifth decline, the 2026 capex increase, or Kuiper timing. The Q&A was AWS-demand-validation heavy and management-friendly; the toughest moment was Mahaney's headcount question.
Management credibility- Positive: the Q2 "several quarters" capacity language was followed by actual acceleration (17.5% → 20.2%) faster than implied — the first time in three quarters the capacity narrative moved in management's favor. Seventh consecutive OI beat; the staged-range pattern is now fully predictable. Backlog disclosed proactively and strengthened ($195B → $200B, plus October deals > all of Q3). Trainium claims gained specificity (multibillion revenue, +150% QoQ, fully subscribed, chip counts). Connect's $1B run rate and Rufus's $10B claim are new quantified disclosures.
- Negative: the severance explanation strains — $1.8B of role eliminations described as "not really financially driven" and "not even really AI-driven, not right now" while simultaneously touting AI productivity and committing to "lean and flat." The qualifier "not right now" leaves the AI-headcount link open. The FTC charge arrived unflagged.
- Consistency issues: ads revenue stated as $17.6B (Jassy) and $17.7B (Olsavsky); "accelerated for the third consecutive quarter" conflicts with the flat 22% print in this record. Alexa+ got no user-count or monetization update despite last quarter's floated subscription/ads optionality — the engagement-only metrics suggest monetization hasn't progressed.
- Under pressure: Jassy's headcount answer was long on culture framing, short on numbers (no role count, no forward headcount guide). Sebastian's disaggregation questions were answered qualitatively — the AI-revenue absolute number remains withheld (seventh quarter), though Trainium's "multibillion-dollar" sizing is a partial crack in the wall.
- Record quality: no Q4 guidance in the transcript at all — either omitted from the call or a transcript gap; must be verified against the press release.
What changed versus the prior quarter- AWS accelerated (17.5% → 20.2%) — the reacceleration the prior packet said had "lost its dated anchor" arrived anyway, with capacity adds (3.8GW, +1GW Q4) as the stated driver. The bear-case timeline concern is deferred.
- AWS margin recovered (32.9% → ~34.5%) as the SBC step-up annualized out; depreciation acknowledged as ongoing but not quantified.
- FCF decline extended ($18.2B → $14.8B TTM, fifth straight) and 2026 capex was guided higher than 2025's ~$125B — the consumption problem worsened even as the demand story improved.
- Charges returned: $4.3B of specials (FTC + severance) after a clean Q2; reported NA margin dropped to 4.5% (6.9% ex-FTC vs. 7.5% in Q2) and International to 2.9% (from 4.1%) — retail margin momentum paused optically, with severance embedded in both.
- Units decelerated slightly (12% → 11%); 3P mix held at 62% but the YoY gain doubled (+200bps).
- Trainium went from "backbone for Anthropic" to a sized business (multibillion, +150% QoQ, fully subscribed) with Rainier live — the biggest single-quarter evidence jump.
- Kuiper got no timeline update (silence after last quarter's slip); Zoox went from testing to public rides in Las Vegas.
- The headcount/cost story moved from background (June memo) to foreground ($1.8B severance charge, public layoff announcement, direct analyst question).
- Anthropic relationship deepened financially: $9.5B investment gain plus 1M-chip Trainium commitment.
Bull case- The AWS thesis is now delivering, not promised: 20.2% on a $132B run rate with $2.1B QoQ adds, backlog $200B plus an October deal pipeline larger than all of Q3, and management claiming monetization keeps pace with capacity adds. Jassy: "I believe that we can continue to grow at a clip like this for a while."
- Trainium is a real, sized, fully subscribed business with a credible broadening path (Trainium3, medium-sized customers) and a 1M-chip Anthropic anchor — custom silicon is becoming a margin and differentiation lever, not just a cost offset.
- AI monetization evidence is accumulating across the stack: Connect $1B run rate, Rufus $10B incremental sales claim, AgentCore at 1M+ downloads with named production customers, Transform at 700K hours saved.
- Advertising at ~$70B annualized with Netflix/Roku/Spotify/SiriusXM distribution and "exceeded expectations" sports upfronts — the DSP is now "fully featured" per Jassy, with video still early.
- Grocery is a new structural leg: $100B+ GMS, perishables in 2,300 cities by year-end with 2x return frequency, plus Whole Foods expansion and the daily-shop format.
- Ex-charges OI of $21.7B shows the underlying earnings power; both charges are presented as non-recurring.
Bear case- FCF $14.8B TTM and falling, against ~$125B 2025 capex that management explicitly says rises in 2026 — no trough, no financing discussion, no capital returns, and depreciation will keep building as the 3.8GW+ of capacity comes online.
- Reported retail margins compressed: NA 4.5% (6.9% ex-FTC vs. 7.5% last quarter) and International 2.9% (from 4.1%) — even ex-charges, NA margin stepped back, and severance costs signal more restructuring expense risk.
- The severance rationale is unstable: if cuts are cultural now, the AI-driven efficiency wave Jassy keeps describing is still ahead — implying further charges and headcount reduction, with the "not right now" qualifier doing heavy lifting.
- Trainium concentration: "a small number of very large customers" (effectively Anthropic) — the +150% QoQ growth is off a captive base; Trainium3's broadening is a preview-stage promise.
- The $9.5B Anthropic gain inflates net income; reported EPS overstates operating performance, and the gain marks a related-party valuation.
- Disclosure debt persists: AI absolute revenue (seventh quarter), ads segment decomposition refused, Alexa+ monetization stalled, Kuiper timing unaddressed after slipping, no Q4 guidance in this record.
- The FTC settlement ($2.5B) is a reminder of regulatory exposure; whether it is truly one-time (vs. further remedies) is untested.
Next-quarter watchlist- Q4 guidance: retrieve the actual ranges from the press release — absent from this transcript. Check whether the OI range is again staged wide, and whether Q4 embeds further severance or FTC-related costs.
- AWS: does ~20% hold against a tougher compare; evidence the October deal cohort (claimed > all of Q3 volume) converts to backlog/revenue; Trainium3 preview on time (end of 2025) and first non-anchor customer names; whether the AI-revenue absolute number finally appears at re:Invent.
- AWS margin: whether ~34.5% holds as Q4's 1GW+ and Rainier depreciation land; any normalized-margin framework.
- FCF/capex: whether $14.8B is the trough; the 2026 capex number when given; any first mention of financing or returns.
- Charges: whether severance is one-and-done or recurring into Q4/2026; headcount numbers behind the $1.8B; FTC settlement finality.
- Retail margins: whether NA ex-charges reclaims 7.5% and International resumes expansion in Q4 peak; units vs. revenue alignment; tariff commentary (notably absent from this call — check whether the pre-buy depletion question resurfaces).
- Rufus: whether the $10B incremental-sales claim gets methodology or updates; any agentic-commerce partnerships with third-party agents.
- Alexa+: user count, international start, and any concrete ads/subscription step — or another quarter of engagement-only metrics.
- Kuiper: commercial-service timing (last guided "late this year or early next year" — now due), launch cadence, capitalization trigger, any revenue disclosure.
- Record quality: reconcile the $17.6B/$17.7B ads discrepancy and the "third consecutive acceleration" claim; confirm Q4 guidance figures.
|
| Jul 31, 2025 | -8.27% | Q2 FY2025 | Read transcript briefingQuarter in one view- Revenue $167.7B, +12% YoY ex-FX, with a $1.5B favorable FX impact (guide had assumed ~−10bps/−$100M — a ~$1.6B FX swing versus plan). Both Jassy and Olsavsky cite $167.7B, resolving last quarter's $155.7B/$165.7B discrepancy in the record.
- Operating income $19.2B, +31% YoY, $1.7B above the high end of the $13–17.5B guide — a sixth consecutive OI beat, and a large one after last quarter's compressed $0.4B beat. The widened, "uncertainty"-staged Q2 range was indeed a low bar.
- TTM FCF fell to $18.2B from $25.9B — a fourth consecutive sequential decline (−$7.7B QoQ). Cash capex jumped to $31.4B (from $24.3B) and Olsavsky says Q2 is "reasonably representative" of the H2 quarterly rate — annualizing to ~$125B, well above the ~$105B implied earlier.
- AWS grew 17.5% ($30.9B, $123B+ run rate) — a slight uptick from 17% but not the hoped-for acceleration; AWS OI fell to $10.2B and margin dropped to 32.9% from a record 39.5% in Q1, with ~half the decline attributed to the seasonal SBC step-up, plus depreciation from capex and FX.
- Retail was the standout: NA margin 7.5% (+190bps YoY), International 4.1% (+320bps), units +12% (reaccelerating from +8%), 3P mix a record 62%.
- Backlog re-disclosed on demand: $195B, +25% YoY (up from $189B/+20% last quarter) — corroboration strengthened.
- Kuiper timeline slipped: commercial service now "later this year or early next year" versus last quarter's "later this year," with rocket-provider delays admitted.
What management is focused on- The tariff defense, now with a sharper edge: Jassy says much of the tariff reporting "has been wrong and misreported," and repeats the H1 evidence — no diminished demand, no broad-scale ASP increases — while conceding the unknowns (where China tariffs settle, what happens when pre-bought/forward-deployed inventory depletes, who absorbs higher costs). Olsavsky notes the Q2 plan "factored in a range of assumptions, not all of which materialized."
- Retail inputs as the margin story: direct-lane share +40% YoY, average package distance −12%, handling touches −15%, higher units per box, 30% more same/next-day items delivered, record Prime speeds. Outbound shipping costs +6% versus units +12% — the cost-to-serve divergence is the core mechanical claim.
- Rural/same-day expansion: same-day/next-day to tens of millions of customers in 4,000+ smaller cities/towns by year-end (1,000+ live today), with early data showing higher shopping frequency and essentials attach.
- AWS capacity constraint reframed again: the single biggest constraint is now named as power, plus chips and server-yield issues. Jassy explicitly abandons near-term resolution: "I don't believe that we will have fully resolved... in a couple of quarters. I think it will take several quarters," though improving each quarter. This is a further softening from Q1's "as the year proceeds" and Q4's H2 promise.
- The AI stack defense against the "AWS is falling behind" narrative (Nowak's question): top-heavy market (frontier models, ChatGPT, coding agents — Cursor, Vercel, Lovable "run significant chunks on AWS"), inference will be 80–90% of cost at scale, Trainium2's 30–40% price-performance as the answer, plus data-gravity argument ("so many more applications and data running in AWS than anywhere else").
- Agentic tooling as the new product frontier: Strands (open-source agent builder), AgentCore (secure serverless agent runtime on Bedrock), Kiro (agentic IDE), AWS Transform (mainframe/VMware/.NET migration agents).
- Internal AI adoption: Jassy's June memo framing — embrace or be shaped; Kiro, Connect, and internal agents positioned as productivity/speed-to-market levers. No headcount or cost-quantification given.
- Capital returns: unmentioned again — the omission streak extends another quarter with FCF now at $18.2B.
Key numbers and quarter mechanics- Segments: NA revenue $100.1B (+11%), OI $7.5B (+$2.5B YoY), margin 7.5% (+190bps). International $36.8B (+11% ex-FX), OI $1.5B (+$1.2B), margin 4.1% (+320bps; ~700bps cumulative over 10 quarters). AWS $30.9B (+17.5%), OI $10.2B, margin 32.9%.
- AWS margin bridge: record 39.5% Q1 → 32.9% Q2; ~half the decline from the seasonal SBC step-up (annual comp cycle — pre-flagged in the Q2 guide), remainder from higher depreciation (capex-driven) and unfavorable FX. Note: Q1's 39.5% is now described as a "record high," reframing last quarter's margin strength as peak rather than trend.
- Units: +12% worldwide — a sharp reacceleration from +8%, now matching ex-FX revenue growth (+12%). Last quarter's units/revenue inversion reversed; the staples-mix concern is deferred, not resolved.
- 3P seller unit mix: 62%, "highest ever," +100bps YoY — broke above the 59–61% oscillation band.
- Advertising: $15.7B, +22% YoY (accelerating from +19%) — ~$63B annualized; 300M+ US ad-supported audience (up from 275M+ claimed last quarter).
- Capex: $31.4B cash capex in Q2, guided as representative for H2 — implies ~$125B annualized versus ~$97B run rate last quarter. AWS/custom silicon/power named as primary drivers.
- FX: +$1.5B favorable in Q2 versus a guided −$100M — a meaningful chunk of the revenue beat is currency, not operations.
- Q3 guidance: revenue $174–179.5B (FX ~+130bps favorable); OI $15.5–20.5B ($5B wide — wider than Q2's $4.5B range). Olsavsky initially misstated the range as "$174 billion to $175 billion" and corrected himself ("$179.5 billion is a typo") — a live-call stumble on the headline guide.
- No tariff charges this quarter: unlike Q1's ~$1B of one-time charges, Q2 shows clean margins; Olsavsky's "not all assumptions materialized" confirms the tariff contingency embedded in the guide didn't hit.
Product and launch scorecard- Kuiper: third launch completed; "impressive" enterprise/government agreements signed pre-launch (no count, no revenue). Timeline slipped to "commercial service later this year or early next year" with rocket-provider delays admitted; Jassy claims most available rocket launches over the next couple of years are secured. The capitalization-trigger timing flagged last quarter is now at risk of sliding into 2026.
- Alexa+: "millions of customers" now have access (up from 100K+ last quarter), rest of US in coming months, international later this year. Jassy claims high ratings, more expansive usage, meaningfully higher call volume — still no retention or monetization metrics. New monetization language: advertising in multiturn conversations and a possible future subscription tier "beyond what there is today." Transcript states non-Prime pricing at $9.99/month — inconsistent with the prior packet's $99/month figure; flag as a record discrepancy (likely $9.99, but unverifiable from this transcript alone).
- Kiro (new): agentic IDE launched in preview; 100K users in first 5 days, "several hundred thousand" using/requesting access in the first couple of weeks. Differentiation claim: spec-driven "vibe coding" plus event-driven agent hooks. Early but the strongest adoption datapoint of any recent AWS launch.
- Trainium2: now "the backbone for Anthropic's newest generation Claude models" and for Bedrock inference; 30–40% price-performance claim restated (still vendor benchmark); Trainium3 confirmed in development. Still no revenue-contribution sizing.
- Bedrock: Claude 4 added and is "the fastest-growing model ever in Bedrock"; Nova is now the second most popular foundation model on Bedrock. AgentCore launched (serverless agent runtime, identity, memory, MCP gateway, observability) — positioned as the fix for enterprise agent-deployment blockage.
- Strands: open-source agent framework with 2,500 GitHub stars and 300K+ downloads — first externally checkable traction metric for the agentic stack.
- AWS Transform: quantified claims — mainframe modernization years→months, VMware conversions up to 80x faster, .NET licensing cost −40%. Self-validated.
- Advertising: Roku partnership (80M CTV households, largest authenticated CTV footprint in the US, exclusive via Amazon DSP) and Disney real-time ad exchange integration (Disney+, ESPN, Hulu) — the strongest DSP distribution evidence to date. Prime Video ads still unsized.
- Retail selection: Nike returns to Amazon retail (headline add); Away, Aveda, Marc Jacobs Fragrances, expanded Saks brands. Perishables pilot: 75% of viewers are first-time perishables shoppers, 20% repeat within a month — unusually concrete pilot metrics.
- Prime Day: biggest ever (record sales, items, Prime sign-ups in the 3 weeks prior); seller best-ever performance; "billions" saved. No dollar figure disclosed, as usual.
- Robotics: 1 millionth robot deployed; DeepFleet AI improves robot travel efficiency 10%.
- Content/sports: NASCAR ~2M viewers/race, youngest NASCAR broadcast audience in a decade+; NBA broadcast crew announced; Denis Villeneuve directing next Bond.
Sell-side read-through- Anmuth (JPMorgan): asked the tariff-absorption question (suppliers/Amazon/consumers) and got the standing "we don't know / haven't seen it yet" answer; his AWS question — is the growth gap vs. #2/#3 demand- or supply-driven — drew Jassy's fullest competitive defense: base effects (#2 is ~65% of AWS's size), security ("adventures at some of these players almost every month"), functionality, and "more demand than capacity."
- Mahaney (Evercore): extracted the backlog number again ($195B, +25%) — disclosure now routine after two quarters of withholding. His capacity-timing question produced the call's most important admission: resolution takes "several quarters," not H2. His Alexa+ monetization question yielded the new advertising/subscription-optionality language.
- Sebastian (Baird): International margin drivers (established countries at US-like margins; 8 emerging countries launched in 5 years at varying investment stages) and the Kuiper question that surfaced the timeline slip and rocket delays.
- Nowak (Morgan Stanley): asked the toughest question directly — the "AWS is falling behind in GenAI" narrative — and got the full stack-by-stack rebuttal plus the inference-economics argument (80–90% of cost at scale). His acceleration question got no commitment ("we don't give guidance by segment") but Jassy's most optimistic framing: migrations resuming + AI production deployments + capacity coming online "in the coming months and quarters."
- Josey (Citi): soft internal-AI question; produced strategy language but no quantified productivity or headcount impact.
- Post (BofA): Q3 guide drivers — Olsavsky cited unit acceleration and Prime Day, "cautiously optimistic," declined any Q4 color. Notably, Olsavsky's initial misstatement of the guide range happened here.
- Cross-read: nobody pressed the AWS margin collapse (39.5% → 32.9%) beyond the prepared bridge; nobody asked about FCF's fourth straight decline or capital returns; nobody challenged the Kuiper slip. The sell-side agenda was AWS-competitive-narrative heavy, and management used it to deliver its most complete AI positioning to date.
Management credibility- Positive: sixth consecutive OI beat, and a $1.7B-above-high-end print validates the prior packet's "staged low bar" read on the widened Q2 range. Backlog disclosed promptly and improved (+20% → +25%). Tariff claims from Q1 held: Q2 direct tariff costs were indeed "not large" — no second charge tranche, clean margins. Units reaccelerated as the input story (speed, in-stock, placement) predicted. Alexa+ scaled from 100K+ to "millions" on schedule.
- Negative — capacity language slipped again: Q4's "really start to relax in the second half of '25" → Q1's "get better as the year proceeds" → Q2's "several quarters" to fully resolve. Each quarter the constraint horizon extends; the H2 acceleration setup the prior packet flagged as "dated and funded" is now explicitly pushed out.
- Kuiper slipped within one quarter: "service later this year" → "later this year or early next year," with supplier (rocket) delays admitted. The accountability gap the prior packet considered closed has partially reopened.
- AWS margin framing: Q1's ~39% was presented without a "record/peak" caveat; this quarter it's retroactively labeled a "record high" as it falls to 32.9%. The SBC step-up was pre-flagged, but the magnitude of the swing was not.
- Record-quality issues persist: Olsavsky misstated his own Q3 revenue guide live ("$175 billion... excuse me, $179.5 billion is a typo"); Alexa+ non-Prime pricing stated as $9.99/month versus the prior quarter's $99/month record — one of the two transcripts is wrong.
- Under pressure: Jassy's Nowak answer was detailed and structured (stack layers, inference economics, data gravity) but still offered no AI revenue number — the "triple-digit, multibillion-dollar" boilerplate survives a sixth quarter. The acceleration answer was optimism without commitment.
What changed versus the prior quarter- OI beat re-expanded dramatically ($0.4B → $1.7B above high end); YoY OI growth rose from +20% to +31%. The pre-excused wide range played out exactly as the prior packet's bear case suspected — as a low bar.
- FCF decline accelerated: $25.9B → $18.2B TTM, fourth straight drop, with capex stepping up to $31.4B/quarter and guided flat for H2. The capex-consumption problem is now the dominant financial mechanic.
- AWS margin inverted: record 39.5% → 32.9%, with depreciation from the capex build now visibly hitting the P&L — the useful-life/accounting mechanics that vanished from discussion last quarter have returned as a real cost line.
- AWS growth ticked up (17% → 17.5%) but the capacity-resolution timeline moved out from "H2" to "several quarters" — the acceleration thesis lost its dated anchor.
- Units reaccelerated (8% → 12%) and matched revenue growth; last quarter's units/revenue inversion and staples-mix worry reversed.
- Tariffs de-escalated as a P&L event: no new charges, no demand attenuation, no broad ASP increases through H1; management went on offense ("much of it wrong and misreported").
- Backlog strengthened ($189B/+20% → $195B/+25%) and disclosure is now normalized.
- Kuiper slipped; Alexa+ scaled; Kiro emerged as a new launch with real early traction; advertising accelerated to 22% with the Roku/Disney deals adding distribution evidence.
- 3P mix broke to a record 62% after years of 59–61% oscillation.
Bull case- The retail engine is compounding: NA 7.5% and International 4.1% margins with units +12%, shipping costs +6%, direct lanes +40%, touches −15% — structural cost-to-serve gains with quantified mechanics, and International has now delivered ~700bps of margin expansion over 10 quarters with established markets at US-like profitability.
- The tariff thesis is winning so far: two quarters of no demand attenuation and no broad ASP increases, no Q2 charge repeat, and the pre-buy strategy validated. If the China-direct structural argument holds, Amazon is relatively advantaged versus traditional retail as tariffs persist.
- AWS demand corroboration keeps building: backlog $195B +25%, new logos across industries (PepsiCo, Airbnb, NASDAQ, LSE, SAP, Warner Bros. Discovery), and management now says capacity comes online "in the coming months and quarters" — if supply lands, 17.5% has room to reaccelerate into 2026 on a $123B base.
- The AI stack is filling in with checkable artifacts: Kiro's 100K-in-5-days, Strands' 300K downloads, Claude 4 as Bedrock's fastest-growing model, Trainium2 as Anthropic's training backbone — more external evidence than any prior quarter.
- Advertising at +22% (~$63B run rate) with Roku exclusivity and Disney integration extends the DSP moat into CTV at scale.
- Alexa+ is scaling (millions of users) with new monetization optionality explicitly floated (ads, future subscription tiers) on top of the Prime value proposition.
Bear case- FCF at $18.2B TTM against a ~$125B annualized capex pace: four consecutive declines, no trough, no financing or capital-returns discussion, and depreciation is now visibly compressing AWS margin (32.9%). The investment cycle's P&L cost has arrived before the revenue acceleration.
- The AWS acceleration story lost its timeline: "several quarters" to resolve constraints means the H2-2025 reacceleration the prior packet treated as a dated setup is off the table; if competitors keep posting faster growth through year-end, the base-effect defense wears thinner each quarter.
- AWS margin fell 660bps sequentially and management offers no normalized target — with H2 capex flat at $31.4B/quarter, depreciation headwinds persist, and the SBC step-up recurs annually.
- The tariff buffer is depleting by construction: Jassy himself flags "what happens when we deplete the prebuys" as an open question — H2 is when the cushion runs out and the absorption question (supplier/Amazon/consumer) gets answered.
- Kuiper slipped again with supplier dependencies (rockets) outside management's control; the capitalization trigger and any revenue contribution slide with it, while launch costs keep expensing.
- Q3's OI guide ($15.5–20.5B) is $5B wide with FX now a tailwind (+130bps) — another staged range, and the mid-point implies decelerating OI growth versus Q2's +31%.
- Disclosure debt unchanged: AI absolute revenue (sixth quarter withheld), Prime Video ads sizing, Alexa+ retention/economics, Kuiper agreement values.
Next-quarter watchlist- AWS: does 17.5% hold, accelerate, or slip; concrete evidence of capacity landing (power procurement, Trainium2 volumes, Grace Blackwell instance availability); whether backlog is re-disclosed and still growing ~25%; any AI revenue absolute number as the boilerplate enters its seventh quarter.
- AWS margin: whether 32.9% is the new base or recovers as SBC annualizes; depreciation trajectory against flat $31.4B quarterly capex; any normalized-margin framework.
- FCF/capex: whether $31.4B holds as the H2 rate or rises; whether FCF bottoms above or below $18B; any first mention of financing or capital returns.
- Tariffs: pre-buy depletion timing; any H2 ASP increases or demand attenuation versus the H1 "none seen" claim; whether a second charge tranche appears; Everyday Essentials mix persistence.
- Q3 print vs. guide: whether the $5B-wide OI range is another staged low bar; FX tailwind realization (~130bps).
- Kuiper: whether "later this year or early next year" firms up; launch cadence against rocket availability; capitalization-trigger timing; any revenue or customer-count disclosure on the signed enterprise/government agreements.
- Alexa+: user count beyond "millions," international rollout start, retention metrics, and any concrete advertising or subscription monetization step; clarification of the $9.99 vs. $99 pricing record.
- Kiro/Strands/AgentCore: conversion of preview traction (100K users, 300K downloads) into GA, paying usage, or named production deployments.
- Advertising: whether 22% growth sustains; any sizing of Prime Video ads or Roku/Disney contribution.
- Record quality: whether the Q3 guide misstatement and Alexa+ pricing discrepancy are addressed; whether Jassy/Olsavsky figures reconcile cleanly next quarter.
|
| May 1, 2025 | -0.12% | Q1 FY2025 | Read transcript briefingQuarter in one view- Revenue $155.7B, +10% YoY ex-FX with a $1.4B FX headwind (Olsavsky). Jassy's opening remarks cite "$165.7 billion" — an unreconciled discrepancy in the transcript; Olsavsky's $155.7B is consistent with the disclosed FX headwind and the prior guide of $151–155.5B. Flag it as a record-quality issue.
- Operating income $18.4B, +20% YoY, ~$400M above the $14–18B guide — a fifth consecutive OI beat (the prior packet counted four), but the smallest beat of the streak ($1.2B, $2.4B, $3.3B prior beats; now $0.4B).
- TTM FCF fell to $25.9B from $36.2B — a third consecutive sequential decline (−$10.3B QoQ). The capex-consumption problem keeps compounding.
- AWS decelerated to 17% ($29.3B, $117B run rate) from 19%, while AWS OI rose to $11.5B, an implied ~39.2% margin — up from ~36.8%, despite the FY2025 useful-life headwinds flagged last quarter. Anmuth pressed as "almost 40%."
- Tariffs dominated: ~$1B of one-time charges (unresolved historical customer returns + costs to receive pull-forward inventory) cut NA margin to 6.3% (7.2% ex-charges) and International to 3.0% (3.7% ex-charges). Management claims no demand attenuation and no meaningful ASP increases yet.
- Two long-standing watch items partially resolved: Kuiper's first production-satellite launch happened this week ("service later this year"), and Alexa+ shipped in Q1 (free with Prime; stated $99/month non-Prime; 100K+ users). AWS backlog was finally disclosed: $189B, +20% YoY, 4.1-year weighted average life.
- Q2 guide: revenue $159–164B (FX ~−10bps), OI $13–17.5B — Olsavsky explicitly admits the range was widened for uncertainty, with the seasonal stock-comp step-up and incremental Kuiper launch costs embedded.
What management is focused on- The tariff defense thesis, highly developed: no demand attenuation seen; some heightened buying that may be pre-stocking; ASPs flat so far; 1P forward buys plus 3P sellers' advance-inbounding; "most sellers just haven't changed pricing yet." Jassy's structural argument: China-direct marketplace sellers face lower total tariffs than retailers sourcing Chinese goods through intermediaries who mark up and rebrand. This is checkable against future quarters.
- Everyday Essentials as the defensive anchor: grew more than 2x the rest of the business; one of every three US units; grocery ex-Whole Foods/Fresh is "over $100B in gross sales last year." Mix is shifting toward staples — management frames this as resilience, not weakness.
- AI/capacity remains the core strategic claim: "as fast as we actually put the capacity in, it's being consumed"; P5 (NVIDIA) and Trainium2 instances landing; motherboards and other componentry "jammed up"; Jassy now expects supply issues to "continue to get better as the year proceeds" — softer language than Q4's "really start to relax in the second half of '25." Olsavsky reaffirms "an increasing amount of capacity in the back half of the year."
- The inference-price mission statement: "for AI to be as successful as we believe it can be, the price of inference needs to come down significantly. We consider this part of our mission." Trainium2's 30–40% price-performance claim is now positioned as the instrument.
- Agentic AI moved from slogan to product claims: Nova Act (browser-action agent, research preview) targeting multi-step accuracy from today's "30% to 60%" to "90-plus percent"; Amazon Q's new agentic CLI coding; Alexa+ positioned as the first "action-oriented" agent.
- Retail cost engine: newly rolled-out inbound architecture, same-day expansion, rural delivery stations, robotics/automation; Q2 focus on managing pull-forward inventory volumes through fulfillment nodes without clogging productivity.
- Notably absent across prepared remarks and six analyst questions: capital returns — the omission streak now extends another quarter with FCF at $25.9B.
Key numbers and quarter mechanics- Segments: NA revenue $92.9B (+8%); International $33.5B (+8% ex-FX); AWS $29.3B (+17%).
- Operating income: NA $5.8B (6.3% margin; 7.2% ex-charges ≈ ~$0.84B of the ~$1B charge); International $1.0B (3.0%; 3.7% ex-charges ≈ ~$0.23B); AWS $11.5B (~39.2% implied margin). The ≈$1.06B implied total charge matches Nowak's "that $1 billion" framing.
- Units: +8% YoY worldwide — a sharp deceleration from +11%, and now below ex-FX revenue growth (+10%). The multi-quarter units-revenue divergence has closed and slightly inverted; watch whether staples mix or higher ASPs drive it.
- 3P unit mix: 61%, flat to Q1 last year — oscillation continues at the high end of the 59–61% band.
- Advertising: $13.9B, +19% YoY (vs. +18% in Q4; reaccelerated) — ~$56B annualized; 275M+ US ad-supported audience claimed; strength attributed to full-funnel portfolio, DSP, clean rooms, and sports properties (NFL, NBA, NASCAR).
- Capex: $24.3B "cash CapEx" in Q1. Comparability flag: Q4's $26.3B included equipment finance leases, so this is not a clean sequential decline. Annualized cash-capex run rate ≈ $97B, versus the ~$105B implied from Q4 commentary.
- Net income: $17.1B including a $3.3B pre-tax non-operating gain from converting part of the Anthropic convertible notes into non-voting preferred stock.
- Q2 guidance: revenue $159–164B; OI $13–17.5B ($4.5B wide). Olsavsky attributes width to general uncertainty; cites stock-comp step-up (annual comp cycle) and added Kuiper launch expenses; direct Q2 tariff costs "not large" because of Q1 pre-buying.
- Accounting silence: no mention this quarter of the useful-life changes (servers 6→5 years; fulfillment 10→13 years; the $920M early-retirement charge) that dominated Q4 mechanics. The prior Q1 guide said those changes were embedded; their effect on segment margins went unexplained here.
Product and launch scorecard- Kuiper: milestone achieved — first production-designed satellites launched "earlier this week" (Jassy calls it "first satellite," Olsavsky "production-designed satellites"). New dated commitment: service to customers "later this year"; Olsavsky says commercialization (capitalization trigger) is planned for later this year with launch costs expensed until then. This closes the accountability gap flagged for two quarters, but the real test moves to service timing and initial revenue.
- Alexa+: shipped, with price, early user count (100K+), and a 500M-device installed base claim — materially better launch evidence than the prior "near future" vagueness. Still no engagement, retention, or conversion metrics; Jassy's evidence is anecdote (restaurant reservations, dinner-guest scenes). International rollout "later this year."
- Trainium2: "laying in capacity in larger quantities with significant appeal and demand"; Anthropic is building its next few training models on Trainium2 on AWS. The 30–40% price-performance claim restated (still vendor benchmark). First revenue-contribution commentary still absent — adoption evidence remains capacity-consumed-on-arrival plus Anthropic.
- Nova family expansion: latest premier model launched "yesterday"; Nova Sonic (speech-to-speech) claimed lower word error / higher win rates; Nova Act research preview for browser-based agentic actions with an accuracy roadmap (30–60% today → 90%+ "with the right building blocks"). Nova customer list updated: Slack, Siemens, Sumo Logic, Coinbase, FanDuel, Glean, Blue Origin — prior launch logos were Palantir/SAP/etc.; "thousands of customers" framing persists without a count. All performance claims remain self-validated.
- Amazon Q: new agentic coding experience in CLI plus a GA integration into a GitHub platform (transcript wording garbled) covering Java 8/11 upgrades, code review, unit testing. Notably, the $260M/4,500 developer-years claim was not repeated this quarter — boilerplate rotated out.
- Bedrock: added Claude 3.7 Sonnet and Llama 4; "first CSP" to GA DeepSeek R1 and Mistral Pixtral Large as fully managed models — speed-to-integrate remains the positioning proof.
- Retail selection: Saks luxury storefront (Dolce&Gabbana, Balmain, Erdem, Giambattista Valli, Jason Wu); brand adds Oura, Michael Kors, The Ordinary. Deal events spanning US/Canada, Europe, and four Middle East markets with $500M+ claimed savings; eleventh Prime Day set for July.
- Advertising product: full-funnel attribution across owned entertainment (Prime Video, Twitch, IMDB, Music, Wondery), live sports, and external publishers (Pinterest, BuzzFeed cited) via DSP and clean rooms. Still no Prime Video ads revenue sizing — the omission persists despite the completed first broadcast year.
- Content footnote: James Bond JV producers named (Amy Pascal, David Heyman) — immaterial to numbers but signals the content slate progressing.
Sell-side read-through- Sandler (Barclays): asked the acceleration-timing question directly — when does AI revenue translate into AWS acceleration, this year or next? Jassy's answer reframed it as purely capacity-gated ("we could be driving more revenue if we had more capacity") without committing to a year. The "as the year proceeds" relaxation language is vaguer than Q4's H2 specificity.
- Sheridan (Goldman): forced the most complete tariff doctrine of the call (forward buys, seller diversity of pricing strategies, China structural argument, six-year supply-chain diversification away from China). Olsavsky's Q2-cost answer yielded the SBC step-up explanation and the incremental Kuiper launch cost disclosure.
- Post (BofA): pressed AWS lumpiness and competitor growth optics. Jassy gave the fullest "lumpiness" mechanics to date — startup product-market-fit unpredictability, enterprise migration sequencing, and successive AI use-case waves (productivity/cost-avoidance → training clusters → chatbots → coding agents like Cursor and Vercel, both "significantly on AWS"). The base-effect defense ("17% on a $117B base") is the standing answer to the Azure/GCP growth-comparison question.
- Anmuth (JPMorgan): margin question extracted real mechanics for once — software/process optimization of server capacity, custom low-cost networking gear, reclaiming power in existing data centers, Graviton — but no normalized-margin target; Olsavsky explicitly warns H2 capacity additions will start weighing. His Alexa follow-up produced the only quantified launch datapoint (100K+ users).
- Nowak (Morgan Stanley): operational focus given (inventory-intake management into fulfillment nodes is the named tariff-era discipline). Extracted two admissions: Q2 tariff costs "not large" because the Q1 pre-buy already happened, and the OI range was deliberately widened for uncertainty.
- Thill (Jefferies): the call's payoff — backlog finally answered by Fildes: $189B, +20% YoY, 4.1-year weighted average life. After two straight quarters of refusal/absence, disclosure arrived the moment the question was asked plainly.
- Cross-read: tariffs crowded the agenda; AWS margin normalization got mechanics but no target; nobody asked capital returns (omission streak extended); nobody pressed the Jassy/Olsavsky revenue-number discrepancy.
Management credibility- Positive: fifth consecutive OI beat on the streak count carried from the prior packet; backlog disclosed on demand after prior lapses; Kuiper actually launched inside the revised window; Alexa+ shipped with price and user count; ex-charge margin disclosure (90bps NA / 70bps Intl) shows willingness to itemize the tariff costs they chose to incur.
- Ongoing withholdings: AI revenue remains "multibillion-dollar run rate, triple-digit growth" for a fifth consecutive quarter — the most durable disclosure debt in the packet. Prime Video ads still unsized. Trainium2's 30–40%, Nova Sonic's error/win rates, Nova Act's accuracy roadmap, and Q's customer claims remain internal benchmarks.
- Consistency slippage: capacity-relaxation language softened from Q4's "really start to relax in the second half of '25" to "continue to get better as the year proceeds," while Olsavsky preserved a back-half capacity commitment — the dated call survives, but enforcement got fuzzier.
- Range-widening admission: Olsavsky explicitly says uncertainty widened the Q2 OI range. That is honest hedging, but it also pre-excuses downside — note the prior quarter's pre-excused guide (FX/leap) landed above its top end, establishing a pattern of conservative staging.
- Record-quality flags: Jassy's "$165.7B" vs. Olsavsky's "$155.7B," and Jassy's singular "first satellite" vs. Olsavsky's plural "production-designed satellites" — small but unwarranted inconsistencies on a scripted call.
- Under pressure: Jassy's tariff answers were specific and falsifiable (1P forward buys, seller pricing dispersion, China-direct tariff arithmetic); the Alexa behavioral answer (no more wake word; incremental capabilities monthly) was concrete; the margin answer gave real drivers but dodged the normalization target.
What changed versus the prior quarter- AWS decelerated (19% → 17%) and run-rate step-up slowed again (+$2B vs. +$5B): the acceleration story is now entirely dependent on H2 capacity delivery.
- OI beat size compressed ($1.2B → $0.4B); YoY OI growth fell from +61% to +20% — the beat streak lives, but with shrinking margin for error.
- FCF declined a third straight quarter ($36.2B → $25.9B TTM) with no trough established.
- Tariffs supplanted DeepSeek as the narrative core; ~$1B charges turned pull-forward inventory from framing (Q4 guidance's FX/leap pre-excuse) into booked cost — and the ex-charges margins show the underlying retail P&L held up (7.2% / 3.7%).
- Both long-standing product overhangs partially resolved: Kuiper launched, Alexa+ launched — versus Q4's silence on both.
- Backlog returned to disclosure ($189B, +20%, 4.1-yr life) after the Q4 lapse.
- Advertising reaccelerated (18% → 19%) while the run-rate framing (~$56B) went unmentioned — pitch quality rose without a new vanity metric.
- Units growth (8%) fell below ex-FX revenue growth (10%) for the first time in the packet history — the units/price divergence inverted; combined with the Everyday Essentials 2x-growth claim, mix is visibly shifting to staples.
- Useful-life/accounting mechanics went from dominant (Q4) to unmentioned (Q1).
- The Q "$260M" boilerplate rotated out after three-plus quarters — replaced by new feature claims.
- NA margin stepped down (8.0% → 6.3%) as Q4 seasonality unwound; International was flat (3.0%) but ex-charges reached 3.7%.
Bull case- Demand corroboration arrived: $189B backlog (+20%, 4.1-year life) finally supports the "we could grow faster with capacity" claim that management has carried for a year. If H2 capacity lands as Olsavsky states, the AWS re-acceleration thesis has a dated, funded setup.
- Tariff exposure is partially engineered: the ~$1B charge represents inventory already secured pre-tariff; Q2 direct tariff cost is "not large"; the China-direct structural argument, if directionally right, means Amazon's marketplace mix weathers trade escalation better than traditional retail — and even spot pre-buys showed up as April strength per Olsavsky.
- Ex-charges margins tell the real story: NA 7.2% and International 3.7% — the cost-to-serve engine (inbound architecture just rolled out, record delivery speeds, highest ever same/next-day volume) held through the tariff maneuvering. The discipline of pulling inventory forward without clogging fulfillment productivity is itself a demonstrated capability.
- AWS margin at ~39% with the useful-life headwind supposedly embedded suggests the efficiency levers (power reclamation, custom networking, Graviton) are real and offsetting the accounting reversal — margin quality is better than the prior packet's bear framing assumed.
- Advertising reaccelerating at 19% on a ~$56B run rate with full-funnel tooling and sports inventory still scaling.
- Product surface broadened: Alexa+ live with pricing and 100K+ users on a 500M-device base; Nova Act offers an agentic differentiation claim; Kuiper's capital-sink phase has an end date ("service later this year" → capitalization switch).
- Trainium2 demand-consumed-on-arrival plus Anthropic training on it is the clearest custom-silicon evidence yet.
Bear case- AWS decelerated to 17% while competitors post higher percentage growth and the concession is now structural: "lumpiness," base effects, and capacity gating. If H2 capacity slips or fails to convert, there is no remaining explanation layer — and Jassy already softened the relaxation language ("as the year proceeds" vs. Q4's H2 promise).
- FCF at $25.9B TTM against ~$24–26B quarterly capex is the third quarterly decline in a row; the prior packet's bear point — that capex is no longer internally covered — stands unrefuted, and management still volunteers nothing on financing or returns.
- The tariff buffer is finite: it worked for Q1 and Q2 ("not large"), but every quarter of tariff persistence burns the pre-buy cushion; the charge structure itself ("historical customer returns unresolved") hints at messiness management hasn't fully quantified.
- Units growth (8%) now lags revenue ex-FX (10%) — if staples mix is doing the work, ASP/deflator dynamics or slowing discretionary demand may be embedded; management's "stocking up" claim for April buying also concedes pull-forward risk into Q2/Q3 demand.
- Q2 OI guide's low end ($13B) plus explicit range-widening plus SBC step-up plus Kuiper expense equals a constructed low-expectation setup — a repeat of the pre-excuse pattern that allows a weak low-end print to be dismissed.
- Inference-price mission cuts both ways: promising lower inference prices as "mission and responsibility" is an open acknowledgment of price deflation in the fastest-growing line — bullish for volume, bearish for the revenue-per-unit Al business still measured in "run rate" boilerplate.
- Disclosure debt persists even after the backlog concession: AI absolute revenue (5th quarter withheld), Prime Video ads sizing, Q external metrics, Alexa+ retention economics.
Next-quarter watchlist- AWS: does 17% stabilize, re-accelerate, or slip; evidence of back-half capacity landing (power procurement, Trainium2 volume, NVIDIA next-gen instances "in the coming months"); whether backlog is re-disclosed and whether it still corroborates at ~20% growth.
- The revenue discrepancy: whether $155.7B vs. Jassy's $165.7B was a misstatement; the Jassy scripted figure entered the public record.
- Tariffs: whether Q2's "not large" holds as pre-buy inventory cycles through; any second tranche of pull-forward charges; ASP/demand evidence vs. management's no-attenuation claim; Everyday Essentials mix persistence.
- Q2 print vs. widened range: whether the $4.5B OI range was genuine uncertainty or another staged low bar; the SBC step-up's actual quantum; incremental Kuiper launch expense trajectory toward the "later this year" capitalization trigger.
- FCF/capex: whether $24.3B "cash CapEx" is reaffirmed as run-rate or rises toward the ~$105B implied 2025 figure; whether FCF bottoms; any financing or capital-returns admission now that the omission streak has extended again.
- Kuiper: service launch confirmation ("later this year"), pricing, and the expensing-to-capitalization switch timing.
- Alexa+: user count beyond 100K, country expansion, retention/engagement metrics, and whether the Prime-free perk moves Prime subscription economics.
- Trainium2/3: first revenue-contribution language; external production deployments beyond Anthropic; Trainium3 preview timing (previously late 2025).
- Nova Act/Sonic: whether agentic-accuracy claims (90%+) get third-party validation; Bedrock model additions cadence.
- Units vs. revenue: whether 8% vs. 10% persists — staples mix shift versus genuine demand softness.
- AI revenue: whether five quarters of "multibillion run rate, triple-digit growth" finally yields an absolute number as base effects force it.
|
| Feb 6, 2025 | -4.05% | Q4 FY2024 | Read transcript briefingQuarter in one view- Revenue $187.8B, +10% YoY (+11% ex-FX). FX was a ~$900M headwind, ~$700M worse than assumed in guidance — the third consecutive quarter of FX-versus-plan noise, and this time it flattered the guide-beating narrative in reverse: management says revenue would have exceeded the top end absent the extra FX hit.
- Operating income $21.2B, +61% YoY, the largest OI quarter ever, $1.2B above the high end of guidance — the fourth consecutive OI beat, though smaller than Q3's $2.4B and Q1's $3.3B.
- TTM FCF adjusted for equipment finance leases fell to $36.2B from $46.1B — a second consecutive sequential decline, now −$9.9B over two quarters, and only +$0.7B YoY. The capex ramp is now decisively consuming cash flow.
- AWS grew 19% YoY ($28.8B), a $115B annualized run rate (from $110B) — flat versus Q3's 19.1%, ending the three-quarter acceleration streak. AWS OI $10.6B (+$3.5B YoY), margin ~36.8% implied, down from ~38%, with the same ~200bps useful-life tailwind embedded.
- Capex: $26.3B in Q4 alone; Olsavsky said that run rate is "reasonably representative" of 2025 — implying ~$105B annualized, a step-change from the ~$75B FY2024 figure given last quarter.
- NA margin jumped to 8.0% ($9.3B OI, +190bps YoY) from 5.9% — a large sequential step-up with no single mechanical bridge, though Q4 is seasonally strong. International margin 3.0% ($1.3B OI, +400bps YoY), down from Q3's 3.6% — the volatility flag from last quarter confirmed again.
- Accounting changes cut both ways: servers/network useful life shortened 6→5 years (−$700M FY2025 OI) plus a $920M Q4 accelerated-depreciation charge from early retirements; fulfillment heavy equipment lengthened 10→13 years (+$900M FY2025 OI). Net ~−$400M FY2025 OI. Notably, the server-life shortening partially reverses the 2024 change that has been flattering AWS margins all year.
What management is focused on- The capex defense, again, now at a bigger number: Jassy repeated the cash-cycle argument (faster growth → upfront datacenter/chips spend) and escalated the rhetoric — AI is "the biggest opportunity since cloud, probably the biggest technology shift and opportunity in business since the internet." The ~$105B implied 2025 rate was delivered matter-of-factly by Olsavsky, with Jassy framing it as "a good sign medium to long-term."
- Capacity constraints got their fullest itemization yet: third-party chips arriving slower "with a lot of midstream changes," hardware yield ramp time, Trainium2 volume landing "over the next couple of quarters," power constraints, and motherboard shortages. New forward claim: constraints "really start to relax in the second half of '25" — a checkable, dated prediction.
- The DeepSeek response is a full thesis: Jassy praised DeepSeek's training/inference techniques, noted AWS added R1 to Bedrock and SageMaker "faster than you saw from others," and argued cheaper inference expands total technology spend (the 2006 S3/EC2 pricing analogy) rather than shrinking it. This is the pre-emptive answer to the "AI capex overbuild" question.
- Cost-to-serve remains the retail margin engine: second consecutive year of lower global per-unit cost to serve; inbound network redesign rolled out "a few months ago" with ideal-building placement up 40%+ YoY ahead of Black Friday; same-day sites +60% in 2024 to 140+ metros; 9B+ units delivered same/next day globally.
- AI inside retail got its most quantified treatment: ~1,000 GenAI applications built or in progress; chatbot +500bps satisfaction; seller detail-page generation; 10% better inventory forecasting / 20% better regional predictions; GenAI in robotics "brains."
- Agentic-commerce defense: Morton's question about third-party agents disrupting the shopping funnel got a "retailers will have terms on how they interact with agents, and we'll be no different" answer — the first hint Amazon intends to gate external agents rather than be disintermediated, plus a Rufus capability pitch still devoid of usage metrics.
Key numbers and quarter mechanics- Revenue: $187.8B, +10% reported / +11% ex-FX. NA $115.6B (+10%); International $43.4B (+9% ex-FX); AWS $28.8B (+19%, $115B run rate).
- Operating income: $21.2B, +61% YoY, +$1.2B above guidance high end. NA OI $9.3B, margin 8.0% (+190bps YoY). International OI $1.3B, margin 3.0% (+400bps YoY) — eighth consecutive quarter of YoY margin improvement in both retail segments. AWS OI $10.6B (+$3.5B YoY).
- Useful-life/accounting: ~200bps of AWS's YoY margin improvement again from the 2024 server-life extension — but now partially unwinding: servers/network 6→5 years from January 2025 (−$700M FY2025 OI), $920M Q4 accelerated-depreciation charge on early-retired equipment, fulfillment heavy equipment 10→13 years (+$900M FY2025 OI). Net FY2025 impact ~−$400M. The stated reason — "increased pace of technology development, particularly in AI/ML" — is an implicit admission that AI server economics depreciate faster than assumed.
- Units: +11% YoY worldwide, again outpacing ex-FX revenue growth only marginally (11% vs 11%) — the units-revenue divergence narrowed versus prior quarters.
- 3P mix: 61% of units for full-year 2024, "highest annual mix ever" — reversing Q3's 60% dip, consistent with the 59–61% oscillation framing.
- Advertising: $17.3B, +18% YoY, $69B annualized run rate, "more than double" the $29B of four years ago. Growth decelerated slightly from 18.8%.
- Capex: $26.3B in Q4 (cash capex + equipment finance leases); Q4 run rate "reasonably representative" of 2025 → ~$105B implied. Majority is AI/AWS.
- FCF: $36.2B TTM adjusted, +$0.7B YoY, −$9.9B vs. Q3 TTM.
- Q1 2025 guidance: revenue $151–155.5B (includes ~$2.1B / 150bps FX headwind and ~$1.5B / 120bps leap-year comp headwind); OI $14–18B (includes the useful-life changes).
- Kuiper: one passing mention — "satellites in the coming months," with the reminder that development costs are expensed until commercial viability. No launch confirmation, no dates, no quantification, despite Q2's stated Q4 launch window.
- Not disclosed: AWS backlog (third lapse in four quarters), GenAI absolute revenue (still "multibillion-dollar run rate, triple-digit growth"), Prime Video ads revenue (first full year completed, still unsized), Rufus metrics, Alexa rebuild timing (not mentioned at all this quarter), capital returns.
Product and launch scorecard- Trainium2: reached GA at re:Invent (December), hitting the "next few weeks" ramp timing given last quarter. First quantified benchmark: "typically 30–40% better price performance than other current GPU-powered instances." Named early testers: Adobe, Databricks, Poolside, Qualcomm. Anthropic's Project Rainier disclosed — a cluster of Trainium2 UltraServers with hundreds of thousands of chips, 5x the compute of the cluster used to train Anthropic's current frontier models. Volume ramp "over the next couple of quarters." Trainium3 preview expected late 2025, Trainium4 in definition. This is the strongest product evidence in the packet — real customers, real benchmarks, real volume timing.
- Nova foundation models: launched in Bedrock; claimed ~75% lower price than other Bedrock models with comparable intelligence; "thousands of customers" including Palantir, SAP, Dentsu, Fortinet, Trellix, Robinhood. First-party model family now exists — a strategic shift worth tracking against Bedrock's neutral-marketplace positioning.
- Amazon Q: the $260M / 4,500 developer-years / 30,000 applications claim repeated for a third consecutive quarter — still internal, still unaudited. New: Q Transform expanded to .NET-to-Linux, VMware-to-EC2, and mainframe migrations, with "early customer testing" claiming >50% time reduction on mainframes. Still no external adoption figures.
- Bedrock: 100+ new models added at re:Invent (Luma AI, Poolside), DeepSeek R1 added quickly post-launch; new features: prompt caching, intelligent prompt routing, model distillation. Customer count still not updated.
- SageMaker: HyperPod's "up to 40%" training-time savings repeated; new cluster-level cost management and workload prioritization; next-gen unified SageMaker launched.
- Core infrastructure: Aurora DSQL (serverless distributed SQL, claimed 4x faster reads/writes than "other popular distributed SQL databases"), S3 Tables (first object store with managed Apache Iceberg), S3 Metadata. New AWS logos: Intuit, PayPal, Norwegian Cruise Line, Northrop Grumman, Guardian Life, Reddit, Japan Airlines, Baker Hughes, Hertz, Chime, Asana.
- Amazon Haul: launched in Q4 (ultra-low-price storefront, the Temu/Shein response); "off to a very strong start" — no metrics.
- Prime Video ads: first full year complete, "quite pleased with the early progress" — five quarters post-launch, still zero revenue sizing.
- Alexa: not mentioned in prepared remarks at all — the "near future" next-gen launch promised last quarter went unaddressed.
- Kuiper: effectively no status update; expensing continues, commercial viability (the capitalization trigger) not achieved.
Sell-side read-through- Mahaney (Evercore) got the call's biggest number: confirmation that Q4's $26.3B capex annualizes into 2025 (~$105B), plus the fullest constraint itemization and the H2 2025 relaxation prediction. The capex answer was given without resistance — a contrast to prior quarters' refusals.
- Sheridan (Goldman) asked the DeepSeek/cost-curve question directly and got Jassy's inference-deflation-expands-demand thesis. Notably, Jassy did not claim DeepSeek changes Amazon's capex plans — the ~$105B stands regardless.
- Anmuth (JPMorgan) pressed AWS margin normalization (mid-20s to high-30s range over two years) and got the most explicit margin admission yet: AI "comes originally with lower margins and a heavy investment load... in the short term, that should be a headwind on margins," with only a long-term promise of "comparable to non-AI business." His UPS question extracted confirmation that UPS is walking away from Amazon volume as lower-margin, and that Amazon will absorb it in its own network.
- Nowak (Morgan Stanley) got Shreveport follow-up: robotics savings confirmed as real but "early days," expansion planned to new and retrofitted facilities, a "next wave" beyond the current half-dozen initiatives — still no quantified savings or timeline to P&L impact.
- Blackledge (TD Cowen) got the delivery-speed economics framing: no diminishing returns observed on faster promises, measured via conversion and downstream purchase frequency; inbound network redesign described as months-old with tuning efficiencies still to come in 2025.
- Morton (MoffettNathanson) asked the quarter's most strategic question — agent-driven disintermediation of the e-commerce funnel — and got a gatekeeping posture ("retailers will have terms... we'll be no different") plus Rufus feature talk with no adoption data.
- Six analysts again. Third consecutive quarter with no capital-returns question — now conspicuous with FCF down to $36.2B. Unasked: Kuiper status (despite the missed Q4 launch window), AWS backlog, Alexa rebuild timing, Prime Video ads sizing, international margin mechanics.
Management credibility- Delivery record extended: fourth consecutive OI beat; Trainium2 GA landed inside the promised window with real benchmarks and named customers; the 3P mix dip reversed as the oscillation framing predicted; international held YoY improvement (eighth straight quarter) even as QoQ volatility persisted.
- New specificity where pressure was applied: capex annualization given plainly, constraints itemized with a dated relaxation call (H2 2025), AWS AI margin headwind admitted explicitly for the first time ("lower margins and a heavy investment load... short-term headwind").
- Credibility flags: (1) the server useful-life reversal is significant — the 6-year assumption that added ~200bps to AWS margins all year is now 5 years with a $920M early-retirement charge, meaning some of 2024's margin was borrowed from 2025; the AI-driven rationale implicitly concedes the original extension was mistimed for AI hardware; (2) GenAI revenue is still "multibillion-dollar run rate, triple-digit growth" — four consecutive quarters without an absolute number; (3) the Q $260M claim is now three-quarters-old boilerplate, still internal-only; (4) Kuiper missed its stated Q4 launch window with zero acknowledgment — the accountability gap from last quarter widened rather than closed; (5) Alexa's "near future" rebuild vanished from the script one quarter after being promised; (6) Prime Video ads remains unsized after a full broadcast year; (7) "30–40% better price performance" (Trainium2) and "4x faster" (Aurora DSQL) and "75% lower price" (Nova) are all vendor benchmarks without third-party validation.
- Consistency: the three-layer stack, cost-to-serve flywheel, capacity-constraint framing, and capex-as-opportunity defense are all stable and now routinized. The H2 2025 constraint-relaxation date and the AI-margin-headwind admission are new, falsifiable commitments — credit for making them.
What changed versus the prior quarter- AWS acceleration ended: 19.1% → 19.0%, and the run-rate step-up slowed ($110B → $115B vs. $105B → $110B). The constraint narrative shifted from "primarily chips" to a five-item list including power and motherboards, with a new dated relaxation claim (H2 2025).
- Capex guidance stepped up hard: ~$75B FY2024 → ~$105B implied 2025, delivered as routine rather than dramatic.
- FCF deteriorated again: $46.1B → $36.2B TTM; YoY growth collapsed from +128% to +2%.
- The useful-life tailwind began reversing: the accounting change that flattered AWS margins all year is now a FY2025 headwind (−$700M servers, +$900M fulfillment equipment, $920M Q4 charge).
- NA margin made a large step-up (5.9% → 8.0%) while international reverted downward (3.6% → 3.0%) — the whipsaw pattern continues in both segments, still without mechanical bridges.
- DeepSeek entered the narrative as both validation (added to Bedrock quickly) and threat-management (inference-deflation thesis); last quarter this didn't exist.
- Amazon Haul launched — the first direct response to ultra-low-price Chinese marketplaces.
- Alexa and Kuiper both disappeared from substantive discussion after being featured (Alexa) or demanded (Kuiper) in prior quarters.
- Units-revenue divergence narrowed (11% vs 11% ex-FX) from 12% vs 11%.
- Advertising decelerated marginally (18.8% → 18%) on a much larger base; the $69B run-rate framing is new.
- Q1 guidance embeds an unusual double headwind disclosure (FX $2.1B + leap year $1.5B) — management is pre-excusing a weak-looking Q1 print.
Bull case- The capex is now fully quantified and demand-justified: ~$105B for 2025 against a stated constraint set (chips, power, motherboards) with management claiming it "could be growing faster" — and a dated H2 2025 relaxation that sets up a potential AWS re-acceleration into 2026.
- Trainium2 is now a real product story: GA on time, 30–40% price-performance claim, Anthropic's Project Rainier at 5x scale, named testers (Adobe, Databricks, Qualcomm), Trainium3 already dated. Custom silicon is the clearest answer to the AI-margin question Sandler raised two quarters ago.
- Retail margin machine keeps delivering: NA at 8.0%, eighth straight quarter of YoY improvement in both segments, second straight year of lower per-unit cost to serve, with inbound redesign, same-day expansion (+60% sites), and robotics (Shreveport scaling) all claimed to have 2025 headroom.
- DeepSeek is arguably net-positive for AWS: cheaper inference expands workload volume (Jassy's argument), Bedrock's model-neutral marketplace captured R1 demand immediately, and Nova at 75% lower price positions Amazon on the right side of inference deflation.
- Advertising at a $69B run rate still growing 18% with Prime Video ads only in year one and full-funnel/attribution tooling just shipped.
- Record OI ($21.2B) with the beat coming despite a $920M accelerated-depreciation charge absorbed in the quarter.
Bear case- The cash math is now stark: FCF down to $36.2B TTM (+2% YoY) against a ~$105B capex year — the gap is financed by the balance sheet, and the "FCF covers capex internally" claim from last quarter is no longer true on these numbers.
- AWS growth stalled at 19% with constraints now including power — a constraint capex alone may not fix on schedule. If H2 2025 relaxation slips, the re-acceleration thesis slips with it, and backlog (unreported again) can't corroborate the demand story.
- Margin quality is deteriorating structurally even as it improves optically: management now admits AI carries lower margins near-term; the 2024 useful-life benefit is reversing into 2025 headwinds; and the $920M early-retirement charge reveals that AI hardware is obsolescing faster than the depreciation schedule assumed — a direct challenge to the "20–30 year asset" defense.
- Q1 guidance is soft: $151–155.5B revenue implies ~5–9% growth (even adjusting the disclosed $3.6B of FX/leap headwinds, ~7–11%), and OI of $14–18B implies a wide range with a possible YoY decline at the low end — management pre-loaded excuses.
- International margin fell back to 3.0% — three quarters, three regimes (0.9% → 3.6% → 3.0%), still no mechanics; the "North America margins" aspiration remains timeline-free.
- Disclosure debt is compounding: GenAI revenue absolute (4 quarters withheld), Kuiper (missed launch window, unacknowledged), Alexa rebuild (promised "near future," now silent), Prime Video ads (full year, unsized), backlog (lapsed again). Each is individually small; collectively they shield the highest-uncertainty bets from scrutiny.
- UPS volume insourcing adds near-term logistics load at exactly the moment capex is peaking.
Next-quarter watchlist- AWS: does 19% hold, re-accelerate, or decelerate in Q1 against the constraint narrative; any progress evidence toward the H2 2025 relaxation (power procurement, Trainium2 volume ramp); whether backlog is re-disclosed; whether GenAI finally gets an absolute number as the base grows.
- Capex/FCF: whether the ~$105B annualized rate is reaffirmed or refined; FCF trajectory from $36.2B — is this the trough; any financing commentary (debt, leases) as the FCF-capex gap widens.
- Useful-life mechanics: how the 6→5-year server change and 10→13-year fulfillment change net out in Q1 segment margins; whether further early-retirement charges recur — a signal of AI hardware obsolescence pace.
- Q1 print versus the pre-excused guide: whether revenue lands above $155.5B-equivalent ex-headwinds or the FX/leap framing was covering genuine softness.
- Kuiper: "coming months" per Olsavsky — launch confirmation, delay, or a third quarter of silence; watch for capitalization-trigger language on commercial viability.
- Alexa: the next-gen rebuild is overdue against last quarter's "near future" — launch, pricing/subscription model, or continued absence.
- Trainium2: first revenue-contribution commentary, external customer production deployments beyond testing, and Trainium3 preview timing (late 2025).
- Nova: adoption beyond the named launch logos; any cannibalization commentary versus third-party Bedrock models.
- Amazon Haul: first metrics or continued "strong start" adjectives; any margin-mix commentary as ultra-low-price volume scales.
- International: whether 3.0% is the new base or the whipsaw continues; any mechanical bridge finally offered.
- Advertising: trajectory from 18%; first Prime Video ads sizing attempt now that a full broadcast year is complete.
- Capital returns: three silent quarters with FCF falling — any analyst finally forcing the question, and the answer.
|
| Oct 31, 2024 | +6.19% | Q3 FY2024 | Read transcript briefingQuarter in one view- Revenue of $158.9B, +11% YoY ex-FX. FX was only a 20bps headwind vs. ~90bps assumed in guidance — a favorable swing of roughly a billion dollars versus plan, meaning part of the "beat" is FX luck rather than demand outperformance. Last quarter it was the reverse ($300M worse than guided).
- Operating income of $17.4B, +56% YoY, the highest quarterly OI ever, and $2.4B above the high end of guidance — the third consecutive OI beat, and the largest of the three (Q1: $3.3B above; Q2: $0.7B above).
- TTM FCF adjusted for equipment finance leases fell to $46.1B (from $51.4B at Q2), still +128% / +$25.9B YoY. First sequential decline in this packet — the capex ramp is now visibly consuming cash flow.
- AWS accelerated a third consecutive quarter: 19.1% YoY (vs. 18.8% Q2, 17.2% Q1), now a $110B annualized run rate (vs. $105B+). AWS OI $10.4B (+$3.5B YoY); margin ~38% as flagged by Anmuth and uncorrected by management — back at/above Q1's implied 37–38% after Q2's "mid-30s%."
- International margin staged a full reversal: OI $1.3B, 3.6% (+390bps YoY) vs. 0.9% in Q2 — the QoQ fragility flagged last quarter inverted within one quarter, with no mechanical bridge offered for either direction.
- NA margin improved to 5.9% ($5.7B OI, +$1.4B YoY) from 5.6% — and notably, Kuiper, the stated Q2 drag, was never mentioned on this call.
- Capex finally quantified: $51.9B YTD, "approximately $75 billion" for FY2024 (first explicit annual number in this packet), and Jassy, unprompted-ish: "I suspect we'll spend more than that in 2025."
- Units accelerated to +12% YoY worldwide (from +11%), still outpacing revenue; trade-down and everyday-essentials mix shift continue; paid Prime membership growth accelerated (no number given).
What management is focused on- Demand-outrunning-capacity is now explicit and sharpened: Q2's "we would like to have more capacity than we already have" became "we have more demand that we could fulfill if we had even more capacity today... it's really primarily chips." Constraint framing is now part of the AWS growth narrative, not just the capex narrative.
- The AI-revenue descriptor advanced for the first time in three quarters: still "multibillion-dollar revenue run rate," but now with qualifiers — "growing at a triple-digit year-over-year percentage" and "growing more than 3 times faster at this stage of its evolution as AWS itself grew." Absolute sizing still withheld.
- A formal capex/ROIC defense: Jassy answered the 2025 capex question with the cash-life-cycle argument (faster demand growth → upfront datacenter/network/chips spend → multi-decade assets, "20 to 30 years"), calling GenAI "a really unusually large maybe once in a lifetime type of opportunity." This is the fullest justification of the spending to date and reads as pre-emptive against a margin-compression narrative.
- International reframed upward: after Q2's 0.9% margin went unexplained, Olsavsky now gives the segment a conceptual destination — "we intend, over the fullness of time... to aim for North America margins." New language, no timeline, no mechanics.
- Cost-to-serve recast as the low-ASP enabler: Olsavsky's "easy to lower prices but much harder to be able to afford to lower prices" framing positions the fulfillment re-architecture (inbound overhaul, 15+ new inbound buildings, 25% better inventory spread, same-day expansion, robotics) as what makes trade-down/essentials mix economically winnable. This is a more strategic articulation than prior quarters' initiative lists.
- Agentic AI as a consumer thesis: Alexa re-architecture on new foundation models ("near future"), with the differentiation claim shifting from answering to "taking actions for customers"; 0.5B devices, "a couple hundred million active endpoints." Rufus expanded to seven more countries; Project Amelia launched for sellers; AI Shopping Guides debuted. Still zero engagement, conversion, or monetization metrics on any of it.
- Robotics got its most detailed treatment: Shreveport 12th-gen FC (25% faster processing, 25% expected peak cost-to-serve improvement), 5–6 new stowing/picking/packing/shipping capabilities, a "primitive building blocks" design philosophy, and a staffed-up robotics-AI hire. Framed explicitly as "early days."
Key numbers and quarter mechanics- Revenue: $158.9B, +11% ex-FX (FX −20bps actual vs. −90bps guided). NA $95.5B (+9%); International $35.9B (+12%); AWS $27.5B (+19.1% ex-FX, $110B run rate).
- Operating income: $17.4B, +56% YoY, +$2.4B above guidance high end. NA OI $5.7B, margin 5.9% (+100bps YoY, +30bps QoQ). International OI $1.3B, margin 3.6% (+390bps YoY, +270bps QoQ), $2.5B YTD, "seventh consecutive quarter of year-over-year operating margin improvement" for both retail segments. AWS OI $10.4B (+$3.5B YoY), margin ~38% (implied; Anmuth's number, unrefuted).
- Useful-life change: quantified identically to Q2 — ~200bps of AWS's YoY margin improvement. Now a standardized disclosure rather than an extracted one.
- Units: +12% YoY worldwide, "pretty similar in North America versus international"; paid Prime membership growth accelerated in the US and globally (unquantified).
- Advertising: $14.3B, +18.8% YoY — the growth-rate disclosure returns after Q2's omission, and confirms deceleration (Q1: +24% ex-FX; Q2: described only as "+$2B YoY").
- Capex: $51.9B YTD (implies ~$21.4B in Q3 alone vs. ~$16.5B implied Q2 — acceleration within the ramp); FY2024 guided to ~$75B; 2025 guided higher ("more than that").
- FCF: $46.1B TTM adjusted, +$25.9B YoY, but −$5.3B vs. Q2's TTM — the first reversal.
- 3P paid-unit mix: 60% in Q3, down sequentially; Jassy framed it as a two-year 59–61% oscillation with everyday essentials skewing 1P. 3P fee levels were not addressed at all — the Q2 "lower than expected seller fees" thread went unmentioned.
- Same-day: 40M+ customers received free same-day delivery in the quarter, +25% YoY.
- Prime events: "largest and most successful" Prime Day and Big Deal Days; >$5B saved across 50M+ deals.
- Headcount: office staff "down slightly year-over-year," flat vs. end of 2023 — a rare workforce quantification, offered as AWS cost-control evidence.
- Not disclosed in this transcript: Q4 guidance numbers (only qualitative "encouraged by the start of the holiday season"), AWS backlog (no update after Q2's restored $156.6B), Kuiper (zero mentions), GenAI absolute revenue, Prime Video ads sizing, Rufus metrics, capital returns, grocery subscription counts (but the $9.99/month price was named for the first time).
Product and launch scorecard- Trainium2: timing converged on schedule — Q1's "H2 2024 and early 2025," Q2's "later this year," now "starting to ramp up in the next few weeks." New evidence of demand: "we've gone back to our manufacturing partners multiple times to produce much more than we'd originally planned." Still no customer names, volumes, or price-performance benchmarks beyond "very compelling."
- Amazon Q: the $260M / 4,500 developer-years / 30,000+ applications internal savings claim is repeated verbatim from Q2 — now a permanent fixture of the script, still internal, still unaudited, no external adoption figures. New claim: "highest reported code acceptance rates in the industry for multiline code suggestions" — unsourced, echoing the Q1 "highest known score" phrasing.
- Bedrock: customer count not updated this quarter (Q1/Q2's "tens of thousands" absent). Model adds: Claude 3.5 Sonnet, Llama 3.2 (Q2 had 3.1), Mistral Large 2, Stability AI. The feature-cadence claim recurred, slightly re-worded: "nearly twice as many ML and GenAI features as the other leading cloud providers combined" over 18 months — unverifiable, and marginally softened from Q2's "2x+."
- SageMaker: HyperPod quantified for the first time — splits training across 1,000+ accelerators, checkpointing/fault-repair saves "up to 40%" training time. "Model builders standardize on SageMaker" repeated without cohort data.
- Infrastructure proofs: Project Ceiba disclosed — NVIDIA chose AWS for its R&D supercomputer, a useful third-party validation. Graviton4: "up to nearly 40% better price performance versus other leading x86 processors." Aurora Limitless Database cited (millions of writes/sec, petabyte scale).
- Prime Video ads: "first broadcast season" just entering after "a very strong showing at upfronts" — four quarters post-launch, still no revenue sizing.
- Alexa/devices: next-gen foundation-model rebuild coming "in the near future"; 0.5B devices, ~200M active endpoints (first disclosed scale figures in this packet). Kindle portfolio fully refreshed (first color Kindle, fastest Paperwhite, pocket model); "early sales significantly outperformed our expectations"; 20B average monthly pages read. Jassy's leadership claim ("opportunity to be the leader in that space") is aspirational; Sebastian's direct monetization question got vision, not economics.
- Pharmacy: ambition expanded with real operating metrics — 95% of first-time US customers served within two business days; 20% of US Prime members within 24 hours; 20 new cities launching next year; "nearly half the US" within-hours capability. This is a step-up from Q2's ">12 cities by year-end" and is checkable against delivery.
- Grocery: the subscription was priced ($9.99/month, Whole Foods/Fresh/local 3P) plus a $0.10/gallon fuel benefit at bp/Amoco/ampm — more specificity than Q2's passing mention, but still no subscriber data and no Fresh V2 decision.
- No metrics: Rufus (geographic expansion only), Project Amelia (described, unmeasured), AI video generator (named, unmeasured).
Sell-side read-through- Anmuth (JPMorgan) finally got the capex number he was refused on Kuiper last quarter: ~$75B for 2024 plus a 2025 "more than that" from Jassy, and the 200bps useful-life re-quantification. He also drew the disclosure that office headcount is down slightly YoY. The 2025 capex admission, delivered casually, is the call's biggest forward-guidance moment.
- Sandler (Barclays) asked the quarter's sharpest question — whether AI datacenter economics replay early-AWS (sub-15% margins, competitive pricing, suboptimal utilization) and how the gap to core ~30%+ closes. Jassy did not dispute the premise; he answered with the capacity-logistics argument (35 regions, ~130 AZs — note the AZ count moved from Q2's ~110), demand signals, and "as the market matures... very healthy margins here in the Generative AI space." Argument by analogy, no numbers. The margin-quality question is now on the record twice (capacity in Q2, margin structure in Q3).
- Nowak (Morgan Stanley) extracted the new international aspiration ("aim for North America margins") but, per pattern, got philosophy rather than targets or mechanics; his robotics question produced the most substantive answer of the call (Shreveport specifics, primitives philosophy, robotics-AI hiring).
- Sheridan (Goldman) got the call's most notable retail admission: "we'll take any short-term degradation in ASP because what we're focused on primarily is free cash flow" — an explicit statement that ASP sacrifice for basket/frequency/share is deliberate strategy, not just macro weather.
- Sebastian (Baird) isolated the 3P mix dip (60%) — the only analyst pressing the unit-mix anomaly — and got the 59–61% band/essentials-1P-skew explanation. His Alexa monetization question was deflected into vision.
- Post (BofA) forced the capacity-constraint admission ("more demand than we could fulfill if we had even more capacity... primarily chips") and the competitive framing (Amazon ~1% of global retail, 80–85% still physical).
- Six analysts again. Notably absent for the second straight quarter: no capital-returns question. Also unasked: Kuiper (after dominating Q2 margin discussion), AWS backlog (the restored series lapses again after one quarter), Q4 guidance bridge, Prime Video ads sizing, seller-fee trajectory. The QA pattern itself — no one re-litigating Q2's unanswered international-margin collapse after it reversed — flatters management's non-answers.
Management credibility- Delivery on forward statements remains the strong suit: third consecutive OI beat and the largest ($2.4B above high end); AWS acceleration promised as "attenuating optimizations" has now compounded for four quarters by management's own count, consistent with the printed sequence (13% → 17.2% → 18.8% → 19.1%); Trainium2 hit its "later this year" window ("next few weeks"); international's Q2 dip did not prove structural, validating the volatility framing they gave (belatedly) this quarter.
- Disclosure improved where it was previously withheld: full-year capex quantified (~$75B) after a year of refusals; the 2025 direction given unprompted; ads growth rate restored after Q2's omission; useful-life quantification now routine; device/endpoint scale figures new.
- Under pressure, performance was mixed: Sandler's margin-structure question was conceded implicitly (premise unchallenged) but answered only by historical analogy; Sebastian's Alexa monetization question got no economics; Nowak's international-margin drivers got framework, not mechanics. Anmuth's AWS-margin-sustainability question got factor lists, not a range.
- Credibility flags: (1) the Q $260M savings claim is now script boilerplate — still internal, unaudited, unvalidated externally; (2) "highest reported code acceptance rates in the industry" and "nearly twice as many ML/GenAI features as the other leading cloud providers combined" remain unsourced superlatives; (3) GenAI revenue has carried the identical "multibillion-dollar run rate" label for three straight quarters — the new triple-digit-growth qualifier without an absolute number implies the base is still small enough to avoid disclosing; (4) Kuiper's complete disappearance one quarter after it was the named NA-margin drag is a real accountability gap — investors were asked to accept an unsized drag in Q2 and are given no status in Q3; (5) the international whipsaw (2.8% → 0.9% → 3.6%) was never mechanically explained in either direction; "aim for North America margins" has no timeline; (6) "~130 AZs" vs. Q2's "~110" — either network expansion or loose number usage, unexplained.
- Consistency: the three-layer AI stack, data-readiness gating, cost-to-serve flywheel, on-premises inversion, and "we like these trends" ASP stance are all stable; this quarter again paired claims with evidence (AZ-scale figures, capex number, Trainium2 timing, device endpoints) more often than with new promises — with the conspicuous exception of GenAI revenue and Kuiper.
What changed versus the prior quarter- AWS accelerated again (18.8% → 19.1%) but with decelerating momentum (+0.3pts vs. +1.6pts in Q2); margin recovered to ~38% from "mid-30s%," with the same ~200bps useful-life assist embedded.
- International margin reversed violently upward (0.9% → 3.6%) and acquired a conceptual destination (NA-like margins) — last quarter's structural-durability worry is now a volatility worry instead.
- Capex crossed into explicit annual guidance: ~$75B for 2024 with 2025 higher — the "no explicit annual number" posture from Q2 ended. Q3 implied spend (~$21.4B) roughly doubles the YoY step-ups.
- FCF inflected: $51.4B → $46.1B TTM — the first decline; the capex-vs-FCF tension flagged as a bear point last quarter is now in the numbers.
- Advertising disclosure was restored (18.8%) and simultaneously confirmed the deceleration that Q2's omission had only implied.
- Kuiper went from named-margin-driver-with-refused-quantification to complete silence.
- GenAI revenue language gained growth qualifiers (triple-digit YoY, 3x early-AWS pace) while the absolute run rate stayed frozen at "multibillion-dollar."
- The consumer story shifted from explanation (trade-down, mix) to strategy (deliberate ASP sacrifice for FCF/baskets; cost-to-serve as the enabler; Prime membership accelerating) — management now claims the divergence rather than merely excusing it.
- 3P mix produced its first anomaly in the packet (60%, essentials skewing 1P); seller-fee momentum was not addressed at all after Q2's "lower than expected" admission.
- Pharmacy quantified delivery capability for the first time (95%/2-day, 20%/24-hr, 20 new cities next year — a timeline expansion vs. Q2's ">12 cities by year-end").
- Devices returned with scale disclosures (Alexa 0.5B devices/~200M endpoints; Kindle 20B monthly pages) after being near-absent in Q2.
- Backlog disclosure lapsed again after Q2's restoration — the series now alternates.
- FX flipped from worse-than-guided (Q2) to better-than-guided (Q3): ±70bps vs. plan — both quarters' "beats/misses vs. guide" carry FX noise in opposite directions.
Bull case- AWS evidence keeps compounding: 19.1% and a third consecutive acceleration, $110B run rate, demand exceeding capacity with chips as the binding constraint, a named supercomputer validation from NVIDIA (Project Ceiba), new enterprise logos across banking/retail/telecom (ANZ, Capital One, Itaú, NAB, Booking.com, Sony, T-Mobile, Toyota), and a newly quantified growth claim for AI revenue (triple-digit, 3x faster than early AWS) — while retaining the structural argument that on-premises migration plus GenAI data architecture still favor cloud capture.
- Margin machine working everywhere at once: record OI with the biggest beat of the year; NA at 5.9% with cost-to-serve programs still claimed to have headroom; International suddenly running at a 3.6% margin with a stated ambition toward NA-level margins; AWS back near ~38%. Multi-segment operating leverage is the cleanest it's looked in this packet.
- The retail strategy now has a coherent economic logic: sacrifice ASP for units, baskets, and Prime acceleration; fund it via inbound re-architecture (25% better inventory spread), same-day scale (40M customers, +25%), and next-gen FCs (25% cost-to-serve improvement at peak) — each element got a measurable metric this quarter, which prior quarters lacked.
- The capacity narrative supports the capex: "we have more demand than we could fulfill if we had even more capacity," Trainium2 manufacturing upsized multiple times, datacenter assets defended as 20–30-year useful — the spending is at least now quantified (~$75B, more next year) rather than open-ended rhetoric.
- New vectors keep adding specifics: pharmacy delivery metrics with a 20-city expansion, Kindle refresh beating expectations, grocery subscription priced, Alexa rebuild "near future" on a 0.5B-device installed base with an action-agent thesis that aligns with where GenAI utility is heading.
- FCF of $46.1B TTM still covers the guided capex trajectory internally, even as it absorbs the first decline.
Bear case- The cash pivot is now visible: FCF fell $5.3B sequentially while capex guidance jumped to ~$75B with 2025 higher; AWS margin near 38% still carries ~200bps of accounting benefit, and depreciation from this spending wave has barely started landing. The Sandler question — whether AI workloads can ever match legacy AWS margins — was answered with analogy, not math, and the pricing-competition premise in his question went undisputed.
- Acceleration arithmetic is thinning: AWS added only 0.3pts this quarter (vs. 1.6pts last); backlog went unreported again after Q2 showed only ~$0.9B sequential growth; GenAI revenue has been "multibillion-dollar run rate" for three consecutive quarters — by management's own triple-digit/3x framing, the absolute base is being kept opaque because it is small relative to a >$75B capex year.
- Revenue quality issue persists and is now策略 rather than accident: units +12% vs. revenue +11% ex-FX, ASP degradation explicitly accepted, everyday-essentials mix structurally dilutive to revenue per unit, and the 3P flywheel ticked backward (60% mix) with seller-fee momentum undisclosed. If essentials margins don't follow the "afford to lower prices" logic as claimed, this is a deliberate margin-expensive share grab.
- International's 3.6% is as unexplained as its 0.9% was — a segment oscillating 270bps QoQ in one direction and +390bps YoY is not yet a dependable margin base; the "North America margins" destination has no path or timeline.
- FX flattered the quarter: actual −20bps vs. −90bps guided (~$1B swing) — repeat of Q2's FX noise in reverse; both quarters' guide-beating require an FX adjustment the guidance math won't show.
- Kuiper vanished without status: a cost big enough to be cited as the Q2 NA-margin explanation got zero Q3 mentions — no launch confirmation, no beta update, no quantification — while long-dated commitments (11-year sports rights) remain unaddressed anywhere in the packet.
- Advertising deceleration is now confirmed on the printed number (24% → 18.8% across two quarters), and Prime Video ads remains unsized four quarters post-launch despite a completed upfronts cycle and first broadcast season.
Next-quarter watchlist- AWS: whether 19.1% holds into Q4 against a tougher compare and the claimed capacity constraint; whether GenAI gets an absolute number (Sandler-style margin questions will escalate without one); whether backlog is re-disclosed and what its sequential trend says about the "more demand than capacity" claim; whether ~38% margin survives as $75B of capex depreciation lands.
- Capex/FCF: any 2025 framework beyond "more than that"; the Q4 capex print versus the ~$23.1B implied FY remainder; FCF trajectory — whether $46.1B TTM was the pivot or a pause. FX assumptions in Q4 guidance after two consecutive misses in opposite directions.
- Kuiper: launches were slated for Q4 (per Q2) — confirmation, delay, or continued silence is itself informative; beta/commercial timing is now overdue by the company's original framing; watch whether it re-enters the margin narrative or only the questions (Anmuth's refused question remains open).
- International: test of 3.6% durability; any mechanical explanation of the whipsaw; concrete milestones toward "North America margins"; UK/Germany/Japan vs. emerging-market detail.
- Retail economics: whether units-minus-revenue divergence narrows in holiday mix; whether the 3P mix dip reverses and seller fees stabilize; response to Sheridan's ASP-for-FCF framing will clarify whether strategy or macro dominates; Prime membership acceleration — any quantification.
- Advertising: trajectory from 18.8%; Prime Video ads revenue first sizing attempt after a completed broadcast season; whether sponsored-products re-acceleration claims hold.
- Alexa: the "near future" next-gen launch — pricing/monetization model (subscription tier?) given Sebastian's unanswered question; whether action-agent claims survive launch reviews.
- Trainium2: GA timing confirmation, first external customer proof points, and whether "ramping in the next few weeks" converts to volume revenue disclosure in Q4.
- Amazon Q: first external adoption datapoint or third-party validation of the now-twice-repeated $260M internal claim.
- Q4/holiday: guidance ranges absent from this transcript; check the press release and subsequent commentary for the revenue/OI envelope and whether management's "encouraged start" framing survives the season.
- Capital returns: two consecutive silent quarters — any re-entry of debt paydown/buyback commentary as OI records accumulate.
|
Corbanu