AMZN (Trans): AWS guides for 40%+ growth over the next two years, aiming for $1tn

Below is Dolphin Research's recap of$Amazon(AMZN.US) FY26 Q3 earnings call. Prepared by Dolphin Research.

I. Key takeaways from the results

1) Q4 guidance

a) Revenue: $240–247bn. At current FX, the headwind is approx. 60bps.

b) OP: $28.5–33.5bn. Management reiterated operating discipline.

c) Drivers: Amazon Leo will begin initial services this year, but capitalization of production and launch costs shifts to Q1 next year (later than prior expectations). Most Leo-related costs will still be expensed in Q4; fuel and linehaul inflation will be higher in Q4 vs. Q3.

2) Capex

a) Q3 cash capex was $75.9bn, largely for AWS and AI. Investment focus remains on capacity and model capabilities.

b) FY26 cash capex to be ~$220bn, with most earmarked for AWS; FY27 will be 'meaningfully higher.' Spending cadence will follow demand visibility.

c) Payback periods continue to shorten. Despite memory price increases and depreciation pressure, AWS investments are set to drive revenue and FCF far beyond what was envisioned two years ago.

3) This quarter's key metrics

a) Total: revenue of $201.9bn (+12% YoY, +13% ex-FX). OP of $28.1bn (+61% YoY), OPM of 13.9%, a record high.

b) Retail: North America revenue $112.9bn (+6% YoY), OP $8.7bn, OPM 7.7%. Intl revenue $43.1bn (+8% YoY ex-FX), OP $1.6bn, OPM 3.7%.

c) Prime Day lens: most large-country Prime Day events shifted to Q2 this year (all in Q3 in 2025). Excluding this, YoY growth for both retail segments in Q3 would be ~400bps higher; global paid units rose 12% YoY.

d) AWS: +39.3% YoY; sequential revenue add of $3.8bn; ARR ~$184bn. OP $17.8bn, OPM 38.7%.

e) Ads: revenue $19.6bn (+12% YoY). Momentum reflects broader engagement and format expansion.

4) Fulfillment costs: shipping expense +11% YoY, below the 12% unit growth. Ex-fuel and linehaul inflation, shipping rose only 8%, ~400bps slower than units, driven by inventory placement, shorter middle-mile, and package consolidation.

II. Details from the call

2.1 Management commentary

1) AWS: growth and strategy

a) Growth accelerated for the sixth straight quarter, the fastest in nearly five years. Backlog reached $578bn, up triple digits YoY and well diversified across customers.

b) Roughly one AI accelerator installed every ~14 seconds, with capacity absorbed almost immediately. Core compute and storage are growing at multi-year highs.

c) AI margins are improving rapidly and outpacing core at the same time period. Inference now accounts for over half of AWS AI revenue.

d) AWS is positioned as a long-term platform, not merely managed GPUs or short-term capacity rentals. ~85% of global IT spend remains on-prem, and we are in a 10–20 year cloud migration where even mainframe and other critical workloads are moving.

e) Following the DuckLabs acquisition, DuckDB will be the unified SQL dialect for AWS storage analytics. This enables zero migration cost across query engines.

2) AWS: Bedrock, agents, and security

a) Fortune 500 spend on Bedrock is up nearly 7x YoY, and tokens served are up nearly 17x YoY. Bedrock is the fastest-growing multi‑billion dollar service in AWS history.

b) Added OpenAI models (with fast usage growth) plus hosted versions of Kimi, GLM, and SpaceX models. Customer data is not shared with model providers.

c) AgentCore now has tens of thousands of customers, and agents deployed in Q3 were 2x the prior three quarters combined. Packaged agents like Kiro, Amazon Connect, AWS Transform, and Quick are scaling quickly.

d) Security service AWS Continuum automatically finds and fixes code vulnerabilities. It will be integrated into Claude Code, Codex, and Kiro, and has drawn the highest customer interest of any new security service to date.

3) AWS: in‑house chips

a) Chip biz grew triple digits YoY, with revenue commitments for Graviton and Trainium both up more than 20x YoY. Post‑training RL and agent tool‑calling often run on CPUs, making AI a tailwind for Graviton.

b) In Q3, in‑house chips saved customers over $2bn. TorchNeuron was open‑sourced so PyTorch code can run on Trainium with no rewrites.

c) Legacy demand remains strong: after Graviton 5 launched, backlog for prior‑gen instances still more than doubled QoQ. Earlier‑gen Trainium capacity commands a premium vs. 1H levels in the near term.

4) Retail: scale and share

a) YTD units up nearly 15% (vs. 30bn+ units sold last year). Over the past 12 months, total GMV added exceeded $90bn, with each quarter's incremental GMV roughly 2x the No.2 U.S. competitor.

b) NA and Intl OPM each improved by nearly 1,000bps over the past few years. Nearly 80% of retail still occurs in physical stores.

5) Retail: delivery, grocery, pharmacy, and AI shopping

a) In Q3, Prime delivered over 1.5bn items same‑day or next‑day. Amazon Now covers 450+ towns, with orders up 10x YoY.

b) Grocery and consumables generated nearly $250bn in GMV over the past 12 months. Monthly same‑day fresh orders by Prime members more than tripled in a year.

c) Pharmacy users grew over 60% YoY and accelerated sequentially through 2026. Repeat customers grew over 50% YoY, and 3‑hour delivery is available in 350+ cities.

d) Over the past 12 months, 400mn+ customers used Alexa for shopping, with interactions up 50% YoY. U.S. AOV is 50% higher, and management sees Alexa as the primary agent for Amazon shopping.

e) In 2026, the fleet of robotic arms (Cardinal, Sparrow, etc.) will double. A foundational model for robotics is also under development.

6) Advertising

a) Q3 growth was impacted by three factors: Prime Day moved to Q2, creating an ~850bps sequential headwind to Q3 YoY growth; Q2 had the first NBA season while Q3 did not, with NBA resuming in Q4. Also, last year's Q3 SEM optimizations are now a tougher comp.

b) Q4 YoY growth is expected to return to the high teens. Management sees durable demand into year‑end.

c) External horizontal agent shopping has had no material impact on ads, as the experience still trails going direct to retailers. Amazon Ads Agent and full‑funnel campaigns were launched, with AI planning, executing, and optimizing spend.

7) Other biz

a) Prime Video's annualized revenue is in the tens of billions and is profitable on a stand‑alone basis. TNF's 2026 season opener delivered the best start ever.

b) 34 new devices were launched. Management said both AWS and retail are on paths toward $1tn annual revenue.

2.2 Q&A

Q: AWS is accelerating YoY but added less sequentially. Why? How should we think about Q4 capacity adds?

A: No direct explanation for the slower sequential add; capacity additions are uneven quarter to quarter, and significant capacity will come online over the next 8–10 quarters. Growth stems from three areas: first, a full‑stack AI capability spanning SageMaker for custom models, Bedrock for frontier models, Strands and AgentCore for custom agents, and packaged agents like Kiro, Connect, Transform, and Quick. Second, continuous capacity expansion with ~8–10 quarters of visibility. Third, in‑house chips: once enterprise inference goes to production, price‑performance matters greatly and Trainium has a clear edge.

AI adoption also drives core compute for RL, post‑training, and tool‑calling, where Graviton delivers 30–40% better price‑performance vs. x86. Management reiterated the view of 40%+ growth in both 2027 and 2028.

Q: How does fuel cost affect Q4 guidance?

A: No specific numbers; fuel and linehaul inflation will be higher in Q4 vs. Q3 and have been embedded in guidance. A fuel surcharge for FBA sellers introduced early this year only recovers part of fuel costs; the rest lands in inbound and outbound retail transport.

Q: Is the $1tn AWS revenue case in 5–10 years top‑down or bottom‑up, and what underpins confidence?

A: Large blocks of capacity are pre‑booked, providing strong visibility for 2027, 2028, and parts of 2029. Blocks of several hundred MW to 1GW must be reserved far in advance, while significant capacity is always reserved for EC2 on‑demand and Bedrock. At current ARR and growth rates, the numbers scale quickly.

Demand remains early: enterprise inference in production is just starting, with constraints in AI talent and inference cost. Driving down inference cost is core to AWS, and a key reason for in‑house chips; some expect ~20% of tokens over the next 1–2 years to go to security, with additional use cases in coding and assistants.

Management believes nearly all enterprise production apps will be rewritten as agents. They cannot be certain about 5–10 years out, but crossing $1tn is 'not far‑fetched.'

Q: Looking to 2027, what are the key robotics milestones for retail?

A: No 2027 milestone provided; robotics impact should be far greater over the next five years than today. Over 1mn robots have been deployed across the fulfillment network, with safety incidents at historic lows. Repetitive tasks will increasingly be handled by robots, though fulfillment centers will still employ many people; last‑100‑yards delivery is also being explored.

Deployment standards require equal or better outcomes on safety, cost, and speed vs. today. Newly designed facilities, including expanded same‑day sites, are highly automated.

Q: Can AWS OPM hold in the 30s–40% range over the next two years?

A: No margin guidance; too many cross‑currents to predict. The 38.7% this quarter was not a fluke, helped by efficiency gains, capacity optimization, fixed‑cost control, and higher in‑house chip mix. Headwinds include: AI margin, while better than core at the same time, is still below core and mixes in as a drag, plus component price increases.

Q: How many households are covered by fresh checkout, and how much does it lift Prime household spend?

A: No coverage or per‑household lift disclosed; same‑day fresh from same‑day facilities has 'cracked the code.' The biggest wallet share gap is in grocery, and grocery spend often begins with fresh. Grocery and consumables are growing ~3x faster than other categories and account for about half of units; customers buying fresh same‑day purchase ~7x more consumables than other customers.

Grocery customers also purchase 4x more non‑grocery items like home and apparel and spend about 2x more. Whole Foods revenue and profit are both outpacing grocery overall, and the urban small‑format 'daily shop' will roll out over the next few years.

Q: How big is Bedrock's opportunity, and how will it expand?

A: There will not be a single model winner; Bedrock can be as large as EC2. Over the next 2–3 years, expect 6–8 roughly comparable models covering 95%+ of workloads, each with strengths; most customers will use multiple models, often mixing within the same app. Customers want breadth of choice, strong price‑performance, robust security and permissions, and the ability to switch anytime.

Bedrock offers closed models like Anthropic and OpenAI, and open‑source options like Kimi and GLM. In Western markets, demand for local open‑source models is strong.

Q: How will you balance selling Trainium externally vs. using it internally?

A: Still evaluating; external sales begin by end‑2027, with the mix re‑assessed annually thereafter. Only two companies meaningfully have both hyperscale cloud and chips to sell; advanced chips are made at the same foundry and allocated by quota. External sales recognize revenue at shipment, skewing near‑term revenue higher.

Keeping chips in‑house to rent via EC2 or Bedrock over 5–6 years creates more absolute revenue and FCF. Management expects external sales to be a big business, but it is fundamentally different from cloud infrastructure.

Q: How much capex is needed to sustain 40%+ growth for two years?

A: No quantification; investment will be 'considerable,' and the company has the capacity to finance it. Management views this as a once‑in‑a‑lifetime opportunity and the biggest tech shift; if they could amortize spend over 25 years it would be easier, but the opportunity does not allow it. They will keep investing to stay ahead, believing this will permanently change revenue and FCF scale.

Q: Any update on payback periods for servers, networking equipment, and data centers?

A: No specific years; payback periods continue to shorten. Payback differs between core and AI, and by funding approach; AWS is also extending asset useful lives. This improves unit economics.

Q: If most of Bedrock runs on Trainium, how does shifting from selling bare metal to selling inference services affect AWS revenue and margins?

A: Most of Bedrock does run on Trainium, and in‑house chips deliver a cost advantage. Trainium's price‑performance allows lower customer prices while maintaining attractive margins. Two factors explain margin differences across clouds: whether DCs are owned/long‑leased vs. colocation or neocloud, where OPM can be ~3,500bps lower; and the use of differentiated hardware.

Amazon's capacity is largely owned or long‑leased, and part of the cost advantage is passed through to customers. This underpins both growth and loyalty.

<End of text>

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure