← back to ranking

AMZN · Amazon.com, Inc.

Specialty Retail · mkt cap $2787.9B · calls: Q1 FY2026 vs Q4 FY2025
67.0 conviction

enthusiasm:40.0 · trend:8 · quantifies:12 · impact:4 · under_radar:0 · credibility:0 · confirmation:3

Enthusiasm latest 10 / prev 9 (rising)

Both calls are dominated by an AI thesis that management quantifies unusually heavily — and the credibility is high because the numbers are tied to real revenue, backlog, and chip economics rather than buzzwords: AWS reaccelerated from 24% to 28%, AI revenue is a disclosed >$15B run rate growing triple digits, backlog jumped from $244B to $364B, and Trainium carries >$225B in commitments plus a claimed margin/CapEx advantage. The story strengthened quarter-over-quarter (rising) as Amazon broke out its custom-silicon business as a top-3 data-center chip maker, added OpenAI models/agents to Bedrock, and disclosed concrete internal productivity proof (5 engineers in 65 days). The main caveat to credibility is the enormous up-front CapEx ($43B in Q1, ~$200B FY26) whose ROIC depends on continued monetization of capacity that management concedes is being installed ahead of revenue.

GROUNDED IMPACT vs CONSENSUS

Grounded on actual base — revenue $716.9B · net income $77.7B · net margin 10.8% · diluted EPS 7.17

Aggregate est. rev uplift: 4.2% · EPS uplift: 10.3% · vs analysts: inline · priced in: medium · confidence: 6/10

ClaimFigureArithmeticRev %EPS %
AWS AI revenue run rate >$15B, triple-digit YoY
revenue		$15B run rate, >100% YoY$15B AI run rate / $716.924B rev = 2.09%; triple-digit growth treats ~$15B as scaling into FY2026; @25% AWS incremental net margin -> $3.75B / $77.67B NI = 4.83% EPS. Overlaps Bedrock/silicon/Trainium claims - counted once in aggregate.2.094.83
AI revenue growing triple digits YoY
revenue · soft		>100% YoYPure growth qualifier; magnitude already captured in the $15B run-rate claim. Not summed separately to avoid double-count.
Bedrock customer spend +170% QoQ
revenue · soft		+170% QoQQoQ growth rate; 'multibillion' run rate but no clean standalone $; a subset of the $15B AWS AI revenue. Excluded from aggregate to avoid double-count.
Trainium revenue commitments (backlog) >$225B, multiyear
revenue		$225B multiyear$225B over ~4-5yr; assume ~15% recognized in FY2026 (early ramp, capacity-gated) = $33.75B / $716.924B = 4.71% rev; @25% margin -> $8.44B / $77.67B = 10.86% EPS. Heavily overlaps AWS backlog & AI run-rate - not added in full to aggregate.4.7110.86
AWS total backlog $364B (excl. >$100B Anthropic)
revenue · soft		$364B (+49% QoQ vs $244B)Total AWS backlog, not AI-only; ~25-30%/yr converts (~$91-109B) but that is largely already in the consensus AWS revenue base. +49% QoQ signals acceleration. Cannot cleanly isolate the AI-incremental slice - soft; used as priced-in evidence, not summed.
Custom silicon/chips run rate >$20B, +40% QoQ
revenue		$20B run rate, +40% QoQ$20B (incl. non-AI Graviton) / $716.924B = 2.79% rev; @25% margin -> $5.0B / $77.67B = 6.44% EPS. '$50B standalone' is a valuation framing, not recognized revenue - ignored. Overlaps AI run-rate & Trainium commitments - not added in full.2.796.44
Trainium CapEx savings + several-hundred-bps margin advantage
cost · soft		tens of $B CapEx/yr + ~300bps op-marginCapEx 'tens of billions' is a cash/FCF item, NOT income - only flows to EPS via lower D&A (~$20B/6yr=$3.3B, x0.79=$2.6B) and only at scale. 'Several hundred bps' (~300bps) inference-margin edge is bottom-line but base unspecified & 'at scale/forward'. FY2026 scenario: ~+$1B after-tax -> 1.29% EPS, largely offset by rising D&A from the $200B CapEx ramp.1.29
Trainium2 ~30% better price-performance vs GPUs, sold out
other · soft		~30% (40% in Q4)Cost/competitiveness + demand ('largely sold out') signal; no standalone $. Feeds the margin-advantage claim above.
Trainium3 30-40% better price-perf vs Trainium2, subscribed
other · soft		30-40%Roadmap/competitiveness + demand signal ('nearly fully subscribed'); no standalone $.
Trainium2 token throughput +4x (FY2025)
other · soft		4xTechnical efficiency metric; no direct $; supports the inference cost/margin advantage.
SageMaker reduces training time up to 40%
productivity · soft		up to 40%Customer-value feature (adoption driver); not an Amazon P&L saving; no standalone $.
AWS Transform saved >1.56M hours of manual effort
productivity · soft		1.56M hours (cumulative)Customers' saved hours, not Amazon's cost; adoption/marketing signal. No Amazon-P&L impact to size.
Internal AI coding: 5 people/65 days vs ~45 people/1 year
productivity · soft		~44 person-years saved on 1 project~44 person-years x ~$300K fully-loaded = ~$13M on a single project (~$10M after-tax). Anecdotal, not generalizable to a company-wide cost line without inventing scope - soft, immaterial at company scale.
Rufus MAU +115%, engagement +400% YoY
engagement · soft		+115% MAU / +400% engagementEngagement metrics; Rufus users ~60% more likely to purchase (Q4), 300M+ users - implies retail conversion lift, but baseline GMV/conversion unknown so cannot size cleanly. Real but soft topline optionality on the retail line.
Q Developer (QRO): developers >2x QoQ, enterprise ~10x
engagement · soft		2x devs / ~10x enterpriseUsage growth for an AWS dev tool; no standalone $; subset of AWS AI revenue.
AWS cash CapEx $43.2B in Q1 (AI-driven)
cost · soft		$43.2B Q1 (~$173B annualized)Spending, NOT an uplift: a rising D&A + near-term margin headwind. The cost that AI revenue must justify; reduces FCF/EPS in the ramp. Sized in bottomline narrative as a drag.
Bedrock customer spend +60% QoQ (Q4 prior call)
revenue · soft		+60% QoQSuperseded by the +170% Q1 figure; same revenue line. Excluded.
Trainium2: 1.4M chips landed, 30-40% better price-perf (Q4)
other · soft		1.4M chipsVolume/competitiveness, prior call; overlaps the current silicon/Trainium claims. No incremental $.
Curo/Kiro developer usage +150% QoQ (Q4)
engagement · soft		+150% QoQUsage growth; no standalone $.
Rufus 300M+ users, ~60% more likely to purchase (Q4)
engagement · soft		300M users, +60% conversion likelihoodConversion-lift signal; base GMV/attach economics undisclosed, so cannot size cleanly. Corroborates Rufus retail topline optionality - soft.
Total CapEx guidance ~$200B FY2026 (AI/AWS)
cost · soft		~$200B (~28% of revenue)FY2026 investment/cost, not uplift: D&A/opex headwind to EPS. This is the key bottom-line risk - the +22% consensus EPS growth requires AI margins (Trainium edge) to absorb this D&A surge.
AWS backlog $244B, +40% YoY / +22% QoQ (Q4 prior)
revenue · soft		$244BPrior-quarter backlog; used only to show acceleration to $364B (+49% QoQ). Not summed.
>1,000 internal AI applications deployed/in build
productivity · soft		1,000+ appsCount; efficiency/productivity signal; no standalone $.
Power capacity +3.99GW in 12mo (1.2GW in Q4)
other · soft		3.99GW (~2x 2022)Capacity enabler for future AI revenue (and a cost) - gates how fast backlog converts; no direct $ uplift.

Assumptions: Tax rate 21%. Incremental net margin: 25% for AWS AI cloud/silicon revenue (AWS op margin ~35%, after-tax ~27%; haircut to 25%); 10.83% company blended for retail/Rufus claims (none sized). Phasing: multiyear backlog/commitments ($225B Trainium, $364B AWS) ramp ~15-30%/yr but are capacity-gated by the $200B CapEx/power build, so most flows to FY2027+; run-rate claims ($15B AI, $20B silicon) treated as scaling into FY2026. Segment mapping: AI revenue claims -> AWS line; Rufus -> retail line (left soft). Aggregate de-duplicates the heavily overlapping AWS-AI disclosures (AI run-rate, Bedrock, silicon, Trainium commitments, backlog all measure overlapping slices) down to ~$30B net AI-attributable FY2026 incremental revenue rather than summing the raw per-claim figures. CapEx 'savings' treated as FCF, not income (only ~$2.6B/yr after-tax D&A effect at scale). Shares assumed ~constant, so EPS uplift % ~= net-income uplift %.

Top line: Hard AI revenue is real but mostly already inside consensus's +14.9% FY2026 revenue growth (+$107B). AWS AI run rate is $15B (2.09% of $716.9B) and custom silicon $20B (2.79%), both triple-digit; de-duplicated AI-attributable FY2026 incremental is ~$30B = +4.2% of revenue - roughly 28% of the consensus revenue add and approximately the entire expected AWS acceleration. The forward signals dwarf in-year recognizable revenue: $364B AWS backlog (+49% QoQ vs $244B), $225B Trainium commitments, and a separate >$100B Anthropic deal not even in the backlog. But FY2026 revenue recognition is gated by capacity/power/CapEx, so the bulk of that backlog lands in FY2027+ (consensus FY2027 rev only +13%), where it is the least priced-in.

Bottom line: Flowing ~$30B incremental AI revenue at a 25% AWS incremental net margin -> ~$7.5B incremental NI (+9.7% of $77.67B); adding a modest at-scale Trainium margin benefit (~300bps, ~+$1B after-tax FY2026) nets ~$8B -> est EPS uplift ~+10.3%, i.e. ~47% of the +22% net-income growth consensus already expects. AI is plausibly half the FY2026 earnings-growth story. The swing risk is cost, not revenue: ~$200B FY2026 CapEx (28% of revenue; $43.2B in Q1 alone) drives a D&A surge that pressures margins, so the +22% EPS only holds if the Trainium price-performance edge (30% vs GPUs; Trainium3 +30-40%) and 'several hundred bps' inference-margin advantage actually land. The 'tens of billions of CapEx saved' is an FCF/valuation positive, not an income-statement saving (~$2.6B/yr after-tax via lower D&A, at scale).

Consensus FY2026: revenue $716.9B->$823.9B = +14.93% (+$107.0B); net income $77.67B->$94.74B = +21.98% (+$17.1B); EPS $7.17->$8.79 = +22.6%. My de-duplicated AI math: +$30B revenue (+4.2%) and ~+$8B NI (+10.3% EPS). Both fit inside consensus - 4.2% is ~28% of the 14.9% revenue add and 10.3% is ~47% of the 22% NI add - so the quantified AI is largely embedded in the AWS re-acceleration already in the numbers, not above it, for FY2026. The genuinely un-modeled upside is in the out-years: $364B backlog (+49% QoQ) + $225B Trainium + >$100B Anthropic point to AWS growth that could top the FY2027 +13% rev / +14% EPS consensus IF capacity executes. That optionality is offset by $200B CapEx margin risk. Net: medium priced-in - inline for FY2026 with asymmetric FY2027+ upside.

QUANTIFICATIONS
AWS AI revenue run rate: over $15 billion (nearly 260x the $58M AWS run rate at the same age) (first 3 years of the AI wave / current (Q1 FY2026), topline)
“In the first 3 years of this AI wave, AWS' AI revenue run rate is over $15 billion, nearly 260x larger.”
AI revenue growth rate: triple digits YoY (Q1 FY2026, topline)
“Our AI revenue is growing triple digits year-over-year.”
Bedrock customer spend growth: 170% QoQ (Q1 FY2026, topline)
“high-performance inference with the leading selection of frontier models in Bedrock, which saw 170% growth in customer spend quarter-over-quarter and processed more tokens in Q1 than all prior years combined.”
Trainium revenue commitments (backlog): over $225 billion (forward / multiyear (as of Q1 FY2026), topline)
“we now have over $225 billion in revenue commitments for Trainium.”
AWS total backlog: $364 billion (excludes the >$100B Anthropic deal) (Q1 FY2026, topline)
“the backlog for Q1 is $364 billion. That does not include the recent deal that we announced with Anthropic for over $100 billion.”
Custom silicon / chips business run rate: over $20B (~$50B standalone), growing triple-digit %, +40% QoQ (Q1 FY2026, topline)
“We saw nearly 40% quarter-over-quarter growth in Q1, and our annual revenue run rate is now over $20 billion and growing triple-digit percentages year-over-year... our annual revenue run rate would be $50 billion.”
Trainium CapEx savings + margin advantage: tens of billions of CapEx/year saved + several hundred bps operating-margin advantage (at scale / forward (Q1 FY2026), bottomline)
“at scale, we expect Trainium will save us tens of billions of dollars of CapEx each year and provide several hundred basis points of operating margin advantage versus relying on others' chips for inference.”
Trainium2 price-performance vs GPUs: ~30% better (40% in Q4 call) (current (Q1 FY2026), bottomline)
“Our Trainium2 chip has about 30% better price performance than comparable GPUs and is largely sold out.”
Trainium3 price-performance vs Trainium2: 30% to 40% better (current (Q1 FY2026), bottomline)
“Trainium3, which just started shipping at the start of 2026 and is 30% to 40% more price performance than Trainium2 is nearly fully subscribed.”
Trainium2 token throughput improvement: 4x (FY2025, bottomline)
“In 2025, we delivered 4x improvements in Trainium2's token throughput.”
SageMaker training-time reduction: up to 40% (current (Q1 FY2026), bottomline)
“model building with SageMaker, which reduces training time by up to 40%”
AWS Transform manual-effort savings: over 1.56 million hours (cumulative (Q1 FY2026), bottomline)
“Customers have used Transform to save over 1.56 million hours of manual effort when migrating and modernizing their workloads.”
Internal AI coding productivity: 5 people in 65 days vs ~40-50 people in ~1 year (current (Q1 FY2026), bottomline)
“normally, that would have taken 40 or 50 people about a year to do, and we took 5 really smart people... and those 5 people rebuilt it in 65 days.”
Rufus monthly active users / engagement: MAU up over 115%, engagement up nearly 400% YoY (Q1 FY2026, topline)
“Monthly active users are up over 115% and engagement is up nearly 400% year-over-year.”
Q Developer (QRO) usage: developers more than doubled QoQ, enterprise usage nearly 10x (Q1 FY2026, both)
“The number of developers using QRO more than doubled quarter-over-quarter and enterprise customer usage increased nearly 10x.”
AWS CapEx (AI-driven): $43.2 billion cash CapEx (Q1 FY2026, bottomline)
“Our cash CapEx was $43.2 billion in Q1. This primarily relates to AWS and generative AI as we invest to support strong customer demand.”
Bedrock customer spend growth (prior call): 60% QoQ (Q4 FY2025, topline)
“Bedrock is now a multibillion-dollar annualized run rate business, and customer spend grew 60% quarter over quarter.”
Trainium2 chips landed: over 1.4 million chips; 30-40% better price-performance than GPUs (Q4 FY2025, both)
“We've landed over 1.4 million Tranium two chips our fastest ramping chip launch ever. Tranium two is 30 to 40% more price performance than comparable GPUs.”
Curo (Kiro) developer usage growth: more than 150% QoQ (Q4 FY2025, both)
“the number of developers using Curo grew more than 150% quarter over quarter.”
Rufus reach / conversion lift: 300M+ users in 2025; ~60% more likely to complete a purchase (FY2025 (Q4 call), topline)
“more than 300 million customers used Rufus... Customers who used Rufus are about 60% more likely to complete a purchase”
Total CapEx guidance (AI/AWS): ~$200 billion (FY2026 (guided in Q4 FY2025), bottomline)
“We expect to invest about $200 billion in capital expenditures across Amazon.com, Inc., but predominantly in AWS, because we have very high demand.”
AWS backlog (prior call): $244 billion, up 40% YoY / 22% QoQ (Q4 FY2025, topline)
“our backlog is $244 billion. That's up 40% year over year. I think it's up 22% quarter over quarter.”
Internal AI applications: over 1,000 AI applications deployed or in build (Q4 FY2025, both)
“We have over a thousand AI applications that we've either deployed or in the process of building”
Power capacity added: 3.99 GW in 12 months; 1.2 GW in Q4; ~2x 2022 footprint (Q4 FY2025 (trailing 12 months), both)
“in the last twelve months, we added 3.99 gigawatts of power. Just for perspective, that's twice what we had in 2022... We added 1.2 gigawatts of power in Q4 just quarter over quarter.”
PAST (realized)
CURRENT (now)
FORWARD (guidance)
PRICED-IN (REFINED)
UNKNOWN

Est. revisions unknown  ·  Fwd P/E 36.2  ·  EV/Sales 3.9x

LLM synthesis failed: ask_json failed after 2 tries: claude exit 1:
COVERAGE — ENTHUSIASM TRAJECTORY + CATALYSTS
RECENT AI CATALYSTS & NEWS
OPTIONS / MARKET STRUCTURE

option liquidity: good

ATM IV
—
TYPICAL BID-ASK
—
OPEN INTEREST
—

proxy inputs — dollar-ADV $11.6B · beta 1.468 · px $258.68

source: proxy (no options chain on FMP)
FMP /stable/ exposes no options-chain endpoint on this key, so ATM IV, bid-ask spread and open interest are unavailable. Liquidity below is a PROXY from dollar-ADV, beta and price level (a stand-in for option depth), not measured option-market data.

CONFIRMATION — INSIDERS · 13F · LANGUAGE
Mixed — insiders selling, institutions adding, management language unknown.
INSIDERS selling 47 open-market sell(s) vs 0 buy(s) — net distribution
INSTITUTIONS (13F) adding as of 2026-03-31: 233 new / 339 closed positions; 3216 increased / 2314 reduced; institutional ownership -2.68pp; -88 net 13F holders
MGMT LANGUAGE n/a unknown language analysis failed: ask_json failed after 2 tries: claude exit 1:
VERBATIM AI QUOTES
“We've never seen a technology grow as rapidly as AI. Amazon is already a leader and companies continue to choose AWS for AI.”
— Andrew Jassy (CEO), Q1 FY2026
“To put our growth in perspective, 3 years after AWS launched, it had a $58 million revenue run rate. In the first 3 years of this AI wave, AWS' AI revenue run rate is over $15 billion, nearly 260x larger.”
— Andrew Jassy (CEO), Q1 FY2026
“Most of the value companies derive from AI will be through agents.”
— Andrew Jassy (CEO), Q1 FY2026
“For our custom AI silicon, we've recently shared very large multiyear, multi-gigawatt Trainium commitments from the 2 leading AI labs in the world in Anthropic and OpenAI as well as an increasing number of companies like Uber betting on Trainium. And we now have over $225 billion in revenue commitments for Trainium.”
— Andrew Jassy (CEO), Q1 FY2026
“As best as we can tell, our custom silicon business is now one of the top 3 data center chip businesses in the world, the speed at which we've gotten here is extraordinary.”
— Andrew Jassy (CEO), Q1 FY2026
“For perspective, at scale, we expect Trainium will save us tens of billions of dollars of CapEx each year and provide several hundred basis points of operating margin advantage versus relying on others' chips for inference.”
— Andrew Jassy (CEO), Q1 FY2026
“Nobody has a better set of chips across AI and CPU workloads than AWS with Trainium and Graviton, and we're unusually well positioned for this AI inflection we're in the early stages of experiencing.”
— Andrew Jassy (CEO), Q1 FY2026
“I do not see a place in any of our businesses or any of the ways that we do work where we're not going to have giant impact on what we do.”
— Andrew Jassy (CEO), Q1 FY2026
“normally, that would have taken 40 or 50 people about a year to do, and we took 5 really smart people, AI forward-thinking people building on agentic coding tools and those 5 people rebuilt it in 65 days. Like that is a very different world of operating.”
— Andrew Jassy (CEO), Q1 FY2026
“Our AI revenue is growing triple digits year-over-year. We're bringing more capacity online to meet high customer demand while also driving meaningful efficiency gains across our installed base.”
— Brian Olsavsky (CFO), Q1 FY2026
“We also see a strong correlation between AI spend and core growth. As customers spend more on AI, we see a corresponding demand increase in core.”
— Brian Olsavsky (CFO), Q1 FY2026
“the biggest reason that AWS continues to gain AI share is our uniquely broad top-to-bottom AI stack functionality.”
— Andrew Jassy (CEO), Q4 FY2025
“Bedrock is now a multibillion-dollar annualized run rate business, and customer spend grew 60% quarter over quarter.”
— Andrew Jassy (CEO), Q4 FY2025
“We've landed over 1.4 million Tranium two chips our fastest ramping chip launch ever. Tranium two is 30 to 40% more price performance than comparable GPUs.”
— Andrew Jassy (CEO), Q4 FY2025
“Looking ahead, the primary way companies will get value from AI is with agents.”
— Andrew Jassy (CEO), Q4 FY2025
“We expect to invest about $200 billion in capital expenditures across Amazon.com, Inc., but predominantly in AWS, because we have very high demand. Customers really want AWS for core and AI workloads. And we are monetizing capacity as fast as we can install it.”
— Andrew Jassy (CEO), Q4 FY2025
“I passionately believe that every customer experience that we know of today is going to be reinvented. With AI, there are gonna be a whole bunch of customer experience none of us ever imagined that are gonna become the norms of how we all operate every day.”
— Andrew Jassy (CEO), Q4 FY2025
ANALYST QUESTIONS ON AI
Q (Q1 FY2026, Eric Sheridan (Goldman Sachs)): What level of investment is needed over the next couple years to scale compute/capacity to meet the revenue backlog, and how does your unique custom-silicon/AI-infrastructure approach position you competitively to build that scale?
A: Jassy: AWS is growing 28% on a $150B run rate because customers choose AWS for AI (broad stack, inference near their data, strong security), and AI growth is also driving core growth via post-training, reinforcement learning and agentic tool usage. Having both Graviton (leading CPU) and Trainium (leading price-performance AI chip) leaves Amazon 'unusually well positioned.' No new capital update — plan largely the same — but views this as a 'once-in-a-lifetime opportunity' and expects to invest a significant amount of capital over the coming years.
Q (Q1 FY2026, Brian Nowak (Morgan Stanley)): What does the AWS backlog look like and how broad is it beyond the big labs? And what are your key milestones for Rufus and Agentic Commerce in 2026?
A: Jassy: Q1 backlog is $364 billion, which does not include the recent >$100B Anthropic deal, with 'reasonable breadth' (not just one or two customers). On Agentic Commerce, he's very bullish; Rufus MAU is up over 115% and engagement over 400% YoY. Third-party horizontal agents are still small (like early search-engine referrals) because they can't get pricing/product info right and lack personalization/shopping history; Amazon aims to make Rufus the best shopping assistant anywhere.
Q (Q1 FY2026, Justin Post (Bank of America)): How big an unlock is having the full suite of OpenAI models on Bedrock, and how focused are you on your own Nova model? Also, on selling racks of Trainium, what's the timing/opportunity given capacity constraints?
A: Jassy: Having all OpenAI models in Bedrock is 'a big deal' because customers want choice (already large AI usage on Anthropic, Llama, Mistral, etc.); the future is stateful APIs/agents, and the Bedrock managed agents built with OpenAI are something 'nobody else has.' On Trainium racks: demand is so high that Amazon must balance allocation, but 'there's a good chance we're going to sell racks over the next couple of years.'
Q (Q1 FY2026, Shweta Khajuria (Wolfe Research)): How are you thinking about rising memory/storage prices and supply-chain inflation's impact on CapEx this year and next? And how do you view the advertising opportunity in a world of Agentic Commerce where agents take the shopping action?
A: Jassy: Memory cost has 'skyrocketed' on insufficient capacity; Amazon secured significant supply early via strategic suppliers and isn't capacity-constrained, and the shortage is pushing on-prem workloads to the cloud since suppliers prioritize large cloud buyers. On ads, he believes advertising will do well in agentic commerce: agentic tools make ad creation far faster/cheaper (more advertisers), and multi-turn agentic conversations create multiple chances to surface organic and sponsored products, plus sponsored prompts.
Q (Q1 FY2026, Colin Sebastian (Baird)): How is incremental AI demand trending between early adopters/large AWS customers versus the broader enterprise base, and where do you see the most opportunity for AI internally across Amazon (product and operating efficiency)?
A: Jassy: AI labs are spending enormous amounts on compute (AI and core), but there's also significant enterprise adoption — largest absolute use is cost-avoidance/productivity (customer service automation, business process automation, fraud), with growing brand-new and reinvented experiences moving to production. Internally, he sees 'giant impact' everywhere; all customer experiences will be reinvented with new interfaces, and AI is already radically changing how Amazon codes, does DevOps, customer service, research, analytics and sales.
Q (Q4 FY2025, Mark Mahaney (Evercore ISI)): How will investors be able to see the strong long-term return on invested capital given this CapEx cycle — what duration, profitability levels, or minimum free-cash-flow guardrails should we expect?
A: Olsavsky: All capacity installed is immediately useful and being put into service with customers, plus a long arc of backlog/commitments (especially AI); AWS ran ~35% operating margin in Q4 (up 40 bps YoY) though margins will fluctuate with AI investment/depreciation, offset by efficiencies. Jassy: Most spend is AI (and faster-than-expected core), it's monetized as fast as installed; Trainium underneath the majority of Bedrock gives both better customer prices and better Amazon economics, so he is 'very confident we're gonna have strong return on invested capital here.'
Q (Q4 FY2025, Douglas Anmuth (JPMorgan)): How is Project Rainier with Anthropic running after a full quarter (500k vs the prior 1M chip figure), and are there financial guardrails/governors around spend in terms of operating-income growth or positive free cash flow?
A: Jassy: Very excited about Trainium growth; customers are thirsty for better price-performance and Trainium2 is 30-40% better than comparable GPUs. Anthropic is training the next Claude on Trainium2 (Project Rainier, ~500k chips, increasing) plus other workloads; Trainium is a fully-subscribed multibillion-dollar run rate. Trainium3 (40% better than T2) expected nearly fully committed by mid-2026, with substantial interest in Trainium4 (2027) and early Trainium5 conversations. He reframed rather than gave guardrails — calling it an 'extraordinarily unusual opportunity' Amazon will 'invest aggressively' in.
Q (Q4 FY2025, Ross Sandler (Barclays)): How is the AI market — currently top-heavy around a few AI-native labs — changing into 2026, and how might you extend the OpenAI relationship to help Amazon's AI efforts on both the retail and AWS sides?
A: Jassy: Describes a 'barbelled' demand market — AI labs (plus a couple runaway apps) spending heavily on one end, enterprises doing productivity/cost-avoidance on the other, with enterprise production workloads in the middle that 'very well may end up being the largest and most durable.' The bulk of demand is still to come as AI talent grows and inference gets cheaper (Trainium's role). On OpenAI: excited about the November agreement ('a big one'), hopes to extend it, but stresses the AI movement will involve thousands of companies, not a couple.
Q (Q4 FY2025, Michael Morton (MoffettNathanson)): How does the agentic shift play out for the retail business and the on-site ads portion, given a possible compression of the funnel as consumers get better answers over time?
A: Jassy: Very optimistic about the agentic shopping experience; it's why Amazon invested in Rufus (300M users in 2025, and Rufus users are ~60% more likely to complete a purchase). Third-party horizontal agents lack shopping history and get product details/pricing wrong, so a value exchange must be found, but he believes many customers will choose a retailer's own great shopping agent because retailers better deliver broad selection, low prices, fast delivery and trust.
Q (Q4 FY2025, Eric Sheridan (Goldman Sachs)): What is the AWS revenue backlog as of Q4, and what's the supply/demand imbalance for internal vs external AI use cases, and how do you close that gap as more capacity comes online in 2026?
A: Jassy: Backlog is $244 billion, up 40% YoY (22% QoQ), with a lot of demand for both AI and core. The vast majority of capacity is consumed by external customers; internally Amazon has over 1,000 AI applications deployed or in build (Rufus, Alexa+, fulfillment forecasting, customer-service chatbots, ads tools, Thursday Night Football defensive alerts). On supply/demand, every provider 'could actually grow faster if we had all the supply' — Amazon added 3.99 GW of power in the last 12 months (twice its entire 2022 footprint) and 1.2 GW in Q4 alone, and expects to double again by 2027.