AWS Reaccelerates Off the Capacity Wall — +20.2% Print, $200B Backlog, Trainium2 Fully Subscribed, and the Print That Retires the Bear Case

Results vs. Consensus

Quality of Beat/Miss

• The AWS print is the highest-quality data point of the year: +20.2% on $33B run rate is mathematically harder than +33% on a $20B run rate (the Azure comp). Backlog of $200B confirms it isn't a one-off pull-forward.
• EPS quality is deceptively muddied: the headline $1.95 is helped by the $9.5B Anthropic non-operating gain (~$5.50/sh of below-the-line tailwind). Underlying operating EPS is actually depressed by the $4.3B in special charges. The cleanest operating EPS reads through is roughly $1.85–$1.90 ex-Anthropic, ex-charges.
• Op-income beat ex-charges is broad-based: NA at 6.9% ex-FTC, AWS margin recovering to 34.6% from 32.9%, International margin expanding YoY ex-severance.
• Advertising acceleration to a third consecutive quarter is the structural read — +18% Q1, +22% Q2, +22% Q3 with breadth of CTV partnership wins (Roku, Disney, Netflix, Spotify, SiriusXM). The line is now a $70B+ annual run rate compounding at +20%+.

Segment Performance

AWS — The +20% Print and What It Means

AWS revenue of $33.0B at +20.2% YoY is the headline of the entire print. The 270-bps sequential acceleration from Q2's +17.5% is the clearest possible confirmation of the supply-side thesis we held through Q1 and Q2: capacity was the binding constraint, demand was real, and as 3.8 GW of new power capacity landed (more than any cloud competitor over the same period), revenue followed. Backlog crossed $200B at quarter end (vs. $195B Q2) — and Jassy added that "October new deals exceed total Q3 deal volume" — meaning Q4-end backlog is positioned to step up materially again.

"AWS is growing at a pace we haven't seen since 2022, reaccelerating to 20.2% year-over-year, our largest growth rate in 11 quarters. As fast as we're bringing in [capacity] right now, we are monetizing it." — Andy Jassy, CEO

Sequentially AWS added $2.1B of revenue and the run rate is now $132B. AWS operating income of $11.4B at 34.6% margin recovered from Q2's compressed 32.9%, even with continued depreciation pressure from new data-center placement-in-service. Critically, Jassy's framing of the bottleneck shifted: "overall in the industry, maybe the bottleneck is power. I think at some point, it may move to chips, but we're bringing in quite a bit of capacity." Translation: the binding chip-supply argument from Q2 has been substantially relieved; AWS now has the chips and is bringing power online faster than competitors.

The Trainium2 disclosure is the AI moat embedding itself in the financials. Fully subscribed. Multibillion-dollar revenue. Up 150% QoQ. Project Rainier — the 500,000-Trainium2-chip cluster — is now powering Anthropic's next Claude generation, scaling to 1M chips by year-end. "The majority of token usage in Bedrock is already running on Trainium." Trainium3 previews end of 2025 with full volumes early 2026 and a claimed ~40% additional price/performance improvement over Trainium2.

Assessment: Three things changed this quarter. (1) The supply-side thesis on AWS converted from hypothesis to proven. (2) Trainium2 stopped being optionality and became a multibillion-dollar revenue line with the highest growth rate in the company. (3) Backlog leading indicators (Q3 close + October new deals) point to acceleration continuing into Q4. The bear case on AWS losing share to Azure / GCP is, for now, retired.

North America — FTC Charge Optically Hides 6.9% Margin

NA revenue of $106.3B was +11% YoY. Reported segment operating income of $4.8B at 4.5% margin understates the underlying profile because of the $2.5B FTC settlement charge (settled this quarter) and a portion of the $1.8B severance charge. Excluding the FTC charge alone, NA op income would have been $7.3B at a 6.9% margin — modestly below Q2's record 7.5% but still a strong year-over-year expansion print. CFO Olsavsky disclosed a "reduction of U.S. inbound lead time by nearly 4 days compared to last year," a structural working-capital improvement that compounds into faster delivery and better in-stock rates.

Same-day perishables is now in 1,000+ U.S. cities, expanding to 2,300 by year-end. The 2× return-rate of perishable-shoppers vs. non-perishable is the most under-appreciated KPI in the company — it directly attacks the grocery-trip frequency moat that Walmart and Costco have historically held. "Add to delivery" button: 80M+ uses since launch. 3-hour delivery rolling out in select U.S. cities. Rural same-day/next-day expansion +60%. 1.3M+ sellers using gen AI listing tools.

Rufus, the AI shopping assistant, hit 250M active customers in 2025 with monthly users +140% YoY and interactions +210% YoY. Critically: customers using Rufus during a shopping trip are "60% more likely to complete a purchase," and Rufus is on track to deliver "$10B+ incremental annualized sales." If accurate, that single feature is the largest measurable AI revenue contribution at any U.S. retailer.

Assessment: The FTC settlement is one-time and now in the rearview mirror. Underlying operating-margin profile is intact. Perishables + Rufus + inbound lead-time compression are three distinct compounding margin drivers that are still in early innings.

International — Quietly Compounding

International revenue of $40.9B was +10% ex-FX with operating income of $1.2B at a 2.9% margin. Olsavsky noted that ex-severance, International margin actually expanded YoY — extending the multi-quarter pattern that was at +320bps in Q2. Established markets (UK / Germany / Japan) approaching U.S. profile; emerging markets (8 countries launched in 5 years) at varying maturity stages, all improving sequentially.

Assessment: No incremental disclosure but the trajectory is intact. International earned its now-decade-long operational respect this year.

Key KPIs

Key Topics & Management Commentary

Overall Management Tone: The most confident and most specific Jassy has sounded in any earnings call we have analyzed. Where Q1 was hopeful and Q2 was guarded, Q3 was triumphant — backed by data points the market has been waiting two quarters for. The headcount-reduction framing ("not financially driven, not AI-driven, culturally driven") is unusually candid. Olsavsky's tone shifted from defensive ("we expect AWS margins to fluctuate") to assertive ("strong returns on invested capital over the long term").

The +20.2% Print — What Changed

The single biggest narrative shift this quarter is the bottleneck reframing. In Q2, Jassy said "the single biggest constraint is power, you also see constraints off and on with chips." In Q3, after 3.8 GW of capacity added in 12 months: "we're bringing in quite a bit of capacity today, overall in the industry, maybe the bottleneck is power. I think at some point, it may move to chips." Translation: chips are no longer the binding constraint at AWS — power is the industry-wide constraint, and AWS has been adding more of it than anyone else (3.8 GW vs. competitor disclosures running at half that). Reacceleration is mechanical: capacity landing → demand monetizing.

"We're bringing in quite a bit of capacity today, overall in the industry, maybe the bottleneck is power. I think at some point, it may move to chips, but we're bringing in quite a bit of capacity. And as fast as we're bringing in right now, we are monetizing it." — Andy Jassy, CEO

Assessment: This is the print where the AWS narrative changes from "watch and wait" to "watch the math." Each new GW of power on a $132B run rate at ~35% margin and triple-digit-growth AI workload mix is incremental high-quality revenue.

Trainium2: From Optionality to Moat

Trainium2 is now a multibillion-dollar AWS revenue line that grew +150% QoQ. Project Rainier — Amazon's massive AI compute cluster spanning multiple U.S. data centers — is built on nearly 500K Trainium2 chips today, scaling to 1M by year-end, and is the platform on which Anthropic is training the next Claude generation. Most consequentially: "the majority of token usage in Bedrock is already running on Trainium." That sentence is the AWS AI gross-margin moat in one line — Bedrock is now the primary inference fabric for AWS AI customers, and the chip stack underneath it is custom silicon Amazon designs and owns.

"It's not simple to be able to build a cluster that has 500,000 plus chips going to 1 million. That's an infrastructure feat that's hard to do at scale." — Andy Jassy, CEO

Trainium3 previews end of 2025; full volumes early 2026; ~40% better price/performance than Trainium2 (which was already 30–40% better than NVIDIA GPUs). Customer base today is "small number of very large customers" — Trainium3 is designed to broaden adoption to mid-market and enterprise.

Assessment: The Trainium / Bedrock combination is now the structural AWS AI moat. NVIDIA partnership remains "as far as I can foresee" and they are buying "very significant amounts" — but the custom-silicon path is monetizing fast enough to materially compound AWS gross margins over the next 18–24 months.

Capex at $125B FY25, Higher in 2026 — And the Math Now Works

Olsavsky formally guided FY25 cash capex to ~$125B (vs. our Q2 model's $120–125B) and stated 2026 will be higher. Q3 alone was $34.2B; YTD $89.9B. The 2026 step-up is consistent with the doubling-by-2027 power capacity commitment. What's different about this disclosure vs. the Q2 capex disclosure: by Q3 the revenue conversion is now visible. AWS sequential revenue +$2.1B, capacity adds 3.8 GW TTM with 1+ GW more in Q4 — capex is no longer "hope and trust"; it is "invest and convert."

Assessment: The capex-FCF compression we flagged in Q2 still holds (TTM FCF $14.8B, down from $18.2B). But the reacceleration print confirms the LTV per dollar of capex is real. We expect FY25 FCF to print $20–30B — well below historical pattern but with backlog of $200B+ as the offset.

Advertising — Third Consecutive Acceleration, Now a CTV Aggregator

Q3 ad revenue $17.7B, +22% YoY — the third consecutive quarter of accelerating growth. Layered onto Q2's Roku and Disney DSP wins: Netflix premium ad inventory direct programmatic access, Spotify (400M+ monthly ad-supported listeners), SiriusXM (160M monthly digital listeners across Pandora + SoundCloud). The 2025–26 upfront "exceeded our own expectations for upfront commitments." Live sports buying is materially up (NBA debut on Prime Video drew 1.25M U.S. viewers per game on opening doubleheader — double-digit increase vs. prior season on cable).

Assessment: Amazon DSP is now the de-facto CTV aggregator alongside The Trade Desk — but with the Prime Video / Twitch / NFL / NBA / NASCAR / The Masters owned-and-operated supply layered in. This is a structurally widening competitive advantage vs. Google and Meta in CTV.

FTC Settlement and Severance — One-Time and Forward-Looking

The $2.5B FTC settlement closes a multi-year overhang. The $1.8B severance reflects role eliminations Jassy explicitly framed as culturally rather than financially driven: "we are committed to operating like the world's largest start-up. And that means removing layers." The severance impacts all three segments via T&I, S&M, and G&A expense lines.

Assessment: Both charges are one-time. The FTC settlement removes a tail risk that has been outstanding since 2022. The severance creates 2026 OpEx tailwinds.

Same-Day Perishables — A Grocery Habit Disruptor

Same-day perishables is in 1,000+ U.S. cities, expanding to 2,300 by year-end. CFO disclosure: "when customers start shopping groceries on Amazon, they are visiting the site more often and returning twice as often as non-perishable shoppers." Jassy reframed Amazon's grocery position: $100B+ annual GMV ex-Whole Foods/Fresh — already a top-3 U.S. grocer. "Daily Shop" small-format urban Whole Foods (3 launched, off to good starts). Combined with the Add-to-Delivery feature (80M+ uses since launch), the Amazon grocery flywheel is structurally accelerating.

Assessment: The grocery perishables move is the single most under-modeled long-term growth driver. If frequency-of-visit doubles, the entire NA economic engine moves up a tier.

Guidance & Outlook

Q4 op-income guide of $21–26B (midpoint $23.5B) is above pre-print Street ~$23.0B and represents +13% YoY at the midpoint. The $5B range is consistent with Q2/Q3 width but starts from a higher base. Revenue guide $206–213B is comfortably above Street; AWS reacceleration is implicit but not explicit.

Implied Q4 op-income trajectory: midpoint $23.5B is +13% YoY. High end $26B is +25%. Low end $21B is +1%. We expect AMZN to print at or above the high end given holiday strength and AWS acceleration.

Street at: ~$23.0B pre-print; will recalibrate to ~$24.0B post-guide.

Guidance style: Confident-with-flexibility. The midpoint is above Street, which is a tonal change from Q1/Q2's midpoint-below-Street pattern.

Analyst Q&A Highlights

AWS Capacity and Trainium Adoption

• Justin Post, Bank of America: Asked about capacity levels and Trainium3 demand. Jassy: 3.8 GW added in 12 months, 1+ GW more in Q4, 2× by 2027. Bottleneck is power industry-wide. Trainium2 fully subscribed, multibillion-$ run rate, +150% QoQ. Trainium3 previewing end of year; "lot of customers, both very large and medium-sized" interested. Assessment: Most data-rich answer of the year. Confirms supply-bottleneck thesis was correct and that Trainium adoption is broadening.
• Brian Nowak, Morgan Stanley: Asked about Trainium3 hurdles and adoption beyond Anthropic. Jassy: 40% better than Trainium2 expected; need to deliver chip + volume + quickly + software ecosystem; Project Rainier proof point with Anthropic builds credibility. Will always have multiple chip options (deep NVIDIA partnership continues). Assessment: Honest answer on what still needs to be proven. The "we will always offer multiple chips" framing is consistent.
• Doug Anmuth, JPMorgan: Asked about Project Rainier architecture and replicating it with Trainium3. Jassy: 500K → 1M chip clusters are "an infrastructure feat that's hard to do at scale" — that infrastructure capability is itself an AWS moat. Other customers interested in large-cluster Trainium for Trainium3. Assessment: The cluster-scale infrastructure capability is itself a competitive advantage we had under-weighted.

Grocery and Headcount

• Mark Mahaney, Evercore: Asked about grocery (whether Amazon needs Fresh stores), and headcount approach. Jassy: $100B+ ex-Whole Foods grocery; perishables same-day "changing the trajectory and size of our grocery business"; weekly stock-up tradition is changing. Headcount: "not really financially driven and it's not even really AI-driven, not right now, at least" — culture-driven (removing layers, ownership). Assessment: The grocery framing is more bullish than any prior quarter. The headcount answer is unusually candid.

Q4 Guide and 2026 Capex

• Multiple analysts asked variants of the 2026 capex question. Olsavsky's only commitment: "we expect [FY25 cash capex] amount will increase in 2026." No specific number. Assessment: 2026 capex of $145–160B is plausible; market will model upward.

What They're NOT Saying

1. 2026 capex specific number: Acknowledged it will increase; no number. Reinvent (December) is the likely venue for clearer framing.
2. AWS Q4 growth rate target: Guide is at the segment level; AWS-specific not committed. Backlog leading indicator suggests +21–22% is achievable.
3. Trainium3 customer pipeline depth: "lots of large and medium customers interested" — but no announced design wins. Reinvent will likely fill this.
4. Anthropic equity-stake updated valuation framework: $9.5B Q3 mark — the implied valuation step-up is not broken out.
5. Tariff cost impact materializing: Q3 commentary minimal. Either tariff impact is genuinely benign or being absorbed across multiple lines such that it's not isolatable.
6. Project Kuiper subscriber metrics: 150+ satellites, >1 Gbps speeds, but no commercial subscriber number.
7. FTC settlement non-financial terms: Settlement amount disclosed; structural changes to Amazon's marketplace practices not detailed.

Market Reaction

• After-hours move (Oct 30): AMZN traded up ~13% in extended hours on the AWS reacceleration print. Jump in market cap exceeded $300B.
• Next-day (Oct 31) early indications: Stock holding gains; multiple sell-side desks moving up estimates; capex-question debate continues but is being out-weighed by AWS print.
• Volume: Highest single-session AMZN volume in 6+ months.
• Analyst reactions (within first 12 hours):
  • Most desks raised PTs by 10–18%, some upgraded to a buy-equivalent rating
  • Bull camp: AWS reacceleration confirms supply thesis; Trainium2 economics; backlog growth into Q4; ad CTV moat widens
  • Bear camp (more muted): 2026 capex step-up is the unmodeled risk; FTC structural remedies still tail risk; AWS margin compression from continued data-center placement-in-service

The +13% reaction is a clean read: the AWS reacceleration print is the catalyst the market has been waiting for since Q1, and the magnitude of the response reflects two quarters of pent-up positioning waiting for confirmation. We expect the move to hold.

Street Perspective

Debate: AWS Reacceleration — Sustainable or One-Off?

Bull view: $200B backlog at quarter-end with October new deals exceeding total Q3 volume; Trainium2 multibillion-dollar +150% QoQ; Project Rainier scaling to 1M chips; 1+ GW added in Q4 alone. Q4 prints +21–22%; FY26 base case is +21–23%.

Bear view: +20% on a $132B run rate is impressive but Azure prints +33%, GCP +28%. AWS share-of-AI may still be eroding even as growth reaccelerates. The 2026 capex step-up implies AWS needs to grow faster still to sustain return-on-capital.

Our take: Backlog accumulation + Trainium2 traction strongly support the bull view. Relative-growth-vs-Azure is base-rate distorted; absolute-dollar-growth in AWS is the largest in the industry. Bull case wins.

Debate: Trainium2 / Bedrock as Structural Moat

Bull view: Bedrock token usage majority on Trainium = AWS captures both compute revenue and silicon margin. Anthropic anchor + Project Rainier scale + Trainium3 in pipeline creates a self-reinforcing flywheel. Differentiated capability NVIDIA-only competitors don't have.

Bear view: Custom silicon hardware is a multi-year bet; Trainium is one product cycle ahead but NVIDIA, AMD, and Microsoft Maia / Google TPU are all closing. Cluster-scale infrastructure (Rainier) is replicable in 2–3 years.

Our take: The "majority of Bedrock token usage on Trainium" disclosure is a data-rich confirmation that the moat is forming faster than the bear case admits. Trainium3 ramping in 1H 2026 widens it.

Debate: 2026 Capex — Headwind or Tailwind?

Bull view: Capex matches backlog accumulation rate. Each $1B of capex is converting to ~$300M of incremental annualized AWS revenue at ~35% margins. ROIC math works.

Bear view: 2026 capex of $145–160B+ on a $50–80B AWS LTM operating income base implies further FCF compression. Even rational investment can drag the multiple if it persists.

Our take: The reacceleration print resolves much of the bear's concern. The capex math now visibly works. Multiple holds.

Model Update Needed

Valuation impact: Net target moves materially higher — AWS reacceleration plus Trainium2 monetization compounds into FY26/27 model. Even after a +13% next-day pop, our fair-value framework still implies upside vs. spot.

Thesis Scorecard Post-Earnings

Overall: Thesis decisively strengthened. Two bears retired (AWS-share, FTC); two bulls confirmed with new data (AWS reaccel, Trainium2 moat).

Action: Upgrade to Outperform. Expect outperformance over the next 12 months even after the post-print pop. Reinvent in December and Q4 print are the next catalysts.