Live: Will Amazon Move Higher After Q3 Earnings After the Bell?

Justin Post (BofA)

Question (2 parts):

1. How constrained are you on AWS capacity? Can you actually meet AI demand?

2. Outside your biggest AI customers, what’s demand like for Trainium (your custom AI chips)?

Answer (Jassy):

1. Amazon added 3.8 GW of power in the last 12 months, expects another >1 GW in Q4, and plans to double AWS power capacity again by 2027. Power is becoming the main bottleneck for the industry, but Amazon is aggressively adding and immediately monetizing capacity.

2. Trainium2 is “fully subscribed,” already a multibillion-dollar business, and grew 150% q/q. Today it’s mostly very large customers (example: Anthropic training Claude on ~500,000 Trainium2 chips, going to ~1 million). Trainium’s advantage is 30–40% better price/performance vs alternatives, which matters as AI moves from experiments to production-scale inference. Trainium3 previews end of ’25 with broader volume in early ’26, and demand interest is not just mega-customers but also “medium-sized” customers.

Takeaway:
AWS capacity ramp is huge and Jassy is saying it’s already being absorbed. Trainium is positioned as cost-per-token weaponry, not just a science project.

Brian Nowak (Morgan Stanley)

Q:
Strategic chip question. For Trainium3 to scale broadly, what hurdles do you still need to clear vs just buying 3rd-party GPUs (like NVIDIA)? Do you risk cannibalizing NVIDIA or do you still need both?

Answer (Jassy):
Amazon will always offer multiple chip options. They’ll keep buying “a lot of NVIDIA” and expect to buy even more in the future. But Amazon is different because it also designs its own chips (Annapurna team):

• Graviton (CPU) already delivers ~40% better price/perf than x86 peers.

• Trainium aims to deliver similar gains vs other GPUs.
  Trainium3 is expected to be ~40% better than Trainium2.
  To win broad adoption, Amazon needs:

• keep delivering new Trainium generations fast and at volume;

• keep maturing the software ecosystem;

• keep building credibility with big real-world proofs like Anthropic/Rainier.

Takeaway:
They’re not positioning it as “replace NVIDIA,” they’re positioning it as “we will be the price/performance leader at scale, and we’ll give you both.”

Doug Anmuth (JPMorgan)

Q:
Explain Project Rainier. What is it architecturally? Why does Anthropic care? Will Rainier expand beyond Anthropic, especially with Trainium3?

Answer (Jassy):
Rainier = enormous Trainium2-based training cluster (500k → 1M chips) built for Anthropic. The differentiation is:

• Infrastructure at that scale is hard (networking, memory fabric, reliability, etc.). AWS can actually stand up million-chip clusters.

• Trainium2’s cost/performance profile.
  He says Rainier is currently Anthropic-specific, but other customers want similar large-scale Trainium clusters, and Trainium3 will broaden that.

Takeaway:
Amazon can now credibly say: “If you’re an AI lab or advanced enterprise and you want frontier-scale training at better economics, we can do that for you, not just for Anthropic.”

Mark Mahaney (Evercore ISI)

Q (2 parts):

1. Grocery: You’ve called same-day perishable delivery a “game changer.” Are you now at true grocery scale without needing a ton more physical Amazon Fresh stores? Have you changed shopper behavior?

2. Headcount: You’ve reduced/flattened. Is that AI-enabled efficiency? Will headcount stay flat?

Answer (Jassy):

1. Grocery:

   • If you combine consumables/household goods plus Whole Foods plus new “daily shop” concept, Amazon is already effectively a top-3 U.S. grocer by gross merchandise sales (> $100B GMS just in the last 12 months for core grocery-like categories).

   • The real unlock is perishables with same-day delivery. Customers can add milk/eggs/etc. to the same cart as other Amazon items and get it in hours.

   • They piloted it in a few markets, adoption was “taken aback”-level strong, and they’ve now rolled it to 1,000 cities and plan ~2,300 by year-end.

   • This is changing frequency (people come back more often, visit the site more often) and could reshape the “weekly stock-up” model. They’ll still experiment with physical formats (Whole Foods, “daily shop”), but the perishable delivery model is the big unlock.

2. Headcount:

   • Recent org changes weren’t mainly about near-term cost or AI job elimination.

   • It’s cultural: Amazon had added layers as it scaled; layers slow decisions.

   • Jassy says the company must run “like the world’s largest start-up”: lean, flat, fast decision-making at the edge, especially now with AI/tech shifting fast.

   • Translation: they’re structurally compressing org charts to move faster.

Takeaway:
Groceries is not just “nice adjacency” anymore; it’s being treated as a core recurring wallet-capture. And headcount discipline is being pitched as speed/ownership, not austerity.

Eric Sheridan (Goldman Sachs)

Q:
Talk about robotics/automation and physical AI across the network. Is this mainly cost takeout, or is it also an enabler to reinvest?

Answer (Jassy):
Amazon already runs >1 million robots in fulfillment. Robotics:

• improves safety,

• boosts productivity,

• increases speed,

• lowers cost to serve.
  He emphasizes a “robots + humans together” model — humans stay at the center doing problem-solving while robots handle repetitive/unsafe motion.
  And yes, long term this supports both better CX (faster, safer, more reliable) and cost efficiency that can be reinvested.

Takeaway:
He’s basically saying: robotics is not replacing the labor model, it’s scaling it. Also implies structural margin tailwind in fulfillment/logistics.

John Blackledge (TD Cowen)

Q:
How do you think about “agentic commerce”? i.e. AI agents buying things for customers, discovering products, etc. How will Amazon play with first-party and third-party agents?

Answer (Jassy):

• He thinks agentic commerce is a huge long-term unlock because it solves the “I don’t know what I want yet, help me narrow” problem — an area where physical retail + a salesperson still has an edge.

• AI agents can guide you through discovery, ask/answer questions, refine preferences in real time.

• Amazon is already doing this with Rufus (AI shopping assistant) and “Buy for Me” (Amazon will even go buy from other merchants on your behalf).

• Over time, Amazon expects to also partner with third-party agents (like how search engines were discovery gateways in early e-commerce).
  BUT: today’s third-party agents have problems (no personalization, wrong prices, bad delivery estimates). Amazon will only partner in ways that fix that and protect CX.

• Long-term, he expects AI agents to increase total e-commerce penetration and believes Amazon wins in that world because of selection, speed, and price.

Takeaway:
He’s positioning Amazon as the default transaction layer in an agent-driven world — even when the agent wasn’t built by Amazon.

Colin Sebastian (Baird)

Q (2 parts):

1. AWS: How much of the AWS acceleration is “core cloud workloads coming back” vs. GenAI workloads? How important are new services like AgentCore in attracting enterprises?

2. Ads: Can you break down what’s driving ad acceleration — core retail ads vs. DSP vs. Prime Video?

Answer (Jassy):
On AWS:

• Growth is broad: AI training, inference, Bedrock (foundation model hosting/inference), SageMaker (build/train your own models), Trainium, and traditional enterprise migrations back into motion.

• He says companies are now serious about “agents,” and that most of the real AI value long term will come from agents automating work.

• Building/operating agents has been too hard, especially at enterprise-grade security/scale.

• That’s why Amazon launched:

  • Strands (to help developers build agents from any model), and

  • AgentCore (infrastructure building blocks: security, runtime, observability, memory, etc., to run agents at scale).

• Customer reaction is strong because there’s “nothing else like it.”

• He explicitly says he believes AWS can “continue to grow at a clip like this for a while.”

On Ads:

• Every ad surface grew.

• Amazon now has a true “full funnel”:

  • top-of-funnel awareness via Prime Video / live sports (NBA on Prime, NFL-style innovation, PGA skins, The Masters rights coming, etc.),

  • down to bottom-of-funnel sponsored product ads at point-of-sale.

• Measurement/targeting are strong because Amazon controls both audience data and commerce conversion.

• Three specific growth engines:

  1. Retail media on Amazon’s core stores — still lots of runway because most global retail is still offline (80–85%), and AI-driven experiences should pull more spend online.

  2. Prime Video & live sports ads — early but already “very large,” with strong upfront demand.

  3. Amazon DSP — now more “fully featured,” plus partnerships (Roku → huge U.S. CTV footprint; Netflix, Spotify, SiriusXM inventory access).

• Net: Ads is scaling across commerce, streaming, live sports, and third-party inventory.

Takeaway:
He’s telling you two secular stories:

• AWS: “AI agents are the next cloud wave, and we own the primitives to build/run them.”

• Ads: “We’re quietly building the next giant cross-channel ad network, not just search-style sponsored listings.”

TL;DR for investors

• AWS: 20%+ growth on $132B run rate, record backlog, doubling power capacity by 2027, Trainium is fully subscribed and scaling to Trainium3 early ’26. Management thinks this pace is sustainable “for a while.”

• AI strategy: Not just GPUs. It’s: (1) custom silicon for price/perf; (2) massive training clusters for frontier model builders; (3) agent platform (Strands/AgentCore); (4) Private cloud capacity expansion at hyperscale.

• Retail core: Grocery same-day perishables in 1,000 cities now, ~2,300 by year-end, rewires weekly spend behavior. Robotics + network optimization improve cost to serve.

• Ads: 22% growth at $17B+ revenue, now spanning commerce, Prime Video live sports, and external DSP inventory — and they’re saying it’s early.

• Culture / Op model: Flattening orgs and cutting layers so they can move like a start-up in an AI arms race.

• Capex: ~$125B cash capex in 2025, increasing in 2026.