Amazon’s AI Capex Super‑Cycle Meets a Faster, Cheaper Retail Engine

AWS moves from cloud leader to capital‑intensive infrastructure giant

Amazon entered 2026 sounding more like a global utilities provider than a traditional internet retailer. Andy Jassy used his prepared remarks to frame the company around a single axis: a vast, capex‑hungry AI and cloud infrastructure build‑out that is already reshaping both the P&L and the strategic story.

AWS revenue rose 24% year on year to $35.6bn in the fourth quarter, its fastest growth in thirteen quarters, pushing the unit to a $142bn annualized run‑rate. Operating income reached $12.5bn, implying a roughly 35% margin even as depreciation from new data centers and chips begins to bite. Jassy stressed that at this scale, 24% growth represents more absolute incremental revenue and capacity than any rival: AWS added $2.6bn of revenue quarter on quarter and, by Brian Olsavsky’s count, more data‑center capacity in 2025 than any other company.

Beneath those headline numbers lies the true pivot: Amazon plans to invest roughly $200bn in capital expenditures “across Amazon.com, Inc., but predominantly in AWS.” The bulk of that spend is being driven by AI infrastructure: custom silicon, networking gear, and power‑hungry data centers. Jassy was blunt that “as fast as we install this capacity, we are monetizing it,” calling the current window an “extraordinarily unusual opportunity to forever change the size of AWS and Amazon.com, Inc. as a whole.”

Investors anxious about returns were offered a few concrete anchor points. AWS’s contractual backlog has swelled to $244bn, up 40% year on year and 22% quarter on quarter. AI‑related services are already at multibillion‑dollar run‑rates: custom chips (Graviton CPUs and Trainium AI accelerators) together exceed $10bn, growing at triple‑digit rates in Trainium’s case, while Amazon Bedrock — the company’s managed foundation‑model platform — is now a multibillion‑dollar business with customer spend up 60% quarter on quarter.

Yet the build‑out will not be painless. AWS margins, at 35%, already incorporate a “headwind from the investments in AI and the depreciation on that capex,” Olsavsky warned, and he reiterated that margins will fluctuate as the cycle progresses. The message was that any compression should be read as the upfront cost of long‑lived infrastructure, mitigated over time by Amazon’s homegrown silicon and networking, which Jassy argues will produce structurally better price‑performance — and thus, better economics — than pure GPU deployments.

Building a full‑stack AI platform: from chips to agents

If 2023 was the year AWS talked about models, 2025 was the year it tried to own the entire AI stack.

At the hardware level, Amazon’s in‑house chips are now central to the narrative. More than 90% of AWS’s top 1,000 customers are using Graviton CPUs, which Jassy said deliver up to 40% better price‑performance than leading x86 processors and account for a multibillion‑dollar run‑rate business growing over 50% annually. On the AI side, Trainium has become the workhorse for generative workloads. Over 1.4m second‑generation Trainium2 chips have been deployed in what Jassy called AWS’s fastest‑ramping chip launch ever, with 30–40% better price‑performance than comparable GPUs and “fully subscribed” demand.

Trainium3, only just launched and promising up to another 40% price‑performance gain versus Trainium2, is already seeing such strong orders that Amazon expects “nearly all” of 2026 supply to be committed by mid‑year. Trainium4 is in development for 2027 and Jassy said the company is “already having conversations” about a fifth generation.

On top of that silicon, Amazon is trying to differentiate with a broad, and increasingly opinionated, AI software stack. Bedrock, which aggregates third‑party and proprietary models, now supports frontier labs such as Anthropic, open‑source options like Mistral and Llama, and Amazon’s own Nova family. The new NovaForge service goes a step further, allowing enterprises to blend their proprietary data into early model checkpoints at the pre‑training stage to create custom “Novas” — or “novellas,” as Jassy dubbed them. He pitched this as akin to teaching a child a foreign language early; the data becomes part of the model’s learning foundation, potentially a structural advantage for complex, domain‑specific agents.

The strategic bet is that the real money in AI will pool around agents rather than standalone chatbots. Jassy sketched a roadmap where enterprises deploy both interactive agents and fully autonomous ones that run for hours or days, scaling out and retaining context. To reduce the friction of building and governing such systems, AWS has rolled out AgentCore, which handles secure integration with compute, data, identity, governance, and monitoring. Early uptake is strong enough that Jassy described it as “unlocking deployments.”

AWS is also dogfooding its own platform with a stable of vertical agents: Curo for coding (developer usage up more than 150% quarter on quarter), Quick for knowledge workers, AWS Transform for legacy software migration, and Amazon Connect agents for call centers. At re:Invent, it layered in “Frontier Agents” for fully autonomous coding, DevOps, and security tasks.

Across this architecture, Amazon’s thesis is that customers will use multiple models and many agents, and that the vendor best able to offer choice at the model layer while tightly integrating chips, networking, and governance will capture disproportionate share. That argument is underpinned by Jassy’s view of demand: an emerging “barbell” where AI labs and a handful of runaway apps are consuming “gobs and gobs of compute” at one end, while at the other, enterprises are starting to get real value from productivity and cost‑avoidance use cases such as customer service automation and fraud detection. The thick middle of enterprise production workloads, he suggested, “very well may end up being the largest and the most durable” part of the market — and has barely started to move.

Stores: everyday essentials, ultrafast delivery, and a subtler AI story

While AWS consumed much of the oxygen, the consumer businesses quietly delivered a more measured, but still robust, performance.

Overall revenue came in at $213.4bn, up 12% excluding foreign exchange. North America grew 10% to $127.1bn; International rose 11% ex‑FX to $50.7bn. Worldwide paid units were up 12%, the fastest quarterly rate of 2025, as Amazon navigated another peak holiday season with what the CFO described as record delivery speeds and lower cost to serve.

Operating income for the group reached $25bn in the quarter, even after absorbing $2.4bn of special charges: $1.1bn for Italian tax and legal settlements, $730m for severance across segments, and $610m in store‑related impairments, primarily in North America. Stripping those out, both North America and International showed margin expansion. North America’s operating margin improved to 9%, from 8% a year earlier, while International reached 2.1%.

The structural shift in retail remains toward everyday essentials and groceries. In the U.S., everyday essentials grew nearly twice as fast as other categories and now account for roughly one in three units sold. Grocery gross sales exceeded $150bn in 2025, leading Jassy to declare Amazon “clearly a large grocer at this point.” Over 150m Americans now rely on Amazon — largely via Whole Foods and online — as a grocery destination, and Jassy highlighted that customers who buy perishables tend to double their overall shopping frequency and add three times as many items to same‑day orders.

The company is leaning into that habit‑forming dynamic. Same‑day is Amazon’s fastest‑growing delivery option, used by nearly 100m U.S. customers last year. Prime members received more than 8bn items same‑ or next‑day in the U.S. in 2025, up over 30% year on year, with groceries and everyday essentials making up half of those items. Amazon can now deliver perishables same‑day in more than 2,300 cities and towns by leveraging its regionalized U.S. fulfillment network, which Olsavsky said continues to drive both speed and cost reductions as inventory is placed closer to demand.

At the extreme end of convenience, Amazon is experimenting with “QuickCommerce.” The “Amazon Now” service promises delivery of thousands of items in around 30 minutes or less and has launched in India, Mexico, and the UAE, with tests underway in parts of the U.S. and UK. Early data from India is striking: Prime members triple their shopping frequency after adopting the service, according to Jassy.

Behind the scenes, Amazon is extracting more productivity from its logistics infrastructure. The U.S. fulfillment network has been further regionalized — with regions increased from eight to ten — and extended to inbound flows. Amazon is also improving box utilization to reduce shipments per order and rolling out more robotics and automation. The company now operates over 1m robots in its network, still “a fraction” of what Jassy believes will be possible, as repetitive tasks are automated and human workers are redeployed to higher‑value roles.

On the storefront itself, the most visible AI experiment is Rufus, Amazon’s in‑house shopping assistant. It is being positioned as a first‑class interface for discovery, price‑tracking, and automated purchasing when items hit target prices. In 2025, more than 300m customers used Rufus, and those users were about 60% more likely to complete a purchase than non‑users. A complementary AI tool, Lens, allows visual search via camera or screenshots and saw usage rise 45% year on year.

Jassy acknowledged that the rise of “horizontal agents” — general‑purpose AI assistants outside Amazon’s control — could compress parts of the traditional e‑commerce funnel as consumers receive more targeted answers. But he pushed the counter‑argument that retailer‑specific agents, armed with complete purchase histories and precise product data, will ultimately win out for many shopping journeys. The challenge now, he suggested, is finding a workable “value exchange” between retailer platforms and third‑party agents that keeps data accurate and economics sustainable.

Advertising and media: AI‑fueled targeting atop a larger audience

Advertising once again outgrew commerce, with fourth‑quarter revenue rising 22% to $21.3bn. Olsavsky said Amazon added more than $12bn of incremental ad revenue in 2025, as the company deepens what it calls a “full funnel” offering spanning on‑site product ads, streaming video, and sports.

Sponsored Products in the core store remains the largest ad format, underpinned by what Jassy described as “trillions of shopping, browsing, and streaming signals” fed into AI and machine learning models to refine relevance. But the real step‑function in reach is coming from Prime Video’s shift to an ad‑supported default. The streaming service now has an average global ad‑supported audience of 315m viewers, up from 200m in early 2024, and is “contributing meaningfully” to overall ad growth.

Live sports continues to act as a tentpole. The fourth season of Thursday Night Football on Prime was its most‑watched yet, averaging more than 15m viewers, up 16% year on year and marking a third consecutive year of double‑digit growth. The Packers–Bears wild card game in particular drew 31.6m viewers, making it the most‑streamed NFL game in history.

Advertising itself is also being re‑engineered by AI agents. Amazon has introduced an “ads agent” to help brands design and optimize campaigns, including targeting, and a “creative agent” that can brainstorm and generate full‑funnel campaigns “from concept to completion” using conversational prompts and Amazon’s retail data. Processes that previously took weeks can now be executed in hours, Jassy said, potentially boosting both ad spend and return on ad spend.

Side bets with heavy lift‑off costs: LEO satellites and Alexa Plus

Beyond the headline segments, Amazon is pushing into two long‑dated infrastructure gambles: low‑Earth orbit satellites and a subscription AI assistant.

Amazon LEO — the company’s satellite constellation aimed at connecting consumers, enterprises, and governments in underserved regions — is moving closer to commercialization. Amazon has launched 180 satellites to date and expects more than 20 launches in 2026 and over 30 in 2027, with commercial service slated for later this year. The flagship enterprise terminal, LEO Ultra, promises download speeds up to 1 Gbps and upload speeds up to 400 Mbps, with private networking that connects directly into AWS and bypasses the public internet. Amazon has already signed “dozens” of commercial agreements, including with AT&T, DIRECTV Latin America, JetBlue, and Australia’s NBN.

In the near term, however, LEO is a cost drag. Olsavsky flagged about $1bn of incremental North American expense in Q1 tied to launch costs, most of which are expensed as incurred. He noted that later in 2026, satellite manufacturing and launch services should begin to be capitalized, easing the P&L impact even as capex rises.

Alexa Plus, meanwhile, represents Amazon’s attempt to reframe its voice assistant as a paid, AI‑augmented service. Now available to all U.S. customers — free for Prime members and $19.99 a month for others — it has been extended to a web chat interface at Alexa.com, a redesigned mobile app, and integrations with third‑party devices such as Samsung TVs and BMW cars. New capabilities include answering Ring doorbells on users’ behalf and deeper hooks into shopping and home management. It is an early test of whether consumers will pay recurring fees for AI companionship and automation layered on top of a commerce relationship.

Guidance: healthy growth, but a more complicated margin story

Amazon’s guidance for the first quarter of 2026 suggests the growth engine remains intact but that headline profitability will be shaped by the timing of its infrastructure projects.

Net sales are forecast at $173.5bn to $178.5bn, with a roughly 180‑basis‑point tailwind from foreign exchange if current rates hold. Operating income is expected to land between $16.5bn and $21.5bn, a range that explicitly bakes in elevated LEO expenses and heavier investment in the International stores business. In overseas markets where competition is intense, Amazon is “working hard to stay sharp on pricing and seller fees,” even if that entails “being more aggressive to meet or beat competitors’ prices” in the short term.

The subtext is that Amazon is deliberately pulling forward costs where it sees either structural advantage — as in AWS AI infrastructure and custom silicon — or high lifetime value, as in ultrafast grocery and QuickCommerce. The company’s full‑year operating cash flow of $139.5bn in 2025, up 20%, and trailing twelve‑month free cash flow of $11.2bn provide some cushion. But investors will now be watching three metrics in tandem: AWS backlog growth, chip and AI service run‑rates, and the trajectory of operating margins as the capex super‑cycle gathers pace.

In the meantime, Amazon’s management is asking shareholders to accept more volatility in near‑term profitability in exchange for what Jassy repeatedly described as a once‑in‑a‑generation infrastructure build‑out — not only of data centers, chips, and satellites, but also of the AI agents and logistics rails that may define how consumers and enterprises transact in the next decade.