TL;DR
- Amazon closed 2025 with revenue of $717 billion, up 12% from $638 billion the year before, with operating income climbing 17% to $80 billion.
- Free cash flow collapsed from $38 billion to $11 billion, driven by a $50.7 billion year-over-year jump in property and equipment spending tied to AI infrastructure.
- Jassy committed to roughly $200 billion in 2026 capex, and pointed to a $100 billion OpenAI agreement plus other unannounced contracts as evidence the spend is not a hunch.
- AWS now runs an AI revenue rate above $15 billion, nearly 260 times larger than AWS itself was at the same point in its life.
- Amazon’s chip business — Graviton, Trainium, Nitro — is generating north of $20 billion in annual revenue and would price out near $50 billion as a standalone.
- Grocery hit $150 billion in gross sales, making Amazon the second-largest grocer in the United States, and Same-Day delivery for perishables grew 40x in a single year.
About the author
David L. Berkowitz, Chief Investment Officer and Financial Advisor
Nearly 40 years of experience — from trading and research at a $250 million hedge fund in the early 1990s, to two decades as a portfolio manager, to teaching thousands of executives and employees how to create shareholder value through EVA and value-based management. Now helping individuals and families become shareholders through disciplined investing, concentrated portfolios, and direct stock ownership.
Why this letter matters
Every spring I read the Amazon shareholder letter the way other people read box scores. It is the cleanest annual look at how one of the most relentless capital allocators in the world is thinking about the next decade. This year’s letter is built around one idea: progress does not move in straight lines, and any leader who pretends it does is selling something.
Andy Jassy borrowed the framing from an indie rock album title and turned “the straight line is a lie” into the through-line of an eight-page argument about why Amazon is willing to crush near-term free cash flow to win the next ten years. Below is what shareholders should actually take away from it.
How big is Amazon’s AI bet, and is it a bubble?
Direct answer: Amazon is spending more on AI infrastructure in one year than most countries spend on defense, and Jassy thinks the bubble talk is wrong.
Asked whether AI is overhyped, whether the industry is in a bubble, and whether the returns will eventually show up, Jassy’s answer in the letter was “no, no, and yes”. The numbers behind that confidence are large enough to test it. AWS now runs an AI revenue rate above $15 billion, nearly 260 times larger than AWS itself was at the same point in its life cycle.
The capex commitment is the part that keeps Wall Street up at night. Jassy reaffirmed a roughly $200 billion capex plan for 2026, most of it pointed at AI data centers. He framed the spend as backed by real demand, including a $100 billion commitment from OpenAI plus other agreements that are signed but not yet public.
The cost is already showing up in the cash flow statement. Free cash flow fell from $38 billion to $11 billion in 2025, driven almost entirely by the $50.7 billion year-over-year increase in property and equipment purchases. Jassy’s bet is that the cash starts coming back in 2027 and 2028 as the new capacity gets monetized.
The critical insight: Amazon is doing exactly what it did with AWS in 2006. Spending heavily on infrastructure that looks expensive today, in front of a demand curve that bears do not yet believe in. If Jassy is right about the curve, the capex is cheap. If he is wrong, this is the most expensive bet in corporate history. There is no middle outcome.
Why does Amazon’s custom silicon matter for the moat?
Direct answer: because owning the chips changes the unit economics of every AI workload Amazon runs, and tightens the gap between Amazon and Nvidia’s pricing power.
Amazon’s three custom silicon lines — Graviton, Trainium, and Nitro — together generate more than $20 billion in annualized revenue, growing at triple-digit rates. Jassy estimated that if the chip business were sold the way Nvidia sells chips, the run rate would land closer to $50 billion.
Trainium2 has largely sold out. Trainium3 is nearly fully subscribed after initial 2026 shipments, and a meaningful portion of Trainium4 is already reserved ahead of broad availability. On the CPU side, AWS’s Graviton chip is now used by 98% of the top 1,000 EC2 customers. Two large customers asked to buy Amazon’s entire 2026 Graviton capacity. Amazon turned them down.
Jassy laid out the economics in a single sentence: scale Trainium across AWS and the company expects to save tens of billions of capex dollars per year, plus pick up several hundred basis points of operating margin on inference workloads versus relying on third-party chips.
The critical insight: this is the same playbook Amazon ran on Intel with Graviton, just compressed into a tighter time window. The companies that own their AI silicon end up controlling cost, supply, and roadmap. Everyone else rents.
The capex math at a glance
| Metric | 2024 | 2025 |
| Total revenue | $638 billion | $717 billion (+12%) |
| Operating income | $69 billion (10.8% margin) | $80 billion (11.2% margin) |
| Free cash flow | $38 billion | $11 billion |
| AWS run rate (Q4) | ~$110 billion | $142 billion |
| AWS AI run rate | ~negligible | $15 billion+ |
| 2026 capex plan | — | ~$200 billion |
Sources: Amazon 2025 shareholder letter and Quartz coverage of full-year financials.
What is the parallel-path strategy in delivery?
Direct answer: Amazon is funding three different ways to make delivery faster at the same time, because waiting to pick the winner is the slowest path of all.
Most companies argue for months about which initiative to back. Jassy used a hockey metaphor — “two greater than zero” — to defend running parallel bets. Amazon has built more than 85 Same-Day Fulfillment Centers across the U.S., carrying its top 90,000 SKUs and delivering more than 500 million same-day units in 2026 so far.
Prime Air, the drone program, now has a scalable design and aims to serve 30 million customers by year-end, with a goal of half a billion deliveries by the end of the decade in under 30 minutes.
The third leg is Amazon Now — micro-fulfillment delivery in under 20 minutes. The service started in India and the UAE, where Amazon now operates more than 360 micro-fulfillment centers and orders are growing 25% month over month. Prime members who try the service triple their shopping frequency.
The critical insight: these are not competing programs. The drones use the Same-Day centers as launch pads. Amazon Now handles the essentials. Prime Air handles the long tail. Amazon is building a layered delivery system, the way Toyota built a layered production system — every piece feeds the next.
How is Amazon expanding where competitors are pulling back?
Direct answer: rural delivery and satellite internet — two markets the rest of the industry treats as too expensive to serve.
Amazon has committed more than $4 billion to expand its rural delivery network, a market most logistics and telecom providers have been quietly retreating from. The result: the average number of monthly Same-Day customers in rural areas nearly doubled in 2025, and the buildout will eventually cover 13,000 ZIP codes spanning 1.2 million square miles.
Then there is Amazon Leo, the low-Earth-orbit satellite network that competes directly with SpaceX’s Starlink. The service currently has more than 200 satellites in orbit and launches commercially in mid-2026. Delta Air Lines, the highest-grossing airline in the world, picked Leo for in-flight Wi-Fi across 500 planes starting in 2028, joining JetBlue, AT&T, Vodafone, Australia’s National Broadband Network, and NASA.
The critical insight: Amazon Leo is not a side project. Stitched into AWS, it becomes the connectivity layer for cloud computing in places that did not have cloud computing before. That is a new pipeline of enterprise customers that did not exist a year ago.
How did Amazon finally crack grocery?
Direct answer: by stopping the search for one big format and instead pushing fresh food into the Same-Day Delivery network it had already built.
Amazon spent two decades buying Whole Foods, launching Amazon Fresh, and tinkering with Go stores. None of it produced the breakout moment. The shift came when the company put perishables — produce, dairy, meat — into the Same-Day pipeline that already moved everything else. Since making that change in early 2025, perishable sales have grown more than 40x and now account for nine of the top ten most-ordered items in Same-Day delivery where available.
Whole Foods now operates over 550 stores, with about 100 more on the way, and a smaller Daily Shop format aimed at quick urban grocery missions. Add it all together and Amazon’s grocery business hit roughly $150 billion in gross sales in 2025, making it the second-largest grocer in the United States.
The critical insight: Amazon did not need a new grocery format. It needed to stop treating grocery as separate from everything else it ships. The win came from removing a wall, not from building one.
What does “going back to the starting line” actually mean?
Direct answer: the willingness to tear down a working product and rebuild it from scratch when the underlying technology shifts.
The new Alexa+ is the cleanest example. Alexa is a massive installed base — 600 million active endpoints across devices, cars, offices, Fire TV, and Prime Video. Generative AI made the old Alexa architecture obsolete. The team rewired the assistant’s brain while keeping the existing service running for hundreds of millions of users. The payoff: customers are talking to Alexa twice as much, completing purchases on devices three times more often, streaming music 25% more, and using smart home features 50% more.
Bedrock, AWS’s inference service, got the same treatment. A team of six engineers, using Amazon’s own agentic coding tool Kiro, rebuilt the Bedrock inference engine from scratch in 76 days and named it Mantle. Jassy noted in the letter that this kind of project would normally take 40 engineers about a year. Mantle now backs the Bedrock service, which processed more tokens in Q1 2026 than in all prior years combined.
The critical insight: the Mantle story is not really about Alexa or Bedrock. It is about Amazon publishing, in a shareholder letter, that AI-assisted coding has compressed the cost of rebuilding production infrastructure by an order of magnitude. Engineering economics just changed, and Amazon is the first hyperscaler to put a number on it.
The bottom line for shareholders
Free cash flow took a brutal hit in 2025. Anyone who screens for cash generation lost interest in Amazon last year. That is exactly the moment Jassy wants long-term holders to lean in.
The argument is straightforward. Amazon’s retail business is approaching $600 billion in revenue, but roughly 80% of global retail still happens in physical stores. AWS runs at a $142 billion annualized rate, but 85% of global IT spend remains on-premises. Both numbers will move toward Amazon over the next decade, and Jassy is willing to spend now to be the one collecting when they do.
Jeff Bezos closed his original 1997 letter with a single sentence: it is all about the long term. Jassy attached that 1997 letter to this year’s edition, as Amazon does every year, and the placement is not decorative. It is an argument that the discipline behind a $200 billion capex plan in 2026 is the same discipline that built AWS out of an internal IT project.
If the straight line is a lie, then long-term shareholders should want a management team that knows how to navigate the curves. This letter is Amazon telling you it does.
Endnotes
-
Amazon CEO Andy Jassy 2025 Letter to Shareholders — The full text of the 2025 annual shareholder letter from Amazon CEO Andy Jassy. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-2025-letter-to-shareholders
-
GeekWire on Jassy’s $200B capex defense — GeekWire’s analysis of Jassy’s 2025 letter and the case for $200 billion in 2026 capex. https://www.geekwire.com/2026/not-on-a-hunch-andy-jassy-defends-amazons-200b-spending-spree/
-
TechCrunch on Amazon’s competitive positioning — TechCrunch coverage of how Jassy’s letter targets Nvidia, Intel, and Starlink. https://techcrunch.com/2026/04/09/amazon-ceo-takes-aim-at-nvidia-intel-starlink-more-in-annual-shareholder-letter/
-
Retail Dive on Jassy’s AI investment defense — Retail Dive’s reporting on Jassy’s defense of AI capex and the rural delivery commitment. https://www.retaildive.com/news/amazon-andy-jassy-defend-billions-ai-investment-shareholder-letter/817058/
-
PYMNTS on AWS power constraints and bubble fears — PYMNTS coverage of Jassy’s response to AI bubble concerns and AWS capacity constraints. https://www.pymnts.com/amazon/2026/amazon-ceo-andy-jassy-rejects-ai-bubble-fears/
-
Quartz on Amazon’s full-year 2025 financials — Quartz reporting on Amazon’s $717 billion revenue, $80 billion operating income, and $11 billion free cash flow. https://qz.com/amazon-jassy-ai-capex-shareholder-letter-040926
-
CIO on AWS chip demand and Mantle rebuild — CIO Magazine on Trainium subscription, Graviton demand, and the 76-day Mantle inference engine rebuild. https://www.cio.com/article/4157494/ai-demand-is-so-high-aws-customers-are-trying-to-buy-out-its-entire-capacity-2.html
-
Converge Digest on Amazon’s full-stack AI strategy — Converge Digest analysis of Amazon’s $200 billion capex tied to AI, AWS, and custom silicon. https://convergedigest.com/amazon-ties-200-billion-2026-capex-plan-to-ai-aws-and-custom-silicon/
-
FreightWaves on Prime Air drone scale-up — FreightWaves coverage of Prime Air’s plan to serve 30 million customers and the role of Same-Day Fulfillment Centers. https://www.freightwaves.com/news/amazon-to-scale-up-drone-delivery-in-2025-ceo-says
-
InvestorsHub on Amazon Leo satellite count and Delta deal — InvestorsHub reporting on Amazon Leo’s 200+ satellites in orbit and the Delta Air Lines contract. https://investorshub.advfn.com/market-news/article/26539/amazon-ceo-jassy-details-ai-spending-plans-forecasts-200b-capex-for-2026
-
Silk Road Nexus on the parallel-path delivery strategy — Substack analysis of how Same-Day Fulfillment Centers, Prime Air, and Amazon Now fit together. https://thesilkroadnexus.substack.com/p/andy-jassys-2025-shareholder-letter
-
Amazon Alexa+ launch announcement — Amazon’s official announcement of Alexa+ and the 600 million Alexa device installed base. https://www.aboutamazon.com/news/devices/new-alexa-generative-artificial-intelligence
-
AWS engineering post on Mantle inference engine — AWS Machine Learning blog on the Mantle inference engine that backs Amazon Bedrock. https://aws.amazon.com/blogs/machine-learning/exploring-the-zero-operator-access-design-of-mantle/
-
Proactive Investors on Trainium economics — Proactive Investors coverage of Jassy’s claim that Trainium will save tens of billions of capex annually. https://www.proactiveinvestors.com/companies/news/1090291/amazon-ceo-andy-jassy-defends-ai-investment-plan-in-annual-letter-to-shareholders-1090291.html
-
Amazon About on Jassy’s reinvention philosophy — Amazon’s editorial on Jassy’s philosophy of going back to the starting line. https://www.aboutamazon.com/news/innovation-at-amazon/andy-jassy-amazon-reinvention
