Amazon’s $200B AI Bet

Amazon is about to spend about $200 billion in 2026, and Andy Jassy is telling investors that this is not a science project but a capacity race for artificial intelligence and cloud computing. He wrote that Amazon has already lined up customer commitments for a substantial portion of that buildout and expects to monetize most of it in 2027 and 2028. (cnbc.com) That number is bigger than the capital spending plans recently laid out by Alphabet and Meta, which CNBC said were about $175 billion to $185 billion and $115 billion to $135 billion for 2026. Amazon’s stock had already sold off after its February forecast, then jumped more than 5% on April 9 after Jassy’s letter gave investors harder numbers behind the spending. (cnbc.com, cnbc.com) Most of the money is going into the plain, expensive parts of artificial intelligence: data centers, chips, and networking gear. In February, Jassy said the spending would go predominantly to Amazon Web Services, the cloud division that rents computing power to other companies. (cnbc.com) The easiest way to read the bet is this: Amazon thinks customers will keep renting more computing power than Amazon can install. Jassy said demand for both traditional cloud workloads and artificial intelligence workloads is high enough that Amazon is “monetizing capacity as fast as we can install it.” (cnbc.com) He added a new revenue marker that Amazon had not previously disclosed: artificial intelligence services inside Amazon Web Services are now running at more than a $15 billion annual pace. Reuters reported that this figure was for the first quarter of 2026, which makes it Amazon’s clearest public proof so far that the artificial intelligence buildout is already producing real sales. (reuters.com, cnbc.com) Amazon also used the letter to show that it does not want to buy every critical part of this boom from Nvidia or other outside suppliers. Jassy said Amazon’s chips business, which includes Graviton, Trainium, and Nitro, is now running at more than $20 billion a year and growing at triple-digit percentages year over year. (aboutamazon.com, reuters.com) That matters inside Amazon Web Services because custom chips can change the math on every server Amazon builds. Jassy said Trainium should eventually save Amazon tens of billions of dollars in annual capital spending on inference, which is the step where a trained model answers your prompt instead of learning from data. (finance.yahoo.com, aboutamazon.com) The management piece in the letter was not random culture talk. Business Insider reported that Jassy used the same document to argue for small teams, fewer layers of management, and more individual ownership, which fits a company trying to build data centers and ship new artificial intelligence products without adding the kind of bureaucracy that slows giant companies down. (businessinsider.com) This is also Amazon answering a question investors have been asking since its free cash flow fell from $38 billion to $11 billion last year while capital spending jumped by $50.7 billion, according to GeekWire’s summary of the shareholder letter. Jassy’s answer is that Amazon saw the same pattern when it built Amazon Web Services in the 2000s: heavy spending first, then a much larger business later. (geekwire.com, aboutamazon.com) So the thread running through all of this is simple: Amazon is trying to own more of the artificial intelligence stack at once. It wants to sell the cloud, design the chips, lock in the customers, and keep the organization lean enough that a company with more than $700 billion in annual revenue can still move like a builder instead of a landlord. (cnbc.com, businessinsider.com, aboutamazon.com)