Amazon plans to invest around $220 billion in 2026, more than the GDP of many countries, and its CEO says even that may not be enough. Andy Jassy, founder of Amazon Web Services and Jeff Bezos’s successor since 2021, argues that demand for artificial intelligence is exceeding the capacity available across the entire industry. And he makes a projection few are willing to say out loud: that Amazon’s cloud business could eventually become a trillion-dollar-a-year revenue operation.
A Bet Measured in Decades
The investment figure increased from the $200 billion Amazon had previously estimated, driven by rising memory prices. According to Jassy, capacity will not be sufficient this year or next, and the demand already committed for 2028 is striking. For a long time, the company believed AWS could reach several hundred billion dollars in revenue. Today, it believes the figure could be at least twice that.
His defense of the spending is based on accounting. A data center has a useful life of thirty years or more and can house five or six generations of servers. Generations after the first are more profitable because the initial investment in land, power, and the building does not have to be repeated. The problem is timing: the money goes out before the facility is ready to generate revenue, putting pressure on cash flow in the short term. Once revenue begins growing faster than incremental investment, he says, returns will become very attractive.
Jassy says he has already seen this movie during the first stage of cloud computing, only in slow motion. He argues that AI margins and returns are following the same curve as AWS’s traditional business at the same stage of development, and are even slightly ahead. In Davos, he added another argument for those watching the growth of Azure and Google Cloud: percentages can be misleading because, in absolute dollars, AWS grew the most, and by a wide margin.
Chips as a Pricing Weapon
The other half of the strategy is silicon. Jassy says he learned the lesson from Intel: when one supplier dominates a market, reducing costs for customers is rarely its priority. That is why Amazon designed Graviton, its own processor, which delivers 40% better price-performance than leading x86 chips. Today, nine out of its thousand largest customers use it significantly.
With AI, Amazon is following the same playbook, although Jassy says it will maintain a deep partnership with Nvidia for as far ahead as he can foresee. Trainium 2 is completely sold out and is already a multibillion-dollar business; Anthropic uses hundreds of thousands of those chips to train its models. Trainium 3 improves another 40% over its predecessor. Bedrock, Amazon’s model platform, already runs mostly on its own chips. For Jassy, anyone building an inference business without proprietary silicon is at a strategic disadvantage, both in lowering prices and maintaining margins.
He also defends Amazon’s own models, the Nova family. They do not need to be at the frontier like those from Anthropic or Google, he explains, but they give Amazon control over cost, development speed, and latency. His thesis is that customers will combine different models depending on each application.
The Barbell Theory
Jassy describes the current adoption of AI as a barbell. At one end are AI laboratories consuming enormous amounts of computing power, along with a few applications that have exploded in popularity, such as ChatGPT. At the other are companies using AI to reduce customer-service costs or automate processes.
The biggest business, he argues, lies in the middle: applications companies already have in production that do not yet use inference. He believes that will become the largest segment in absolute terms and that those workloads will run close to where each company’s data is stored. According to him, much of that data already lives on AWS.
OpenAI, Rival and Partner
AI-agent commerce is one area where Amazon overlaps with the laboratories. OpenAI and others want people to shop through their chatbots, and OpenAI has announced that it will enter the advertising business. Jassy is optimistic but sets limits. Third-party agents do not know a user’s purchasing history and frequently get prices and product information wrong. For the model to work, he says, there has to be a fair exchange of value between agents and merchants. In the meantime, Amazon is betting on Rufus, its shopping assistant, and Jassy points out that Amazon represents only about 1% of global retail.
The relationship with OpenAI is ambiguous. Amazon recently signed a significant cloud agreement with the company, and Jassy does not comment on rumors of a possible investment. He simply says that both teams wanted to work together and that he hopes to deepen the relationship.
Fewer Layers, More Owners
With around 1.6 million employees, Amazon made cuts that the market attributed to AI. Jassy rejects that interpretation. He says rapid growth had multiplied management layers and taken decision-making power away from the people who should have had it, and that his goal is for Amazon to operate like the world’s largest startup.
On the future of employment, he acknowledges that AI will perform much of the work involved in programming, customer service, research, and analysis, and that even creating a spreadsheet will change. But he believes those jobs will continue to exist, that every task will begin from a more advanced starting point, and that new professions will emerge.
On energy, another global bottleneck, he points out that Amazon has signed nuclear agreements and was the largest corporate buyer of renewable energy in each of the past five years. Regarding Donald Trump’s request that technology companies generate their own power, he says it is too early to assess, but that the company has always been prepared. And in discussions with governments outside the United States, he warns that some recent regulation ultimately harms consumers in those countries.
Beyond the Cloud
AI is also spreading across the rest of the Amazon empire. The company has the third-largest grocery business in the United States, with $150 billion in sales. It expects to exceed 500 million drone deliveries before the end of the decade. And it launched Kiro, an agent-based coding service, along with agents capable of reading an entire company’s codebase and independently working through its backlog of bugs. For Jassy, AI is the most transformative technology he expects to see in his lifetime.
