Amazon's Andy Jassy on a $220 Billion AI Buildout and the Demand Behind It

Jassy opened by saying that even with roughly $220 billion of capital spending this year, Amazon is leaving a meaningful amount of demand on the table. He said the company always keeps some capacity for on-demand customers, but 2026 is obviously sold out, 2027 is largely sold out, and a good chunk of 2028 is already sold out.

He acknowledged that financial markets question whether there is enough money in the world to finance such buildouts. He answered that everything people do today, including stock trading, banking, and healthcare, runs on data centers, and that Amazon's job is to add as much capacity as it can.

Jassy also noted that frontier AI labs are willing to pay almost any price for power, cooling and chips. He joked that Amazon might buy out its older customers and sell to the new ones, since they will pay whatever is asked.

Jassy compared Amazon's position with 2021, when he took over as CEO. He said total revenue is about 80% higher, AWS is almost three times bigger, and ads revenue is about two and a half times bigger.

He described Amazon's mission as making customers' lives better and easier every day, which he said unifies work across businesses including Prime Video, the Amazon Leo satellites, the Zoox ride-hailing service, healthcare and robotics.

Jassy explained that Amazon doubled the size of its fulfillment network, which had taken 25 years to build, in about 18 months. The team knew that moving this fast would be harder and more expensive than moving slowly, and that the result would not be as optimal, but customers and sellers needed the capacity, especially when physical stores were closed during the pandemic.

He said Amazon decided, as it almost always does, to side with customers, and then accepted some cleanup afterward. He noted that the company had not foreseen exactly how the network would turn out, but that good metrics and a truth-telling culture helped it fix the design. Regionalizing the nationally flat fulfillment network later cut transportation distance, lowering cost to serve while speeding delivery.

Ben Horowitz noted that AWS began with a build-it-and-they-will-come logic, since Amazon was building infrastructure before knowing whether outsiders would rent it. Jassy said the company had internal signals that its product teams were not moving as quickly as it wanted.

Those teams spent 70 to 80% of their time on infrastructure that did not differentiate their products and only about 20% on the features that did. Amazon began building high-scale, low-cost building blocks such as compute, storage, database and messaging, and later offered them to outside customers.

Jassy said early AWS grew quickly and reached about a $60 million annual run rate after three years. In contrast, in the first three years of the generative AI wave, Amazon's AI business in AWS has an annual run rate of over $25 billion.

He said this is several hundred times faster and that Amazon had never seen anything like it. He added that the question then becomes how much capital expenditure to invest to pursue the opportunity, and he recalled that 15 years earlier the company found a telecom company's $50 billion capex hard to believe.

Jassy said data centers were probably the last thing anyone would have expected to become a political issue when AWS started, but that the United States in particular has a strongly anti-data-center politics. He pointed out that stock markets, banking, healthcare, entertainment, shopping, education and ride-hailing all run on data centers, even before AI.

He described AI as the biggest technology transformation of the current era, after the internet, cloud and mobile. He compared the needed buildout to the highways and bridges constructed in the 1950s and said the country cannot afford not to build data centers, or else a country such as China could establish leadership and norms there.

On water use, he said golf courses consume about 200 times more water than all of Amazon's data centers combined, and he cited the small share of time backup generators run as a counterpoint to popular claims.

Jassy said that in St. Joseph County, Indiana, Amazon has paid about $3 billion in tax, and that the communities get that money and can spend it on things that really matter to them. He also said Amazon committed another $1 billion over the next five years to the communities where it has data center presence, through programs such as Built Together. He framed the buildout as a public-private partnership that must keep being worked on with those communities, comparing its importance to the country's future with the way highways and bridges built earlier shaped the country's economic prosperity.

He framed the buildout as a public-private partnership that must be maintained with communities, comparing its importance to the country's future with the economic value of the highways and bridges built earlier.

Jassy said there is only so much data center capacity, power and memory, and he did not think the shortage would end in 2027. He noted that 2026 is sold out, 2027 is largely sold out, and a good chunk of 2028 is already sold out.

Asked whether Amazon should sell to new AI labs at higher prices than older customers, Jassy said the company does not frame it that way. He said that Amazon is not going to squeeze its customers and will make sure they can scale. He described memory as a real trade-off between high-bandwidth memory for AI accelerators and standard memory for the core business.

He also said that during the pandemic Amazon proactively helped customers optimize costs, accepting short-term revenue loss for long-term trust.

Jassy called the Annapurna Labs acquisition one of the best in Amazon's history. He said Amazon did not buy the team specifically for AI chips. The company wanted to offload the work EC2 servers were doing for virtualization, networking, security and storage, so it could use nearly 100% of each server for customer compute.

That offload technology, called Nitro, ended up in virtually every EC2 server. Amazon bought the team because it was talented at building chips and could probably build others.

Jassy explained that Intel held nearly all CPU share, so it was hard for Amazon to get attention on price performance from a partner that focused on other features to gain more share. Customers moving on-premises workloads to the cloud wanted better price performance, so Amazon had the Annapurna team build a CPU chip.

The resulting Graviton is used by 99% of the top 1,000 EC2 customers and is about 40% more price performant than other x86 processors. A few years ago, Amazon saw a similar pattern in AI and spun out a separate Annapurna team to build the Trainium AI accelerator, which Jassy said is about 30 to 40% more price performant than comparable processors.

Jassy described a Bedrock team that rebuilt its inference engine with six engineers in 76 days, work he said would have taken 40 people a year before. He also cited a prototype built by four or five people over a weekend that was 80% shippable by Sunday.

He said specialization grew as companies scaled, with product managers owning a single slice of a product, but that AI calls for more crossfunctional people because models can carry work about 80% of the way on many functions. He predicted that businesses will have far fewer humans doing what they do today, although not zero, with new jobs emerging in AI, robotics and space.

Jassy said that teams will be much smaller and that every team will start at a more advanced stage than in the past, which lets companies invent more expansively and quickly. He said most startups starting today are being built agentically and AI-native from the start.

He said established companies will have to learn to build at that same pace. Some will manage it, but those with large businesses and loyal customers may move slowly and become more vulnerable to startups.

Jassy predicted companies will need to rewrite almost all their code to be agentically capable, allowing agents to update different pieces of code, and that constraints will need to be built in. He said recent public worry has focused on fantastical scenarios such as evil botnets or malicious robots, which he said are possible but are not the most immediate risk.

The more immediate concern, he said, is cybersecurity: models find far more vulnerabilities than humans previously did. He argued that applications will need to respond to security threats in minutes rather than the long timeline of finding, classifying, patching and deploying fixes that most companies follow today.

Jassy described three ways companies can use AI: cutting costs, inventing something new, and rethinking a successful product from the beginning. He said some venture firms buy older companies to cut costs, which Amazon does not find very interesting, while building something new is exciting.

He said streamlined organizations and small teams can build the second category of new experiences. He warned that companies treating AI mainly as incremental improvement to existing experiences will likely not have winning experiences three to five years from now. He pointed to Alexa and to other consumer businesses such as Prime Video, Fire TV and Prime Music, which he said must be reimagined as several teams build fully new versions.

Jassy said that companies must decide whether to build a completely different experience, such as something like a new app, or to advance quickly enough to add elements into the existing experience and change it bit by bit. He said every one of these experiences has to be different in the next few years to be the successful leader.

Horowitz compared this with the old engineering saying that everything is up for discussion until the first line of code is written. Jassy said big transformations require change management, and that infrastructure leaders proud of their on-premises work resisted moving to the cloud, making a complete restart intellectually and emotionally hard.

Jassy said his 22-year-old child was 19 when AI began to take off, and that he learned quickly you cannot tell young people not to use AI in education. He suggested universities would be wise not to fight it. He said people can tell when something has been written by AI because the pieces do not fit together quite right.

His advice was that young people cannot run away from AI or wish it away. They need to learn about it, educate themselves, use it a lot, and try things outside their comfort zone. He added that AI does not replace critical reasoning and thinking, and noted that neither of his children were developers yet used AI to build applications.