Andy Jassy

Revenue at Amazon Web Services jumped 37% to $42.2 billion in the second quarter ended June 30, the fastest growth rate since 2021. But CapEx spend on building out infrastructure to meet AI-driven demand  is set to rise to $220 billion for this year.

Total sales for Amazon as a whole rose 20% year-on-year to $200.6 billion, compared with $167.7 billion. Net income increased to $62.6 billion from $18.2 billion.  AWS operating income was $16.6 billion, compared with $10.2 billion in second quarter 2025.

AWS is “booming”, according to Amazon CEO Andy Jassy:

AWS is now $169 billion annualized revenue run rate business, which for perspective, would place it 24th in the Fortune 500 list if it was a stand-alone company. Our chips business now has an annual revenue run rate of over $25 billion, growing triple-digit percentages year-over-year. Our AI revenue run rate climbed significantly quarter-over-quarter and is now also over $25 billion, growing triple-digit percentages year-over-year.

Customers choose AWS because we offer the broadest capabilities, they want their AI inference to reside near their other applications and data and more of it resides in AWS than anywhere else and because AWS has the strongest security and operational performance.

Growth is coming from both AI and non-AI users, the latter referred to as core, he adds: 

Growth in AI drives core because post-training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators. This is an advantage for AWS as our Graviton chip is the strongest CPU chip offering up to 30% to 40% better price performance than other options. You need a place to store this AI data and to run vector databases which are also emblematic of a meaningful edge for AWS because we have the broadest and most capable functionality by a fair bit in these core infrastructure areas.

New AWS agreements during the quarter included Warner Bros. Discovery, Vodafone, Siemens Energy, Ryanair, Pinterest, Snowflake, Moody’s, Danske Bank, WNBA, Pennymac, Fiserv, WPP Enterprise Solutions, Vonage, Recursive, fal, Chai Discovery, Odyssey, TwelveLabs, Reactor, OpenRouter, Dash0, New York State Office of Information Technology Services, State of Iowa, University of South Florida, and The University of Utah.

Model behavior 

In common with an emerging theme among the hyperscalers, Jassy also posited that a multi-model approach is increasingly favored by customers, rather than strict adherence to one of the big frontier model providers:

As we’ve been saying for 18 months now, technically competent companies are going to build their own foundation models, not the really big frontier models, but smaller models that leverage their proprietary data. There is no easier service for this than our SageMaker AI service.

Customers also need a high-performance, cost-effective inference service, and that’s what Amazon Bedrock provides. Bedrock not only provides the best selection of leading models as superior performance and with the governance and security controls the companies need, it’s also continuing to grow incredibly quickly.

It’s the same story when it comes to the rise of agents, he suggested:

In addition to leading model building and inference services, customers need easier ways to build, run and leverage agents. For example, even if you’ve built an agent, you have a lot of muck to worry about. A production agent needs somewhere secure to run, memory, so it holds context, and identity, so it can act on a user’s behalf, tools and data to connect to and a way to watch what it’s doing once real traffic hits. Stitching all that together reliably is hard and it stalled many production deployments. It’s why we built Bedrock AgentCore.

While companies will construct their own purpose-built agents from the ground up, most will also use turnkey agentic services. Coding agents are a good example and there are several successful ones, including Claude Code, Codex and our own spec-driven Kiro, which is up to 50% more cost effective than others and tripled in usage quarter-over-quarter. Another of these agentic services is Amazon Quick, an intelligent AI work companion that helps you manage, search and automate your digital workload across e-mail, calendar, local or cloud files and custom workflows. Unlike other offerings in this space, Quick also lets you manage across leading SaaS tools like Slack, Salesforce, Jira, Teams and ServiceNow. 

Spend, spend, spend

As for that CapEx increase, a development that can still turn Wall Street hostile on a whim, Jassy was defiant, insisting:

At this level of spend and higher, we have clear line of sight to strong financial returns. I’ll explain why. There are 2 major parts of the investment, the data centers and the servers and networking equipment that go into them. These are different capital cycles.

Data center capital is spent starting two years before we can put servers into them to start monetizing. Once a data center opens with servers plugged in, we start generating significant revenue right away and then get to monetize these data centers for 30-plus years without having to spend that start-up capital again.

Servers and networking equipment operate on a shorter cycle. We typically purchase these a few months before putting them into service, so we have strong visibility into customer demand before we trigger the spend. If the demand isn’t there, we won’t spend the capital. 

He added:

For servers and networking equipment, on average, it takes a little less than 3 years to break even on that investment. The servers currently have a useful life of at least 5 to six years. And most of our AI capacity these days is being contracted for at least five-year terms. That means that we’re driving significant free cash flow on the servers and networking equipment in the two to three years after we break even.

It’s also worth noting that AWS has a strong track record of pulling forward breakevens on server equipment, where we’ve already made meaningful progress and finding ways to extend the useful life of this equipment without sacrificing customer experience. So for our data centers, which have 30-plus year useful lives, we should get at least five to six generations of server economics, like I explained earlier, with subsequent generations after the first having even better overall economics because we don’t have to repeat that upfront data center investment.

This has implications for budgeting, he notes:

in the short term, when demand is necessitating so many data centers being built simultaneously in advance of when we can start monetizing them, we’ll spend a lot of CapEx and encounter free cash flow headwinds until these data centers come online, can be monetized and we get a few years into these servers being utilized.

But as we get a few years out, the revenue growth outpaces the incremental CapEx growth, which will happen at some point. The resulting revenue, free cash flow and return on invested capital is very compelling. We’ve done this before in the first era of cloud computing just over a longer time horizon where demand built more gradually than it has in AI but we see the margins and returns in AI tracking what we saw with core at the same point of evolution, actually a little ahead. 

And if that wasn’t enough to finally get the message through to short-termists on Wall Street, he added for good measure a hint that there will be more spending to come:

We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from a prior estimate of about $200 billion. But even at that amount, we will still not have enough capacity to meet all the demand we have in 2026. And I believe this dynamic will also be true in 2027 too. In fact, the demand we already have for 2028 is striking. 

My take

It’s a long game.

It really shouldn’t be as difficult for the red-suspender brigade to get their greedy little heads around as it seems, should it?