July 31, 2026. Amazon reported second quarter results on Thursday, and the headline number was AWS growing 36.7 percent year over year, its fastest in 18 quarters, on a $169 billion annualized run rate that would place it 24th on the Fortune 500 as a standalone company. The more useful line came a few paragraphs later, when CEO Andy Jassy named what is actually stopping companies from putting AI agents into production. It is not the model.
What Amazon reported
- AWS accelerated. Revenue growth reached 36.7 percent year over year, the fastest rate in 18 quarters, at a $169 billion annualized run rate.
- Two AI businesses crossed $25 billion each. Amazon's custom chips business is now above a $25 billion annual run rate, and its AI revenue run rate is also above $25 billion. Both are growing at triple digit percentages year over year.
- AI is pulling non AI workloads with it. Jassy said post training, reinforcement learning and agent tool use run mostly on CPUs rather than AI accelerators, so growth in AI is driving growth in core compute, storage and vector databases.
- The agent stack got infrastructure, not intelligence. Bedrock AgentCore added Policies for deterministic controls over what agents can do, Payments so agents can transact, Web Search for grounding, and a new Harness to assemble it faster.
- Coding agents are a real line item. Amazon's spec driven agent Kiro is described as up to 50 percent more cost effective than others, with usage tripling quarter over quarter.
The quote worth pinning to your wall
Describing why agent projects stall, Jassy listed the problem precisely: even after you have built an agent, a production agent needs somewhere secure to run, memory so it holds context, an identity so it can act on a user's behalf, tools and data to connect to, and a way to watch what it is doing once real traffic hits. Stitching all of that together reliably is hard, and in his words it has stalled many production deployments.
Read that list again. Not one item on it is about model quality. The head of the largest cloud business on earth, reporting record AI revenue, is saying the bottleneck is plumbing. That matches what we see on almost every rescue project that lands on our desk: the demo worked, and then nobody could say where the agent runs, what it remembers, whose credentials it uses, or what it did last Tuesday.
What it means for operators
You do not need AWS to apply this. Take any automation you already run in GoHighLevel, n8n, Zapier or Make and answer Jassy's five questions in writing.
Where does it run, and who can turn it off? Name the account and the person. What does it remember between runs, and for how long? An agent with no memory repeats itself and an agent with unbounded memory drifts. Whose identity does it act under? A shared login is not an identity, and it makes every audit impossible. What can it reach? List the connected apps and delete the ones nobody can justify. And how would you know it broke? If the answer is a client complaint, you do not have observability, you have luck.
Five answers per automation, in a shared document, with a named owner on each. That register is the single highest return hour of work available to most agencies right now, and it costs nothing. If you would rather have it designed and built properly, that is exactly what our AI automation and n8n development work covers.
The pattern behind the quarter is simple enough. The models are good enough that the money has moved to everything around them. Your budget should move the same way.
Frequently Asked Questions
AWS revenue grew 36.7 percent year over year, which Amazon described as its fastest growth in 18 quarters. CEO Andy Jassy said AWS is now a $169 billion annualized revenue run rate business, large enough to rank 24th on the Fortune 500 if it were a standalone company.
He said that even after an agent is built, a production agent needs somewhere secure to run, memory so it holds context, an identity so it can act on a user's behalf, tools and data to connect to, and a way to watch what it is doing once real traffic hits. Stitching those together reliably is hard, and he said it has stalled many production deployments.
No. The five requirements are platform independent. Whether your automations run in GoHighLevel, n8n, Zapier or Make, you can write down where each one runs, what it remembers, which identity it acts under, which apps it can reach, and how you would find out if it broke. Most small teams have never written that down, which is why failures surface as client complaints.
Jassy said both crossed a $25 billion annual revenue run rate in the quarter, and that both are growing at triple digit percentages year over year. He also noted that AI growth is driving core infrastructure growth, because post training, reinforcement learning and agent tool use run largely on CPUs rather than AI accelerators.