Amazon says AWS has passed a $15 billion AI run rate

Amazon says AWS now has an AI revenue run rate above $15 billion.

Fact: In a Q1 commentary published on April 30, Amazon said AWS’s AI revenue run rate is over $15 billion, almost 260 times larger than AWS was three years after launch. CEO Andy Jassy pointed to four drivers: a broader model and agent platform, inference close to customer data in AWS, the breadth of adjacent cloud services, and security plus operational performance.

Amazon also gave operational signals. Bedrock customer spend grew 170 percent quarter over quarter and processed more tokens in Q1 than in all previous years combined. OpenAI models are moving into Bedrock, with GPT-5.4 available and GPT-5.5 coming. Amazon has also started a preview of Amazon Bedrock Managed Agents, Powered by OpenAI, positioned as a stateful runtime for production agent applications.

The agent layer is the leadership point. Strands has been downloaded more than 25 million times, with downloads up threefold quarter over quarter. AgentCore is, according to Amazon, being used to deploy an agent as frequently as every 10 seconds. Kiro, Transform, Connect and Quick are being positioned as ready-made agents for coding, migrations, customer work, operations and knowledge work.

Assessment: This is not just another cloud vendor saying AI is growing. Amazon is trying to make AWS the place where agents run because the data, security model, network and legacy applications already sit there. For CIOs, the implication is straightforward: AI architecture is becoming cloud architecture. The agent platform choice decides where data moves, how access is controlled, what gets logged, and how hard it will be to switch later.

It is also a cost issue. As AI moves from chat pilots into inference, agents and automated workflows, the consumption pattern starts to look more like production workload than software licensing. CFOs and CIOs should therefore track token use, runtime, data access and agent actions together, not as separate budget lines. Bedrock’s growth shows how quickly cost can move from experiment to operations.

The source matters: this is Amazon’s own presentation of Q1 and the AWS strategy. The numbers are official, but the framing is Amazon’s. The practical conclusion still holds: enterprises should not buy agent features as isolated tools. They should require an architecture map, data boundary, logging model, role design, kill switch and exit plan before agents get access to email, files, Slack, code, customer data or ERP.