Amazon

Before Amazon could build thousands of AI applications, it had to solve a problem most enterprises ignore until it’s too late. At scale, inference costs compound into billions.

Inference is the process by which a trained AI model applies its learning to make predictions.

Companies like Nvidia dominate this market because they manufacture chips that execute AI inference.

This is, of course, a huge investment for any Fortune 500 company, so Amazon built its own chips - Trainium and Graviton.

Trainium handles AI model training and inference at a lower cost than Nvidia equivalents. Graviton powers general-purpose computing across AWS, including agentic AI workloads. Together with Nitro (the infrastructure layer), the custom chip business has reached an annual revenue run rate of over $20 billion, per CEO Andy Jassy.

Controlling your compute layer changes what’s economically viable to automate. Workflows that would be too expensive to run on third-party infrastructure become viable when you own the cost structure.

CEO has said Trainium is expected to save Amazon’s annual capital expenditure and provide an operating margin advantage. Good structural change in how Amazon prices its own AI ambitions.

Now all enterprises can’t build their chips, but the lesson here is your compute agreements determine your AI economics. Negotiate hard.

Amazon is asking most of its engineering teams to triple their code release velocity.

Curated, high-performing teams are expected to reach 10x output, using Amazon’s internal tools for its engineers:

Amazon tracks monthly active users per tool, a metric called Value Deriving Events (VDEs) to measure how frequently engineers generate useful outputs, and weekly production deployments per engineer.

To be honest, not everything worked as planned. Reports from early 2026 describe engineers spending more time reviewing and correcting AI-generated code than writing their own. Some gamed VDE scores by running tools on unnecessary tasks to inflate their numbers.

This is Goodhart’s Law applied to AI adoption. When a measure becomes a target, it stops being a good measure.

Amazon pulled back from mandating specific tools and shifted toward building a centralized learning platform for best practices. One of their core AI engineering principles is “AI-native is not AI-exclusive.” AI handles the repetitive tasks while humans focus on architectural decisions and edge cases.

Amazon has deployed over 1 million robots across its global fulfillment network. Today, AI coordinates the fleet rather than each robot executing a fixed routine.

Some examples and outcomes:

Individual automation produces incremental gains. The structural advantage is when AI orchestrates a fleet of specialized systems simultaneously.

“The company might need fewer people doing some of the jobs that are being done today,” the CEO told employees in a June 2025 internal memo.

Thousands of corporate roles have been cut since 2022, with AI among the contributing factors. Amazon named the implication early and communicated it directly, rather than letting it surface as a surprise during a restructuring.

It led to multiple tools and use cases in the knowledge work:

Amazon’s next phase is agentic AI at enterprise scale.

In March 2025, AWS created a dedicated agent-focused group and launched Bedrock AgentCore, enabling secure deployment of AI agents that handle multi-step tasks autonomously.

The company is also committed to investing heavily in agentic AI development through AWS Marketplace.

The direction is consistent across all the companies we’ve covered in our newsletter. Enterprises are moving from AI that assists humans on discrete tasks to AI that owns multi-step workflows end-to-end.

Build your measurement infrastructure before you roll out tools. Amazon’s VDE tracking, weekly deployment velocity, and NPS by tool existed before Kiro launched. When adoption problems surfaced, Amazon had the data to diagnose them. Enterprises that deploy tools without measurement systems have no feedback loop.

Move from single-task automation to coordinated systems. Proteus handles picking. DeepFleet coordinates the fleet. Project Eluna advises the operations manager. Each layer on its own is useful. Together, they compound. So which of your AI tools are talking to each other and which are running in silos?

Your adoption metric is probably measuring the wrong thing. Amazon learned this the hard way when engineers gamed VDE scores. Usage frequency tells you people are opening the tool. It doesn’t tell you whether the tool is producing better outcomes. Before rolling out any AI program, define what a successful output looks like, not just what a successful session looks like.

Sources
Officially from Amazon: 1, 2, 3, 4, 5, 6
Coverage from CNBC, New York Times, Business Insider, People Matters