On March 24, 2026, ARM CEO Rene Haas stepped onto a San Francisco stage and broke 35 years of company tradition. He announced ARM’s first in-house silicon and declared that agentic computing is driving a fourfold increase in CPU demand. Then he added something that hit me hard: “We may be under-calling that number. I think the demand is higher than we think it is.”

He’s right. But not entirely for the reasons he speculated.

ARM’s model accounts for what agents do externally: orchestrating accelerators, exchanging data, managing API traffic across distributed systems, and trading coin. That’s a real and significant demand signal. But there is a second signal that doesn’t appear in their model, because it isn’t about agents acting on the world. It’s about agents acting on themselves.

I call it agent self-care.

Agent self-care is the autonomous investment agents make in testing, validation, deployment infrastructure, and documentation — not because they are explicitly instructed to, but because it emerges from how they are prompted and how they self-correct. It arises. And it has infrastructure economics implications that nobody has yet accounted for — including ARM.

I know this because I’ve been measuring it in git data. For the past several months I’ve been using CommitPulse — an open-source git analytics tool I built — to analyze the engineering composition of codebases ranging from my own agentic development portfolio to some of the most significant open source projects on GitHub: OpenClaw, WordPress, Rails, OpenRocket, and Supabase. Across five independent codebases, in different languages, different domains, and different team structures, the data shows the same pattern. And in October 2025, every one of them shows the same inflection point.