The Pipes Are Mapped. There’s a Floor Above Them Nobody’s Watching.

I spent the last year+ watching smart infrastructure investors map the agentic AI layer. The routing companies. The fiber backbones being upgraded at two and a half times their historical rate. The content delivery networks repositioning themselves as compute delivery networks. The chip companies licensing specialized inference silicon for always-on workloads.

They are going to do well. The thesis is sound.

Cisco confirmed that a single AI agent generates roughly 25 times more network traffic than a chatbot session. And unlike a chatbot that opens, answers, and closes, agent traffic is continuous. It does not stop. Gartner projects 40% of enterprise applications will embed task-specific AI agents by the end of this year, up from less than 5% not long ago. That is procurement already in motion, not a forecast.

Nearly $700 billion in cloud infrastructure spending this year is explicitly pivoting from building and training models to running them continuously at scale. You do not build that kind of capacity for a tool people use occasionally. You build it for agents that never turn off. If you own the pipes that carry that traffic, you collect.

So the infrastructure play is real. I am not arguing otherwise.

I am however arguing there is a floor above it that is not getting the same attention.

The question that matters next is not who carries the traffic. It is who shapes the intent before anything moves at all.

A new kind of customer is emerging. Not the consumer buying a product. Not the business buying software. The agent, acting as buyer, negotiator, and decision-maker on behalf of a human who set the goal and stepped away. The infrastructure investors can see the traffic volumes. They cannot see the intent layer sitting above those volumes, deciding what gets routed and why.

I keep hearing the same conversation from CTOs, from founders pitching, from people I interview in my research. Everybody is talking about which AI vendor to pick. Which model scores best on benchmarks. That question is already one layer too low. The question underneath it is: who holds the working memory of how your organization makes decisions? Under what contract terms? What happens to that memory at renewal?

The infrastructure companies will collect tolls on 25x traffic growth. That is a very good business. But the companies that accumulate memory and context, the ones that learn how you actually operate over years of continuous interaction, will have something the pipe companies never get: pricing power that compounds.

The pipes carry the traffic. Memory shapes it.

Both layers will make money. But they are not the same bet. And only one of them is getting the scrutiny it deserves.

What would change in your investment thesis if you underwrote the memory layer separately from the infrastructure layer?

Harry Glorikian is the author of The Invisible Interface (Ideapress Publishing / Simon & Schuster, June 2026)