Here’s the thing about the AI gold rush: everyone’s watching the wrong miners.
Right now, the entire market is fixated on who’s going to win the AI model wars. OpenAI versus Google versus Anthropic versus whoever’s launching next week. It’s like watching people argue about which website will dominate the internet while completely ignoring the fact that someone’s making bank selling the routers that connect them all.
That someone was Cisco during the dot-com era. They didn’t have to pick the winning website—they just sold the infrastructure every website needed. Their stock went up about 3,400% in five years while Pets.com and a thousand other “sure things” went to zero.
The same thing is happening right now with AI, except there’s a twist that most investors haven’t figured out yet.
The Infrastructure Bill Just Got Insane
Here’s where it gets interesting: AI is about to consume way more infrastructure than anyone’s pricing in.
A chatbot? That’s cute. You ask it a question, it spits back an answer, done. Maybe a few hundred tokens processed. But an AI agent? That’s a different beast entirely. Tell an agent to “grow our market share by 15% this quarter” and it doesn’t just answer—it works. It researches competitors, pulls internal data, drafts campaigns, tests messages, coordinates with other agents, revises, and keeps iterating until the job’s done.
We’re talking 20 to 30 times more compute, memory, networking, and cooling than a simple chatbot exchange. Not 20% more. Twenty to thirty times more.
And this isn’t some sci-fi scenario. More than half of major enterprises already have AI agents running in production right now.
The Six Tollbooths
Think of the AI economy as a superhighway. Every agent task has to pass through six tollbooths, and someone’s collecting at each one:
1. Compute – The chips that actually run the models
2. Memory – Agents need context to remember what they’re doing
3. Networking – Agents talk to databases, APIs, and each other constantly
4. Thermal Management – Dense AI racks generate insane heat
5. Power – Agents run 24/7, so you need reliable electricity
6. Real Estate – All this stuff has to live somewhere
A chatbot taps all six. An agent pounds them.
The Numbers Don’t Lie
Google’s processing 16 billion tokens per minute—up 60% last quarter. Nvidia’s CEO says inference compute demand is already 100 times higher than expected. Hyperscalers are spending like crazy on AI infrastructure. Memory demand is surging. Cooling backlogs are expanding. Power companies are signing long-term deals with cloud giants.
The tollbooth companies aren’t hoping this demand shows up. They’re reporting it every quarter.
The Play
Here’s the beautiful part: it doesn’t matter which AI model wins. OpenAI could dominate, or Google could, or some startup nobody’s heard of yet. Doesn’t matter.
Whoever wins still needs compute. Still needs memory. Still needs networking. Still needs cooling. Still needs power.
The infrastructure companies get paid regardless of who crosses the finish line first.
That’s the real AI story. And most investors are still watching the wrong race.
