Data centers are already straining. GPUs sit above 70% utilization, a number that stops looking abstract the moment you realize it reflects actual paying customers running actual workloads. Gavin Baker, the CIO of Atreides Management, just put numbers to what infrastructure people have been whispering about for months: 500,000 users run agentic AI systems today. If that scales to 100 million, the grid breaks.

The math is brutal. Half a million users represents 0.1% of global population. A hundred million pushes that to 1.3%. Sounds small until you remember that each user doesn't consume the same compute as a Google search. Agentic systems run continuously, making decisions, iterating, burning electricity in ways we've never seen before at scale.

The bottleneck nobody wants to name

Power is the real constraint, not chips. TSMC can squeeze out more wafers, but the physical infrastructure to deliver electricity to data centers is already overbooked. Companies like EdgeConneX are dumping billions into Texas expansion, which tells you exactly how desperate this has become. Baker doesn't expect meaningful relief on the power side until 2027 or 2028.

That gap matters. Early commercial deployments of GPT-5.2, Grok 4.20, and Codex 5.3 are already live. This isn't research consumption anymore. It's revenue-generating workloads tied to actual economic activity. The 70% utilization figure isn't a vanity metric. It's a warning.

Space compute and what comes next

Some companies are exploring orbital infrastructure, deploying compute in space where solar power and thermal conditions beat anything Earth-based. It sounds like science fiction until you realize it's a genuine solution being discussed in serious rooms.

For anyone tracking AI infrastructure as an investment angle, three metrics matter: TSMC's wafer capacity roadmap, power queue timelines at major data center hubs, and how fast agentic AI adoption moves from early adopters into mainstream enterprise. That third one is the variable that either breaks the grid or forces a pause in deployment. Baker's question wasn't academic. It was a warning dressed as a thought experiment.

This material is informational only and should not be taken as investment advice or financial guidance.