AI Agents Don't Sleep. That Changes the Power Math.

A post went viral this week. 2.7 million views on X. Not a new model launch, not a fundraise announcement. A setup guide.

Nick Spisak published instructions for chaining three open-source tools into what he calls a one-person AI company. Paperclip manages the org chart and assigns work. Gstack (from Y Combinator CEO Garry Tan) gives your agents 15 specialist roles. Autoresearch (from Andrej Karpathy) runs autonomous experiments overnight: 12 per hour, roughly 100 by morning.

Setup takes ten minutes. The instructions are copy-paste terminal commands. And the recommended workflow is: start a sprint, go to sleep, review what shipped when you wake up.

That last part is what matters for power planning.

The Load Profile Shift

Datacenter power models already account for inference growth. Goldman Sachs, IEA, McKinsey, everyone. That isn't new. What IS new is the shape of the demand curve.

Consumer AI usage today looks like web traffic: peaks during business hours, drops overnight, spikes on weekdays. It's bursty. Predictable. Familiar to anyone who's planned capacity for a SaaS product.

Autonomous agents don't follow that pattern. Spisak's recommended setup runs 15 agents through planning, code generation, browser testing, and deployment in continuous loops. Overnight. Unsupervised. For hours. Karpathy's autoresearch runs experiment after experiment on a fixed five-minute cycle, generating and evaluating code until you tell it to stop.

This isn't a chatbot conversation. It's a batch workload. And batch workloads run 24/7.

Why That Distinction Matters

Bursty inference demand can be served with flexible capacity. Solar-plus-battery in Texas can handle afternoon peaks. Demand response programs can shave spikes. The intermittency challenge is manageable when your load follows the sun.

Flat, continuous inference demand is a different problem. It looks like baseload. And baseload favors dispatchable generation: natural gas, nuclear, geothermal. The resources that run at 2 AM the same as 2 PM.

The 40 GW of datacenter-designated gas capacity currently in the Texas pipeline was justified primarily by training workloads, which are inherently continuous. If inference demand shifts from bursty consumer chat toward continuous autonomous agent workloads, that same capacity justification gets stronger, not weaker.

This matters because the political case for datacenter power often hinges on utilization. Critics ask why build generation capacity for a load that's only heavy 12 hours a day. Autonomous agent workloads weaken that argument considerably.

The Scale Question

How many people will actually run autonomous agents overnight? Nobody knows yet. The tools exist today, they're free and open source, and a post explaining how to use them just hit 2.7 million views. That's the supply side.

The demand side is harder to project. Most readers won't install anything. But the ones who do won't use AI the way they used to. A developer who previously made 50 API calls during a workday will now dispatch agents that make thousands of calls overnight while they sleep. Per-user consumption jumps by orders of magnitude.

Anthropic and OpenAI are already seeing this. Claude Code and similar agentic tools generate dramatically longer sessions with more API calls per user than chat interfaces. The per-session token consumption for agentic use cases can be 10-50x higher than conversational use. Scale that across every agent framework gaining traction right now (CrewAI, AutoGen, LangGraph, OpenHands, and dozens more), and you get a meaningful shift in how inference infrastructure needs to be provisioned.

Not just more capacity. Different capacity. Capacity that runs all night.

What to Watch

Three signals will tell you how fast this shift is happening:

API pricing changes. If Anthropic and OpenAI start offering off-peak inference pricing (cheaper tokens at night), that confirms they're seeing load flatten toward 24/7 patterns and want to manage it.

Inference-optimized facilities. Training clusters and inference farms have different hardware profiles. Watch for announcements of facilities specifically designed for high-throughput, continuous inference rather than massive parallel training.

Agent framework adoption. GitHub stars and npm downloads for agent orchestration tools are the leading indicator. When CrewAI or LangGraph hit the same adoption curve as React or Next.js, the load profile shift will be well underway.

The Takeaway

The argument for datacenter power buildout doesn't rest on AI agents. Training demand alone justifies the pipeline. But the shift from bursty chat to continuous autonomous workloads changes the demand *shape* in ways that favor the exact generation resources Texas is building.

2.7 million people just read instructions for running AI agents overnight. Most won't act on it. Enough will that the overnight inference load starts looking like something worth planning for.

The agents don't sleep. Plan accordingly.