Should we Leapfrog the Terrestrial Data Center Build-out?

I've been using agents pretty heavily for coding and building, I tried to estimate my usage, I think on average I'm using about 100 W of AI (a typical chat user would be more like less than 0.1 W constant draw). This is calculated from the number of tokens I spend per month on Anthropic and OpenAI (I have the max accounts which I fairly fully utilize, there's always a few agents doing stuff for me these days). Of course it's pretty hard to have good numbers for this unless you have access to internal numbers at one of the labs, but I think something between 20 and 200 W is fairly defensible.

How fast does this grow? For a heavy user, like me, I think it will double roughly ever year. This is based on my own usage curves and from what's reported more broadly. The vendors report that developer usage in terms of tokens is increasing by roughly 2x to 4x a year (Claude code, OpenRouter, Gemini, M365 Copilot, etc.), with the number of joules per token dropping every year thanks to efficiency improvements -- call it 1.5-2x/yr, with occasional bigger jumps.

Usage patterns are now very different from the chat-based era. I now routinely have agents work essentially without supervision for up to 24 hours. For the work that I do, there has been a strong transition to delegation in the last four or five months -- so I'm no longer capped by how quickly I can type. Agents spawn their own subagents, and so on, and so I don't really see usage number being limited by any human-based rate-limiting process, and we're left only with the availability of compute.

So what if usage keeps compounding like this?

Starting from 100 W in 2026:

The low case says that in 2035, there's roughly a microwave oven worth of compute power for every heavy user. This is a microwave oven that's always running. The high case seems rediculous, but these Moore's law type things often are, and sometimes they're right.

Then diffusion starts multiplying it

Diffusion is an interesting process. A technology grows very slowly while some bottleneck isn't resolved. Then capability, price, usability crosses a threshold and adoption takes off. Growth remains fast until it reaches the next bottleneck. This results in a logistic curve.

OpenAI claims >5M weekly Codex users in 2026, and a Microsoft/GitHub study talks about 3.2 million Copilot users.

I am going to go with 3 million heavy-users in 2026 as an useful starting point, saturatomg toward a ceiling of roughly a billion knowledge workers.

For calibration: workplace AI use in the US has been diffusing fast ... daily use at work went from 4% to 15% of employees in two years. The 2x column is not crazy: OpenAI reports active Codex users grew more than 5x in the first half of 2026

Compute Power Law

The product of the heavy users and their usage gives us the electricty draw we should expect in the couple of decades. This is plotted below.

Let's look at the crossings off the middle curve (our median case). Heavy agentic users alone, not training (see below), exceed all of today's data centers around 2031, all of US electricity around 2033, and world electricity around 2036. Can this happen? Maybe!

I've ignored training in this 0th order model ... let's assume that it takes a fixed percentage of the inference power. Really there are some caveats -- training runs need to probably be much more centralized than inference. Also with diffusion, the percetage of compute on training will probably go down.

What needs to be built?

Two enormous industrial systems have to rise up for this to happen. Chip factories and electrically powered data centers.

First, we need way more chips. GPUs, TPUs, ASICS, HBMs, lasers, modulators, and so on. We can build more fabs and facilities. This requires a staggering investment and improved supply chain, but it is at least a familiar kind of industrial problem: we build more factories, learn faster, improve yield, and amortize cost over more units.

Powered data centers are harder to build and not just for physics reasons (I'll get to that as well). US grid growth is too slow. Gas turbines are backordered past 2030; large transformers are complicated and take years to build. These can be also be solved with more factories. Gigawatt-scale campuses have now become a political issue, across the political spectrum for a variety of reasons. Even if in principle we can build this, it will take time. Time-to-power-up becomes really valuable.

Space

An important aspect of the above curve is that at some scale, the sensible place to collect the energy and run the next unit of computation is no longer the surface of the planet. Some meaningful fraction of compute then has to move off-world. This may come 10 years from now, it may come 30 years from now. But it's coming!

I'll think more more deeply about space compute at some point, but at this point, I can't find a basic physics reason that prevents it from growing essentially without limit for centuries.

So we will have to transition to space compute in the near future. Maybe before we have fusion, and maybe even before we have quantum computers.

So why not just do it now?

AI was used to research this post, for brainstorming, and to generate the graphics.