← All Notes

New York just banned large data center construction for a year. First state to do it. Won't be the last.

Governor Hochul calls it a pause to develop "consistent standards." Translation: we don't want your power consumption, your water usage, or your heat. Build somewhere else.

This is how it begins. Not with a grand policy debate about AI's future. With local resistance to infrastructure. With electricity bills and aquifer depletion and neighbors who don't want a 50-megawatt facility humming next door.

The AI industry is panicking. They see regulatory capture, competitive disadvantage, the end of American leadership. They're not wrong to worry. But they're missing the actual story.

This might be the bottleneck that matters.

The Physics Problem Nobody Wants to Discuss

We've spent two years obsessing over model capabilities, training efficiency, inference optimization. All of it assumes one thing: you can plug in and draw power.

That assumption just broke.

You cannot run frontier AI without massive compute. You cannot run massive compute without massive energy. And you cannot draw massive energy without someone, somewhere, paying the cost. New York just said: not here, not now, not until we understand what we're trading.

Other states will say yes. Texas, Wyoming, states with cheap land and fewer regulatory hurdles. The compute will move. The industry will adapt. Data centers will cluster where they're welcome, and the geography of AI will reshape around energy access and political will.

But this is a temporary solution to a permanent problem. Every location has limits. Every grid has capacity constraints. Every community will eventually ask: what are we getting for what we're giving up?

The real question isn't whether New York kills AI development. It's whether Earth can sustainably host the compute requirements we're building toward.

AI Latency Changes the Equation

Here's what most people miss: AI inference is not like traditional compute workloads.

When you load a webpage, latency matters enormously. Every millisecond of delay costs conversions, engagement, revenue. Your CDN needs to be geographically close. Your database needs to respond instantly.

AI is different. Inference already takes seconds, sometimes longer. Users expect it. A language model thinking for three seconds feels normal. A reasoning model taking thirty seconds is acceptable for complex tasks.

That latency tolerance creates new possibilities.

The round trip to low Earth orbit is roughly 5-10 milliseconds at light speed. Double it for processing overhead, add packet switching, and you're still well under 100ms. For most AI workloads, that's noise. Unnoticeable. Irrelevant.

Which means orbit becomes viable.

Not today. Not next year. But as launch costs continue dropping and as ground-based resistance continues rising, the economics shift. Solar power is free and unlimited in space. Cooling is a radiation problem, not a water problem. No neighbors to complain. No local grid to overwhelm.

You build the infrastructure once, and it scales.

The Orbital Compute Thesis

I run AI products in production. I know what compute costs look like. I know what happens when you try to scale inference across geographic regions with different energy prices, different latency profiles, different regulatory environments.

The complexity is brutal. Every optimization you make in one dimension creates tradeoffs in another.

Orbital compute solves several problems simultaneously:

The technical challenges are real. Radiation hardening, thermal management, maintenance and repair, initial capital costs. But these are engineering problems, not physics problems. We know how to solve them. It's a question of when the economics make sense.

New York's moratorium accelerates that timeline.

Every jurisdiction that says no to ground-based data centers makes the orbital alternative more attractive. Every state that imposes energy taxes or water restrictions or environmental reviews shifts the calculation. The more friction on the ground, the more compelling the case for orbit.

Infrastructure Precedes Civilization

Here's the pattern that repeats throughout history: infrastructure comes first, then people follow.

Railroads opened the American West. Highways created suburbia. Fiber optic cables built the internet economy. The infrastructure enables economic activity, and economic activity creates reasons for humans to be present.

If we build massive compute infrastructure in orbit—and I believe we will—humans will follow. Not immediately. Not en masse. But the presence of valuable, maintained infrastructure creates jobs, creates economic zones, creates reasons to stay longer and build more.

Orbital colonies in fifty years isn't science fiction. It's the logical endpoint of decisions we're making right now about where to put compute.

You want to understand the future of AI? Stop watching model benchmarks. Start watching where the data centers go.

The Choice We're Not Making

New York thinks it's buying time to develop standards. What it's actually doing is selecting out of the next phase of infrastructure development.

The compute will go somewhere. The question is whether it stays on Earth or whether we start building up.

I know which way I'd bet.