The Data Center
Question

For most of history, finding a drug meant a chemist with a hunch and a decade. Then protein folding — a problem that had eaten 50 years of biology — got solved by a model. AlphaFold turned structure prediction from a PhD thesis into an API call.

It kept going. Isomorphic Labs shipped IsoDDE in February 2026, a unified drug-design engine that roughly doubled AlphaFold 3's accuracy on the hardest cases — 50% versus 23.3% where the target looks nothing like anything in the training data. Those are exactly the targets nobody could drug before. The company now has 17 active programs across oncology, immunology, and cardiovascular disease, with its first AI-designed cancer drug tracking into Phase 1 trials by the end of 2026.

Here's the part that matters for this page: that pipeline is compute-bound, not idea-bound. Optimized training runs for protein-folding models now use 2,080 H100 GPUs to finish in ten hours what originally took a week. Every one of those GPUs lives in a building with a substation and a cooling plant. When you ask whether we should build data centers, you are also asking how fast we get to run the experiment that finds the thing that kills the cancer.

Every era has one industry that decides who sets the rules. Steel did it once. Oil did it for most of a century. Right now it's compute — because compute is what turns into weapons research, drug pipelines, code, logistics, intelligence analysis, and the models that everyone else ends up renting.

America is currently ahead, and the margin is not small. US and allied manufacturing capacity for advanced AI processor dies runs an estimated 35 to 38 times China's, quality-adjusted. Huawei's yields sit around 5–20% against Nvidia's 60–80%. Absent chip exports, US installed compute capacity in 2026 would be more than ten times China's. On spending: America's four largest hyperscalers plan roughly $650 billion this year, while Alibaba — China's most aggressive builder — committed $53 billion across three years.

But a lead in chips only becomes a lead in capability if somebody plugs them in. A GPU in a warehouse is a paperweight. The data center is the part where the advantage becomes real — and it's the part that has to be built here, on our grid, under our law, by people we can subpoena. You cannot offshore a substation.

The numbers, plainly. Federal debt held by the public reaches roughly 101% of GDP in 2026. The FY2026 deficit runs about $1.9 trillion, 5.8% of GDP. Interest alone costs $1.0 trillion this year — which is more than the $885 billion for national defense and more than the $708 billion for Medicaid. CBO has interest reaching $2.1 trillion by 2036.

There are exactly three ways out of a debt-to-GDP ratio like that: raise taxes enough to close a 5.8%-of-GDP gap, cut spending enough to close it, or make the denominator bigger. Every country that has actually escaped a debt load like this did the third one. You grow.

And growth is what this buildout is currently producing. AI-related capital expenditure contributed roughly 1.1 percentage points to GDP growth in the first half of 2025 — outpacing the American consumer as an engine of expansion. By Q1 2026, investment in AI data centers, hardware, and networking amounted to 1.4% of US GDP, up from 0.7%, and roughly half of headline GDP growth traced to categories where the AI buildout is the dominant source of new demand.

That's a real, live, measurable thing pulling in the right direction on the worst number in American public finance. It is not the whole answer. It is the only part of the answer currently showing up in the data.

The drug pipeline gets the headlines, but the machines are general. The same racks that fold proteins also run hurricane track models that give the Gulf Coast an extra day of warning, screen candidate battery chemistries and superconductors without touching a beaker, model plasma confinement for fusion, plan the transmission upgrades the grid has been deferring since the 1990s, and tutor a kid in a rural school district who doesn't have a calculus teacher.

None of that is speculative technology. It's existing technology waiting on capacity. The bottleneck between "we could run that experiment" and "we ran it" is almost always a queue for GPUs. Every megawatt that comes online shortens somebody's queue.

This is the difference between a scarcity mindset and a growth mindset, and it's worth being precise about it. Scarcity says the pie is fixed, so the fight is over slices. Growth says we have repeatedly, historically, and recently made the pie bigger — and the mechanism for making it bigger has always been building physical infrastructure that lets people do more with the same day.

The most common mistake in this whole debate is treating "don't build it" as the neutral option — the safe default where nothing changes and nothing is risked. That's not one of the choices on the table.

If a town says no, the campus goes to the next county. If a state says no, it goes to the next state. If America says no, it goes somewhere with worse labor law, dirtier power, and a government that would very much like to read what's on those drives. The compute gets built. The only open question is who gets the jobs, the tax base, the leverage, and the say.

Which is exactly why the rest of this page is not a sales pitch. Saying yes to the technology and yes to any deal offered are completely different positions, and the gap between them is where a community either gets rich or gets used. So: here is every argument, both directions, with the numbers attached.