The two arguments over every data center
Two speeches about the AI buildout, side by side and measured against one spreadsheet: national champion, or a machine for moving money out of a county.
Todd Ruffner-Schoenfeld
Editor in chief. A knack for the fine print, and likes it.
The two arguments, side by side
Where the cost-shift is real
Map both against the same sheet and the extraction case has the harder evidence in one place: the mid-Atlantic. PJM’s independent market monitor, Monitoring Analytics, attributed $29.4 billion of the capacity charges across the last four auctions, about 46 percent, to existing and forecast data-center load. That is a real transfer, from a private customer’s private demand onto everybody’s bill, and it is why New York paused and why Virginia wrote the GS-5 tariff to wall households off. But the same monitor’s number does not prove that every rate increase in Florida or Texas is the same story. Those states have their own utilities, their own fuel costs, their own weather. The extraction speech is strongest inside PJM, where the hard numbers are. It thins out the further you carry it from there.
The strongest case for building
The champion sentence is also real, in one narrow respect. Training clusters and inference farms are lumpy things to build, and a two-year permitting delay really is a two-year hole in a product cycle. National-security officials say this in classified briefings and then, more softly, in public. They are not only lobbying. What does not follow is that every 80-acre rezoning in a dry county is a carrier strike group. A moratorium that is really a one-year planning pause, like New York’s, is not a ban on building intelligence. Treating the two as the same thing is how the champion speech stays undefeated: it answers a fight nobody is having. The honest version of its case survives; the version that calls a comprehensive-plan update a surrender to Beijing does not.
The chips still come from overseas
Foreign capital and foreign silicon sit in the middle of both speeches, and neither likes to mention it. American firms own most of the famous campuses, which is what the “we built it here” line is pointing at. But the chips that make the campus worth building are another matter. The advanced GPUs come from TSMC, fabricated and packaged largely in Taiwan; the high-bandwidth memory stacked next to them comes mostly from SK hynix and Samsung in South Korea. A “we built it here” talking point can be true about the concrete slab and false about the rack sitting on it. When an incentive shows up on a county agenda, the question worth asking is which one it was written for: the building, or the machine that could be unplugged and shipped to whichever county bids next.
How to read both
Neither speech is a lie, and that is what makes them dangerous. Each is a true sentence wearing a much larger coat. A county does not have to pick a team to protect itself; it only has to make each speech show the smaller, checkable claim underneath the big one, and decide on that. The largest thing both speeches talk past is not rhetorical but mechanical: before any of this gets built, it has to wait in a line nobody on stage mentions. Part 6 is about that line.
- Monitoring Analytics (PJM’s independent market monitor): about $29.4 billion, roughly 46 percent, of capacity-market charges across the last four auctions attributed to existing and forecast data-center load. Utility Dive.
- Advanced AI GPUs are fabricated and packaged largely by TSMC in Taiwan; high-bandwidth memory comes mostly from SK hynix and Samsung in South Korea, so the hardware supply is concentrated overseas even when the campus is U.S.-owned. CNBC.
- The national-security framing of compute and export controls, and its limits, in the Congressional Research Service’s analysis of AI chip and data-center policy. Congressional Research Service.
- New York’s one-year statewide pause (Executive Order 62) and Virginia’s GS-5 large-load tariff as regulator responses to the cost-shift argument. Office of the Governor of New York.
