Forty-Nine Megawatts

On July 14, Governor Kathy Hochul signed an executive order making New York the first state in the country to pause the construction of new hyperscale data centers. The moratorium covers new facilities with an electrical demand of 50 megawatts or more, freezes state environmental permits for up to a year, and includes a promise to build a nation-leading framework for grid reliability, electricity costs, water, noise, and land use before anything else is approved. Thirty-nine more are somewhere in the application pipeline, which tells you how fast this was moving before somebody hit the brakes.

I think this is a TERRIBLE idea, but there's SOME logic behind it, so let's dive in.

Back in May, I wrote about Loudoun County, the most permissive data center jurisdiction in America, flipping to moratorium proposals in about eighteen months. When the friendliest county in the friendliest state turns hostile, it is no surprise that these things appear more often in other districts and states. The grid genuinely cannot deliver power where people want to put the load, and ratepayers are really eating costs they never agreed to. (I SHOULD NOTE: the reason the rates are going up is NOT that the data centers are raising the cost! It's because these power companies see the opportunity, take it, and are heavily regulated. A one-year pause to write actual rules, in a state where thirty-nine applications were stacked up like planes over LaGuardia, is a defensible thing for a governor to do. The same week, for what it's worth, Microsoft was cutting the ribbon on $3.3 billion in Wisconsin, so nobody should mistake this for a national trend yet.

But let's talk about the guidance - it's JUST at 50 megawatts. How'd they choose that?

Every threshold becomes a spec sheet

France has a rule that at 50 employees, a firm crosses into a different regulatory universe: works councils, union delegates, profit-sharing obligations, formal restructuring plans. The result is one of the most famous pictures in empirical economics. Plot the distribution of French firms by headcount, and there is a pileup at exactly 49 and a crater just past it. Firms do not grow in a straight line. They stop at it, split in two, spin off subsidiaries, push work to contractors, whatever it takes to stay a hair under. The economists who studied it, Garicano, Lelarge, and Van Reenen, found the distortion was big enough to measure in fractions of national output. Not because anyone cheated, but because the line was published, and published lines get engineered to.

This is not a French quirk. Federal law caps trucks at 80,000 pounds gross weight, so an enormous amount of American freight rolls at 79,900 pounds. Coastal shipping fleets around the world are full of vessels built to slide just under whatever tonnage triggers the next tier of crewing and inspection rules. Buildings stop one floor short of the elevator code. Wherever a regulation names a number, an industry grows a callus at that number. The market never argues with a threshold. It reads as a product requirements document.

So here is my prediction for New York: a wave of 49-megawatt data centers, with facilities designed to the line the way French firms are staffed to it. Campuses that are legally three separate 45-megawatt projects with three separate applications and, remarkably, three separate fences. Developers did not stop wanting to build in New York on July 14. They started reading the order for its edges that afternoon, billing by the hour.

The accidental architecture

But here's the funny bit... a 49-megawatt data center is a good idea!

A fleet of smaller facilities, spread across a state instead of piled onto one substation, sited where the grid has actual headroom or next to their own generation, closer to the people and data they serve: that is what the physics has been begging for the entire time. The gigawatt campus exists because it is convenient for the operator, not because the workload demands it. Training wants tight coupling; the inference and agent workloads that are actually growing mostly do not care, and demand is on track to nearly triple by 2035, whether we build it in three monoliths or three hundred sheds. New York, in trying to stop data centers, may have accidentally written the first state-level specification for distributed compute. The moratorium is the mandate.

THAT SAID, twenty 49-megawatt buildings running software that assumes one big building is not distributed computing. It's a monolith with extra steps and worse latency. The hard part was never pouring smaller slabs; it was treating computers spread across a state as one coherent system, moving the work to where the capacity and the data already are, instead of hauling everything to a headquarters that no longer exists. The real estate industry will complete the 49-megawatt building in about six months. The software model that makes a hundred of them worth more than the sum of their parts is the actual project, and just dealing with land/power/shell is not going to get you there.

New York thinks it drew a boundary. What was published was a design constraint, and the one thing this industry reliably does with a design constraint is build to it, at volume, faster than the regulator can schedule the follow-up meeting. The question worth watching is not whether the moratorium survives the year. It's what gets built at forty-nine.

Want to run compute where the power and data already live, instead of where the substation used to be? Check out Expanso. Or don't. Who am I to tell you what to do?

NOTE: I'm currently writing a book based on what I have seen about the real-world challenges of data preparation for machine learning, focusing on operational, compliance, and cost. I'd love to hear your thoughts!