Data centers need the grid — so we have to make the grid better

Almost none of this post is Lubershane’s. He writes an introduction of about ten paragraphs and then cross-posts an essay by Tim Hade in full, and unlike other guest sections in this publication there is no handover back at the end, so everything from the first subheading onward is Hade’s, including his own numbered source list. Hade’s starting point is that the constraint on AI in 2026 is neither chips nor talent but the grid, and that what the grid keeps running out of is not generation but slack: spare transformer capacity, spare substation bays, spare thermal margin when a single element is out of service, spare time in equipment supply chains. That converts the question from how to build more power plants into how to build more power system capacity fast, and the two have different answers. It is a calendar problem rather than an ideological one, he argues: data centers are built on roughly an 18-month clock and bulk power equipment on a 50- to 80-month one.

The seductive answer is to stop waiting and bring your own power. Hade’s objection is not that co-location fails but that it is misdescribed. A grid connection is not a wire; it is access to a system supplying voltage stiffness, frequency stability, high fault current and inertia, so stepping behind the meter changes the physics and not merely the ownership. Data centers make that harder than most loads would, being dense in power electronics, drawing non-linear current, ramping fast and designed to be intolerant of disturbance, so an islanded system has to absorb harmonics, step loads, a lower and softer fault current that can blind protective devices, and the risk of one transient tripping everything at once. All solvable, none trivial. Announced megawatts are also not delivered megawatts, especially where the project is really three projects: a data center, a power plant, and the microgrid controls tying them together. Nor is it an unregulated lane, and he reads a federal rejection of an expanded co-located arrangement at a nuclear plant as a warning that the rules are being written in real time. His conclusion is deliberately bounded: co-location will work in some places, for firms that can execute power-plant-grade projects, and cannot be the only plan.

The load is not a fixed object either, and that is where the argument turns. Training clusters are large, concentrated and indifferent to latency, so they can be sited where land and transmission already are. Inference cannot be, because it has to sit near users, which pulls compute toward metro areas, distribution substations and congested feeders, exactly where equipment bottlenecks and permitting friction are worst. Inference is also spikier than training, and spiky load at a constrained node is not a job for slow thermal plants but for fast, controllable resources. So the practical question becomes how to energize a large load sooner without breaking reliability, and the answer forming in the market is to split first power from ultimate firm service: a load that accepts an enforceable operating envelope can be connected while upgrades catch up. But an envelope only works if flexibility exists on the other side of the contract, and a curtailment obligation is not by itself a strategy. Hence the toolkit of demand response, grid-enhancing technologies and distributed batteries, of which he rates batteries highest because they are modular, fast and placeable. Placeable is the load-bearing word: for interconnection, location is destiny, and a megawatt of flexibility at the constrained substation is worth more than ten far away. The scale then comes from large commercial buildings, the overlooked middle between millions of homes and a handful of megaprojects.

His last move is that the binding constraint is not technical at all. We know how to curtail load, deploy batteries and instrument lines; what is missing is whether institutions can value, verify and accept flexibility as a substitute for steel, with capacity accreditation for aggregated portfolios the hardest part. Federal rules have opened a market pathway for aggregations of distributed resources, but implementation is uneven and the difficult pieces, coordination with distribution utilities, telemetry standards and accreditation itself, are unfinished. The political argument is where he ends: the coalition for using the existing grid better is broader than the coalition for new corridors, and a hyperscaler funding batteries across local commercial buildings converts a permitting fight into a visible local benefit. The closing frame is a choice between fortified private islands, each solving its own problem, and a better shared grid built partly with the capital already pouring into AI.