How 230-kW AI racks are shifting data center capital to switchgear and rate design

Siemens Grid Software put a figure on the constraint during a webinar recapped by Utility Dive: AI racks drawing 230 kilowatts, planning advice to work in scenarios rather than settle on a single design case, and a pointer to Dominion Energy's Northern Virginia experience of coordinating early with large loads. The recap does not put a megawatt threshold on when that coordination starts, or describe what the sequence asks of a developer, which is the part anyone modelling a schedule would want. What it does establish is that the vendor selling the electrical gear and the utility connecting the load were both spending their time on how a connection gets sequenced rather than on how many megawatts to add.

The engineering half of the same problem is older than the AI hall. A DCD opinion piece draws a line between the failure modes an engineer can name and the interactions between healthy systems that millisecond load swings can set off. Redundancy playbooks are written for the transformer that fails and the breaker that trips. The second category is what happens when nothing fails and the load still moves faster than the controls assume, and another feed, another switch, another layer of redundancy built to the same assumption does not cover it.

From a developer's seat, coordinating early means the electrical design arrives before the interconnection request is filed. That puts design work and capital into a project sooner and stretches the interval in which money is committed to an asset that cannot yet deliver anything, which is a cost measured in months rather than in megawatts. Scenario-based planning is the same shift seen from the equipment side: the design case stops being a single number, so the specification grows, and what has to be bought against it grows with it.

Designing to scenarios rather than one case implies power trains sized above the average the hall will draw, and the equipment that absorbs the oversizing is switchgear and cooling rather than turbines. It also moves the buyer of record for reliability inside the fence, to the operator who has to hold the load steady rather than to the developer who signs the interconnection agreement.

For an allocator sorting the pipeline, the cut that may matter most runs between assets that need a queue position and assets that do not. Equipment capacity, cooling and headroom on lines already in the ground can be underwritten against a delivery date; a generation project carries a regulatory timetable it does not control. That distinction, rather than the headline megawatt count, is closer to what the current announcements actually price.

Vertiv's plan for its Nové Mesto nad Váhom site in Slovakia is the clearest read on where that money goes: a 22,000-square-meter expansion on a two-year buildout, adding switchgear and liquid cooling lines, with hiring set for 2027 through 2029. Both product lines treat power after it reaches the site, switchgear conditioning and switching it while liquid cooling removes heat that air cannot move at these densities, and neither line needs a queue position to be built.

Staffing production through 2029 assumes the density problem stays a procurement problem for years, which means the halls already standing have to be re-engineered for loads their original designs did not contemplate, alongside the ones still on drawing boards. If that assumption holds, the equipment cycle runs longer than the capex wave currently paying for it, and a Slovak plant is a position on the retrofit market as much as on new construction. Vertiv is not the only company able to read the same number. It is one of the few that has put square meters behind the reading.

Washington's answer to the same constraint arrived as a $1.9 billion package. SPARK buys 23 gigawatts of capacity on transmission already in the ground, and finding capacity on the existing system is a schedule as much as a quantity, which moves the scarce item from the wire to the right to use it. Headroom on a line is shared, and who may draw on it, and in what order, is settled through a tariff rather than an interconnection study. That is why the grid's accounting for AI load is being argued in rate cases rather than in generation queues.

Michigan's ratemaking push covers the revenue side, and its simpler half is a longer filing calendar, more time between rate cases, which slows how quickly capital compounds into customer bills. The half that changes the arithmetic attaches earnings to a measured result rather than to the size of the asset base. Regulated utilities have sold investors on spending that earns. A performance standard asks for a different justification for growth, and it puts a price on the reliability the AI buildout keeps asking the grid to produce.

Taken together, the two packages describe a system asked to carry more load through the same wires and to be paid on different terms for doing it. Neither adds a generator. PWD's count of the pipeline makes the schedule concrete: an 87 percent rural generation share leaves county boards and interconnection dockets, not land, as the binding constraint on the next tranche of clean capacity. The wires are being asked to do more because the alternative, new supply sited, permitted and connected, is slower than the load.

New supply still has to explain what happened to Oklo. FERC kept the company's 750-megawatt Virginia project out of PJM's interconnection queue on procedural grounds, leaving a three-technology plant at least 18 months behind and handing five similar complaints the same reasoning. Nothing in that is a finding about the technology. It is a finding about the timetable, and the timetable is what a data center financed against a construction calendar needs. A project with a Virginia address and large loads within reach of it gave up time to process.

Skanska's $84 million Georgia award prices at $343 a square foot, which buys a shell and a schedule, and the announcement names no power arrangement — unremarkable in a contractor's release, and still a reminder that a signed construction contract is not evidence a site will energize on time. In Northern Virginia, Prince William County shrank its by-right data center zoning overlay by two thirds, pushing more of the region's pipeline into discretionary review. Siting is getting slower and the electrical design inside the fence is getting harder, and only one of those yields to a purchase order.

The commitments attached to the rest of the pipeline are thin in the same way. Lirion Power's deal to acquire a 77-megawatt Irish wind portfolio from GIP carries no purchase price, offtake counterparty or completion date. ACEN's sale of 10 percent of a 250-megawatt Indian solar project names neither buyer nor valuation, which leaves the implied price per megawatt unstated. Capacity is the cheap number to publish. The counterparty, the price and the energization date are the expensive ones, and the far end of the trade shows the same asymmetry: Google's Project Suncatcher test is four chips and one kilowatt, with radiation tolerance the test that matters and mission life the number orbital compute has to earn before it competes with anything on the ground.

Where capital has been raised against a date rather than a drawing, the terms say so. Nscale's $3.36 billion comes as notes converting to ordinary shares at listing, a wager on a public price rather than on a data center fleet 5 percent switched on. Vocus's 4,000-kilometer Brisbane-Darwin fiber lands in 2030 with no tenant, cost or contract named. Low Carbon's 500-megawatt Oxfordshire plan arrives without a buyer, and Vattenfall's 429 megawatts in England without a site, a consent status or a counterparty.

Vertiv's Slovak lines will be tooled and staffed on the premise that reliability is a product with a buyer, and the plant's hiring window runs to 2029. Oklo's Virginia project is at least 18 months behind, with five complaints waiting on the same reasoning. Capital that needs a dated return can take the first of those and not the second, which is why the reliability layer drew the money. Whether the utilities connecting those racks may sell reliability in the first place is what the Michigan package would settle, and the five complaints will show how far the procedural reasoning travels.