AI Is a Physical-Infrastructure Problem First

Public conversation treats large-scale AI as a software achievement that happens to need buildings. For planners, the dependency runs the other way: the physical layer beneath the compute is decided first, and it caps what the software above it can do.

The order most conversations use is reversed

The dominant framing of AI is a software framing: model architectures, training runs, inference costs, the pace of new releases. That framing is accurate about where the capability comes from. It is misleading about where the constraint lives.

For anyone responsible for siting, powering, or approving large-scale compute in Wisconsin and the broader Midwest, the practical sequence is close to the opposite of the public one. Before a model is trained or served at meaningful scale, the location hosting it has to resolve a set of physical questions: whether power can be delivered to the site, whether the heat can be rejected, whether water is available and permitted, whether the land supports the build, and whether all of it can be completed on a timeline the project can absorb. These are settled, or not, well before any software decision matters. They are the ceiling. The compute is what fits underneath.

Naming that ceiling is useful precisely because it separates plans that are physically bounded from plans that are asserted. A roadmap can be sophisticated about models and still be structurally optimistic about the site, because it assumes the physical layer will keep pace with the software layer. It generally will not, because the two advance on different clocks.

What the ceiling is made of

The physical layer is not a single constraint but a small set of them, each with its own floor.

The first and usually most binding is deliverable power. It is not enough for generation to exist somewhere on the system; the capacity has to be connected and deliverable to the specific site. Data centers are the single largest driver of projected load growth across the regional system through 2035, and the interconnection and transmission work required to serve new large loads carries lead times measured in years. That work (studies, approvals, substation construction, transmission upgrades) runs on the grid’s calendar, not the developer’s, and it frequently dominates the overall schedule.

The second is cooling. Compute at AI density produces heat that has to be moved off the equipment continuously and rejected somewhere. The cooling method chosen is not a detail bolted on at the end; it shapes the electrical load and the water requirement of the entire facility.

The third is water, where evaporative heat rejection is used, along with the land the campus occupies and the permitting each of these triggers. None of these is exotic. What makes them a ceiling rather than a checklist is that they take years to establish and cannot be compressed by the speed of the software they are meant to serve.

Efficiency does not raise the ceiling

A natural objection is that AI is getting more efficient, so the physical demand should ease over time. The historical record and the current dynamic point the other way.

Over roughly the last two decades, Wisconsin’s total electricity use actually declined, on the order of nine percent, as efficiency gains and industrial plant closures reduced consumption. That is the pattern most planners internalized: efficiency lowers demand. Under AI growth, that relationship does not hold. When compute becomes more efficient, the efficiency is reinvested into more compute rather than banked as reduced consumption, so total demand continues to rise. The ceiling does not lift as the technology improves; the pressure against it increases.

This is why efficiency improvements, while real and worth pursuing, are not a substitute for physical capacity. They change how much useful work a given megawatt performs. They do not remove the need to deliver the megawatts.

The ceiling is one interdependent system

The physical layer also resists being optimized one piece at a time, because power, cooling, and water are coupled. Moving one moves the others.

Closed-loop cooling reduces on-site water consumption but increases electricity use. Liquid cooling improves thermal performance at high density, but it does not reduce water consumption if the heat is ultimately rejected through an evaporative tower. Procuring renewable energy addresses the carbon profile of the electricity but does not, on its own, remove the need for firm deliverable capacity, transmission, or backup. Each of these is a real choice with real merit; none of them is free of consequences elsewhere in the system.

The implication is that the ceiling is set by the physical plant considered as a whole, not by its most favorable single component. A plan that optimizes for low water use, or for renewable supply, or for high-density cooling, without accounting for what that choice does to the other two, has not actually lowered its ceiling. It has moved the constraint somewhere less visible.

What this means for planners

The reframing is straightforward to state and consequential in practice. The physical feasibility of a location is not a downstream engineering concern to be resolved once the compute plan is set. It is the upstream boundary that the compute plan should be built inside.

In practice that means establishing, early, what a site can actually support (how much power can be delivered and by when, how heat will be rejected and at what electrical and water cost, and what the full permitting and construction timeline looks like) and treating projected compute as bounded by those answers rather than driving them. It also means reading efficiency and procurement choices as trade-offs within one system, not as independent wins.

None of this argues against AI development. It argues for sequencing that development against the physical realities that determine whether it can be built and operated where it is proposed. The projects most likely to be delivered on something close to their stated terms are those that treated the physical ceiling as the first question rather than the last.

Closing thought

The public story of AI will remain a software story, because that is where the visible progress happens. The buildable story is a physical one, decided in interconnection queues, cooling design, water permits, and multi-year construction schedules that the software curve does not accelerate.

For institutions weighing large-load development, the open question is where the physical ceiling of a given location should be established in the process, and how compute projections should be tested against it before, rather than after, commitments are made. We would welcome perspective from utility planners, developers, and public-sector reviewers on where in their own processes that boundary is currently set, and whether it is being set early enough to be useful.

Source: the Wisconsin AI Infrastructure Readiness Brief, AI infrastructure reality — the physical layer as the binding limit (power delivery, cooling, water, land, multi-year timelines), data centers as the single largest driver of MISO load growth through 2035, the ~9% decline in Wisconsin electricity use over roughly two decades, and the interdependence of power, cooling, and water choices.