The Forecast Nobody Stress-Tested: Demand Overestimation as a Failure Mode

When large AI infrastructure projects fail, the cause is rarely the technology. More often it is a demand forecast that no one pressure-tested: a projection that was optimistic, unexamined, and load-bearing. This failure mode is boring, knowable, and largely avoidable, which is precisely why it deserves attention.

The failure is undramatic, which is why it persists

Public conversation about AI infrastructure tends to concentrate on the hard, visible constraints: whether there is enough power, whether transformers and switchgear can be procured in time, whether cooling and water can meet the thermal load. Those constraints are real. But the single most common way these projects disappoint is quieter than any of them, and it sits upstream of all the engineering.

It is a demand forecast that was never stress-tested. Investment gets scaled to a projected load. Growth then arrives more slowly than the projection assumed, or takes a different shape entirely, and infrastructure sized for demand that has not materialized becomes difficult to justify and expensive to carry. Utilities have cautioned publicly against over-building generation that could become a stranded asset if the current pace of AI expansion slows.

What makes this failure mode distinctive is that nothing appears to break. There is no catastrophic outage, no supplier collapse, no permitting reversal. The project is engineered competently and built well. The error was made much earlier, in a planning assumption that everyone treated as settled and no one owned. That is why it is worth naming directly: a risk that looks resolved at the outset is easy to carry all the way to commissioning before its cost becomes visible.

Contracted demand versus anticipated demand

The preventive discipline is straightforward to state, if not always to practice: scale investment to contracted demand rather than to anticipated demand.

The distinction is not semantic. Anticipated demand is the number in the projection. It is the load a developer, tenant, or planner expects will eventually be there. Contracted demand is the load someone has actually committed to take, under an agreement, with consequences for both sides. The two are frequently very different, and the gap between them is where over-building happens.

The illustrative case is deliberately simple. One does not build a 300 MW plant to serve a 100 MW initial need on the expectation that the remaining 200 MW of load will arrive later. The initial need is real and contracted; the balance is a forecast. Committing capital to the forecast, rather than to the commitment, is the essence of demand overestimation. It converts an optimistic assumption into fixed infrastructure that must be paid for whether or not the assumption holds.

None of this argues against planning for growth. It argues for a particular sequencing of it: building to what is committed, while keeping a defined and low-cost path to expand as commitment materializes, rather than pre-building to a projection and hoping demand rises to meet the concrete already poured.

The uncertainty underneath the forecast

The reason contracted demand deserves this weight is that the demand signal itself is unusually young and unusually uncertain. Infrastructure decisions are made on a multi-decade horizon. The generation, substations, and transmission that serve a large load are financed and depreciated over thirty years or more. The AI demand curve driving today’s projections is a small fraction of that age, and its durability is genuinely unsettled.

A consumer advocate framed the problem plainly, asking how anyone can be confident that “we’re going to be talking about the same thing three years from now.” The question is not rhetorical or dismissive; it is a precise statement of the mismatch. A thirty-year asset is being justified by a demand expectation that may not be legible even three years out. When the horizon of the commitment vastly exceeds the horizon of the confidence behind it, the prudent response is to narrow the commitment to what is actually known, not to widen the forecast to fill the asset.

This is what separates demand overestimation from ordinary business risk. Every forecast can be wrong. The specific hazard here is a structural asymmetry between how long the infrastructure lasts and how briefly the demand behind it can be seen with any clarity.

Structuring commitment: step-in and conversion clauses

If the discipline is to build to contracted demand while preserving a path to grow, the mechanism that makes it workable is contractual rather than technical. Step-in and conversion clauses are the instruments most often discussed for this purpose.

Broadly, these provisions tie the scale and timing of investment to actual uptake. They allow a project to expand, convert, or reassign capacity as demand is confirmed, and they distribute the exposure created when a forecast and a build diverge. The details vary by arrangement, and structuring them well is not trivial. But their function is consistent: to keep commitment aligned with materialized demand rather than with a projection, so that the party benefiting from growth is also the party carrying the risk that growth does not come.

The Wisconsin AI Infrastructure Readiness Brief treats these clauses as part of the preventive discipline around demand overestimation, alongside the contracted-demand principle itself. They are qualitative mechanisms in the Brief (named as instruments, not quantified) and they are best understood as the connective tissue that lets a build stay right-sized without foreclosing future scale.

What this asks of planners

For regional planners, utilities, and decision-makers, demand overestimation reframes a familiar question. The issue is less “how much demand is coming” and more “how much of the projected demand is actually committed, and how is the remainder being treated.” A projection carried into a build as though it were a commitment is the failure mode in its purest form.

That reframing points toward a few durable practices: distinguishing contracted from anticipated load explicitly in planning documents; sizing initial investment to the former; using step-in and conversion provisions to govern the path to the latter; and holding the horizon mismatch in view, so that the length of an asset’s life is weighed against the brevity of the demand confidence behind it. These are planning disciplines, not engineering ones, which is consistent with where the failure originates.

Takeaway: The most common way a large AI infrastructure project fails is not on its technology but on an un-stress-tested demand forecast: an optimistic projection treated as a commitment and built into fixed, long-lived infrastructure. The preventive discipline is to scale to contracted demand, preserve a defined path to grow, and structure commitment so it tracks materialized rather than projected load.

As the AI demand signal continues to evolve, how should contracted-demand thresholds and step-in provisions be structured so that committed load, rather than projected load, sets the scale of what a region actually builds?

Source: the Wisconsin AI Infrastructure Readiness Brief, execution and risk — demand overestimation as a documented failure mode, scaling to contracted rather than anticipated demand, and step-in and conversion clauses as qualitative instruments. The 100 MW / 300 MW case is illustrative, not a reported project.