Stranded-Cost Risk Is the Quiet Exposure in Large-Load Planning

Much of the debate over large new loads focuses on whether the power can be built in time. A quieter question sits underneath it: if the load the infrastructure was built for never fully arrives, who still owns the asset, and who pays for it.

Most of the attention on AI-scale development goes to whether generation, transmission, and sites can be delivered fast enough to meet demand. That is the visible risk. The less visible one runs in the opposite direction. Infrastructure is long-lived and largely irreversible, while the demand that justifies it is a forecast, and forecasts for a technology moving this quickly are uncertain by nature. When the two diverge, the physical asset does not adjust to match. It remains on the system, and its remaining cost has to be recovered from someone. We treat this as a cost-exposure problem rather than a judgment on any particular project; the mechanics below hold regardless of who the applicant is.

The exposure begins where a forecast ends

Utilities and their regulators have been explicit that demand overestimation is a live concern in this cycle. The caution is against over-building generation and network capacity that becomes a stranded asset if the current growth in AI compute slows or relocates. One consumer advocate framed the underlying uncertainty plainly, asking how anyone can be confident “we’re going to be talking about the same thing three years from now.” That is not skepticism about whether demand is real. It is a recognition that the confidence interval around a multi-decade asset built for a fast-moving load is wide, and that the cost of being wrong does not fall on the forecast. It falls on whoever owns and pays for the asset.

A billion-dollar reminder already on the books

Wisconsin does not have to imagine this exposure. It is already carrying a version of it. Ratepayers in the state currently owe on the order of $1 billion in remaining cost on generating plants that have already been retired: assets that no longer produce power but whose unpaid balance is still being collected from customers. This is what “stranded cost” means in concrete terms. A commitment entered under one set of assumptions outlived both the plant and the conditions that justified building it, and the remaining balance did not disappear when the plant did. It is being paid down by whoever remains on the system.

That figure is useful precisely because it is not hypothetical. Every proposed allocation for new large-load infrastructure is, implicitly, being weighed against that memory. The question underneath the technical filings is direct: if a commitment made today does not hold, who carries what is left, and for how long.

The mismatch is a mismatch of time

The structural reason stranded-cost risk is elevated in the current build-out is a divergence between how long commitments last and how long assets last. Supply arrangements for large new loads are frequently written as relatively short or flexible agreements, on the order of five to ten years. The generation and transmission built to serve them carry operating lives of thirty years or more. A five-to-ten-year commitment placed against a thirty-year-plus asset leaves roughly two decades of asset life uncovered by the agreement that justified the investment.

During those uncovered years, the outcome depends entirely on what the load does. If it stays, ramps as expected, and renews, the mismatch is invisible. If it changes its operating profile, consumes less than projected, or departs when its term ends, the assets it prompted still have to be paid for, and the party left holding the unrecovered balance is generally not the party that left. This is the mechanism by which a private forecasting error becomes a shared cost.

Where the cost lands when a load leaves

The distribution of that exposure is the heart of the matter. When a large customer underperforms or exits, the infrastructure remains rate-based or otherwise owed, and the remaining balance reverts to the broader base of customers who were already on the system. Those customers did not make the forecast, did not benefit from the load, and inherit the residual cost in the same way the earlier retired-plant balance was inherited. The concern is not that any single project is imprudent. It is that the accounting can look balanced at full utilization and look very different under stress, and the stress case is borne by parties who had no role in the original bet.

The disciplines that limit the exposure

None of this argues against building. It argues for building in a way that keeps the exposure with the party that created it and sized to demand that is real rather than projected. Several disciplines recur in the current planning discussion.

The first is scaling investment to contracted demand rather than to forecast demand. The illustrative version regulators have used is straightforward: do not build a 300 MW plant for a 100 MW initial need on the assumption that the remaining growth will arrive. Building to what is committed, and phasing further investment as further demand is contracted, keeps the uncovered gap small.

The second is contract design. Staff at the Public Service Commission have pushed toward longer contract terms that better match the life of the assets being built, narrowing the two-decade gap described above. Longer terms shift more of the asset’s life onto the commitment that justified it.

The third is financial assurance. Step-in and conversion clauses, along with financial guarantees, are mechanisms for deciding in advance what happens to an asset and its remaining cost if the original load does not hold: who assumes the commitment, how the resource is repurposed toward general system use, and what security stands behind the obligation. These provisions do not eliminate the risk that a forecast is wrong. They determine, before the fact, where the consequences of a wrong forecast land.

A related timing exposure

Stranded cost is the exposure when demand arrives short. A related timing risk runs the other way: if a load is ready before the grid upgrades that support it are complete, it may have to curtail or delay its ramp. Both are failures of alignment between long-lead infrastructure and the demand it serves, and both reward the same discipline, matching the pace and scale of irreversible investment to demand that is actually contracted and actually deliverable, rather than to a forecast on either side of the gap.

The planning question

For anyone evaluating whether a region can serve a large new load, the readiness of the power supply is only part of the assessment. The other part is what happens if the demand behind it does not fully materialize. The same infrastructure can represent prudent investment or a future stranded balance depending on how the commitment is structured, how long it runs, and what assurance stands behind it. Treating that downside as a design input, rather than a risk discovered late, is what separates a durable plan from one that transfers its errors to customers who never made them.

Takeaway: Infrastructure outlives the forecasts that justify it. The central stranded-cost question is not whether a large load can be served, but who owns the asset, and who pays the remainder, if the load never fully arrives.

For a given large load, how much of the supporting asset’s operating life is covered by a firm commitment, and what mechanism decides where the remaining cost lands if that commitment ends early?

Source: the Wisconsin AI Infrastructure Readiness Brief, execution and risk — stranded-cost risk and demand overestimation: the ~$1 billion in remaining cost on already-retired Wisconsin plants, 5–10 year supply contracts against 30+ year asset lives, unrecovered balances reverting to remaining customers, scaling to contracted rather than forecast demand, PSC staff positions on longer terms and financial guarantees, step-in and conversion clauses, and the timing mismatch when a load is ready before its grid upgrades. The 100 MW / 300 MW case is illustrative, not a reported project.