If the forecasted demand for data centers fails to materialize, utilities could be left with expensive, stranded assets. Without explicit protections, the costs of this overbuild could be passed on to residential and commercial ratepayers, creating significant political and financial risk.
Utilities have firm commitments for 110 gigawatts of data center power capacity, while demand forecasts only predict a need for an additional 50 gigawatts by 2030. This significant discrepancy, based on simple math, points to a potential overbuild and future oversupply in the market.
Despite forecasts of massive energy demand growth from AI data centers, forward power curves are flat and natural gas futures are downward sloping. This suggests that sophisticated energy traders do not believe the bullish demand narrative and are not pricing in a future supply crunch.
After massive cost overruns on traditional nuclear projects, no utility will build a Small Modular Reactor (SMR) alone. The only viable path forward is for a tech giant to provide both a purchase agreement for the power and direct equity investment in the SMR manufacturer to fund capital expenditures.
After PIMCO's highly profitable $2 billion gain on a loan to a Meta data center, other private credit lenders are piling into the space. This fierce competition is driving down rates and weakening investor protections like covenants, a classic sign of a frothy market nearing its peak.
Tech giants like Meta are using off-balance sheet vehicles to finance data centers, paying a significant premium over their own borrowing costs. This structure could be designed to avoid depreciation on the income statement or, more critically, to retain the option to abandon the asset if technology changes.
While a new gas plant's cost has soared to $3,000 per KW, the data center it powers costs $40,000 per KW. For tech giants, paying a huge premium to secure a dedicated power source is an insignificant rounding error, explaining their willingness to pay far above-market rates for electricity.
Contrary to doomsday scenarios, the existing U.S. power grid has enough latent capacity to handle a massive influx of AI demand. For example, the Texas grid could power a full year's worth of new NVIDIA chip production running 24/7, failing for only about 40-50 peak hours.