Thirty-Three Billion Dollars and a Grid That May Not Need It

SoftBank is building a 9.2 GW natural gas power plant in Portsmouth, Ohio. The price tag is $33 billion. It would be the largest single power plant in American history, sited on the grounds of a former uranium enrichment facility. The stated purpose: baseload power for AI data centers across the Midwest.

The announcement arrived alongside research from Duke University’s Nicholas Institute showing that the existing U.S. grid can absorb 98 GW of new flexible load with average annual curtailment of less than 44 hours per year.

These two data points represent a fundamental disagreement about how the country should power its next decade of computing infrastructure.

The plant. The Portsmouth project is part of a $36 billion first-phase Japanese investment package under a bilateral trade deal, alongside a $2.1 billion Texas crude oil export terminal and a $600 million synthetic diamond factory in Georgia. SB Energy, SoftBank’s energy subsidiary, is involved in the project. Construction timelines have not been disclosed.

For a facility of this scale, the commitment is to burning natural gas for decades, locked in before the first shovel turns dirt.

The study. Tyler Norris, a J.B. Duke Fellow, and Dalia Patino-Echeverri, an associate professor at Duke’s Nicholas School, published research examining how much new data center load the existing grid can absorb. Their central finding: 98 GW of flexible load can be integrated with curtailment averaging less than 44 hours per year. Those curtailment windows fall within the discharge duration of commercially available lithium-iron-phosphate battery systems.

A separate analysis published by Latitude Media modeled the long-term grid effects. If 20 percent of data center demand is flexible (meaning 80 percent remains firm), new natural gas construction drops by nearly 20 percent and average electricity prices fall by up to $3 per megawatt-hour. At 50 percent flexibility, nearly half of all new gas capacity becomes unnecessary.

The mechanism is straightforward. Flexible loads consume power when it is cheap and abundant (solar midday, windy nights) and curtail briefly when the grid is stressed. Batteries provide the bridge.

The queue. If developer interest were the constraint, this would be a different conversation. A study led by two energy lawyers, published this week and covered by pv magazine USA, documents the causes of lagging battery deployment in the Eastern United States compared to CAISO and ERCOT territories. The researchers identify 15 specific policy fixes across four categories: market design, interconnection process, procurement rules, and planning methodology. Their analysis estimates that a combined solar and storage buildout in PJM alone could save $178 billion by 2035 with 54 GW of storage.

Texas has deployed nearly ten times more storage than the eastern regions combined, not because of superior technology or cheaper batteries, but because of fewer structural barriers. The finding is direct: eastern battery deployment lags because of regulatory architecture, not economics.

The reliability question. NERC, the continent’s grid reliability authority, is preparing its own response. E&E News reported on February 18 that NERC will publish formal data center interconnection recommendations in the first quarter of 2026. The concern is specific: rapid swings in power consumption during AI model training can destabilize frequency and voltage, potentially damaging generator equipment.

This distinction is relevant to the capacity debate. A 9.2 GW gas plant provides constant output. It does not modulate with the variable demand profile of GPU clusters cycling between training runs and idle states. Batteries do. A battery system dispatches in milliseconds. A gas turbine takes minutes to ramp. If NERC’s forthcoming rules require data centers to manage their own load volatility, batteries become not just economically competitive but operationally required.

The cost comparison. A direct dollar-for-dollar comparison between a gas plant and battery storage is imperfect. Gas plants generate energy continuously; batteries shift it. But the Duke study’s central insight is that data center load does not need continuous new generation. It needs flexibility. And flexibility costs less than baseload when the grid already has surplus capacity in most hours.

The cost gap compounds over time. Gas plants carry fuel costs for their entire operating life. Batteries do not. Domestic natural gas prices are subject to upward pressure from rising LNG exports and growing industrial demand. Every increase in gas prices widens the gap between the two options.

The political context. None of this analysis accounts for the political logic of the Portsmouth plant. The project sits on the grounds of a former uranium enrichment facility. It carries the imprimatur of a bilateral trade deal with Japan. These are not factors that appear in levelized cost models, but they influence what gets built.

The Duke study, the eastern deployment analysis, and NERC’s forthcoming recommendations all point in the same direction: batteries and flexible load management can accommodate data center growth without new large-scale gas plants. The pipeline of storage projects exists. The financing exists. In much of the Eastern United States, the regulatory framework has not caught up.

What gets built next will depend on which set of facts carries more weight.

Sources