How to Evaluate a Proposed AI Power Project

A reported $22.3 billion gas-power project near Encinal, Texas has become a useful test case for reading AI-infrastructure announcements. Korean reporting described a possible 6.3-gigawatt project that could serve data centers and form part of South Korea's wider US investment discussions. But South Korea's industry ministry said consultations were still under way and did not confirm the reported details. That places the project in a different category from a financed, permitted power plant.

The distinction matters because AI demand is changing how data-center plans are marketed. A multi-gigawatt number implies scarce computing capacity, industrial policy and a route around grid bottlenecks. None of those implications, however, proves that a site has a customer, a contract or an operating date. The responsible way to assess a proposal is to separate the scale being discussed from the commitments that make scale executable.

First, establish the project stage

A project can be politically selected, commercially negotiated, financed, permitted and constructed at different times. The Korea Herald's report says the Encinal project was being considered under a $350 billion investment commitment, while also reporting that its final amount and capacity could change. Reuters likewise reported that the trade ministry said US-investment details had not been finalized.

Those facts support a narrow description: a proposed project under negotiation. They do not establish a final capital commitment, a signed engineering contract, a completed interconnection agreement or a construction notice. Readers should be wary when a preliminary selection is written as if turbines have already been ordered.

The number that matters is contracted demand

A 6.3-GW plant needs a reason to be built in phases. For an AI campus, the essential evidence is a credible load: a named buyer, a power-purchase agreement, a data-center construction plan, or another binding commitment that specifies how much electricity will be used and when. Without it, a project is exposed to a familiar development risk: several prospective customers may reserve capacity while only some build.

This is especially important in Texas. The state is examining whether large-load projects have enough evidence behind their requests. Governor Abbott's data-center audit directive asks for ownership, power and water needs, incentives, generation plans and local impacts. A dedicated generator may change the grid path, but it does not remove the need for reliable fuel, backup, cooling, permits or a buyer that can support long-term financing.

Treat the delivery model as a commercial choice

Reports have described two broad possibilities: selling into ERCOT or supplying a nearby load directly. These are not interchangeable. A merchant generator depends on power-market prices and grid arrangements. A direct-supply project depends heavily on the credit, operating schedule and reliability requirements of its customer. A hybrid model can add flexibility, but it also has to resolve physical connection, operating and regulatory questions.

The proposed gas configuration has a real attraction: firm generation can be developed on a different timetable from a major transmission expansion. Its tradeoffs are equally real. Equipment lead times, pipeline performance in extreme conditions, water, emissions, maintenance reserve and fuel-price exposure all affect whether a plan can meet the availability expected by a large computing operation. No project-specific technical package should be inferred from a headline capacity figure.

What would materially change the assessment

Three disclosures would move this story beyond a negotiation. First, an official bilateral announcement should identify the investment structure: who supplies equity, debt, guarantees or other support, and what approvals remain. Second, project documents should identify the sponsor, phase schedule, major equipment and permits. Third, a disclosed customer or PPA should state the contracted load and the delivery model.

Until that evidence appears, the Encinal reports are most useful as an illustration of the new AI power constraint. Chips and buildings are not enough; dependable electricity is now part of the computing product. That makes proposed power projects worth tracking, but it also makes precision essential. A large announced number is a starting point for diligence, not its conclusion.