Kai ChenGlobal Energy Monitor counted more than 97 GW of proposed U.S. gas projects earmarked for data centers in its 2025 pipeline. PJM Interconnection has proposed that large campuses bring their own generation or curtail demand to help prevent a large-scale outage. Yet a campus can switch to its own plant without warning the grid operator.
Key takeaways
- Global Energy Monitor counted more than 97 GW of proposed U.S. gas projects explicitly earmarked for data centers in the 2025 pipeline, versus 4 GW in 2024.
- In July 2025, OpenAI and Oracle announced plans for an additional 4.5 GW of U.S. data-center capacity, taking their stated total above 5 GW.
- SemiAnalysis reported in September that xAI’s Colossus 2 had reached 200 MW in six months.
- Google signed demand-response agreements with five U.S. utilities covering up to 1 GW of data-center load.
- Air-permit documents tied to 11 U.S. data-center campuses indicated potential annual greenhouse-gas emissions above 129 million tons.
Compute builders are treating firm power as a deployment input
In July 2025, OpenAI and Oracle announced plans for another 4.5 GW of U.S. data-center capacity, taking their stated total above 5 GW. In September, SemiAnalysis reported that xAI’s Colossus 2 had reached 200 MW in six months and was on track for gigawatt scale. OpenAI and Oracle set out planned capacity, while xAI had reached 200 MW; each needed power in place before it could put compute to work.
Other builders are choosing a more direct route to supply. CoreWeave and Poolside planned a natural-gas-powered campus on a 500-acre West Texas ranch. Servers, cooling and electrical supply have to work together; a missing subsystem limits what a campus can deliver. A generator that lets developers commission capacity has a different economic job from one reserved for outages.
The gas pipeline measures intent, not delivered capacity
The 2025 tally was up from 4 GW in 2024, as Wired reported. That count measures proposed capacity; it does not establish how many plants were built, how many run routinely, or whether each would operate behind the meter.
Developers pursuing onsite plants to bypass an overloaded grid still face permitting and supply-chain constraints before they can energize a campus. Gas also carries a cost beyond construction: air-permit documents for projects tied to 11 U.S. data-center campuses indicated potential annual emissions exceeding 129 million tons of greenhouse gases.
A campus can solve its outage problem and create a grid event
Reuters documented data centers making “unannounced disconnections” when switching to local generators and reported that the behavior may contribute to cascading outages. When a large campus islands, the grid abruptly loses that load; when it reconnects, the grid must supply it again. Reuters did not measure how often campuses switch at once.
A utility needs to know how much load a campus can remove, how quickly it can move, and how it will return. The campus needs to know who may call for curtailment and under what conditions. Redundant equipment protects compute only as well as the decisions that operate it; it does not give the grid operator advance notice.
PJM Interconnection’s proposed framework makes the allocation of responsibility explicit: large data centers would bring their own generation or curtail use to help prevent a large-scale outage. PJM has proposed those alternatives; the proposal alone cannot show which campuses have the controls to follow through.
Google shows that flexibility can compete with generation
Google signed demand-response agreements with five U.S. utilities covering up to 1 GW in total. The agreements allow data-center consumption to fall during peak grid periods; the 1 GW is contracted scope, not proof that Google has curtailed that amount. Utilities can request lower campus demand at peaks rather than rely solely on added generation.
Frequently asked questions
Has PJM adopted its proposed data-center reliability framework?
The piece describes PJM’s framework as a proposal and provides no final adoption date, approved tariff language, or implementation schedule. Its status beyond the proposal is not established in the evidence.
Which AI data centers have controls that can coordinate islanding, curtailment and return to grid supply?
The evidence does not identify any specific campus with documented controls or operating rules for those actions. It says PJM’s proposal alone cannot show which campuses can follow through.
How much of the more-than-97-GW gas pipeline has actually been built or is operating?
The evidence does not provide a construction or operating total. The 97-GW figure counts proposed projects, not delivered capacity or routine generation.
Did Google actually reduce 1 GW of data-center demand under its utility agreements?
No such performance result is supplied. The 1 GW is the maximum contracted scope across five utility agreements, not evidence of an executed curtailment.
Proposed U.S. gas capacity explicitly earmarked for data centers
| Year | Proposed capacity |
|---|---|
| 2024 | 4 GW |
| 2025 | More than 97 GW |
Enchanted Rock sits near the market for dispatchable onsite capacity, but no documented AI-data-center customer, contract, or operating footprint ties it to this buildout. For any provider, the more informative evidence would be a committed megawatt together with rules for dispatch, notice and return to grid supply. A megawatt only its owner can switch is not the same product as a megawatt the grid can plan around.