Texas and Ohio Fast-Track Gas Plants for AI Data Centers, but the Emissions Math Is Still Unsettled

Texas and Ohio are accelerating gas-fired power projects for AI data centers, despite estimates linking the national buildout to 662 million tons of annual emissions. The figure appeared across Google News after an environmental report compared that maximum permitted pollution with 140 million vehicles operating for one year.

That comparison is striking, but it needs context. The estimate covers 74 planned gas plants across the United States, not only projects in Texas and Ohio. It also reflects permit limits or other full-operation assumptions, rather than a forecast that every proposed turbine will run continuously.

The real story is bigger than one emissions calculation. AI developers are placing generation beside data centers so they can secure electricity without waiting for crowded regional grids. That strategy promises faster construction and clearer cost allocation, but it can also reduce public oversight and lock new computing capacity to fossil fuels.

What Texas and Ohio Are Actually Accelerating

The change is not simply that data centers need more electricity. Developers are increasingly building dedicated power systems around the computing sites themselves.

The Environmental Integrity Project identified at least 74 gas-fired plants planned to serve data centers directly. Its gas plant analysis counted 71 new plants and three expansions with a combined capacity of 143 gigawatts.

These facilities use a structure commonly called behind-the-meter generation. The term describes a power source connected primarily to one customer, without sending most of its electricity through the wider utility grid.

A developer gains more control under this model. It can coordinate power construction with data halls, cooling equipment, and server installation instead of waiting for a utility interconnection.

Texas has the largest share of the projects in the report. Researchers counted 32 proposed plants there, followed by 10 in Ohio, seven in Pennsylvania, and four each in West Virginia and Wyoming.

The Texas proposals span several regions and development strategies. Some would supply individual cloud or social-media campuses. Others belong to enormous mixed-energy complexes marketed to multiple AI tenants.

The Texas plants could release more than 287 million tons of greenhouse gases annually at their permitted limits, according to the report. That estimate is equivalent to the annual emissions attributed to 61 million gasoline-powered vehicles.

Texas is attractive because it combines available land, natural gas infrastructure, renewable generation, and a competitive electricity market. Local governments have also pursued data centers as sources of construction spending and taxable property.

Yet the state grid faces an extraordinary queue. ERCOT, the organization operating most of the Texas grid, had received requests representing about 439 gigawatts of potential new demand by May 2026.

Approximately 89 percent of that requested load came from data centers, according to figures cited in a Texas project review. Many requests will not become operating projects, but even a fraction would alter electricity planning.

Ohio presents a different version of the same rush. It already hosts major campuses operated by Amazon Web Services, Google, Meta, and other infrastructure providers.

The state has access to established transmission corridors and the broader PJM electricity market. It also has industrial sites that can accommodate large facilities.

One prominent proposal centers on the former Portsmouth Gaseous Diffusion Plant near Piketon. Federal officials announced a public-private development there in March 2026.

The plan includes a data center with a targeted capacity of 10 gigawatts and up to 10 gigawatts of new generation. The Department of Energy said 9.2 gigawatts could come from natural gas.

The project also includes grid-connected generation and transmission construction. SoftBank affiliate SB Energy and AEP Ohio are involved in the power infrastructure.

This is not a minor backup arrangement. A 9.2-gigawatt gas fleet would resemble a major regional generation system built around one technology campus.

Together, the Texas and Ohio projects show how AI infrastructure is changing the definition of a data center. A campus can now combine computing, industrial-scale cooling, pipelines, turbines, substations, and transmission equipment.

That combination makes the permitting question central. Regulators are no longer reviewing only warehouses filled with servers. They are reviewing privately directed energy complexes with consequences far beyond the property line.

Why the AI Power Race Is Moving Off the Grid

AI companies are turning to dedicated gas generation because electricity delivery has become a constraint on computing growth.

Training and serving large AI models requires dense clusters of accelerators. These chips draw electricity continuously and produce heat that cooling systems must remove.

A conventional data center can wait years for transmission upgrades, new substations, or permission to connect a very large load. The delay can undermine an AI company’s ability to deploy already-purchased chips.

Gas turbines offer a different schedule. Developers can place generation near a pipeline and connect it directly to the campus, reducing dependence on a completed utility interconnection.

That does not make the projects instant. Turbines, transformers, and other electrical equipment face their own supply constraints. Developers must also obtain air permits and arrange fuel delivery.

However, control matters almost as much as total construction time. A company coordinating the data center and power plant can manage both schedules instead of relying on separate utility investments.

The urgency comes from a sharp change in national electricity demand. Data centers consumed about 176 terawatt-hours in 2023, or 4.4 percent of total U.S. electricity.

A 2025 update from Lawrence Berkeley National Laboratory estimates that data centers could reach 11.8 percent of U.S. electricity consumption by 2030. Its scenario range runs from 9.5 percent to 15.3 percent.

Those projections depend on chip shipments, server utilization, cooling efficiency, and the pace of data center construction. Still, the energy demand update makes the direction clear.

The previous efficiency model for cloud computing is under pressure. During the 2010s, providers consolidated workloads into larger facilities and improved server utilization.

AI reverses part of that pattern. Companies are adding accelerators faster than efficiency gains can offset their total electricity consumption.

The load profile also matters. AI data centers want dependable power across every hour, even when wind or solar output falls.

Battery storage can shift renewable electricity across part of a day. It cannot always cover extended periods without sun or wind at the scale these campuses demand.

Nuclear developers are proposing reactors for some sites, but new nuclear projects generally require longer development periods. Existing reactors offer another route, although available capacity is limited and politically contested.

Gas therefore occupies the gap between AI companies’ deployment schedules and the slower expansion of transmission, storage, and firm carbon-free generation. Developers often call it a bridge fuel, but a bridge can remain in service for decades.

Federal policy is reinforcing the trend. In March 2026, President Donald Trump encouraged technology companies to add power generation beside data centers.

Supporters argue that dedicated generation prevents existing customers from paying for infrastructure serving private computing campuses. Texas Governor Greg Abbott has similarly called for developers to bear the costs associated with their new demand.

That argument has practical force. Households and established businesses should not automatically finance substations or transmission lines built for speculative data center proposals.

Behind-the-meter plants, however, do not isolate every cost. Large projects can raise regional demand for gas, turbines, construction labor, and transmission equipment.

A private plant may also need grid support when turbines fail or enter maintenance. Regulators must decide how much backup service the owner should fund.

This is why the Google News framing around emissions captures only one part of the dispute. The same infrastructure decision affects reliability, utility rates, land use, and the allocation of financial risk.

The Fast Track Trades Grid Delays for Public Oversight

Dedicated generation can shorten one approval path, but it also shifts decisions away from the institutions that normally test regional need and cost.

A utility seeking to build a large power plant usually must explain how the project serves its customers. Regulators can examine alternatives, projected demand, construction costs, and the effect on rates.

A plant serving a private data center can face a narrower review. State environmental agencies still evaluate air permits, but that process does not necessarily answer broader electricity-planning questions.

An air regulator may determine whether a turbine meets applicable pollution limits. It may not decide whether storage, transmission, demand flexibility, or cleaner generation would serve the campus more effectively.

That division creates a policy gap. No single proceeding necessarily evaluates the full computing complex, its energy alternatives, and its effects on neighboring communities.

The Portsmouth project illustrates the speed and ambition behind the new approach. Federal officials presented the former Ohio uranium site as a platform for AI infrastructure and reindustrialization.

The government said the project would include major private investment, generation, and transmission upgrades. Officials also said the associated infrastructure would not raise customer rates.

Those claims remain important tests, not settled outcomes. Construction costs can change, data center tenants can alter their plans, and grid-support arrangements can evolve after initial announcements.

The Ohio campus announcement also prompted opposition from residents concerned about the environmental and financial scale of mega data centers.

Texas communities are raising similar questions. Some proposed gas plants would operate near areas that already struggle with unhealthy ground-level ozone.

Ozone forms when nitrogen oxides and volatile organic compounds react in sunlight. It can aggravate asthma, reduce lung function, and worsen other respiratory conditions.

The Environmental Integrity Project estimated that the 74 plants could release 159,142 tons of health-damaging air pollutants annually at their assessed levels. That total includes 44,281 tons of nitrogen oxides and 32,684 tons of fine particulate matter.

These pollutants create local effects that a national carbon comparison cannot describe. Greenhouse gases accumulate globally, while nitrogen oxides and fine particles can burden people living close to a plant.

Location therefore matters as much as the national total. The report found that 188 schools sit within three miles of the proposed plants for which locations were available.

It also reported that 40 percent of residents within three miles of 67 assessed projects lived in low-income households. That finding does not establish the health impact of any individual plant.

It does show why communities are demanding site-specific analysis. A plant placed in an already polluted region does not begin from a clean baseline.

West Texas raises another monitoring problem. It contains many proposed facilities but has fewer air-quality monitors than the state’s largest urban regions.

A permit model can estimate emissions, yet residents need operating data after turbines start. Continuous monitoring and public reporting determine whether modeled protections work in practice.

The central opponent in this story is therefore not Texas against Ohio. It is deployment speed against transparent energy planning.

Developers view time as a competitive asset. Local communities experience the same schedule as a compressed opportunity to understand health, water, noise, and financial consequences.

Both positions can be true. AI companies need credible power plans, and residents need meaningful participation before industrial infrastructure reshapes their surroundings.

The 140 Million Vehicle Comparison Has a Major Caveat

The headline emissions estimate is a warning about permitted scale, not a prediction that every proposed plant will operate at its maximum limit.

The Environmental Integrity Project calculated nearly 662 million tons of annual greenhouse gas pollution across the 74 facilities. It compared that figure with 140 million cars and trucks driven for a year.

The estimate is broadly consistent with the Environmental Protection Agency’s conversion method. The EPA says a typical passenger vehicle emits about 4.6 metric tons of carbon dioxide annually.

Multiplying 140 million vehicles by that value produces roughly 644 million metric tons. Differences can arise from rounding, vehicle definitions, and the greenhouse gases included.

The EPA also cautions that its vehicle emissions estimate is an approximation. Actual pollution varies with fuel economy, annual mileage, and fuel type.

More importantly, an air permit establishes an allowable ceiling. It does not necessarily forecast a plant’s annual fuel use.

Permit models often assume high operating hours and conservative conditions. Turbines can run less because of maintenance, changing demand, fuel costs, or project delays.

Williams Companies, which is developing three Ohio plants connected with Meta facilities, challenged the use of permit ceilings as expected emissions. A company representative said its modeling produced figures potentially two-thirds below the permitted amounts.

That response appeared in a permit analysis covering several data center projects. It identifies the strongest limitation in the 662-million-ton headline.

The caveat does not erase the issue. Data centers differ from conventional grid plants because their loads can remain relatively steady around the clock.

A traditional power plant may reduce output when electricity demand falls or cheaper generators become available. A dedicated data center plant can serve a customer seeking continuous operation.

Energy researcher Jonathan Koomey has argued that efficient grid-connected gas plants might emit 40 to 50 percent of their permitted limits. Dedicated data center plants could operate closer to modeled levels because their loads fluctuate less.

The final emissions outcome will depend on several variables. These include plant efficiency, utilization, renewable energy supply, storage, grid imports, and the pace of computing deployment.

The type of turbine is especially important. Combined-cycle plants capture heat from a gas turbine to generate additional electricity, improving fuel efficiency.

Simple-cycle turbines can start quickly but generally burn more fuel for each unit of electricity. Temporary or refurbished turbines can add another layer of uncertainty.

Methane leakage also sits outside many power-plant headline figures. Natural gas is primarily methane, which can escape during production, processing, and transportation.

A permit focused on combustion at the campus does not necessarily represent the full fuel-cycle climate impact. Conversely, a project that operates below its permit limit can produce much less pollution than the maximum estimate.

Project completion is another uncertainty. ERCOT’s enormous connection queue demonstrates how proposed demand can exceed the number of facilities ultimately built.

Some developers submit multiple requests while evaluating locations. Others reduce project sizes or abandon campuses when financing, equipment, or customer commitments fail to materialize.

The EIP report itself describes planned projects, not 74 completed plants running today. Readers encountering the story through Google News should not confuse potential emissions with an observed annual inventory.

A defensible interpretation sits between dismissal and certainty. The report measures the pollution capacity being authorized around AI infrastructure.

That capacity matters because permits establish what operators can legally emit. Once turbines, pipelines, and gas contracts exist, changing to a cleaner system can require additional investment.

The number should therefore trigger scrutiny, not be treated as destiny. Regulators and developers can reduce the outcome through enforceable limits, efficient equipment, cleaner power, and transparent operating data.

Gas Solves the Schedule Problem but Deepens the Climate Conflict

AI companies are using gas to protect deployment schedules while making their climate commitments harder to meet.

Large technology companies have announced targets involving renewable energy, carbon neutrality, or round-the-clock carbon-free electricity. Their exact commitments differ, so they should not be grouped into one promise.

The underlying conflict is shared. AI infrastructure adds major electricity demand before the clean generation and transmission needed to serve it can arrive.

Companies can purchase renewable energy equal to their annual consumption. That accounting approach does not mean every data center runs on carbon-free electricity during every hour.

A campus may consume gas-fired power at night while its owner supports solar production during the day or in another region. Annual matching can balance certificates without eliminating hourly fossil generation.

Dedicated gas plants make the physical relationship more visible. The data center and its emissions source can occupy the same property or connect through private infrastructure.

Microsoft has said it uses a portfolio approach to power and might consider onsite infrastructure where grid constraints slow deployment. The company has also continued investing in carbon-free electricity.

Crusoe has described gas as a bridge rather than a destination. That language leaves two unanswered questions: how long the bridge lasts and what replaces it.

A modern gas plant commonly operates for decades. Investors will usually expect extensive use before retiring equipment that required large upfront spending.

Future conversion is possible but not automatic. Turbines marketed as hydrogen-ready still need low-emissions hydrogen, suitable pipelines, storage, and verified operating performance.

Carbon capture offers another proposed pathway. It adds equipment, energy consumption, cost, and the need to transport and store captured carbon dioxide.

Neither option should be credited before it operates at the required scale. A permit mentioning future compatibility does not guarantee lower emissions.

Renewables and batteries can still reduce gas consumption. Texas has abundant wind and solar resources, while large campuses can use batteries to manage short-term variability.

Data centers can also shift some computing activity across time or location. Not every inference request can wait, but training jobs and flexible workloads can sometimes respond to grid conditions.

Those methods become more valuable when companies publish hourly electricity use and carbon intensity. Annual sustainability reports often hide when and where emissions occur.

The conflict also extends beyond corporate reporting. Governments want AI investment, construction jobs, and control over strategic computing infrastructure.

Federal officials presented the Ohio project as part of a contest for AI leadership. That framing encourages agencies to measure success through gigawatts built and servers installed.

Communities may use a different scorecard. They will watch electricity bills, air pollution, water demand, noise, tax agreements, and permanent employment.

Fast deployment can succeed under the first scorecard while failing under the second. Durable policy must reconcile both.

This is where the Google News headline risks simplifying the debate. The choice is not limited to building gas plants or abandoning AI.

The meaningful choices concern plant efficiency, permit conditions, public disclosure, grid contributions, clean-energy schedules, and who pays when projects change.

Developers could accept enforceable annual emissions limits below theoretical permit ceilings. They could fund independent monitoring and publish hourly generation data.

States could require large campuses to provide flexible load or firm capacity during grid emergencies. They could also make developers responsible for backup and transmission costs.

Communities could receive clearer timelines and cumulative pollution assessments. Reviewing each turbine separately can miss the combined effect of several projects in one region.

The fastest proposal is not necessarily the fastest reliable system. A project delayed by local opposition, equipment shortages, or litigation can lose the schedule advantage that made gas attractive.

Three Signals Will Show Whether the Forecast Becomes Reality

The next phase will be determined by operating limits, construction progress, and the power mix used by completed campuses.

The first signal is how many of the 74 proposed plants receive final permits and begin construction. A long project list does not equal an operating fleet.

Watch Texas closely because it contains 32 of the projects. County decisions, state air permits, equipment deliveries, and financing disclosures will reveal how much of the queue is credible.

The Portsmouth project offers a similarly important test in Ohio. Its proposed 9.2 gigawatts of gas generation represents a large share of the state’s announced AI power expansion.

If major projects shrink or stall, the 662-million-ton maximum will overstate the likely buildout. If construction begins across many sites, the estimate will become more relevant.

The second signal is whether permits contain enforceable annual limits below theoretical maximum operation. Developers already argue that actual emissions will be lower.

That claim can be tested through permit language and operating reports. Voluntary projections offer less protection than legal limits tied to fuel use or annual emissions.

Public access matters as well. Hourly generation, turbine utilization, and measured nitrogen oxide releases would let regulators and researchers compare actual performance with modeled assumptions.

If operating data stay unavailable, the debate will continue to rely on competing estimates. Transparent records would strengthen or weaken the headline comparison with evidence.

The third signal is how much carbon-free electricity reaches these campuses before gas infrastructure becomes the default. That includes new renewable generation, storage, nuclear contracts, and usable transmission.

A gas plant paired with rapidly expanding clean supply can run fewer hours over time. A plant serving a constant load without enforceable transition milestones can become a long-lived emissions source.

Companies should publish site-level electricity information rather than relying only on companywide renewable purchases. Customers increasingly depend on cloud infrastructure for AI workloads, so its energy profile enters their own emissions calculations.

Readers should also separate three numbers when this story returns to Google News: permitted emissions, developer forecasts, and measured annual emissions. They answer different questions and should never be presented as interchangeable.

The permit number shows authorized pollution capacity. The developer forecast reflects planned operations. The measured figure records what happened after construction.

For technology buyers and developers, the issue is no longer abstract climate accounting. Infrastructure choices influence cloud availability, regional regulation, energy costs, and the credibility of corporate sustainability claims.

For residents, the decisive evidence will arrive closer to home. Air-monitor readings, utility filings, construction notices, and permit conditions will show whether promised safeguards survive the rush to build.

The 140-million-vehicle comparison has done its job if it prompts careful examination rather than passive repetition. The next step is to track which plants become real, how often they run, and what enforceable plan replaces their gas generation. Will AI companies publish enough site-level evidence for the public to judge that tradeoff?