HARC says Texas data centers could push smog-forming NOx back to 2004 levels by 2031

What HARC found

The Houston Advanced Research Center (HARC), an independent nonprofit research group in The Woodlands, Texas, published an interactive assessment on 20 October 2026 of how the state's data-centre boom could affect its air. Its conclusion is that the way new data centres get their electricity will decide whether Texas keeps or loses two decades of pollution cuts, and that in several plausible cases it loses most of them by 2031.

Texas has a lot to lose. According to HARC's press release, the state's power sector cut nitrogen oxides (NOx) emissions by 68 percent, sulphur dioxide by 84 percent and fine particulate matter (PM2.5) by 60 percent between 2002 and 2024, even as its population grew by millions and its economy roughly doubled. HARC credits pollution controls, fuel switching, tailpipe standards and the retirement of coal plants.

HARC also says Texas data centres are expected to add more electricity demand in the next six years than the whole grid added in the last 20. Its analysis asks what that demand does to five pollutants (NOx, sulphur dioxide, PM2.5, volatile organic compounds and carbon dioxide) under several ways of supplying the power.

The grid-mix numbers

The scenario that gets the headlines assumes every new data centre is served by the grid's projected fuel mix. In that case HARC's research page says 2031 NOx from the power and industrial sectors would rise 136 percent and approach its 2004 volume, VOCs would exceed their 2002 volume by 32 percent, and PM2.5 would climb back to its 2006 volume. Sulphur dioxide would rise 18 percent but stay below its 2002 level. The press release frames the same scenario against 2025: power-sector NOx up 136 percent, with PM2.5 and CO2 each up about 90 percent.

NOx matters because it helps form ground-level ozone, the main ingredient of smog, and PM2.5 is fine enough to reach deep into the lungs. Both are already a problem in the state's largest cities.

"Our analysis is cause for concern. There is a good chance that our children and grandchildren will breathe dirtier air than we did," HARC chief executive John Hall said in the release. The Texas Tribune notes that Houston, Dallas-Fort Worth, San Antonio and Austin often struggle to meet federal air standards already.

On-site gas is the biggest swing

Because grid connections in Texas can take years, many developers are planning their own power plants on site. HARC finds that choice outweighs everything else, including how fast data centres grow. NOx from data-centre power could be 2 to 10 times higher if all new demand were met by on-site combustion instead of grid supply paired with solar and wind, depending on the demand forecast. For VOCs the range is 2 to 165 times.

Under the highest demand forecast, on-site gas engines would emit more NOx each year than all sources in Texas emit today, the release says. Pollution controls would cut that, but even controlled engines would emit about two to three times as much as grid power. Power-sector CO2 would reach about 251 million tonnes in 2031 under the most conservative forecast, roughly 38 percent above 2025, and about 900 million tonnes under the highest one, more than Texas's total emissions from all sources in 2024.

The cleanest option HARC tested is a mix of about one-third grid power with the rest from on-site solar and wind. That cuts NOx and CO2 by roughly two-thirds compared with grid power alone. But HARC is explicit that even this case leaves emissions above a world with no new data centres, because the added demand is so large.

How the model was built

The methodology is published alongside the dashboard. HARC takes three forecasts of data-centre capacity: the grid operator ERCOT's adjusted large-load figures, figures from transmission service providers, and the BAXTEL data-centre database. It assumes every facility runs around the clock at 80 percent of capacity, with about 40 percent of electricity going to cooling.

For the power mix it simulated one representative 1-gigawatt data centre with the National Renewable Energy Laboratory's REopt tool and scaled the result linearly to the whole state. Grid emissions factors come from the EPA's eGRID data and ERCOT generation by fuel, projected forward with a damping factor; on-site gas factors come from Texas Commission on Environmental Quality rules and EPA documents, modelled with and without post-combustion controls. The baseline assumes no new data centres and continues the recent trend in state point-source emissions. In total HARC produced 77 modelled trajectories across the five pollutants and three forecasts.

Why Texas is the test case

Texas is where the build-out is most concentrated. The Verge notes the state leads the country with 26 gigawatts of existing and under-construction data-centre capacity, and cites a Wood Mackenzie report from March projecting four times more proposed capacity on on-site power in Texas than in any other state.

The Texas Tribune counts at least 288 planned data-centre projects in the state and reports that ERCOT received more than 500 requests to connect large electricity users in two years, against about two dozen the year before. It also cites the Environmental Integrity Project's count of at least 32 planned gas-fired power plants, more than any other state, and reports that Governor Greg Abbott has temporarily halted new data-centre grid connections pending an audit.

"We need to treat data centers as the industrial facilities they are," Carlos Gamarra, HARC's vice president for energy, air and IT, told the Tribune, arguing they should meet the same standards that produced Texas's progress. He also suggested operators disclose their electricity use, power sources and emissions before communities decide on new sites.

What the report does not say

HARC is careful about its limits. Each scenario assumes all new demand uses a single kind of supply, while real projects will mix approaches, so actual outcomes should fall between the lines. Scaling one 1-gigawatt facility to the state ignores differences in design and efficiency. The BAXTEL forecast assumes projects open on their developers' schedules, and delays or cancellations are common. Running at 80 percent capacity all year may also be a high assumption for many sites.

It is also a model of emissions, not of air quality at street level. It does not estimate ozone concentrations, health effects or deaths. And news coverage has blurred the scenarios at times, so readers should check figures against HARC's own pages. "These conclusions may never come to pass," the report says.

What it means for AI's footprint

Most debate about AI's environmental cost has centred on carbon and water. HARC's report adds local air pollution, and it ties the size of that cost to decisions made now about turbines, engines and grid connections. It fits a pattern we have covered: Berget AI's open calculator estimates the same AI request emits several times more CO2 in Texas than in Sweden because of the grid behind it, and our explainer on AI data-centre electricity argues projections like these should be read as scenarios with their assumptions attached.

The pressure to build fast is rising, not falling. NVIDIA this week pitched DSX POD prefabricated AI factories it says can be running in 14 days, and gigawatt sites like Meta's Prometheus in Ohio show the scale involved. HARC's numbers suggest that how those sites are powered, not just how many are built, will decide what the people living near them breathe.