As artificial intelligence data centers consume exponentially more electricity, researchers have identified an unexpected solution: capturing and permanently storing the carbon emissions from the power plants that fuel them underground. A new study from Rice University shows that geological formations beneath the United States could absorb the vast majority of AI data center emissions, potentially transforming how the industry addresses its environmental footprint while still meeting explosive power demands.

How Much Power Will AI Data Centers Actually Need?

The numbers are staggering. According to the Rice study, U.S. data center power capacity could more than quadruple in just five years, growing from 40 gigawatts in 2025 to 169 gigawatts by 2030. To put that in perspective, a gigawatt is enough electricity to power roughly one million homes. This explosion is driven almost entirely by artificial intelligence workloads, which are expected to make up roughly half of all data center computing by 2030, compared to about a quarter today.

The challenge isn't just building more data centers. It's powering them reliably. Graphics processing units (GPUs), the specialized chips that train and run AI models, consume enormous amounts of electricity. A single rack of AI servers can use 20 to 40 kilowatts of power, or even exceed 80 to 100 kilowatts in high-end GPU clusters, compared to just 4 to 8 kilowatts for traditional data center racks. This tenfold increase in power density is straining electrical grids across the country.

Without proper regulation of emissions, carbon dioxide produced by fossil fuel power plants supplying electricity to data centers could grow from 90 million metric tons in 2025 to more than 404 million metric tons by 2030. That's equivalent to the annual emissions of roughly 87 million cars.

Where Can All That Carbon Actually Go?

This is where the Rice study offers surprising hope. Researchers analyzed publicly available data on announced U.S. data centers, including their energy sources, locations, and projected power capacity. They then examined whether the resulting carbon emissions could be captured and stored underground in saline aquifers, which are porous rock formations saturated with saltwater deep beneath the surface.

The findings are encouraging. The team estimates that 34 states have enough saline aquifer storage capacity to store more than 100 years of projected data center-related carbon dioxide emissions beyond 2030. In 2025 alone, aquifers could store an estimated 59 million metric tons of data center-related carbon dioxide, or about 66 percent of the sector's emissions. By 2030, that storage capacity could grow to 299 million metric tons, or about 74 percent of projected data center-related emissions.

When out-of-state storage options are included, researchers found that more than 90 percent of data center-related carbon dioxide emissions could potentially be mitigated through carbon capture and storage. The states with the strongest data center growth, including Texas, Virginia, Pennsylvania, Ohio, Arizona, Colorado, Utah, and Illinois, would benefit most from this approach.

"Data centers are becoming one of the defining energy challenges of the AI era. The question is not only whether we can build enough computing infrastructure, but whether we can power it in a way that is reliable, affordable and compatible with decarbonization goals," said Hon Chung Lau, an adjunct professor in the Department of Chemical and Biomolecular Engineering at Rice University.

Hon Chung Lau, Adjunct Professor, Rice University

Why Power Reliability Matters More Than You'd Think

The power demands of AI data centers aren't just about raw consumption. They're also about consistency and speed. Training a large language model, the type of AI system behind tools like ChatGPT, can take several weeks or even months. Even a brief power interruption lasting just a few milliseconds can interrupt GPU synchronization, corrupt training processes, and cause expensive computational losses that require restarting large portions of the workload.

This is why energy storage systems have evolved from optional backup solutions into critical infrastructure. Modern AI data centers increasingly deploy battery systems in two complementary ways:

• Centralized Storage: Large battery systems installed at utility substations or facility substations, often providing capacities ranging from tens to hundreds of megawatt-hours with discharge durations between 2 and 4 hours, smoothing sudden demand spikes and reducing stress on the electrical grid.
• Rack-Level Storage: Smaller high-power batteries placed close to individual AI racks and GPU cabinets, capable of delivering extremely high discharge rates and responding in microseconds or milliseconds to prevent voltage drops during sudden power surges.
• Lithium Battery Technology: LiFePO4 (lithium iron phosphate) batteries have become the preferred choice due to their high energy density, long cycle life, excellent round-trip efficiency, fast charging capability, and superior thermal stability for mission-critical infrastructure.

Together, these systems create a resilient power architecture capable of supporting continuous AI operations while managing both long-duration energy needs and instantaneous power fluctuations.

How to Prepare for the Data Center Power Boom

• Assess Regional Geology: States and utilities should evaluate their saline aquifer storage capacity and begin planning carbon capture infrastructure in regions with strong data center growth, particularly Texas, Virginia, Pennsylvania, and Ohio.
• Invest in Grid Modernization: Electrical utilities need to upgrade substations, high-voltage transmission lines, and smart systems to manage the massive power demands that will arrive over the next five years, preventing delays to data center projects.
• Diversify Energy Sources: Data center operators should combine natural gas-fired power plants with renewable energy sources like solar and wind, paired with battery storage systems, to meet reliability requirements while reducing emissions.
• Plan for Cooling Infrastructure: Beyond power, data centers must deploy advanced cooling systems including direct liquid cooling, immersion cooling, and rear-door heat exchangers to manage the enormous heat generated by GPU clusters.

The scale of the challenge is enormous. Global data center electricity consumption is projected to increase from around 600 terawatt-hours in 2026 to more than 1,000 terawatt-hours by 2030, roughly equivalent to Japan's current total electricity consumption. Data centers will account for nearly half of all electricity-demand growth between now and 2030.

Yet the Rice study suggests that the infrastructure to address this challenge already exists beneath our feet. The geology is there; what's needed now is the investment, planning, and policy frameworks to deploy carbon capture and storage at scale alongside the massive buildout of AI computing infrastructure.

"It does show that the geology exists to make a meaningful impact, especially in states where data center growth is strongest," noted Hon Chung Lau.

Hon Chung Lau, Adjunct Professor, Rice University

The next five years will determine whether the AI industry can power its explosive growth responsibly. The tools exist. The question now is whether companies, utilities, and governments will deploy them fast enough.