On February 24, 2026, Eric Masanet, Mellichamp Chair of Sustainability Science for Emerging Technologies at the Bren School, sat before the House Subcommittee on Investigations and Oversight to address one of the fastest-moving questions in U.S. energy policy: how much power and water will the AI boom actually require, and what will it cost the communities hosting it?
His answer was not a forecast. It was a diagnosis. After twenty years studying data center energy systems, including co-leadership of the first national data center energy study commissioned by Congress in 2007, Masanet told lawmakers that the U.S. cannot confidently answer the most basic infrastructure questions about its own AI build-out. The reason is straightforward: the data simply isn't there.
Masanet's testimony arrived at a consequential moment. Lawrence Berkeley National Laboratory projects that under a high-growth scenario, U.S. data centers could account for up to 12 percent of national electricity use by 2028, up from 1.8 percent in 2018. That is nearly half the electricity consumed by the entire U.S. industrial sector last year. Yet unlike power, buildings, manufacturing, and transportation, the data center sector has no standardized federal program for collecting and publishing the operational data analysts need to model its real-world impacts.
For the Bren School, Masanet's appearance is part of a broader pattern of faculty research moving directly into state and federal policy. Earlier this fall, Bren economist Kyle Meng advised Governor Newsom on California's cap-and-trade reauthorization. Masanet's congressional testimony, paired with a recent address to the California State Assembly, extends that footprint to the question now defining U.S. industrial policy: how to power artificial intelligence without flying blind.
What did Masanet tell Congress about AI and data centers?
Masanet framed his testimony around four blind spots that prevent analysts, policymakers, and local communities from making informed decisions about data center growth.
Operators report too little, too late
Most U.S. data center operators do not publicly report the energy and water demands of individual facilities. Apple discloses electricity use, Meta discloses electricity and water withdrawals, and Google discloses water withdrawals and consumption. Beyond those three, operators typically aggregate data at the company or regional level, masking the footprint of specific sites. Some report nothing at all.
When operators do disclose, they generally do so in annual sustainability reports covering the prior year. By the time analysts can use the numbers, they are months out of date. In a sector where new builds are coming online with novel cooling systems, behind-the-meter generation, and rack designs unlike anything that existed even two years ago, calibrating models to last year's data leaves a wide margin of error.
Nondisclosure agreements obscure new builds
Many real-time project details are not publicly disclosed because of nondisclosure agreements between operators, utilities, and local jurisdictions. That opacity affects more than analysts. It also affects the communities being asked to host the projects, who often cannot independently assess what a planned facility will draw from their grid or their watershed before approving it.
What is the "net" effect of AI on energy and water use?
One of the most consequential gaps concerns the downstream effects of AI applications themselves. While data centers add measurable load to the grid, the AI tools they run may either reduce or increase national energy and water use depending on how they are deployed.
This is not merely bluster. A recent Bren student innovation project, MediMRF, uses computer vision and robotic sorters at hospital loading docks to automatically classify waste streams before compaction. Misclassified waste is typically incinerated or autoclaved, both of which are energy-intensive disposal pathways. The team's environmental modeling shows that proper classification reduces carbon emissions by 65.4 percent and eliminates the release of carcinogenic air pollutants. The promise is real. The problem, on the larger scale, is that the evidence base behind most AI environmental benefit claims is not yet there.
Masanet pointed to a recent independent analysis concluding that many corporate claims about AI's environmental benefits are not adequately supported by evidence. Closing that gap, he argued, requires consensus-based standards for measuring application-level impacts, not marketing assertions.
Why open data is an innovation policy, not just an oversight policy
Masanet's case to Congress was not simply a regulatory argument. It was an innovation argument.
High-quality open datasets, he told the committee, catalyze the independent development of new models, software, and forecasts. He listed the kinds of insights they unlock: promising areas for technology R&D, opportunities for new policy designs, new grid optimization approaches, and improved understanding of how to reduce energy and water impacts.
This is why, Masanet noted, the United States has long invested in standardized data collection for every major sector of strategic national importance. Power, buildings, manufacturing, transportation. Every major sector except data centers.
How could Congress close the data center data gap?
Masanet's recommendations drew on established precedents rather than novel mandates. The U.S. Energy Information Administration and U.S. Census Bureau have collected, anonymized, and published data on every other major energy-consuming sector for decades. The data center sector is the conspicuous exception.
He pointed to the European Commission's Energy Efficiency Directive, under which many U.S. operators already report detailed facility-level data for their European operations. The reporting infrastructure exists. The operators are already using it. What is missing is a domestic equivalent.
He also highlighted the global cement industry's "Getting the Numbers Right" initiative as an example of an industry-led data sharing program that protects trade secrets while providing the public evidence base regulators and analysts need.
What does this mean for the Bren School?
Masanet's testimony reflects a research agenda that has shaped national and international understanding of digital infrastructure for two decades. His 2020 Science paper recalibrating global data center energy estimates corrected widely cited figures that had overstated growth, and his 2024 Joule commentary, "To better understand AI's growing energy use, analysts need a data revolution," laid the groundwork for the recommendations he carried to Capitol Hill.
The throughline for Bren is the kind of analyst this moment requires. Closing the data gap Masanet described will demand a generation of researchers fluent in energy systems, environmental impact assessment, and large-scale data, working at the intersection where federal AI policy is now being written. The school's programs in corporate sustainability, innovation, and environmental data science, are training that workforce. Masanet's appearance before Congress is what it looks like when that training meets the policy questions of the decade.
What's next?
Masanet's congressional testimony is one of two recent appearances bringing his data center research into public conversation. He also delivered an hourlong keynote to the UCSB Community of Practice for AI, examining the same questions in greater technical depth for a campus audience. The Bren School will cover that talk in a forthcoming article.
