The Data Center Subsidies Are Going to the Wrong Buildings

Public opinion on data centers is not settled. It is forming, and it is moving in one direction.

Pew surveyed 8,512 American adults in January, the first time it has ever polled this question. Three quarters have heard about data centers. On the environment, 39 percent say mostly bad against 4 percent mostly good. On home energy costs, 38 to 6. On quality of life for people nearby, 30 to 6. Then it flips: on local jobs, 25 percent say mostly good against 15 percent bad, and on local tax revenue, 23 to 12. Roughly one in five say they are not sure in every category.

Here is the finding that should concern anyone in this industry. People who have heard a lot about data centers are more negative than people who have heard a little, in every single category. Two thirds of those who have heard a lot say bad for home energy costs, against 42 percent of those who have heard a little. Sixty three versus 48 on the environment. Fifty one versus 35 on quality of life.

Familiarity usually softens opposition on technology questions. Here it does the opposite. That is my read of the pattern, not a finding in the survey, but it is the read that matters for anyone trying to site a facility in a county that has started paying attention.

What the best available research actually shows

Brookings researchers Dany Bahar and Greg Wright published in May and updated on August 10 with an expanded sample: roughly 1,500 US data center facilities plus 52 announced-but-canceled projects, linked to BLS county employment and wage data from 2003 to 2024.

The method is the important part. They compare counties that got a built facility against counties where one was announced and then canceled. That controls for site selection, because data center counties were already growing faster than comparison counties before any facility arrived. The authors say plainly that industry-sponsored studies which skip that control overstate the jobs impact.

What they found: data centers do create local jobs. Data processing employment rises 56 percent over the first decade, telecommunications 43 percent, working out to roughly 100 to 200 jobs in a typical county. Wages are unchanged. Home prices rise 2 to 5 percent.

Facility type matters in a way that is almost entirely absent from the policy debate. Hyperscale campuses, the Amazon and Google and Microsoft and Meta facilities, produce the telecommunications gains. Colocation facilities, the landlord model leasing space to remote tenants, do not.

The finding that should reorganize the debate

In hyperscale counties, state incentives amount to about 2 percent of total construction investment. Siting is driven by power, land and fiber, not tax breaks. In colocation counties, incentives are 62 percent of total investment.

The subsidies matter most for exactly the facilities that generate the fewest jobs.

For scale: Virginia's data center sales tax exemption cost an estimated $1.6 billion in fiscal 2025. Good Jobs First, in a study titled Money Lost to the Cloud, put the average at $1.95 million per job across eleven major projects, with the highest at $6.4 million per job. It recommends capping public subsidies at $50,000 per job. Twenty seven states run these exemption programs.

Disclosure, because it belongs in the piece and not in a footnote: Amazon, Google, Meta and Microsoft are general unrestricted donors to Brookings. The authors state the findings are not influenced by that, and they also disclose using an AI tool for data analysis and manuscript preparation, reviewed and revised by them.

Water: the correction that has to be in this conversation

Congressional Research Service report R48646 is the reference. American data centers directly used about 17 billion gallons of water in 2023, roughly three tenths of one percent of public water supply nationally. Municipal systems supply an estimated 97 percent of on-site needs. A large facility can directly consume as much water in a day as about 2,600 households.

But roughly 80 percent of a data center's total water footprint is indirect, meaning water consumed at the power plants generating its electricity rather than at the facility. Most of the viral water numbers combine direct and indirect and present the total as facility consumption. Both are real water. They are not the same claim, and conflating them makes the industry's rebuttal too easy.

The CRS explicitly declines to frame this as a single national water problem, because impact is intensely local. A facility in Arizona drawing on Colorado River supply is a fundamentally different situation from one in Ohio.

Power, and the number ratepayers actually feel

More than 4,500 active US data centers consume about 176 terawatt hours annually, roughly 4.4 percent of US electricity. That is a 2023 figure and is almost certainly higher now, with more than 700 additional facilities under construction across 40 states.

In the PJM region, which serves 65 million people across 13 states, capacity costs jumped from $2.2 billion to $14.7 billion in a single year, with IEEFA attributing nearly two thirds of that increase to data centers. That describes the 2024 auction. PJM has since cleared $16.1 billion and then a record $16.4 billion for 2027-2028, with the market monitor putting data centers at 38 to 40 percent of that latest auction and roughly 46 percent across the last four combined.

The policy question

There is a version of this debate where you are either for data centers or against them. That version is useless.

The useful version is narrower and more answerable. If hyperscale facilities site based on power, land and fiber rather than incentives, and they are the ones producing the employment gains, then the incentive programs are not buying the jobs. They are subsidizing the facility type that produces the fewest of them, at an average cost of nearly $2 million per job.

That is not an argument against data centers. It is an argument about where 27 states are spending money, and it is one that people who want these facilities built should want resolved, because the fastest way to lose siting approval in a county is for residents to conclude they were sold something that was not true.

One thing worth noting from last week's Wyoming primary: Eric Barlow won the Republican gubernatorial nomination running, among other things, on putting data center siting decisions in the hands of local governments. That is a market signal.

Full episode, with the polling, the research, and the water and power breakdowns: https://podcasts.apple.com/us/podcast/trump-claims-hormuz-as-us-territory-angie-nixon-stuns/id1626987640?i=1000784589643