Power Concentration in AI: Data Centers and Public Oversight

AI & Governance

Jun 27

David E. Nicholson · Originally published June 27, 2026 · Revised October 2026

Thesis

Over the last few years, artificial intelligence has transformed the technology sector.

As a result, major companies and new entrants compete fiercely for computing capacity, infrastructure contracts, investment, talent, and customers.

But within this emerging and expanding domain, a small number of companies are dedicating vast sums of money to build data centers. In turn, these few companies are placing enormous demands on existing electricity systems.

Yet, oversight of these activities is dispersed among institutions with a variety of legal responsibilities.

Federal, State, and local regulatory arrangements are not adequately equipped to handle the scale and pace of data-center development. This is evidenced by a lack of regulatory authority and expertise, an inability to assess the cumulative impacts of these projects, and a lack of information for the public.

Evidence

1. Concentrated capital spending

Four companies—Amazon, Alphabet, Microsoft, and Meta—outlined roughly $720 billion to $745 billion in capital expenditures for 2026. Much of this buildout is driven by AI, although the totals include spending on other company operations. These are projections, not dollars already spent. [1]

The point is not that every dollar goes into a U.S. data center. It is that a few companies can make investment commitments on this scale. The resulting projects arrive before utility commissions and local governments as requests for electricity service, new lines, land-use approvals, and other infrastructure.

Figure 1. Estimated 2026 capital expenditure by company

US$ billions · companywide forecasts

Amazon ~$220B

Alphabet $195–205B

Microsoft ~$175B

Meta $130–145B

0 125 250

Combined outlook: approximately $720–745 billion. Bars use approximate midpoints where outlooks are ranges. Fiscal periods and lease treatment differ. Not all spending is AI-related, in the United States, or on data centers. [1]

2. Electricity demand concentrates locally

Data centers made up 24% of Virginia Power’s electricity sales in 2023, 26% in 2024, and 28% in 2025. Dominion reports that the concentration of facilities around Loudoun County requires substantial new transmission and generation investment. [2][3]

These percentages cover data centers of all kinds, not AI computing alone. They describe one utility, not all electricity used in Virginia. The local effect is still significant: demand from one customer category now represents more than a quarter of that utility’s sales.

Figure 2. Data centers as a share of Virginia Power electricity sales

Percent of the utility’s electricity sales · 2023–2025

2023 24%

2024 26%

2025 28%

0% 10% 20% 30%

These are sales by one regulated utility, not all electricity consumed in Virginia. The category includes non-AI data centers. [2][3]

3. Gaps in oversight

The oversight evidence is less tidy than the spending figures, but it is more direct. Local governments report inadequate policies; Virginia’s legislative auditors found uncertainty about some powers; a Minnesota court found a specific failure to consider cumulative effects; and Virginia researchers documented widespread confidentiality agreements. [4][5][6][7]

These findings do not show that every agency lacks expertise. They show that some jurisdictions are making decisions without established rules, sufficient analysis, or information available for public debate. A 2026 Federation of American Scientists report also identified knowledge gaps and uneven local policy capacity. [8]

Exhibit 3. Documented problems in local review and public information

Source	Documented finding	What it establishes
North Texas (2026) [4]	72% of 31 respondents disagreed that they had clear local data-center evaluation policies; 78% said existing infrastructure would need major upgrades.	Survey evidence of local readiness gaps.
Virginia JLARC (2024) [5]	Recommended clarifying local authority to require water-use estimates and sound modeling; noted inadequate planning near some neighborhoods.	Gaps or uncertainty in specific legal tools.
Minnesota Court of Appeals (2026) [6]	Reversed Faribault’s decision not to require further environmental study; the record did not adequately consider cumulative potential effects.	Documented failure in one project review.
Virginia NDA study (2025) [7]	Researchers found at least one data-center NDA in 25 of 31 localities examined.	Limits on public information may arise; secrecy of approvals was not established.

Counterargument

Data centers can bring construction work and local tax revenue. Regulators are not inactive. Virginia’s State Corporation Commission approved a separate rate class for very large electricity customers, effective in 2027, with minimum payment requirements intended to protect other ratepayers. Courts can overturn inadequate decisions, as happened in Minnesota. [5][6][9]

These are real checks. Their existence also shows that regulation is changing while development is underway. The evidence supports failures and uneven readiness in particular places, rather than a claim that every local authority is incapable.

Conclusion

The concentration of capital is one part of the story. A few companies can finance data centers at a speed that changes what utilities and local governments have to plan for. Decisions are made project by project, often across agencies that do not have the same information or authority.

The evidence here identifies three problems: policies that have not kept up, incomplete assessment of cumulative effects, and limits on public information. The question is not whether all data centers should be stopped. It is whether public authorities have the legal powers, technical support, and disclosure requirements to judge the full costs and risks before approving major projects.

The broader Power Concentration in AI — Live Indicators page tracks related measurements over time. This brief draws on a smaller set of evidence.

Evidence and limits

The three exhibits measure different things. Company expenditure forecasts indicate the scale of investment, not direct control over U.S. electricity. Virginia Power figures reflect actual utility sales, but include non-AI data centers. The North Texas survey has 31 respondents and is not a national sample. The Minnesota ruling concerns one project; the NDA study establishes the presence of agreements, not secret approvals. The strongest broad claim is that regulatory readiness is uneven, not that every authority lacks expertise.

Method

I selected three questions: how much capital is being committed, what it means for electricity infrastructure, and where oversight has fallen short. Figures were taken from company-reporting analyses, utility filings, a government survey, a legislative investigation, a court opinion, and public-records research. Forecasts and actual observations are identified separately. Evidence that regulators are taking action is included as counterevidence.

AI use

The argument is mine, and the curation of evidence is mine. I used ChatGPT and Claude for research and editorial help, including identifying sources and counterevidence. I used ChatGPT to structure the post and I take responsibility for the final argument and the sources used.

Sources

David Nicholson