AI Data Centers Hit a Trust Wall That NDAs Won't Fix

On October 3, Amazon Web Services chief executive Matt Garman buried a concession inside a long blog post defending data centers: the company no longer uses nondisclosure agreements when it negotiates with government agencies over new projects. Microsoft promised the same in January, then went further and said it would stop seeking local tax breaks. On the surface, that looks like the AI infrastructure industry finally opening the curtains.

Then, on October 9, three U.S. senators published the findings of a nearly yearlong investigation that explains why dropping an NDA is not the same as earning consent. Senators Elizabeth Warren, Chris Van Hollen and Richard Blumenthal interviewed and gathered written answers from seven of the largest data center operators, including Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix, and concluded that confidentiality agreements were not an edge case but a routine operating procedure, requested from utilities, landowners, commercial partners and in some cases public officials. The report, Power and Profits: How the AI Data Center Boom Costs Households and Communities, is the clearest evidence yet that AI's physical buildout has a legitimacy problem that no single policy tweak will solve.

The Secrecy Was the Product, Not a Side Effect

The most damaging part of the Senate report is not that companies used NDAs. It is the stated reason. According to the investigation, several of the hyperscalers "explicitly acknowledged that the goal of requesting NDAs with public officials is to limit public scrutiny by preventing the local community from learning about the project". Meta told investigators that confidentiality during development "increases efficiency and speed by enabling all stakeholders to remain focused on the needs of a proposed project", a sentence that defines local residents as something other than stakeholders. Environmental activist Erin Brockovich told TechCrunch that the number one complaint she hears about data centers is transparency, with a recurring pattern: projects announced after permits are already secured, developers who do not return calls, and officials who signed NDAs before their neighbors knew a project was under consideration. Academic work cited in the Senate report found that 25 of 31 Virginia communities with existing, approved or proposed data centers had nondisclosure agreements with local officials. In a system like that, publishing a press release later does not undo the fact that the decision had already been made.

Why Dropping NDAs Only Fixes Half the Problem

The pledges announced so far are narrower than they sound. Microsoft said in January it would stop seeking NDAs with local governments, but confirmed it would keep using them with state utility regulators and power companies. Amazon's October change covers government agencies going forward and does not revisit past agreements or bind its contractors. Google and Meta declined to commit to ending NDAs with local governments at all. Every company surveyed except Microsoft said it continues to seek every tax break available, including sales tax exemptions on chips and equipment, and none of the seven could provide comprehensive figures on the full-time jobs their sites actually create.

The money question is the harder one. All four Big Tech companies told the Senate they pay for the electricity they use, and on paper they do. The dispute is about shared infrastructure: new transmission lines and grid upgrades that serve everyone but would not be needed "but for" the data center. The companies broadly opposed a "but-for" cost allocation standard and argued those costs should be spread across all ratepayers, according to the report. In Louisiana, a public utility commission consultant estimated that new generation driven primarily by one hyperscaler's campus could cost the average ratepayer as much as $150 per year, while the company characterized the finding as inaccurate speculation.

Anthropic's chief executive Dario Amodei has called the backlash "fundamentally a crisis of trust", and that framing is useful because it predicts the outcome. When opponents are presented with the industry's own numbers, the most common response, as writer Jasmine Sun observed, is simply "I don't believe them". A company that spent years negotiating behind NDAs has no accumulated credibility to spend on a rebuttal.

The Real Constraint Is Power, Permits and Politics

The concurrency of the Senate report and Amazon's announcement lands at a moment when the buildout is already colliding with local government. More than 100 data center moratoriums are being considered across the United States, and New York imposed a one-year pause on permits for large facilities. At least 15 states have weighed pauses, and more than 100 localities have approved their own. Data Center Watch counted roughly 75 projects worth about $130 billion blocked or delayed in the first quarter of 2026 alone. An Economist/YouGov poll in late August found that 63 percent of Americans would oppose a data center in their own community.

At the federal level, the Ratepayer Protection Act, which would have created a framework states could adopt to make large facilities pay more of their grid costs, passed the House 417-3 before Senate Democrats blocked it in a 57-43 vote on September 30. Meanwhile, electricity prices in states with the heaviest concentration of data centers are running well above the national average, which is exactly the political fuel that keeps moratoriums moving. This is the same demand shock that broke the AI infrastructure funding market this month, when an Nvidia-backed operator pulled its $5.5 billion Australian listing hours before pricing in what became an AI funding reality check.

Builder Impact: Compute Is Now a Political Input

For teams shipping AI products, the practical lesson is that the marginal constraint on your inference budget is increasingly a county planning meeting, not a supply chain. Four implications follow.

First, treat siting and permit risk as a first-class supply risk. Capacity that was announced is not capacity that will be energized; moratoriums delay timelines by years even when they are framed as temporary studies. Any architecture that assumes a specific region comes online on schedule is exposed.

Second, expect cost pass-through to take longer and cost more. When utilities cannot shift grid-upgrade costs onto developers, they shift them onto ratepayers, which invites political correction, which invites further delay. Stable per-token economics increasingly depend on energy contracts, not just silicon.

Third, diversification is now a governance argument, not just an engineering one. Multi-vendor and multi-region routing protects against a single gatekeeper of power or a single hostile jurisdiction. The same logic that pushed teams onto multiple model providers now applies to compute regions.

What to Watch Next

The Ratepayer Protection Act is likely to return in some form after the midterms, and its reintroduction will be the clearest signal on how far federal cost-allocation rules can go. Watch whether Google and Meta move on NDA pledges now that two of their biggest competitors have. Watch state-level moratorium counts, which are the leading indicator of where new capacity will actually be built.

AI did not hit a public-relations problem. It hit the limits of a land, power and permitting system that was never designed for this rate of construction, and it is being asked to earn consent it spent years avoiding.

AI assisted with research and drafting. Factual claims are reviewed by an editor.