The Constraints Washington Is Removing Were Never the Problem

Texas grid operators are holding connection requests for 466,497 megawatts of data-center power, more than five times the most electricity the state's grid has ever delivered ^[1]. The Justice Department has filed to shield OpenAI from copyright liability, arguing that training on others' work is fair use and that licensing fees would hand foreign adversaries an advantage ^[2]. Those two facts share the same mismatch. Over the past year Washington has dismantled a series of legal and regulatory constraints on AI: copyright, safety rules, local zoning. Each one was never the thing actually blocking the buildout. On copyright, the DOJ's filing frames licensing fees as a vulnerability ^[2]. On safety, a June national-security memorandum ordered the military to accelerate AI adoption for warfighting and gave the defense secretary 90 days to rewrite autonomous-weapons policy ^[3]. On zoning, the administration designated data centers critical national-security infrastructure ^[4] and pushed FERC to order grid operators to speed up connections for large power users ^[5]. Three prongs, three constraints removed, none of them the bottleneck. The bottleneck is physical and it is local. OpenAI itself has warned the White House that the country needs 100 gigawatts of new power a year, pointing out that China added 429 gigawatts last year to America's 51 ^[6]. Roughly 500 communities have imposed bans or moratoriums on data centers ^[7]. New York enacted the first statewide moratorium ^[8]. The House introduced a bill to ban large AI data centers on federal land outright ^[9]. A federal judge blocked the Pentagon's supply-chain designation of Anthropic ^[10]. Bank analysts now say the constraint has shifted from chips and energy to local utility regulation, zoning, and permitting ^[11]. The buildout does proceed where neighbors don't fight it: Meta's largest-ever data center is going up in rural Louisiana, Oracle secured $16.3 billion in Michigan, Amazon pulled 28 permits in Virginia ^[12]. That is the tell. It succeeds where local friction is low, not where federal override prevails. The military prong is doing something worse than missing the target. The Pentagon designated Anthropic a supply-chain risk after the company barred its model from autonomous weapons and mass surveillance, and barred other contractors from working with the firm ^[13]. The defense secretary made the standard explicit.

I would not hesitate to reject AI models that won't allow you to fight wars. — Pete Hegseth

OpenAI took the deal instead. 1.5 million users pledged to delete the app, uninstalls surged 295%, and the company's head of robotics resigned over lethal-autonomy concerns ^[14]. Altman later conceded the rollout was mishandled.

We believe our agreement with the Pentagon creates a workable path for responsible national security uses of AI while making clear our red lines: no domestic surveillance and no autonomous weapons. — OpenAI

And the coercion that set all this off was itself blocked by a federal judge ^[10]. The prong meant to weaponize the labs is corroding them. This week, faced with a wall it cannot override, the administration has turned to delegitimizing the people who built it. Trump framed opposition to data centers as a choice between progress and poverty.

If you want to go through poverty, crime and squalor, I would say don’t approve data centers. — Donald Trump

House Republicans are urging an investigation into whether China is funding the domestic opposition ^[15]. And the administration has pointed to suspected Chinese bot farms, some 200,000 accounts, as the force amplifying American resistance to data centers ^[16]. When you can't solve the problem, you brand the people pointing at it as foreign-backed.