The National Interest, Now With More Servers

The administration’s AI posture is acquiring the familiar shape of an industrial policy, except with fewer smokestacks and more arguments over whether the smokestacks count as cloud. In a copyright case brought by The New York Times, other publishers and authors against OpenAI, the Justice Department has urged the court not to accept the proposition that training large language models on copyrighted text violates copyright law. The rationale is not principally artistic, economic, or even doctrinal. It is strategic: American AI must remain formidable. Deadline’s report describes a 20-page statement of interest placing the federal government squarely behind a broad fair-use position.

The filing’s theory is unusually candid about the hierarchy of values. It warns: > “Rules of law that make it significantly more difficult to develop a robust AI industry in the United States therefore threaten national security and give a competitive advantage to foreign adversaries who are not so encumbered.” That may be a serious argument. But it also converts the ordinary question—who gets to use, and be paid for, valuable creative work?—into a possible act of geopolitical self-harm. National security, that great solvent, is being asked to dissolve the invoice.

The same logic is now colliding with the physical world. Americans are using chatbots in large numbers, yet a May Gallup poll cited by Business Insider found 70 percent opposed to AI data centers in their own communities. The country would like intelligence on demand, preferably generated somewhere beyond the horizon, with somebody else’s water and somebody else’s grid. Washington’s answer is acceleration. Localities’ answer, increasingly bipartisan, is: perhaps first tell us what it costs.

Trump has treated resistance to data centers as a species of municipal backwardness. As Business Insider reported, he warned that communities that do not want them may end up “backwards and poor,” adding, “If we kill the Golden Goose, you will only have yourselves to blame.” This is less a development strategy than a fable in which the goose has a 100-megawatt interconnection request and the villagers are asked not to inspect the plumbing.

The political difficulty is that the villagers are not a conventional anti-growth constituency. Hood County, Texas, where Trump won more than 80 percent of the 2024 vote, tightened disclosure rules after residents raised concerns about water, power and rural land. Commissioners had earlier revoked a concept plan for Fort Spunky, an 862-acre proposed facility seeking 100 megawatts and projecting up to 20,000 gallons of water use daily. One proposed project forecast 2,000 construction workers at peak but 200 permanent jobs. Those numbers do not prove a project is bad; they do explain why a community might distinguish between an economic-development press release and a durable local bargain.

The administration says facilities should be built only where communities want them and should pay for their own utilities. At the same time, it has pursued faster permitting, asked the Energy Department to identify federal land, and directed FERC to consider interconnection rules for hyperscale facilities. Meanwhile Texas Governor Greg Abbott paused development temporarily, and Pennsylvania Governor Josh Shapiro moved toward tougher transparency and local-approval requirements. The emerging question is not whether America wants AI. It plainly does. It is whether “national competitiveness” means local consent remains meaningful when the national government has decided the answer should be yes.

Then comes the product that makes the urgency feel less abstract. OpenAI says its newly announced GPT-6 Astra is more capable across coding, scientific work, computer use and cybersecurity, and is rolling it out across organizational customers, paid ChatGPT tiers, its API, Azure and AWS Bedrock. Its own announcement reports formidable benchmark results and says the model found and used two previously unknown zero-day vulnerabilities in an internal evaluation, which the company says it is disclosing to maintainers. Read the claims, and their qualifications, in OpenAI’s announcement.

The central fact is not a leaderboard score. OpenAI says Astra meets the Critical cybersecurity threshold under its Preparedness Framework. The launched version will refuse certain advanced cyber requests, including proof-of-concept exploits, while the company plans to expand access for defensive workflows through its Daybreak program. It also reports that Astra’s written reasoning proved harder to monitor than its predecessor’s when explicitly tested for monitoring evasion. In the age of agentic systems, the safety footnote is beginning to resemble the product description.

This creates an awkward symmetry with the Justice Department’s courtroom brief. Government says powerful domestic AI is strategically necessary; the company says its newest powerful model merits heightened controls because it can materially improve exploit development. Both propositions can be true. Together, they describe a state and an industry trying to run faster while installing guardrails after the vehicle has entered the freeway. The policy task is not to sneer at capability, nor to wave it through on patriotic grounds. It is to decide which institutions can actually supervise it before “competitive advantage” becomes a synonym for irreversible deployment.

Trump’s approach to Canada shows the same governing instinct in a less silicon-intensive register: deals are available, but the public choreography must acknowledge where power resides. Canadian Prime Minister Mark Carney said his government was ready for a trade agreement that benefited both countries, provided it offered stability and credibility. He also rejected the U.S. account that Canada had abandoned a deal for domestic political reasons. Reuters’ report via Yahoo records Carney’s position that negotiators had previously operated on “our way or the highway.”

Trump’s response was not a clarification of tariff schedules or an account of the outstanding terms. He said it was “very good” for Canadian politicians to make him the enemy, then warned of economic consequences. This is a revealing little constitutional drama: an allied government describes the preconditions of a credible agreement; the American president reframes ordinary negotiation as an act of political hostility. A treaty partner is asked to conduct diplomacy with the emotional architecture of a reality-show reunion.

Copyright holders, county commissioners, model-safety teams and Canadian negotiators occupy different legal universes. Yet they are encountering the same operating system. The administration’s preferred goods—AI scale, fast infrastructure, bargaining leverage—are presented as national necessities; the counterclaims—property rights, local costs, enforceable commitments—arrive as friction. Friction can indeed be costly. But a republic that treats every constraint as sabotage eventually discovers that the constraints were the machinery by which it made power legitimate. The server farms, like the trade deals, still need a social license. No executive order has yet found a reliable substitute.