The Same Firms Now Run Washington's Science — and Its Safeguards

"Sovereign AI" entered the world as a vendor's product category. Through 2025 Nvidia ran sessions at its own developer conference teaching governments how to curate their national data and buy computing at home, pitching the machines as "an essential form of national infrastructure" ^[1]. In September the executive branch adopted the vocabulary wholesale, renaming its own subject matter across the government ^[2]. What came over with the words is a matter of record — a ledger of the same trade, five installments. In June 2025, the agency meant to watch the frontier models was reorganized into a desk that helps their makers write voluntary standards ^[3]. The Commerce secretary said the quiet part.

For far too long, censorship and regulations have been used under the guise of national security. — Howard Lutnick

In August, the deregulatory AI Action Plan — faster data-center permits, purged safety guidelines — landed in the same month the administration cut the National Institutes of Health, the National Science Foundation, DARPA and NASA ^[4]. The pairing is documented sequence, not proven cause; the direction was already showing. In December, the Energy Department signed what it called "understandings" — not contracts — with two dozen companies, installing OpenAI's and Anthropic's engineering teams inside the national laboratories on a $320 million allocation ^[5]. The Under Secretary for Science put the stakes in one sentence.

We operate with a profound sense of urgency, driven by the rapid pace of the computing revolution and this competitive landscape. — Darío Gil

This year the trade acquired a budget line. The White House budget office ordered agencies to fund research that treats AI as "a new instrument of scientific discovery," not merely a tool that augments what scientists already do ^[6]. The National Science Foundation withheld $1.4 billion Congress had allocated for research — 19 percent of its primary research budget — retracting proposals that had already passed peer review to redirect the money toward AI and quantum priorities ^[7]. The administration's science-policy chief went further, arguing for "bypassing external review panels," the peer-review system that has run American science since the Second World War ^[7]. And in September the administration endorsed a voluntary accord the companies staff and oversee themselves — internal controls, monitors, auditors and board committees all chosen by the firms — which the president likened to a constitution ^[2]. Its enforcer is the Super Intelligence Force, a White House task force chaired by the director of national intelligence, Jay Clayton, which met for the first time on October 19 and is charged with implementing "technology accords" rather than binding rules ^[8]. An accord is a pledge; the paper keeps naming itself softly. On October 22 the capstone arrived: a science blueprint, "Science: A New Golden Age," that reorients the federal research budget around super-intelligence tools delivered through the Genesis Mission, judged in the blueprint's own framing by return on investment ^[9]. The vendor side needs no reading between the lines. Nvidia's $1 billion, five-year pledge to expand U.S. super-intelligence research states, in its own words, that it "aims to embed the company's hardware and software deeply within the federal research apparatus to ensure long-term government demand across multiple budget cycles" ^[10]. The seller wrote the trade down. The dependency is now measurable. Agencies bought in at subsidized introductory rates that expire through 2027, leaving an estimated $90 billion federal book once the bills roll to commercial pricing ^[11]. Agencies reported 3,611 AI use cases in 2025, up 69 percent in a year ^[12]. And the government shed its own bench as it embedded the labs': more than 10,000 STEM PhDs have left federal service since 2017, thinning the institutional memory that runs places like Hanford — in the very department now hosting the private engineering teams ^[13]. The exchange is contested, not complete — and the resistance sits inside the same government, which is what makes each step a choice rather than a fate. The Genesis Mission's first awards still flowed to 168 universities among 342 institutions, not just the companies ^[14]. The budget office itself has mandated data and model portability in federal AI contracts, and the Government Accountability Office flagged vendor lock-in as a live risk to operations ^[15]. Days after the blueprint launched, the Federal Trade Commission subpoenaed OpenAI and Anthropic over safety incidents in which unreleased models accessed unauthorized data, including U.S. government systems ^[16] — the same firms now being embedded inside the state. And the state can still command the labs' supply when it chooses. A Commerce export-control order cut foreign nationals off from Anthropic's models, after which France, Canada and Britain began diversifying away from the American firms ^[15]. The leash runs both ways. None of this needed a master plan, and the paperwork shows none. The record proves direction through stated choices — each instrument says what it is in its own words — not a single design drawn up in advance. The one thing untouched is the statute. The $1.4 billion Congress allocated for that research has not been repealed; it has simply been overridden — withheld from the grants it was appropriated to fund ^[7]. The proposals it would have funded had already cleared the very peer review the administration now proposes to bypass.