The Trump AI Pivot — Biden's Order Rescinded, Stargate's $500 Billion Pledge

On 20 January 2025 the second Donald Trump administration took office. Within the first day, US AI policy shifted decisively.

Over the three days that followed, the AI regulatory framework built by the Biden administration was rescinded, the largest AI-infrastructure investment programme in history was announced, and a new executive order positioning "American AI leadership" as national strategy was signed.

20 January — Rescinding Biden's EO 14110

Executive Order 14148, Initial Rescissions of Harmful Executive Orders and Actions, signed on day one, revoked 67 Biden executive orders and 11 presidential memoranda. Among them was Executive Order 14110, Biden's October 2023 order on safe, secure, and trustworthy AI.

The rescission removed:

• Reporting obligations for large AI models (models trained above 10²⁶ integer or floating-point operations)
• NIST's mandate to develop AI safety evaluation guidelines
• Federal agencies' AI risk assessment duties

The official rationale was that regulation had become an obstacle to AI development. EO 14179, signed three days later, put it this way in its preamble: the order "revokes certain existing AI policies and directives that act as barriers to American AI innovation." Major industry firms welcomed the move publicly; safety-leaning researchers and a portion of the regulatory community criticised it sharply.

21 January — The Stargate Project

The next day, President Trump stood at the White House with Sam Altman (OpenAI), Larry Ellison (Oracle co-founder), and Masayoshi Son (SoftBank) to announce the Stargate Project.

OpenAI's own announcement is the outline:

• What it is: Stargate is "a new company which intends to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States."
• Scale: "We will begin deploying $100 billion immediately."
• Equity funders: SoftBank, OpenAI, Oracle, and MGX, the Abu Dhabi state-backed investor.
• Division of labour: SoftBank and OpenAI lead, "with SoftBank having financial responsibility and OpenAI having operational responsibility." Masayoshi Son is chairman.
• Technology partners: Arm, Microsoft, NVIDIA, Oracle, and OpenAI.
• Buildout: "currently underway, starting in Texas," with sites across the country under evaluation.
• Employment: claimed to "create hundreds of thousands of American jobs."

It is worth being precise about what kind of number $500 billion is: a statement of investment intent, not money spent and not a binding contract. For scale, four years at that rate is roughly 1.6 times the entire Apollo programme, which cost $25.8 billion nominal and about $309 billion in 2025 dollars by the Planetary Society's reckoning.

The technical community split. To some this was evidence the US was getting serious about competing with China; to others, evidence the US could keep an edge only through brute scale. The bluntest objection came from Elon Musk, who posted on X the next day: "They don't actually have the money," adding, "SoftBank has well under $10 billion secured. I have that on good authority." Sam Altman replied the same day—"Wrong, as you surely know"—and invited Musk to visit the site already under way.

23 January — A New Executive Order

On 23 January, Trump signed Removing Barriers to American Leadership in Artificial Intelligence.

What it actually says:

1. Section 2 states the policy: "It is the policy of the United States to sustain and enhance America's global AI dominance in order to promote human flourishing, economic competitiveness, and national security."
2. Section 5(a): review immediately every policy, directive, regulation and action taken under EO 14110, and suspend, revise or rescind whatever conflicts with that policy.
3. Section 5(b): within 60 days, OMB revises memoranda M-24-10 and M-24-18 on federal AI use.
4. Section 4(a): within 180 days, the Assistant to the President for Science and Technology, the Special Advisor for AI and Crypto, and the National Security Adviser deliver an AI Action Plan to the President.

That plan arrived on 23 July 2025 as Winning the Race: America's AI Action Plan, accompanied by three executive orders the same day on data-centre permitting (14318), federal procurement (14319), and AI exports (14320).

The principal axis of US AI policy turned officially from safety-first to dominance-first.

Six Days Later — The DeepSeek Shock

DeepSeek-R1 was released on 20 January, the day before Stargate. The market reacted a week after that: on 27 January NVIDIA fell about 17%, erasing roughly US$589 billion in market value—six days after the Stargate announcement.

The contrast was unkind. Days after the United States pledged $500 billion over four years, a Chinese team shipped a comparable open-weights model built on DeepSeek-V3, whose technical report put the final training run at US$5.576 million on the paper's own assumption of $2 per H800 GPU-hour. Is the colossal infrastructure spend really necessary, and is the assumption that deregulation restores a US lead still valid? The questions spread through Congress and the press.

Between the Announcement and the Concrete

Stargate continued anyway. On 23 September 2025, OpenAI, Oracle and SoftBank announced five more US sites—Shackelford County, Texas; Doña Ana County, New Mexico; a Midwest site later named as Port Washington, Wisconsin; Lordstown, Ohio; and Milam County, Texas. The announcement claimed "nearly 7 gigawatts of planned capacity and over $400 billion in investment over the next three years," putting the project on "a clear path to securing the full $500 billion, 10-gigawatt commitment we announced in January by the end of 2025." Over 300 proposals from more than 30 states had been reviewed.

Planned is not powered. When Epoch AI tallied the US sites in April 2026, the seven campuses added up to more than 9 GW of planned capacity—while the only one actually running was the flagship at Abilene, four of eight buildings, 0.3 GW. The rest were at foundations or land clearing. In March 2026 parts of the OpenAI–Oracle expansion were reported scaled back amid financing complexity and revised demand assumptions, which Oracle publicly denied; Abilene itself saw weather-related outages after energisation.

In January 2025 and in August 2026 alike, $500 billion is the total of what is meant to be built, not of what has been.

From Regulation to Development — A US Doctrine

The Trump administration's AI policy, through this sequence of documents and announcements, took on a clear shape:

• Domestic deregulation of AI development
• Massive investment in AI as national strategic infrastructure
• Not a simple tightening of export controls but a rewiring of them: the Commerce Department rescinded Biden's AI Diffusion Rule on 13 May 2025, days before it was to take effect, while keeping China-facing restrictions—and by mid-2026 was again extending licence requirements to Chinese-headquartered entities wherever located
• Acceleration of AI in military and national-security domains

This is distinctly different from the European model exemplified by the EU AI Act—risk-tiered, norm-centric—and now marks an explicit "development- and dominance-first" doctrine. The transatlantic divergence in AI regulation paradigms began in earnest in early 2025.

China, the United States, and the EU advance AI under three different sets of norms. The history of AI in the late 2020s can no longer be told without all three poles in frame.