Custody Without Ownership

No more AI. SI It’s better. — Elon Musk

As a description of the White House Accord on Super Intelligence, Elon Musk's verdict was fair. The accord is voluntary, "morally binding" only in the president's own phrasing, and its central promise of independent audits was never defined ^[1][2]. But the joke audited a single paper. Around it, across the same year, something harder was being assembled: a layer of control over the frontier AI companies that no one story captured. The stakes had been set in plainer terms.

Whoever wins SI, whoever wins superintelligence, wins. — Donald Trump

What got built across that year was not ownership. It was custody. The first act was the quietest. In June an export-control directive — the same legal machinery that governs chips — forced Anthropic to abruptly switch two of its systems off for foreign nationals ^[3]. Model access now travels under the same rules as silicon: a federal directive can shut it off. The second is a command structure. A Super Intelligence Force co-chaired by the Director of National Intelligence and the Pentagon's chief technology officer, with the labs required to report AI incidents up that chain ^[4][5]. The third points outward. Washington and Beijing opened a two-power channel: an incident hotline and an exchange of testing practice before models are released, styled openly on the arms-control machinery of the 1980s ^[6][7][8]. The perimeter was aimed at the rival. The strange part is that ownership was actually proposed. Anthropic's chief executive offered the deed himself.

my view here is it has always been very strange that this technology is being built by a private company. — Dario Amodei

Washington took the keys instead. The refusals came in pairs, often in the same documents and the same days. The immunity denial sits in the same dialogue that produced the hotline: Washington embraced the two-power channel with Beijing and refused the labs' request for immunity from liability ^[6]. The administration now commanding incident reports had already hollowed out its own safety shop — the Commerce Department's AI-safety center stopped publishing its research, its staff redirected to report informally to the White House, a redirection critics warn politicizes its bioweapons and cybersecurity evaluations ^[4]. The accord's "independent audits" never got a definition of who counts as independent, so private standard-setters wrote their own certification to fill the gap ^[2]. The measuring stick stayed with the industry being policed. Inside seventy-two hours in November the divide showed itself three times. The probe into Nvidia's chief executive went public ^[9]. Two days later, on November 10, the Super Intelligence Force launched ^[5], and TSMC's board approved up to $1 billion into Anthropic and OpenAI ^[10]. A state command, a private investment, and a counterintelligence investigation, inside three days. Anthropic is meanwhile preparing a float estimated at $1.4 to $1.5 trillion, with its own chip suppliers buying in ^[11][10]. The labs stayed monetizable. Outsiders read the shape early.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. — Anthropic

The Huang investigation is the perimeter reaching a person, and it warrants care. The FBI alleges an intelligence-linked handler and roughly $7.5 billion in restricted chip shipments ^[9]. The White House did not order the inquiry; the president was still publicly defending Huang in the story's first days ^[12]. The allegation remains unproven, and the company's position is flat.

no reason to believe that either national security or Nvidia secrets were compromised — Nvidia

Two wry notes. Markets steadied within days ^[13]. And Nvidia was still selling "sovereign AI" as a product the same week.

Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence. — Nvidia Inc.

The next bad day divides cleanly. The incident report goes up the new chain to the Director of National Intelligence's desk, while the liability stays with the shareholders ^[6][5][11]. Nothing in the record yet shows whether this custody layer is a stage toward ownership or its replacement. The first real use of the machinery is the thing to watch.