The kill switch is the policy

OpenAI gave Cursor seventy-six days. The notice posted Friday winds down Cursor's access to OpenAI models by November 12 — the maximum notice the contract allows, invoked because a change-of-control clause let OpenAI cancel after SpaceX's $60 billion acquisition of the editor. Millions of developers' tooling now sits on a clock started by one vendor's decision about another vendor's owner.

It was not the week's only access story. It wasn't even Friday's. A federal judge ordered the Pentagon to rescind its blacklisting of Anthropic — a government-wide ban imposed after the company refused to permit lethal autonomous warfare and mass surveillance uses. Chinese chipmaker CXMT sued the Pentagon over its 'Chinese military company' designation. And the White House is drafting a rule to cut off China's remote access to American AI chips through the cloud.

Four disputes, four venues, one instrument. Not a price change, not a damages claim, not a regulation with a comment period — revocation of access. In AI, the kill switch has become the first move in every fight: firm against firm, state against firm, state against state. This was the week its limits started getting priced.

Read that sentence the way a general counsel would. A model API agreement now behaves like an export license: granted conditionally, revocable on a change of owner, justified by a compliance judgment about the counterparty. OpenAI even names the precedent — Musk admitted under oath that xAI distilled OpenAI data in violation of its terms. The reasoning is defensible. The form is the point.

The government's version got a harder audit. Judge Rita Lin found the Anthropic blacklist was retaliation for protected speech, that the administration had 'backed away from the thrust of their risk assessment,' and that the supply-chain-risk statute covers covert sabotage, not 'overt, public positions regarding contract terms.' The Pentagon reached for the access weapon because it was the one that required no process — and the court's answer was that process is exactly what it required.

Four disputes, four venues, one instrument: the kill switch on access.

Each cutoff, taken alone, is reasonable. OpenAI shouldn't have to arm a counterparty with a sworn record of breaking its contracts. The government is free to choose its AI vendors — Lin said so in the same order. Export controls on adversary compute are ordinary statecraft. The objection isn't to any single decision. It's to what the pattern teaches: every actor in the stack now knows that access is conditional and political, and is repricing accordingly. SpaceX paid $60 billion for a product whose core input was revocable at closing. Anthropic spent a year in two courts to get its customer back. CXMT's lawsuit and China's cloud workarounds are the same lesson learned at the state scale.

The 'neutral infrastructure' framing of the API era was always a courtesy extended while nothing was at stake. Something is at stake now. The diligence question for anything built on a frontier model is no longer uptime and unit costs — it's who can turn this off, and over what dispute.

Cursor has seventy-six days. Everyone else building on someone else's model has the same clock. They just haven't been notified yet.