In the Coming AI Cold War, the Country That Owns Its Memory Wins

Two weeks ago the US government ordered Anthropic to cut off its newest model for anyone who wasn’t American. Three days after launch, no notice, no appeal. We called it a warning. The ink was barely dry.

On Friday OpenAI shipped GPT-5.6, the most powerful model it has built, and almost no one can use it. The top tier, Sol, went to about twenty companies the US government approved by name, cleared, per a staff memo, “customer by customer,” with no published criteria. It is the first time an American lab has launched a frontier model from behind a government access desk. Even OpenAI is uneasy, saying it does not believe this kind of “government access process should become the long-term default.”

Both sides are pulling back

Two cutoffs in a month is not a glitch. It is the opening of an AI cold war, and both sides are pulling back. In April, China blocked Meta’s two billion dollar purchase of Manus, one of the best AI agents. Washington gates its models, Beijing guards its own, and the most powerful AI is becoming something you are permitted to use, not something you can buy.

Plenty of serious voices call this overblown. Foreign Affairs dubs the AI race a myth, and others argue export controls mostly push rivals to build their own. Maybe. But you do not need to call it a war to see the exposure. The best models on earth now ship by government permission, and that permission may never come.

The best models just stay home

Here is how it plays out. Not with the machines going dark, but with the best ones quietly staying home. AI is starting to look less like software and more like fighter jets: too powerful to sell to anyone who might one day point it back at you. The frontier stays with the two countries that build it, and everyone else becomes a spectator over a tool they have already wired into hospitals, ministries, and payrolls.

You will still have a model

You will not be left without any AI. The need is too large and too profitable to leave unmet. Vendors in the countries that sit outside both blocs will build for the customers the giants are no longer allowed to serve, and even Washington and Beijing will likely keep releasing something abroad: older models, capped versions, whatever is safe to let go. The frontier may be closed to you but a capable model will not be. The real question was never whether you will have a model. It is whether you can move your work onto it the day your current one is pulled.

The model is rented. The memory is yours.

Here is what that memory actually is. Call it your context: the accumulated record of how you work, every task you have run, every document, decision, correction, and preference, built up over years until the AI stops guessing and starts getting things right the first time. One person’s context is a sharper assistant. A company’s is its institutional knowledge. A country’s is a real share of its productivity, the reason work that once took a month now takes a morning.

That is why a weaker model with deep context beats a stronger one without it. A frontier model that knows nothing about your work starts every task cold. A capable model that holds years of your context already has the audience, the constraints, the house style, the thousand small judgments that decide whether the output ships or gets thrown away. Most real work is not a hard puzzle. It is those small calls, and the model with the context to make them wins.

So a model cutoff is only a cliff if your context lived inside the model that vanished. Own your context, and the cutoff is a detour: connect it to the next available model and keep going. A morning lost, not a year.

Now run it across a country

A nation whose people’s context is locked inside two foreign vendors is one decision in Washington or Beijing away from disaster. The dependence is already deep: of the national sovereign-AI projects one recent index tracked, roughly seven in ten lean on a foreign partner, most of them American. A nation whose context is portable shrugs the same blow off, for a rounding error of what a frontier lab costs.

Make portability a right

Governments already do this elsewhere. You keep your phone number when you switch carriers because a regulator decided that letting one company hold it hostage was not acceptable. A phone number was worth protecting. The context a country runs its work on is worth far more. Europe is already moving on this, with one EU official putting the goal plainly: make sure “nobody has a kill switch” over the continent’s AI. The responsible move is to make context portable by right:

• Citizens own their AI context and can export it in an open format
• Public institutions can move their context between providers
• The right to leave is written into every AI contract the government signs

None of this picks a winner among models or favors one lab. It guarantees that a country can always move its work from one model to the next, so no single vendor, and no foreign government, can trap it.

The case for Canada

Canada will not out-build the frontier labs, and it does not have to. It can make sure that when the next model goes dark, and there will be a next one, every Canadian still holds the context they spent years building. We are starting to make that case in Ottawa. Every country outside the US and China should be too.

Rent the model. Own the memory. It is the one part of this you control, and the window is open right now.