Own the Context, Rent the Intelligence

The Kill Switch Was Real

Three weeks ago I argued that the durable advantage in corporate AI wouldn’t come from having the best model. It would come from having the best memory — a compounding, structured record of how your organisation actually thinks, held somewhere you control. I called the problem the Dory problem: we wanted Mike Ross, the associate with total recall, and we got Dory, brilliant in the moment and reset by lunchtime.

I wrote it as a bet. A contrarian one, I thought — the whole industry was still measuring itself in benchmark points and context windows, and here I was saying the model was about to become the least interesting part of the stack.

I did not expect the argument to be settled this quickly. In the three weeks since, two things happened that I would have filed under “maybe, in a year.” Neither of them was mine. Both of them made the bet look less like a bet.

The first thing: the people selling the models told you not to depend on them

When you make a memory-over-model argument, the obvious rebuttal is “well, the model companies would say otherwise.” So it’s worth noticing when they don’t.

Satya Nadella published a short essay in June with a line I keep coming back to: you can offload a task, or even a job, but you can never offload your learning. His framing is that every company now runs on two kinds of capital — the human kind (judgement, relationships, pattern recognition) and what he calls token capital (the AI capability you build and own). The opportunity, he says, “is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound.” And the test of whether you actually own that loop: can you swap out a generalist model without losing the company veteran built into your system?

That is the memory-over-model bet, restated by the CEO of a company whose most obvious incentive is to sell you a model and keep you there.

Then Microsoft put money behind it. A fortnight later they announced Microsoft Frontier Company — a new $2.5 billion operating business, six thousand engineers, built to embed inside customers and turn their proprietary knowledge into systems that compound from within. The pitch, in their own commercial-chief’s words: “no company should be dependent upon any one model or any one model’s harness.” Model diversity by design. Your intelligence, protected — explicitly not used to train models in ways that commoditise what makes you different.

I don’t cite this because I think Microsoft has the answer. I cite it because when the incumbent bets a business this size on own-your-context-and-keep-the-model-swappable, the thing I described as a contrarian bet is now the consensus product thesis. The disagreement in the market isn’t whether the loop is the asset. It’s who gets to own it — and where it lives.

The second thing: the risk I was hedging against stopped being hypothetical

The uncomfortable half of the memory-over-model argument was always the downside case. If your institutional knowledge lives inside a single vendor’s harness, you are one decision away from losing access to your own thinking. I made that argument in the abstract, the way you make arguments about risks you don’t expect to see demonstrated on a Tuesday.

Then it was demonstrated on a Tuesday.

In the middle of June, the US government restricted foreign access to Anthropic’s frontier models — Fable and the more capable Mythos — over cybersecurity concerns. The restriction was broad enough that Anthropic disabled access globally, including, reportedly, for some of its own staff. A frontier model that thousands of companies had built on was, for eighteen days, simply switched off. It came back only after closed-door negotiations whose terms remain private. And the fix that brought it back — new classifiers, broadened guardrails — comes with an admission from Anthropic itself that it now blocks more legitimate work than before.

Sit with the shape of that. Not “the model got more expensive.” Not “the model got worse at a benchmark.” The model became unavailable, overnight, by a decision made somewhere you had no seat, and came back on terms you can’t read, behaving differently than it did the week before. If your company’s memory — your decisions, your context, your accumulated judgement — lived inside that model’s harness, your recourse was to wait and hope.

Every enterprise customer watching that happen learned the same lesson, whether they wanted to or not. In Europe it’s already turning into a policy reflex: build homegrown AI so the thing you depend on can’t be pulled out from under you without warning. That’s not sovereignty as a slogan. That’s sovereignty as a purchasing requirement.

What this actually means (and what I’d do about it)

Put the two together and the design principle writes itself: own the context, rent the intelligence.

Rent the intelligence, because the model is going to keep changing hands, changing price, changing rules, and — as we just saw — changing availability. Renting is fine. Renting is correct. You don’t want to own a depreciating frontier model any more than you want to own the power station.

But own the context. Your organisation’s memory — the decisions and the reasons behind them, the briefs that build on the last nine briefs, the hard-won “we tried that and here’s why it didn’t work” — should live in a form you can read, move, and keep, independent of whichever model is cheapest and smartest this quarter. For us that’s plain text files, versioned, structured by a schema we control, maintained by the AI but owned by the company. Unglamorous on purpose. The whole point is that it survives the next model upgrade, the next price change, and the next Tuesday.

I’m not describing this from the sidelines. At Janus Digital we’ve been operating on this conviction for months — long enough that the last three weeks read less like news and more like confirmation of something we’d already committed to. We built a system for it. We call it the Prime Radiant: the company’s memory held as structured, version-controlled text that the AI maintains and we own, governed by a rulebook rather than a vendor. It started inside our AI Office, and it’s now rolling out across the company one function at a time — each department gets its own instance, its own accumulating context, the same portable substrate underneath. None of that was a reaction to the Fable ban or to Microsoft’s announcement; both simply landed on the side of a bet we’d already placed. The discipline that felt slightly paranoid when we started — keep the memory in our hands, treat every model as a tenant we can evict — is the discipline the whole market is now reaching for.

And there’s a test you can run today, borrowed from Nadella: if you switched off your current model provider tomorrow, would you keep the company veteran — the accumulated expertise your AI has built up — or would it walk out the door with the vendor? If the answer is “it walks,” you don’t control your AI. Your vendor does. You just haven’t been sent the invoice for that yet.

The one thing I got wrong

I said inference would become like electricity in eighteen to twenty-four months — infrastructure-level, essentially free, universally available — and that the organisations that win would be the ones with the best memory, not the best model. I still believe the destination.

What I got wrong was the timeline on the risk. I wrote as though we had eighteen months to get our memory out of other people’s harnesses. June said otherwise. The commoditisation of the model is arriving on schedule. The fragility of depending on one is already here.

You don’t have a year to start owning your context. You have until the next Tuesday.