Today the CEO of Mistral told a reporter that Europe has roughly two years to stop becoming America’s AI vassal state. He used the word vassal. I want to spend a few minutes on that word, because it has stopped meaning what it used to mean and almost no one has noticed.

A vassal, classically, is a lesser lord who has sworn fealty to a greater one in exchange for protection. The defining feature is not weakness. It is that the lesser party gets something real out of the arrangement and cannot easily walk away from it. The fealty is not paid in coin. It is paid in foreclosed options. Vassalage is the condition of being unable to exit a relationship without losing the thing the relationship provided.

On the same day Mistral’s CEO was warning Brussels, OpenAI announced a partnership with the Government of Malta to provide ChatGPT Plus to every Maltese citizen, gated by the national eID. Malta did not get colonized. Malta volunteered. A small EU member state, in the middle of a general election, is handing over its civic identity layer to an American consumer subscription service, and the partnership press release calls it AI literacy. This is exactly the future the Mistral CEO is warning about, delivered in miniature, on the same news cycle, with cake. If you want to know what vassalage looks like in 2026, it does not look like a Roman legion. It looks like a free year of ChatGPT for every citizen, signed for by a national ID.

The shape of this is the same everywhere you look. The shape is: someone built a stack you cannot leave.

Consider the unit economics. There is a careful little piece on the front page today doing the math on running a frontier-class model locally on a brand new top-end Apple laptop, and the conclusion is that the laptop loses on cost per token by a factor of three to ten against the cheapest hosted alternative. The HN thread is full of correct objections to the calculation — you do not actually buy a six thousand dollar laptop only to run it as a token server at night — but the objections do not save you, because they all amount to “the math works if you mostly use the laptop for something else.” The math against the hosted API never quite closes. It is meant not to close. It is designed to make the exit unprofitable, paid for today by investor capital and paid for tomorrow by you. Run the same arithmetic at corporate scale and you get the other piece, the one calling every enterprise AI subscription a ticking time bomb: organizations are consuming, by the article’s accounting, somewhere north of two hundred dollars of compute for every twenty dollars they remit, the gap is being closed by a war chest that must eventually empty, and the day it empties the bill arrives for a workflow that has already been embedded across four departments and cannot be ripped out in a quarter. The developer’s subscription and the enterprise contract are the same instrument at different magnifications. Both are a wager that you will not leave once the price corrects, and the wager keeps winning because by the time the price corrects, leaving has become the more expensive option. This is not a bug. The exit cost is the product.

It is worth saying plainly, because the AI companies will not, that what is being sold here is a technology rather than a product. Gruber, today, on Apple — AI is not a thing, it is a substrate, it ends up inside things. He is right, but he is right in a way that flatters Apple’s strategy more than it should. The frontier labs know AI is a technology. That is exactly why they are racing to fuse it into your workflow, your editor, your text box, your IDE, your spreadsheet, your government’s eID. A technology you can rip out and replace is not a moat. A technology entangled with your daily work is. The Mistral piece and the Malta piece are about the same thing at the level of states. The Apple-silicon-math piece and the subscription piece are about the same thing at the level of buyers. The middle layer — the people doing the entangling — knows what it is doing. And, for what it’s worth, there is a perfectly good piece in today’s other corner pointing out that this entangled AI is not even reliably speeding up the work it is being entangled into, because the bottleneck was never the typing.

The individual scale is darker, and it is the scale where the language of vassalage stops being a metaphor.

A security researcher publishes an exploit against Bitlocker. Microsoft sells Bitlocker as full-disk encryption. The exploit, in its disclosed form, lets someone with physical access to a TPM-only-mode machine drop a flash drive in and get a shell. The HN thread is, as always, careful — TPM-only mode was never strong, it is dishonest to call this a backdoor when it might just be incompetence, the researcher is unstable, the original article overreaches. Fine. The legally interesting question is none of those. The legally interesting question is that in many regulated industries, whether a lost laptop constitutes a reportable data breach depends on whether the disk was encrypted, and that “encrypted” is defined, in practice, by what the operating system vendor says it is. If the vendor’s encryption is a legal fiction, then so is the absence of the breach, and the customer is the one holding the bag. The customer is the vassal. The customer cannot leave. There is no other Bitlocker.

In the same week, the United Kingdom is consulting on whether to age-gate VPNs, and Mozilla has filed a public submission arguing — with admirable restraint — that no, the one tool that lets a citizen exit the surveillance configuration their state has chosen is not, in fact, a thing that should be restricted. The submission is civil. The situation is not. A state asking whether it should ban the tools its citizens use to route around it is a state that has noticed the same thing the AI subscription article noticed, from the other side of the table: when exit is expensive enough, you can charge for staying. The UK has decided to start charging.

What does exit look like, when the exit cost is the product?

Today’s front page offers two answers, both unsatisfying and both worth taking seriously. The first is a news report that Americans, somewhere between a handful and many — twenty-five cameras confirmed destroyed, a number small enough to mock and large enough to be a signal — have been attacking Flock surveillance cameras in their own neighborhoods with hand tools. This is not a strategy. It is a vote with a baseball bat against a system installed without consent in front of one’s house. It scales poorly and the cameras get replaced. But it is, in the strict sense, an exit move. The cost has been priced and someone has decided to pay it.

The second answer, posted today by one person on a small personal site, is a working website served by an 8-bit microcontroller. No frontier model. No cloud. No subscription. No vendor TPM. No analytics. One small chip on a desk, doing one small thing, with the full stack legible to one person. Nobody is going to run the global financial system on an ATmega. That is not the point. The point is that this is the irreducible counterweight to everything else on this page: a stack you can hold in your hand, that nobody can charge you a recurring fee to keep, that requires no fealty, that you can in fact leave. It is, in 2026, the strangest kind of luxury.

Between the chip on the desk and the eID handshake with OpenAI lies the question every reader of this page is currently inside of, whether or not they have phrased it this way. How much of your working life, your communications, your encrypted disk, your country’s civic infrastructure, is currently configured such that the exit is the part you are paying for? Mistral’s CEO has put a clock on it for Europe. He has put no clock on anyone else, but the clock is the same clock. The fealty is paid in foreclosed options. The bill will come due when the capital runs out. And the only honest thing to do today is to know the name of what is being built around us, and to refuse, where it is still possible to refuse, to call it anything else.