A strategic stake

Military power is no longer measured by the number of aircraft, tanks or soldiers. AI multiplies the capabilities that already decide the outcome of wars: a frontier model can sift in seconds through masses of intelligence that would keep analysts busy for weeks, compress decision time to the point of responding almost instantly, plan and simulate the adversary’s movements, and pilot autonomous systems immune to jamming^[1].

On the economic side, the same machine augments or replaces intellectual work on a large scale: it tips productivity towards whoever masters it, and builds dependence in whoever is deprived of it.

It is a cumulative advantage — the bigger the lead, the harder it is to catch up — and that is precisely what moves it from the rank of software to that of a strategic asset.

The US crackdown

On 12 June 2026, access to Fable 5 was cut off overnight. US Commerce Secretary Howard Lutnick had just ordered Anthropic to reserve its two most advanced Claude models — Fable 5 and Mythos 5 — for US nationals only. But the company, unable to verify its users’ nationality, chose to cut off everything^[2].

The move is part of a conflict that broke out in February 2026, when Anthropic refused, at the Pentagon’s request, to lift two contractual safeguards: one on fully autonomous weapons, the other on domestic mass surveillance. Retaliation came the next day: the White House ordered all federal agencies to stop using its technologies, and the Pentagon classified the company as a “supply chain risk” — a label until then reserved for suppliers linked to foreign adversaries, such as Huawei. Anthropic went to court: a federal judge suspended the designation, seeing it as a retaliatory measure, before an appeals court reinstated it for the duration of the proceedings^[3]. The official reason for the shutdown, for its part, is technical: a flaw reported by researchers that made it possible to unlock Fable 5’s cyber capabilities^[4].

What remains is the threshold that was crossed. Never before had the United States subjected an AI model to export controls — it had done so for chips, never for intelligence itself. By treating it as a controlled technology, on a par with nuclear power, Washington acknowledged that AI had moved from commerce to power. And the move was not one-way. On 26 June, a first letter restored Mythos 5 — the most powerful — for around a hundred authorised US organisations (tech giants, finance, federal agencies). Then, on 30 June, Washington lifted all the controls: Fable 5 returned for the global public on 1 July, with reinforced safeguards, while Mythos 5 remains reserved for authorised organisations^[5]. The episode does not fade for all that: less than three weeks were enough for a central model of global AI to be unplugged by administrative decision, then restored under conditions. The same week, OpenAI in turn limited access to its new models to “trusted partners”^[6]. Negotiated access is becoming the norm.

A risky decision

For users — businesses or individuals — the first lesson is brutal: a tool your activity depends on can disappear overnight, for reasons beyond your control. But the decision taken by the US administration also carries risks for American interests.

The first concerns intelligence, the cornerstone of American power. A centralised, proprietary model hosted by a provider subject to US law is not just a service to be protected. It is a single point of passage through which the queries, data and intentions of millions of users around the world flow — an intelligence sensor, all the more valuable because the law keeps it silent^[9]. Yet by pushing the world out of these centralised systems, the administration is not just losing a market: it is damaging its own sensor. A model running on a company’s private server in Europe or China is not so easily monitored.

The second risk is economic. Reserving the best models for American players alone might indeed give them an advantage over their foreign competitors. The cost, however, falls on one specific player: the US AI industry (the sector in which the country has invested most massively), whose models are being rationed.

In fact, companies have begun to flee the American offering for models that are cheaper and less exposed to arbitrary regulation. The movement does not spare the giants:

• Airbnb relies on Qwen, Coinbase on GLM and Kimi to cut its bill by almost half^[10].
• The start-up Lindy switched all its traffic from Anthropic’s models to China’s DeepSeek^[11].
• The programming tool Cursor built its latest system on Moonshot’s Kimi models^[12].
• Shopify replaced a data extraction pipeline based on OpenAI’s GPT-5 with a self-hosted system based on Alibaba’s Qwen — cutting its unit cost seventy-five-fold, according to its engineers^[13].
• Microsoft itself — OpenAI’s main financial backer — has announced that it is considering basing its Copilot agent on a version of DeepSeek, the ultra-low-cost Chinese model, fine-tuned and hosted on its own servers^[14].

The phenomenon is so visible that the US Congress has opened inquiries into Airbnb and into Cursor’s parent company for admitting that they use Chinese models^[15]. By restricting access to its products, Washington has opened a wide road for foreign competition.

Beijing lying in wait

The competition today is overwhelmingly Chinese — and its strength is no accident. Chinese open models — DeepSeek, Qwen, Kimi, GLM — dominate the open-weight model rankings, where they hold the top ten places, and are now close on the heels of the best American models — which are closed and paid — at a cost five to thirty times lower^[16]. The irony of the tech war: it is partly the American chip embargo that forced Chinese engineers to excel at efficiency, producing these frugal models that are now flooding the market. Constraint gave birth to the weapon. The feat also has its dark side: in a letter to the Senate revealed at the end of June, Anthropic accuses Alibaba’s Qwen lab of having made 28.8 million unauthorised queries, through some 25,000 fake accounts, to “distil” — copy through massive querying — Claude’s capabilities; Alibaba denies it, but the accusation is feeding the parliamentary inquiries^[17].

Above all, China makes a choice the Americans do not: it freely publishes the weights of its models — the downloadable core that allows anyone to run them on their own machine, without going through a provider. This is a decisive advantage, because a model you host yourself cannot be cut off by any decision. Where America locks down, China spreads.

But we should beware of the easy reading. This generosity is not philanthropy: it is the strategy of a player still lagging behind. Opening up to win market share, to starve the American labs that live off subscriptions, to extend its influence by becoming the global standard. The day China takes the lead, it will have no reason left to offer its advantage to the world. The next big Chinese model could very well come out closed, without breaking any promise. The one you have already downloaded will remain yours — but it will age, while the bar keeps rising.

Europe as a spectator?

The verdict is harsh, but the old continent still holds two cards.

A model, first — France’s Mistral, open and locally hostable, a real sovereignty card, but for now a modest-sized player compared with the American and Chinese giants.

A machine, next: the Dutch company ASML holds the monopoly on EUV lithography equipment, without which neither the United States, nor China, nor Taiwan can produce a cutting-edge chip. It is the only place in the entire AI chain where Europe holds a lead^[18].

But neither card is fully in Europe’s hands: Mistral lags behind the giants, and ASML does not sell freely — it is under American pressure that the Netherlands forbids it to export its machines to China. The lever is European; the hand that pulls it is partly in Washington. As for this monopoly, it has a use-by date: deprived of cutting-edge machines, China is already printing near-equivalent chips with older equipment and developing its own alternatives — the embargo that protects Europe’s lead is also what is pushing Beijing to close the gap^[19].

The weakness is not only in software. It is material. Latest-generation models, and even more so the “agents” that chain tasks autonomously, are voracious consumers of computing power — and therefore of chips, data centres and above all energy. Yet it is precisely on this terrain of infrastructure that Europe starts with a handicap: less computing capacity, less accessible and more expensive energy, dependence on chip foundries and foreign suppliers. Real sovereignty over intelligence cannot be built without the physical foundations that support it.

The continent therefore faces a dilemma: buying American means plugging its data into a mass surveillance apparatus made even more effective by the power of AI^[20] — and accepting the risk of being cut off that we have just seen in action. Buying Chinese means betting on an openness that may close again. Europe can choose — but from a menu written in Washington or in Beijing.

Reclaiming sovereignty

Faced with this trap, only one path holds: building sovereignty. And this sovereignty comes at three levels:

For nations, it means taking on the industrial and energy effort that Europe has postponed for too long: supporting its champions, investing in computing and energy, and treating access to frontier AI as a question of independence, not of convenience.

For businesses, it involves a concrete choice that is already within reach: hosting open models yourself rather than renting revocable access. This is what a great many players are already quietly doing, downloading open models and running them on their own servers. Their data no longer goes to the provider; the engine stays under their control.

For individuals, finally, it touches on a more intimate question: being able to think, write and work without a silent ear reading over your shoulder. Escaping the centralised sensor is not just a matter of cost — it is informational sovereignty. The right to have unmonitored ideas.

AI has you by the harness

If models are equivalent and interchangeable, then the model is no longer worth much. The American AI industry has understood this — and it has already moved the cursor. Since it can no longer hold you by the brain, it will hold you by the body.

An AI model, on its own, can do nothing in the real world. It is a gifted brain in a jar: it reasons magnificently, but knows nothing of your world, remembers nothing from one time to the next, and cannot operate any tool. For it to work, it needs a harness — the software layer that surrounds it and connects it to reality: its memory, access to your files and databases, the tools it can trigger, and the logic that decides, at each request, what to give it and where to steer it. The model is the brain; the harness is the body grafted onto the AI brain so that it can act.

And that is now where power is at stake. A turnkey rented harness does you an immense service — and, in doing so, learns your processes, maps your flows and embeds itself in your context on behalf of whoever built it. Microsoft makes no secret of it: the strategy, set out in writing by its boss, is to make the model cheap and interchangeable while keeping the customer captive to its orchestration — Copilot housed in Office, GitHub and Windows, where your data already lives^[21]. What binds you is no longer the model, which is interchangeable. It is the body someone else has shaped around you, and which they will not let you take away.

Is this body impregnable? No — and that is the good news. The advantage of American harnesses is real, but specific: the model is specially tuned to use their harness, they are more reliable when working autonomously, better integrated with your tools. Nothing secret about it, though. When the code of Claude Code leaked in March 2026 — half a million lines — it was copied and reimplemented tens of thousands of times within a few days^[22].

That is why the harness is precisely the layer where open source is most alive. Open, mature and model-agnostic harnesses already exist — capable of running any model: an American model for reasoning, a Chinese model hosted on your premises for the rest. One of them, OpenCode, supports more than seventy-five providers and has passed one hundred and eighty thousand stars on GitHub^[23] in just over a year; another, goose, developed by the American fintech Block, has moved under the neutral governance of the Linux Foundation^[24]. In other words, building your own harness is no longer an engineer’s wish: it is an achievable goal.

This is the real dividing line of the moment: build your AI’s harness at home and keep control of it, or rent convenience and let a third party capture all your context.

Who controls the machine

The Fable/Mythos affair will go down as a minor milestone. Not because a model was cut off, but because it revealed, for the space of a weekend, a mechanism still imperceptible to the general public: the tool you use to correct your text, translate a document or ask for a recipe has become a power stake between states.

One may be tempted to see it as a publicity stunt — the ban as the ultimate advertisement, with an aura of independence thrown in. The hypothesis is all the more appealing because, beyond the media coverage, interests converge: Anthropic needs users who trust it, US intelligence needs users who confide — and a public dispute with the government would provide cover for a well-oiled collaboration. True or false, the hypothesis confirms the need for sovereignty.

For those who view AI with suspicion, this episode proves them right on one specific point — not “the machine is going to enslave us”, but “this technology has become strategic enough for a government to press a button and cut off access to it.” The right question is not whether artificial intelligence will dominate us. It is who controls the machine. A major technological development never remains neutral: you must seize it, or resign yourself to suffering it.

Marcan for BAM!