Last week, we talked about Kimi as an inflection point, but much of what we said had to do with the technological question (is it a good model?) or the financial one (can this model disrupt Anthropic's growth? The short answer is no).

But what interests me most today is the philosophical and political question. The dynamic between two visions of tech is far more visible now that it is clear China is not just a copy of the US model, but a much broader political vision.

Among the various analyses, one caught my attention: Dean Ball's post on X. Ball, who runs strategic foresight at OpenAI, made no secret of his fear, because progress is coming much faster than anyone hoped. But it was the part of his post where he talks about communist AI, set against the progressive, personal AI the United States is building, that raised a lot of questions for me.

Some time ago, at HEAD, I had the opportunity to give a talk on AI in cinema, which has been my lifelong passion. I believe I have seen nearly every film that deals with the subject.

From the obscure B movie to the American series that never came out in France. Forgotten masterpieces, notably 2010, the sequel to 2001, which revisits the question of AI in an extremely interesting form.

💡

That presentation covered 100 films and series, in 10 subgenres, that have shaped us. I have linked it at the end for members.

The theory I am developing is that a large part of our vision of AI comes from the films we watched when we were younger.

And the series that best defines AI itself, the LLMs, is obviously Star Trek: The Next Generation (Star Trek: Voyager a little too).

This idea of a central onboard computer that organizes life aboard the ship, answers each crew member personally while protecting their privacy.

Like many, I was fascinated by this machine, whose voice was borrowed from Majel Barrett, Gene Roddenberry's wife. As I often say, for anyone working in AI, never having watched Star Trek: The Next Generation is professional misconduct, so brilliantly does it tackle human-machine interaction and the ethics of AI.

But two things are worth remembering.

The first is that this computer is deployed inside an environment, a ship, with a military-style structure. The captain can open access to the archives of every crew member's prompts.
The second, obviously, is that Star Trek imagines a hyper-progressive future where money no longer exists, where brands no longer exist, and where technology is built in a dynamic of progress. Not always, but the dark side of Star Trek (Section 31) was never properly developed.

💡

In a way, Star Trek is the best implementation of what a communist AI could be.

But let's return for a moment to Dean Ball's post, which deserves a careful read. The term communist AI does not come out of nowhere: it echoes a conversation between Reid Hoffman and Peter Thiel, who had already explained that AI was communist and crypto libertarian.

His observations are interesting because they are dispassionate (though slanted in his employer's favor), unlike those of American political actors who need to pile it on.

The first point : in his view, Kimi is a genuine frontier model. He dismisses the distillation explanation and places it, in agentic use, at the level of the best public models of the first quarter of 2026. His only reservation: the model burns a lot of tokens, and nothing proves it is actually cheap to run. This is interesting, because it means that models, if they are expensive, cannot be the foundation of an AI for everyone.

The second point: he says he is surprised that the Chinese state lets models of this caliber out as open weights. Essentially, he says, the Party does not believe in AGI and underestimates what is at stake. Beyond that, American GPU restrictions deprive Chinese companies of the compute they need to serve customers around the world: opening the weights means letting others pay for inference, so that what you do have can go first to training models. On the company side, the math is simpler: behind the frontier, they know almost no one would pay for sub-frontier Chinese APIs.

💡

That, incidentally, is Mistral's new problem: it has ended up sandwiched.

Third point, and this is the heart of the text: open models are, in his view, inherently decelerationist. If the best models become free, it becomes impossible to justify tens of billions in private investment to train the next generation. That is the key point for the two American market leaders. It always takes more money to reach AGI as fast as possible.

Fourth point: the probable consequence of a world dominated by open weights is what he calls AI communism. AI stops being a market product and becomes a public good, a digital public infrastructure provided by the state. He deems this future dystopian, while acknowledging that open-weight advocates admit they want to head in that direction.

In his eyes this is a major risk, because one could very well imagine an alternative world in which the rest of the planet connects to Chinese AI services, some of them free, feeding those services with fresh data and new contexts.

China would become H. G. Wells's world brain, serving (and influencing) the rest of the world.

Finally, he anticipates that Washington will create regulatory risk around Chinese models. Not an outright ban, but friction that would make the cost of using these models less competitive. The problem is that a majority of American startups already use these models, and as we have seen, the cyber world is preparing to use them en masse to get around Fable's blocks.

💡

A petition has in fact been launched to protect access to these open-weight models.

Building profitable cognitive colonies

The difference between the US and China ought to be clear, but with the United States demanding a form of strict obedience, as we saw between Anthropic and the Department of War, it is becoming less and less clear to Europeans. If Chinese and American models differ in their vision of deployment, their vision remains the same: building profitable cognitive colonies.

For now, the risk is mostly economic; the arrival of frontier models is not so much a risk to infrastructure building as to the revenues of the two leaders. And those risks are being brushed aside, let's be honest.

On one side, the seamless integration of Anthropic's and OpenAI's services continues to give them an edge. On the other, the sovereign on-prem infrastructure offering is only just getting started, and the solutions for standing up your own clusters running Kimi, Qwen or DeepSeek at full potential still cost an arm and a leg. As we said before, as long as models are not designed for direct distribution to enterprises, integrating them will remain more expensive.

What is interesting is to see that the dynamic China is accused of could be put in place in the United States. If Chinese AI dumping succeeded, then the American labs, self-declared champions of the market, could ask the state for protection.

That is exactly what Ball proposes in his text.

The situation mirrors, point for point, that of Chinese electric cars versus Europe. Europe made the electric transition mandatory, then slapped tariffs on affordable Chinese vehicles to protect its carmakers. It defends the goal and, at the same time, the industry struggling to reach it, and the consumer pays the difference.

The United States is heading down the same road with intelligence. AI is proclaimed a tool of progress and abundance, but American industry is in no position to deliver that abundance at an accessible cost, and is getting ready to push aside the one distributing it for free.

That narrative has already begun. DeepSeek is already banned on US federal government devices, restricted in several states, and a coalition of attorneys general is calling for a national ban. Beijing, for its part, denounces the politicization of trade and technology issues. France is aligning with Europe and has decided not to use Chinese models, which are also criticized by the Rassemblement national and, more quietly, by every political party (though let's be clear, for now no political party has any clear doctrine on the use of Chinese AI in the 2027 campaign).

But wanting to cut off access to Chinese models carries two costs.

It forces companies to use more expensive products. This is Chamath Palihapitiya's famous $800 barrel: he compares forcing American companies to pay fifty to a hundred times the market price to a state that would only allow oil to be bought at $800 a barrel when the free market sells it for 80.
And discrimination by origin, without formal procedure, sits uneasily with international trade commitments, unless one invokes the national security exception, which has become the skeleton key of the era.

That leaves the fundamental question, the one that really interests us: why the word communism, and why now. That is the subject of the second part, which is reserved for our members.

Join our community and become a member. You get access to our monitoring and to all our bonuses. Here are 4 quotes from the messages you have sent us. Thank you again for your support

"This newsletter reminds me of a reality we tend to forget: technology decisions are now geopolitical choices." CTO of a French industrial group

"With every issue, I come away with a fresh angle, a reference I did not know, or a unique analytical framework." European investor, Brussels

"Tariq has this unique ability to think against the grain of the times, to ask the real questions before they become urgent." Former French government adviser