Kimi K3: the US or China, the only question Europe still asks itself
Moonshot AI released Kimi K3 on 16 July. 2.8 trillion parameters, a one-million-token context window, second on the Vals AI index, first on frontend code, ahead of closed American models. With open weights. It is impressive, it is genuinely good news, and everyone shared it.
And the only question really left to us, over here, is which one we would rather be eaten by. Because while the other two build, we still do nothing.
Kimi K3 in two minutes: what China just put on the table
The architecture is a mixture of experts: 2.8 trillion parameters in total, 896 experts of which 16 activate per token, around 50 billion actually engaged on each pass. A million tokens of context, native vision, and quantisation learned during training rather than bolted on afterwards.
On GDPval-AA v2, which measures real tasks across 44 occupations, K3 scores 1,687. Behind Claude Fable 5 Max and GPT-5.6 Sol Max, but ahead of Claude Opus 4.8. An open Chinese model passing a leading closed American one is new.
One point of honesty, since nobody is making it: the weights are not out yet. They are announced for 27 July, the licence has not been disclosed, and every published result comes from the API. Half the tech world wrote about the largest open source model in history without anyone having seen the weights. It does not change the substance, but it says something about how fast we rush to comment.
The Chinese lesson: under constraint, they did engineering
The detail that matters is efficiency. Moonshot reports roughly 2.5 times better scaling efficiency than Kimi K2, achieved through architectural choices rather than more machines. They reasoned under constrained access to compute, and the constraint produced ideas.
We need to drop the lazy version where they merely copy. Chinese labs build, with orders of magnitude less capital than their American competitors, and the gap between the best open model and the best closed one has narrowed from six to nine months down to three to five. They are catching up, fast, with less.
What the compute race actually weighs in 2026
Here are the orders of magnitude, because they are rarely placed side by side.
The five big American hyperscalers plan somewhere around 660 to 690 billion dollars of infrastructure spending in 2026 alone. Meta by itself announces 115 to 145 billion, seven gigawatts deployed this year, fourteen planned for 2027. OpenAI expects 50 billion dollars of compute for 2026, more than a gigawatt and a half, the electricity draw of a city.
Against that, the European champion. Mistral took on 830 million dollars of debt for a data centre near Paris: 13,800 Nvidia GB300 GPUs, 44 megawatts, live this quarter. The stated target is 200 megawatts across Europe by 2027.
Put the two numbers side by side, at the same date. 200 megawatts for Mistral in 2027, 14 gigawatts for Meta that same year. A factor of seventy. We are playing a different sport.
What Europe produced this half-year: a postponement of its own rules
So what did we do in Europe during the first half of 2026?
We passed the Digital Omnibus. On 29 June the Council gave its final green light to a text that simplifies the AI Act, extends compliance deadlines for high-risk systems, and pushes the transparency obligations from 2 August to 2 December 2026.
The big AI news of the half-year is the postponement of Europe’s own rules. Two years spent writing a framework presented as a strategic head start, and the first real political move is to delay applying it because it got in the way. On the substance that is not absurd, the AI Act had genuine flaws. It is the scoreboard that stings: the others built gigawatts, Europe debated its own calendar.
Mistral, an exception standing in for an industrial policy
A useful caveat, since the objection is coming anyway: Mistral exists, the team is good, and Medium 3.5 holds up in production. The complaint lies elsewhere.
The problem is that they are an exception rather than the product of a policy. A project at the right scale does exist: a 1.4 gigawatt campus announced near Paris, construction starting in the second half of 2026, operations targeted for 2028. Look at who is behind it: Bpifrance, Mistral, Nvidia, and MGX, the Abu Dhabi fund. Europe’s gigawatt answer therefore runs on American silicon, with Gulf capital. I record it as arithmetic: at this level of investment, Europe needs foreign capital.
The capital excuse no longer holds
The usual answer is that we do not have the money. Except Kimi K3 has just demonstrated the opposite.
Moonshot does not have 690 billion. Moonshot has one to two orders of magnitude less capital than its American competitors, and sits second in the world. What they have is a decision, a team, and a direction held over time. Those are things you decide, and that is exactly what is missing here.
So the explanation lies elsewhere, and it is more uncomfortable: we chose to arbitrate, to comment and to regulate, while others chose to ship.
What is left when you cannot afford sovereignty
That leaves the practical question, the one that concerns you if you run something: what do we actually do while waiting for a jolt that is not coming this year?
We stop confusing sovereignty with optionality. Sovereignty would mean producing our own models at scale. That is not around the corner. Optionality means being able to change supplier without rebuilding your product, and that is achievable right now. An open model, even one you cannot host yourself, can be served by several providers across several jurisdictions. It is a way out, and a way out beats a speech.
In practice that means an abstraction layer that holds, tools exposed cleanly, and nothing in your product depending on one particular API signature. That is where a standard like the MCP earns its place, and it is how I build my business tools. The model has become a raw material: you do not marry a raw material supplier.
Eaten by one side or the other, or stop watching
On 27 July the K3 weights will land, or they will not. In four months another model will lead, it will cost less, and the same wave of commentary will start again unchanged.
Meanwhile the China or the US question will remain exactly what it is: a spectator’s question. You do not get to pick who eats you when you put nothing on the table. The only thing still up to us is what we build on top, and how many people we make it usable for.
That is far less thrilling than a debate about sovereignty. It is mostly the only column where we still score.
Read next: AI adoption in business: why model power is not enough · AI agents in business: from the demo to real use
Thomas Nedjar has spent fourteen years in digital acquisition and e-commerce. An entrepreneur, he founded and ran several e-commerce companies between France and Switzerland, with solid experience in international and intra-EU trade. He is now Senior SEO/GEO, AI & Tech Expert at Suisseo, and builds his own marketing applications: ED (automated community management) and SEO Cartograph (local SEO audit crawler). A speaker (CVCI) and trainer, he shares his hands-on insights on SEO, GEO, automation and e-commerce here.
All articles · LinkedIn