Open models and the chance for small countries

On 16 July two events occurred that are worth looking at together.

Bloomberg reported that Google’s flagship model, Gemini 3.5 Pro, is months behind schedule: its coding capability fell short of internal expectations. Alphabet’s shares fell around 4.4%, erasing roughly $200 billion of market value in a day. At its conference in May, Google had said the model was already in internal use and would ship in June. It did not.

The same day, China’s Moonshot released Kimi K3: 2.8 trillion parameters, natively multimodal, a one-million-token context window. Pricing of $3 and $15 per million input and output tokens. First place in the Frontend Code Arena evaluation, fourth among all frontier models on independent testing. And crucially: open weights, promised for publication by 27 July.

One company with near-unlimited resources could not ship a model on time. Another shipped a comparable one and gave it to everybody.

What “open weights” means

The difference here is not technical but proprietary.

When you use a closed model, you are renting access. The provider can change the price, change the rules, restrict a region, close an account, or simply discontinue the service. Your product stands on a foundation somebody else owns.

When the weights are open, you can download the model, run it on your own servers, and operate without external permission. Nobody will cut you off for political reasons, double your price, or ask you to explain why you need the technology.

For large countries this is a question of cost. For small ones it is a question of principle.

Why it matters here specifically

Kazakhstan has declared 2026 the year of digitalisation and artificial intelligence. A Digital Code has been in force since 11 July, and a law on artificial intelligence since July. Data Center Valley is being built at Ekibastuz, a digital hub at Alatau. The state is putting serious money into computing infrastructure.

The question those plans rarely settle: what exactly will run on that capacity, and on whose terms.

If state services, courts, medicine and schools are built on a rented model, the country acquires a new kind of dependency — less visible than an oil pipeline, but built the same way. The supplier sits in another jurisdiction and answers to its export controls and its political decisions. Open models remove that risk: downloaded weights cannot be revoked.

There is a financial side too. Kimi K3’s price — $3 per million input tokens — is a fraction of what the American market leaders charge. For a Kazakh startup or newsroom, the difference between renting intelligence and owning it is measured in orders of magnitude.

I am obliged to state the other side. Open weights from China are not a neutral technology. The model was trained on particular data, with particular assumptions about what may be discussed. Running it yourself gains you independence from an American supplier and inherits you the constraints of a Chinese developer. That is exchanging one dependency for another, not escaping dependency. The difference is that downloaded weights can be inspected, fine-tuned and studied — a closed API cannot.

A third event: the regulator retreated

And one more item of the same kind, less noticed.

The European Union has postponed the most contested part of its AI Act. Obligations for high-risk systems — biometrics, critical infrastructure, education, employment, migration, border control — have moved from 2 August 2026 to 2 December 2027. That is a seventeen-month delay. For systems embedded in regulated products, to August 2028.

From 2 August 2026 the lighter requirements do take effect: disclosing that you are talking to a chatbot, marking synthetic content and labelling deepfakes.

Europe has been presented for years as the model of AI regulation. Anyone citing that model now has to explain the delay as well. The reason is plain: industry said the rules could not be met on time, and the rules moved.

For countries writing their own regulation — Kazakhstan among them — there is a useful lesson. What should be copied is not the text of the European law but its fate: even the EU lacked the weight to make the market meet its strictest rules on schedule.

What I think

The week showed three things at once.

Technological leadership has stopped being guaranteed: a company with near-limitless resources missed a release, and the market billed it $200 billion immediately. That is a healthy sign — it means promises are actually being checked.

Frontier capability has stopped being the monopoly of a few companies in one country. A frontier-grade model can now be downloaded. For small economies that is the most important event of the year: for the first time there is a real option to own intelligence rather than rent it.

And regulators retreat where they meet a real industry. Anyone who believes a well-drafted law settles a problem should keep that in mind.

My view: countries like Kazakhstan should build their state systems on open weights — not because those models are better, but because they cannot be taken away. Computing centres without control over the models running on them are a tenancy with a high rent and a landlord’s right to terminate.

Technological sovereignty is not measured in data-centre megawatts. It is measured by whether you can keep working if, tomorrow, your supplier decides you may not.