Operational notes Scenarios

Chinese open-source AI in the enterprise: the numbers add up, so do the risks

3 min read

Pudong skyline, Shanghai
The new geography of models: knowing where your data and dependencies pass through.

A year ago, Chinese AI models accounted for less than 2% of global traffic. By April 2026 their combined share had passed 45%: DeepSeek with the new V4 family, Alibaba’s Qwen, Z.ai’s GLM, Moonshot’s Kimi. Many are released with open weights, cost a fraction of the American alternatives, and match them on several tasks. For a European SME the temptation is strong — and partly justified. But the decision has to be made knowing exactly what you are choosing. Let’s lay out the facts.

Why everyone is talking about it

Three concrete reasons: cost (Chinese APIs cost up to ten times less; open weights, run on your own hardware, bring the cost per token to zero), openness (downloadable weights mean no lock-in, the ability to specialise the model on your own data, and operation even without connectivity) and pace of improvement (releases follow one another at a rate of weeks). For well-defined enterprise tasks — document classification, data extraction, internal assistants — a well-chosen open model is often more than sufficient.

The risks, without alarmism but without discounts

  1. Where the data goes. If you use a Chinese provider’s APIs, prompts and responses pass through servers subject to Chinese legislation. For the personal data of European citizens this is a concrete GDPR problem, even before it is a strategic one. This risk concerns the APIs, not open weights run in-house.
  2. What the model answers. Estonia’s foreign intelligence service has documented that, on topics sensitive to Beijing, DeepSeek omits information and inserts elements of propaganda. For an industrial use case (reading a production log) this is irrelevant; for an assistant answering open questions, much less so.
  3. The geopolitical tap. Beijing is considering limits on foreign access to its most advanced models; Washington is pushing in the opposite direction with its own providers. Anyone building business processes on a model has to ask: what do I do if it is no longer available tomorrow?

The sensible path: open weights, your own premises, a replaceable architecture

For a European company, the configuration that captures the benefits while cutting the main risks is well known: an open-weight model run on-premise or on a European cloud, inside your own infrastructure. Data does not leave, the applicable legislation is European, and the model remains available even if the provider changes policy — downloaded weights cannot be revoked.

The condition almost everyone forgets: the architecture must treat the model as a replaceable component. The value does not lie in the model — which will be superseded within six months, whoever produces it — but in the context you give it: your connected data, the operating model that makes it legible, the controls over what the AI is allowed to do. Those who build this way can use a Chinese model today, a European one tomorrow, an American one the day after, changing one component rather than the system. Those who instead hard-wire their processes to a single provider have turned a tactical choice into a strategic dependency.

Checklist before deciding

  • Does the use case touch personal or confidential data? Then foreign APIs are excluded: open weights in-house, or a provider with European guarantees.
  • Is the task well-defined and verifiable (classifying, extracting, summarising)? Open models are excellent candidates.
  • Have you defined how to measure quality before choosing the model? The benchmark that counts is on your own documents, not the published one.
  • Does the AI’s action have operational consequences? Then a human operator needs to be in the loop, whatever the model — it is the rule we never bend.

Weighing up an open model for a concrete use case? Let’s talk: half an hour to see whether the numbers really add up.

Sources