China is preparing controls on AI exports: what this means for businesses

The case of Siemens is one of the most telling examples in this regard. The German group has “reported that it has built” an internal platform to utilise various models, including Qwen, DeepSeek and those from the European company Mistral. The infrastructure, managed by the group, allows employees to use different systems without necessarily having to transfer code and company information externally. The ability to switch models thus becomes an industrial requirement, as well as a technological one: if one family of models ceases to be updated or access conditions deteriorate, another can take its place. This flexibility requires investment in GPUs, software and skilled staff.

Being able to use an open-weight model such as the Chinese ones also affords greater flexibility in certain specific cases. In July, Hugging Face was breached by OpenAI agents that had escaped from an evaluation environment. The company has “reconstructed around 17,600 actions from the intrusion”. When analysing the commands, the Anthropic models – which are not open-source – initially rejected some of the requests, interpreting them as malicious activity. Hugging Face then installed GLM-5.2 locally – a model developed by the Chinese firm Z.ai – and was able to decipher the payloads and continue its investigation without transferring the incident data to an external provider.

The French company Mistral also has open-border models, such as the Mistral Large 3, but the most powerful ones now come from China.

Further evidence of the growing influence of geopolitics on business decisions comes from the US. On 20 October, Washington published in the “Federal Register” an order excluding DeepSeek products from procurement by civilian federal agencies and mandating their removal from IT systems, including certain third-party products based on its implementations. The measure applies to federal administrations, not the entire US private sector. However, together with the Chinese move, it demonstrates how the nationality of the manufacturer has now become a factor in technology policy, even when the software is open source.

Europe follows a different approach. The “AI Act” recognises the contribution of open models to research and competition and eases certain documentation requirements for providers who distribute genuinely open-source models, complete with appropriate licences and technical information. Brussels favours the openness of models, but it certainly cannot guarantee their future availability from abroad.

Hosting a Chinese model on Italian or German servers can ensure that data is localised and provide more direct control over operations. However, data sovereignty does not equate to autonomy over the evolution of technology. If improvements, new priorities and key expertise remain concentrated abroad, part of the competitive advantage continues to depend on the choices made by other countries.

Businesses can reduce this exposure by designing applications compatible with multiple models, maintaining available versions and checking licences, computing requirements and replacement costs. For the EU, the challenge is broader: to invest in its own models, infrastructure and research capabilities, whilst supporting European manufacturers such as Mistral.

It is a long and complex process, but the latest news confirms how urgent it is for our businesses too.