Europe’s AI Independence Is Not a Download

Europe wants more control over its technological future. Joe Tsai has an offer.

Speaking at Wave by Vento in Turin, Alibaba’s chairman argued that open-source AI is Europe’s path to technological independence. It is an appealing message for European businesses that do not want their industrial data, operating costs and access to critical technology dictated from elsewhere.

But before Europe accepts the promise, it should ask what is actually in the package. Is it getting the knowledge and tools to develop AI independently, or permission to operate a model somebody else built?

Those are different propositions. Calling both of them open source does not make them the same.

The South China Morning Post reported Tsai’s remarks. SCMP is owned by Alibaba, not an official Chinese government outlet. That ownership deserves disclosure, not a leap to the conclusion that Beijing scripted the coverage.

Alibaba has an obvious commercial interest in this conversation. It develops Qwen models and sells cloud and AI services. That does not make Tsai wrong. It does mean Europe should examine the product as carefully as the pitch.

Give Tsai Credit, Then Examine the Claim

The fuller account from The Next Web is more nuanced than a headline about escaping American dependence. Tsai warned against relying entirely on either American or Chinese technology. He advocated European computing infrastructure and using Europe’s industrial data to adapt models locally.

That is sensible advice. A manufacturer can have good reasons to keep sensitive data inside its own environment. Downloadable models create options that an API-only service does not provide.

A European company running a model on infrastructure it controls can gain substantial operational autonomy. It need not send its prompts to the original developer merely because that developer trained the model.

The difficulty comes when those benefits become a sweeping promise of technological independence. TNW reports that Tsai compared open-weight models to a recipe users can improve. The distinction matters: the trained model is the product of a recipe. Its release does not necessarily disclose the whole recipe.

Open Weights Are Not the Whole Process

Model weights are the learned parameters produced by training. With the necessary software, hardware and legal rights, users can run the model and often fine-tune it for their own purposes.

Those freedoms are valuable. They do not necessarily reveal the complete training corpus, how it was selected and filtered, the code used to train the model, or the full procedures used to shape its behavior afterward.

Alibaba’s own Qwen documentation distinguishes proprietary models from open-weight versions. Its original Qwen3 announcement described eight open-weight models released under Apache 2.0.

A permissive license answers important questions about what users may do with the released materials. It does not establish that all the materials needed to reproduce the development process have been released.

Nor should we claim Alibaba has disclosed nothing. Its Qwen3 technical report provides substantial information about architecture, training and evaluation. The release announcement describes approximately 36 trillion training tokens and explains stages of pretraining and post-training.

The curtain is partly open. That is different from removing it.

The Open Source Initiative’s Open Source AI Definition 1.0 offers a more demanding standard. It calls for model parameters, complete training and inference source code, and sufficiently detailed training-data information to allow a skilled person to build a substantially equivalent system.

It does not require every original training item to be freely downloadable. It accommodates unshareable data, while requiring descriptions of provenance, selection, characteristics and processing, along with information about obtainable datasets.

That definition is a specific standard, not a universally accepted settlement of the debate. But it exposes how much can be missing when the word “open” refers principally to weights.

Even regulatory language can blur the distinction. The European Commission’s AI Act guidance describes an open-source exemption based on licensing and availability of parameters, architecture and usage information. It acknowledges that this does not necessarily disclose training data. Regulatory treatment is not a certificate of sovereignty.

Who Can Build What Comes Next?

We need to separate control over deployment from the capacity to develop technology.

Can a European organization operate a model without contacting its original supplier? Can it investigate limitations and change the system meaningfully? Can it maintain and advance its capabilities? Can it replace the model without rebuilding its business around another supplier?

Open weights help, particularly with the first question. The others depend on what else is available and what capabilities the user possesses.

An existing locally hosted model does not stop working because its developer releases a successor or changes strategy. Europe is not automatically condemned to wait for the next Qwen release. Its researchers and companies can adapt models, develop alternatives and build their own systems.

But those possibilities require expertise, investment, data and compute. Downloading weights does not supply them. Neither does publishing every line of training code.

Transparency and sovereignty are related, but they are not interchangeable. Full disclosure gives people more means to exercise independence. They still need the capacity to use those means.

Alibaba understands the value of that capacity. Its own recap of Tsai’s appearance describes a full-stack strategy across computing infrastructure, model development and applications.

My reading is straightforward: Widespread adoption of Alibaba’s models can create commercial opportunities and technical influence, even where users host the models themselves. Developers learn its tools, build around its architectures and may consider its related services.

That is an ecosystem strategy. It is not proof of government orchestration, and it is not identical to contractual lock-in. A permissively licensed local model gives users real choices. Europe should recognize both the freedom it gains and the supplier’s interests.

The Same Test for Everyone

This scrutiny cannot stop at China’s border.

The December 2025 Foundation Model Transparency Index scored Alibaba 26 out of 100 and DeepSeek 32, despite their open-weight releases. Those are dated transparency scores, not ratings of model quality, security or present performance. The researchers also found substantial opacity across the industry.

An American or European address does not make a development process transparent.

There are examples of broader openness. Switzerland’s Apertus, developed through its public research institutions, discloses training data, code, weights and methods. American nonprofit Ai2’s Olmo 3 also releases training data, code and weights.

These projects demonstrate that more complete disclosure is possible. They do not prove parity with the strongest commercial models, guarantee safety or eliminate infrastructure dependencies. Switzerland is European, but it is not an EU member. Precision matters here, too.

Europe does not have to reject Chinese models, worship American ones or insist that everything be built domestically. It should use technology that serves its needs while examining licenses, provenance, disclosure and realistic replacement options.

Imported open-weight models can contribute to sovereignty. So can genuinely open research, local infrastructure and sustained investment in people who can develop the next generation.

Tsai is offering a real increase in operational freedom with an inflated promise of technological sovereignty. Europe should take the useful technology without accepting the entire sales pitch.

Europe can download a model. It cannot download technological sovereignty. That has to be built.