Two visions, two strategies
The BBC report highlights a fundamental divergence: while American giants — OpenAI, Anthropic, Google — keep their most powerful models behind paid APIs and proprietary walls, China has bet the opposite way. DeepSeek, Alibaba's Qwen, and 01.AI's Yi release their weights, architecture, sometimes even their training data.
This isn't generosity. It's strategy.
Open source as a geopolitical weapon
For Beijing, opening models serves multiple objectives. First, circumventing US sanctions on advanced chips: by democratizing access, China ensures its AI ecosystem doesn't depend on a single provider. Second, creating reverse dependencies: when thousands of developers worldwide integrate Qwen into their products, they become a silent lobby against technological decoupling.
Chinese open models have another advantage: they enable local fine-tuning without sending sensitive data to American servers. For Europe, India, the Middle East, this is a powerful sovereignty argument.
The American lockdown: security or panic?
On the American side, closure is justified by security. Closed models are easier to monitor, align, and red-team. But the economic subtext is obvious: an open model doesn't generate API revenue. OpenAI can't charge $20/month for ChatGPT Plus if Llama 4 does the same job for free.
The problem: this artificial scarcity strategy works as long as American models remain significantly better. But the gap is closing. DeepSeek R1 rivals GPT-4 on several benchmarks. Qwen 2.5 excels in multilingual tasks. And Meta, caught between both camps, plays the balancing act with Llama.
The real winners
Ironically, the big winners of this war are neither American nor Chinese. They're independent developers, African startups, academic researchers — all those who can't afford $0.03 per token but can run a 7B parameter model on a gaming GPU.
China understood this: in AI, control comes through adoption, not lockdown.
What this means for us
For the European tech ecosystem, the choice is difficult. Relying on closed American models means accepting technological dependence and a growing bill. Adopting open Chinese models means gaining technical autonomy but exposing yourself to geopolitical risks.
The solution? Probably Mistral and European models trying to chart a third way. But let's be honest: with budgets 100x smaller than the giants, the path is steep.
AI in 2026 won't just be defined by model performance, but by who controls the conditions of their use.