Last month I rode on the back of a friend’s scooter to the office in the Hsinchu Science Park. Being a weekday, we joined a river of similar vehicles flowing toward a group of buildings housing TSMC (Taiwan Semiconductor Manufacturing Company Limited), one of the world’s largest chipmakers.

Around us were riders in jeans, hoodies, trainers, long gloves and helmets – TSMC employees heading to work. Nobody looked rushed or special. That was when I realised that AI does not live in a distant cloud. AI lives down the street, in those white-and-glass buildings. It struck me how ordinary something so enormous could feel here.

Taiwan plays an important role in the semiconductor supply chain by manufacturing chips that power AI systems worldwide. That role is significant, but it is not the whole chain: chip design and manufacturing are distinct stages. US-based NVIDIA, led by Taiwan-born Jensen Huang, designs AI chips, while factories in Taiwan manufacture them, turning those designs into billions of tiny transistors and complex circuits on silicon wafers.

Chips produced in Taiwan support AI models developed in turn back in the US States, China, and elsewhere. But producing the hardware does not determine which AI services consumers choose, or how much control they have over those services.

According to the Taiwan Network Information Centre’s 2025 Taiwan Internet Report, 43% of Taiwanese internet users have also used generative AI. But when Taiwanese users open ChatGPT, Gemini, Claude or DeepSeek, what shapes their choices beyond the chips underneath? And what do those choices reveal about AI sovereignty for this small island?

In Taiwan, geopolitical concerns around AI have inevitably need to be taken into consideration given the nature of the country’s relationship with China. In 2025, for example, Taiwan's Ministry of Digital Affairs announced restrictions on the use of DeepSeek's open-source AIs by government agencies, citing information security concerns and potential risks involving data transmission and leakage when using the models.

This is fairly typical of the government's risk-averse approach to deploying technology in official or institutional settings. Individual citizens, however, remain free to use Chinese AI services should they so wish, and given the important of language and cultural context when it comes to AI adoption, you might have thought these would be popular.

After all, consumer tech choices tend to depend largely on how naturally a tool fits into people’s existing routines, and in Taiwan we use traditional Chinese characters to write, speak Taiwanese Mandarin in conversation (albeit with local expressions), and depend on a lot region-specific social and institutional knowledge to go about our daily business. And yet a 2026 Taiwan survey showed that ChatGPT and Gemini are currently the most popular options in Taiwan for everyday AI assistance.

Why is this? The main reason seems to be that ChatGPT became widely known among Taiwanese people because it launched before competing brands, like Anthropic's Claude. It therefore introduced many ordinary users to generative AI for the first time, and they’ve stuck with it ever since. Subsequently, Gemini has also gained market share because of its tight integration with Google’s existing ecosystem, including Android, which has made it easy to adopt and attractive to users.

Things might be changing though. The survey also found out that consumers can be fickle. People on the free tiers of these tools would rather switch to another free service than pay to stick with the AI they’re currently using. And as AI use grows Taiwanese users are not just being asked to pay by the market incumbents, who are having to find ways to make their services profitable, but they are beginning to realise that most international AI models have overwhelmingly English-language training data sets with what Chinese-language material they are trained on dominated by Simplified Chinese. Traditional Chinese interactions with these models primarily rely on post-prompt/response translation which can introduce confusion and inaccuracy on top of the already present problems of hallucination and bias.

One project designed to address this is the Trustworthy AI Dialogue Engine (TAIDE), which has retrained several open-source foundation models on Taiwanese data and Traditional Chinese language material, a development that it seems to me is only going to become more important as AI capabilities migrate, as they are beginning to, from global clouds to local data centres or even individual devices (Apple’s latest operating system, Mac OS27, now comes with a local LLM built in!).

Sitting on the backseat of the scooter, I keep coming back to the thought that where AI comes from shapes the way it understands and responds to us, which shapes us in return. The next time you enter your prompt in a chatbox, ask yourself (or your AI) where that data is going, and where the responses are coming from. The answer you’ll arrive at will be about more than AI model performance and brand. It will be about data governance, technology independence and sovereignty; about who learns from the things we write and what we learn from the things the AI writes back. In Taiwan, we’re becoming more and more aware of this. Having built the chips that built the current AI boom; we’re now learning to read the rest of the AI map.