In the second part of my piece on foundation models, I explore the geopolitics of AI — particularly the race between the US and China — and the vast, largely uncharted world of languages no foundation model has ever touched.
The US vs. China: A Race With Very High Stakes
Now for the geopolitics. Because foundation models are not just a technology product. They are, increasingly, a matter of national strategy.
The United States got there first. OpenAI, Anthropic, Google DeepMind, and Meta have been building and releasing the most powerful models in the world. The US government, acutely aware of what is at stake, has imposed export controls on high-end chips — specifically NVIDIA’s most advanced GPUs — to prevent China from acquiring the hardware needed to train competing models. The strategy has had some success. American labs still hold a meaningful lead in raw model capability.
But here is the twist. Being denied the best chips did not slow Chinese researchers down the way Washington hoped. It forced them to innovate differently.
DeepSeek’s January 2025 release shocked Silicon Valley not because it was slightly good, but because it reportedly cost a small fraction of what OpenAI spends, used less powerful chips, and still performed at a comparable level. Chinese engineers, forced to do more with less, developed architectural tricks — like Mixture of Experts, where only a fraction of the model activates for any given query — that dramatically cut both training and running costs.
Getting blocked from the best hardware turned out to be, inadvertently, a lesson in efficiency.
And DeepSeek is far from alone. Alibaba’s Qwen series is widely used by developers worldwide — including many in the West. Baidu’s ERNIE, ByteDance, Moonshot AI’s Kimi — a whole constellation of Chinese foundation models is rising, backed by government mandates, access to data from 1.4 billion people, and an explicit national goal of AI dominance. As of early 2026, China’s best models trail US leaders by just 2.7% on key benchmarks — down from a much larger gap just two years ago. China also holds a remarkable 74.2% of global AI patents. The race is very much on.
What this means in practice is a world splitting into two AI ecosystems, each with its own foundation models, its own values baked in, its own data, and its own geopolitical loyalties. Companies, governments, and citizens are increasingly going to have to choose which infrastructure they build on top of — and those choices will have consequences far beyond which chatbot answers their emails.
The Language Problem Nobody Wants to Talk About
And here is the uncomfortable truth underneath all of this technological triumphalism. Foundation models, for all their power, speak a very particular dialect of human civilisation — and it is mostly English.
There are over 7,000 languages spoken in the world. Current AI foundation models are meaningfully trained on around 100 of them. Fewer than 20 languages have the kind of large digital text corpus that a foundation model genuinely needs to learn from. English, despite being spoken by less than 20% of the global population, dominates nearly half of all content on the web — and therefore dominates what these models know, how they think, and what they can do well.
What does this mean in practice? It means that if you are a Swahili speaker — and Swahili has 200 million speakers, which is not a small community — AI tools work poorly for you, not because the language is unimportant, but because enough of it has not been digitised in a form these models can consume. It means that across Africa’s roughly 2,000 languages, almost none have meaningful AI representation. Stanford researchers have confirmed what many already suspected: AI is creating a new digital divide, one where access to powerful tools tracks almost perfectly with whether you are lucky enough to speak a high-resource language.
Japanese is an instructive case, because it is neither small nor obscure. Japan has the third-largest economy in the world, one of the most digitally sophisticated societies on earth, and a rich literary and technical tradition.
Yet Japanese has long struggled to get proper representation in foundation models, partly because of its writing system complexity and partly because English simply drowns it out in training data. Japan’s national research institute, AIST, only released its first dedicated Japanese speech foundation models — named “Izanami” and “Kushinada” after Japanese mythology — in 2025, using 60,000 hours of Japanese speech data described as the largest ever assembled for the purpose. That Japan needed a dedicated national effort to build basic speech AI in its own language tells you something important about the default assumptions built into this technology.
There are practical solutions being explored. Small Language Models (SLMs) — much cheaper to train, requiring far less data — are emerging as a more realistic path for low-resource languages than trying to squeeze into the training set of a trillion-parameter giant. Nigeria launched its first multilingual LLM in 2024, training it on five low-resource languages. Grassroots organisations like Masakhane are doing painstaking community-driven work to build NLP resources for African languages. Sakana AI in Japan is building models rooted in Japanese characters from the ground up.
But these are sprints in a marathon.
The languages that do not get foundation model support in the next few years risk falling into a self-reinforcing trap: no model means no tools, no tools means less digital activity, and less digital activity means even less data to train future models. For some communities, the window may be shorter than we think.