Malaysia has spent two years saying it wants sovereign artificial intelligence: models trained on Malaysian data, owned and run inside the country rather than rented from a foreign giant. Most of that work sits on one American website. On Thursday, a different American giant bought it.

Table of Contents

What Nvidia actually bought

Nvidia confirmed on Thursday that it will acquire Hugging Face, the platform that hosts most of the world's open AI models, for close to US$13 billion. CNBC reported the chipmaker will pay about US$11.9 billion to Hugging Face shareholders and another US$1 billion in equity to keep its staff. It is Nvidia's second-largest purchase ever, behind the roughly US$20 billion it spent on Groq's assets late last year, and it is expected to close in the first half of 2027.

Hugging Face is not a household name, but it is the closest thing open-source AI has to a home. More than 18 million developers use it, it stores over 3 million models, and about 200,000 companies build on it. When a research lab releases a model anyone can download and run, this is usually where it lands.

Chart of what Nvidia bought with its US$13 billion Hugging Face deal: 18 million developers, 3 million models, 200,000 companies

Why this reaches Malaysia

Malaysia's own AI runs through the same address. MaLLaM, the Malaysian large language model built by local startup Mesolitica and trained from scratch on about 90 billion Malaysian-context tokens, is published on Hugging Face. So is the work of Malaysia AI, a non-profit that keeps two dozen open models there covering Malay speech, translation and text. These are the building blocks the country points to when it argues it does not have to depend on Silicon Valley.

Open-source code on a laptop screen, the kind of software the developers on Hugging Face build

That is the awkward part. The government has been pushing hard the other way. Budget 2026 told regulator MCMC to build sovereign AI infrastructure that keeps national data inside Malaysian systems. The AI untuk Rakyat programme aims to put free AI tools in front of 100,000 young Malaysians. A US$250 million deal with Britain's Arm is meant to seed a homegrown chip-design industry. The plan is more control, not less. Yet the shelf those Malaysian models sit on will soon be owned by the same company that already supplies the chips most AI models are trained on.

What changes, and what does not

For now, very little changes in practice. Hugging Face's open models stay downloadable, and Nvidia has every reason to keep the platform open, because it sells more chips when more people build AI. MaLLaM does not stop working because its landlord changed.

The longer question is leverage. Owning the place the world downloads its models from, and the chips it runs them on, hands one company an unusually complete view of the open-AI supply chain. None of this traps MaLLaM, because open weights can be downloaded, mirrored and run on other hardware. But defaults have gravity, and the default home of open AI now belongs to the largest AI chipmaker. For a country trying to build something it can genuinely call its own, "our model, on their platform, on their silicon" is a thinner kind of independence than the announcements suggest. It is a reason for Malaysia to think harder about where its models are hosted, not only who wrote them. It is the same dependence question raised by Malaysia's large but low-margin chip trade, moved up the stack to AI.

The tools are still free and still Malaysian. The ground they stand on just moved, toward the same company Malaysia is already buying its chips from.

Image(s) courtesy of Winston Chen and Luca Bravo on Unsplash.