Sovereign AI in Southeast Asia: Strategic Necessity or Expensive Duplication?
"Sovereign AI" is on every government agenda in Southeast Asia. The idea is simple to state: a country should have its own AI computing capacity and its own models, under its own control. Whether that is wise, achievable or affordable is one of the liveliest arguments in technology policy.
This article does not take a side on the policy question. It sets out what the region is doing, the strongest case for it, the strongest case against, and then the part we can speak to directly: what "sovereign" should mean when an ordinary organisation is buying AI infrastructure.
What the region is doing
| Country | Notable initiative | Layer |
|---|---|---|
| ๐ธ๐ฌ Singapore | SEA-LION open-source model for Southeast Asian languages; national supercomputers such as ASPIRE 2B | Models and compute |
| ๐ฒ๐พ Malaysia | ILMU, a language model built for Bahasa Malaysia, launched in August 2025 | Model built from the foundation up |
| ๐ฎ๐ฉ Indonesia | Sahabat-AI, adapting models for Bahasa Indonesia, Javanese and Sundanese | Local-language models |
| ๐ป๐ณ Vietnam | An AI law passed in December 2025; domestic compute through FPT's AI Factory | Law and compute |
| ๐น๐ญ Thailand | National supercomputing and a fast-growing data centre build-out | Compute |
Sources: ISEAS Fulcrum on sovereign AI models, ISEAS Fulcrum on mixed AI stacks, OpenGov Asia
The case for
Supporters generally make three arguments.
| Argument | In plain terms |
|---|---|
| ๐ฃ๏ธ Language and culture | Models trained mostly on English and Chinese serve Southeast Asian languages and contexts poorly. Someone has to build for Bahasa, Thai, Vietnamese, Tagalog and the rest. |
| ๐ Data protection | Health, financial and government data may need to be processed in-country, under local law. |
| ๐ก๏ธ Resilience | No government wants essential services to depend entirely on infrastructure that another country's decisions could switch off. |
These are not abstract. The language argument in particular is hard to dispute: regional models exist because global ones left a gap. And Singapore's national systems are already being used to train Southeast Asia-focused models.
Source: Avasant on AI sovereignty
The case against
Critics do not usually dispute the goals. They question whether the approach achieves them.
| Criticism | In plain terms |
|---|---|
| ๐ The scale gap | National clusters are small beside private fleets. One consultancy notes reports that a single US company bought about 485,000 GPUs in one year. |
| ๐ Dependence moves, it does not vanish | A domestic data centre still runs on foreign chips and foreign software. As one industry essay puts it, "location is not control". |
| ๐ด Utilisation | Some commentators argue that publicly funded clusters often sit under-used, because running them well is a specialist job that grants do not cover. |
| ๐งฉ Fragmentation | Several small national projects may each be too small to matter, where a shared regional effort might not be. |
A paper from a Boston Consulting Group institute puts the sceptical position in its title: for most countries, AI sovereignty is an illusion, while resilience is real. It argues that sustaining sovereign control over entire layers of the technology stack is exceptionally resource-intensive.
Analysts at Singapore's ISEAS make a related point about the region's habit of combining American, Chinese and local technology: mixed stacks are pragmatic, but they do not automatically reduce dependence, and may replace one obvious dependency with several less visible ones.
Sources: BCG Henderson Institute, Scale on the myths and realities of sovereign AI, ISEAS Fulcrum
A middle position
Between "build everything" and "it is all an illusion" sits a third view. Researchers at the National University of Singapore suggest that for Southeast Asia, a more useful idea of AI sovereignty is not self-sufficiency but coordinated resilience: being able to take part in global AI ecosystems without dependence hardening into subordination. On that view the question is not whether to depend on others, but how to keep choices open.
Source: NUS AI Institute
The two views side by side
| Question | ๐ข Supporters say | ๐ก Critics say |
|---|---|---|
| Does local compute give control? | Yes, over data and operations | Only partly; chips and software are still foreign |
| Are national models worth building? | Yes, global models neglect local languages | Adapting open models may achieve the same for less |
| Is the spending justified? | Yes, as strategic infrastructure like roads or power | Only if the capacity is actually used |
| National or regional? | National, because priorities and laws differ | Regional, because scale matters |
These are genuine disagreements among informed people, and reasonable readers will weigh them differently.
What "sovereign" should mean for an enterprise buyer
Here we can be more practical. The word is now used in sales material for servers, clouds and data centres, usually without saying which kind of control is on offer. It helps to break it into layers.
| Layer | The question | Owning hardware helps? |
|---|---|---|
| ๐ Data residency | Where does my data physically sit? | โ Yes |
| ๐งโ๐ป Operational control | Who can log in and administer the system? | โ Yes, if your own staff run it |
| โ๏ธ Legal jurisdiction | Whose laws can compel access? | ๐ก Depends on who operates it |
| ๐ง Model control | Can I keep running the model if a provider withdraws it? | โ Yes, with open-weight models on your own hardware |
| ๐ง Supply chain | Can I get parts, support and replacements? | โ No. This stays dependent on foreign suppliers and export rules. |
Two practical conclusions follow. First, when a vendor says "sovereign", ask which of these five layers they mean. Usually it is the first. Second, running open-weight models on hardware you control delivers real independence on four of the five layers, at a scale far below a national programme. A few GPU servers can do it.
The fifth layer is a reminder that nobody in this region is fully independent. Advanced AI hardware is subject to export documentation wherever it is installed, which we cover in Why Your AI Infrastructure Supplier Should Understand Export Compliance.
This article is general commentary, not legal or policy advice.




