Key facts
- Definition: Sovereign AI is the degree of control that stays within a country over who gets computing time, which models run, where data goes and who can take over when a supplier leaves
- India’s chips: Providers to India’s public AI programme buy chips abroad; only the US and China have data centres with AI chips from a domestic vendor (November 2024 mapping of nine public cloud providers)
- India’s subsidy: More than 45,000 GPUs enrolled by September 2026; standard subsidy up to 40% of computing costs for eligible portal users
- Malaysia: YTL AI Labs’ coding assistant, launched in October at Universiti Malaya, uses a model developed with China’s Z.ai; YTL says data is processed in Malaysia, but its terms (clause 2.4) let it change the models the assistant connects to
- Indonesia: Indosat hosts Sahabat-AI models in Indonesia, but choosing a third-party model forwards the text to that provider; usage and billing records stay in an Indonesian database
- Continuity: India’s UMANG tender (January 2021) required the outgoing supplier to keep the assistant running and hand over models, conversation flows and credentials; intellectual property stayed with the agency
Can AI be sovereign if the chips come from abroad?
The companies supplying India’s public AI programme buy their chips abroad. India decides whose work it will subsidise on those chips.
In August, the government told Parliament that subsidised access to graphics processing units, or GPUs, was part of India’s sovereign AI programme. Researchers can train models on them. Providers still rely on overseas suppliers for replacements.
Zoe Jay Hawkins and her colleagues mapped the AI chips offered by nine public cloud providers in November 2024. Only in the US and China did they find data centres equipped with AI chips from a domestic vendor.
To see how much control stays in a country, look at who allocates computing time, who chooses the model, who decides where data goes and who can take over when a supplier leaves.
Who gets to use those chips?
By September 2026, India had enrolled more than 45,000 GPUs in its subsidised computing programme. The standard subsidy for eligible portal users covered up to 40 per cent of their computing costs, the government told Parliament.
Most of the money committed by 8 October backed the development of foundation models, which can be adapted to many tasks. ThePrint’s Vrinda Tulsian found that nine organisations building such models accounted for 16 of the 304 approvals on the portal and 83 per cent of the subsidy money. Prathosh A.P., an associate professor at the Indian Institute of Science, told ThePrint that academic teams struggled to afford even the subsidised rates.
Can Malaysia build sovereign AI with a Chinese model partner?
In October, Malaysia’s YTL AI Labs launched a coding assistant at Universiti Malaya. YTL says it developed the model behind the tool with China’s Z.ai. For a software team handling a client’s code, the decision to use the assistant may turn on where that code goes.
A programmer trying to fix an error might type only a few words. The assistant can send more than that to the model, including the contents of a file, changes the programmer has made and the error message on screen. YTL says it processes all data sent through its own model interface in Malaysia.
The same team has a different choice on Amazon’s cloud service, which offers eligible customers a model from the same Chinese family. Amazon lets buyers route requests within the US or across its global network. Neither setting lets a buyer specify that the code must be processed in Malaysia.
In August, Z.ai released its GLM-5.3 model’s weights, the numerical values the model learned during training. Under Z.ai’s licence, businesses can download this public version and run it commercially on their own servers, though some large providers must first pass Z.ai’s security review. The launch announcement leaves open who holds the files for the version YTL serves, who can update it and whether YTL could keep it running if the partnership ended.
A Malaysian team can therefore choose domestic processing through YTL’s interface even when the model was developed with a Chinese partner. If the team wants to keep a model it has tested, it depends on YTL to keep offering it. Clause 2.4 of YTL’s terms lets the company change which models the assistant connects to.
Does using a local AI platform keep a document in the country?
An Indonesian business can buy AI services from a local telecoms group and still send a document abroad when it asks for a summary. Indosat offers one account for a catalogue that includes Sahabat-AI, a family of language models built for Indonesian and regional languages such as Javanese. Indosat lists a version it hosts in Indonesia alongside models served by outside providers.
The words the business sends for a summary and the record used to bill for the work can go to different places. Indosat says the models it hosts process those words on its infrastructure in Indonesia. If the customer selects a third-party model, Indosat forwards the text to that provider. Some providers route requests wherever they have capacity, so the document may be processed abroad.
The usage record identifies the model, the time of the request and what it cost. Indosat keeps usage and billing records in an Indonesian database. A bill kept in Indonesia can therefore cover a document processed elsewhere. An Indonesian business that needs its documents processed at home can restrict which models its staff and applications use.
If the supplier leaves, who can keep the service running?
India’s UMANG app lets people access government services. When the National e-Governance Division sought a supplier for the app’s chat and voice assistant in January 2021, it specified how a replacement could take over. The outgoing supplier had to keep the assistant running during the transition. The agency or its chosen successor was to receive the current models, the conversation flows that guide a user through a request, and the credentials needed to operate the system.
The tender also specified that the agency would hold the intellectual property rights to models and training data developed for the project. Under those terms, the agency could appoint a new operator and give it both the files and the right to use them.
Sovereign AI is the degree of control that remains within a country over who gets computing time, which models run, where data goes and who can take over a service when a supplier leaves.