Sovereign AI in 12 months: Reality behind promise

Bangladesh’s ambition to establish sovereign AI capabilities within the next 12 months marks a significant policy commitment, with implications for its technological autonomy, national security and international partnerships. However, translating this ambition into measurable outcomes will require more than a political declaration. Three decades of reporting on telecommunications and information technology have reinforced a fundamental lesson: the credibility of a policy commitment ultimately rests on the government’s capacity to implement it.

The announcement emerged from the AI and Innovation Dialogue, jointly organised by the Prime Minister’s Office and the Ministry of Foreign Affairs on 8 October.

The event was attended by two state ministers for foreign affairs and Rehan Asif Asad, the prime minister’s adviser on telecommunications and ICT. Neither the Posts and Telecommunications Division nor the ICT Division was a co-organiser. This suggests that the government sees AI less as a purely technological matter than as an instrument of diplomacy—and that is the right way to approach it.

AI is no longer merely a technological issue. It is also a matter of diplomacy and national security. Where will the chips come from? Which countries will Bangladesh partner with on technology? What position will it take in global AI governance? Ultimately, questions such as these will be settled at the foreign policy table.

Yet one crucial question remains: who will be responsible for implementation? Turning a diplomatic ambition into reality will require coordinated action by the ICT Division, the Posts and Telecommunications Division, the Bangladesh Telecommunication Regulatory Commission (BTRC), the Power Division and several other institutions. How they will work together will be the first real test of this announcement.

We have seen grand promises before. During the era of the “Digital Bangladesh” slogan, infrastructure was built, but many projects later lay idle because of inadequate maintenance and a lack of operational capacity.

The promise of sovereign AI deserves a welcome. But it also demands some difficult questions. What, exactly, do we intend to achieve in 12 months? How will we get there? And what obstacles lie along the way?

Define sovereign AI first
If sovereign AI means developing our own chips, building our own frontier models and relying entirely on domestic technology, then even achieving it in 12 years—not to mention 12 months—would be doubtful. Outside the United States and China, very few countries are sovereign in this sense. Europe has not managed it either. For Bangladesh, such an ambition would therefore be unrealistic and misguided.

A more practical definition would be sovereignty through control at different levels. Computing infrastructure would be located within the country. Citizens’ data would remain subject to Bangladeshi law. A home-grown Bengali-language model would be developed on the basis of open-weight models. And that model would be put to work in public services.

It is worth explaining what an open-weight model is. During training, an AI model learns from vast amounts of data, encoding what it has learnt in millions or billions of numerical values known as weights, or parameters. In simple terms, these weights form the model’s “brain”.

For closed models such as ChatGPT or Claude, that brain remains on the servers of the companies that develop them. Users send their queries over the internet, and the responses are generated on infrastructure that may be located abroad.

With open-weight models, by contrast, the weights can be downloaded publicly. Familiar examples include Meta’s Llama, China’s DeepSeek and Alibaba’s Qwen, France’s Mistral, and Google’s Gemma. This makes it possible to run an entire model in a domestic data centre and further train or adapt it using local languages and data.

But open-weight does not necessarily mean fully open-source. The data used to train a model are generally not disclosed, and licences may impose various restrictions.

Moreover, the underlying model was developed elsewhere, so its assumptions, perspectives and biases may travel with it.

The World Bank is making a broadly similar argument. Its World Development Report 2026, published on 4 August this year, examines AI's potential for developing countries. It outlines a three-stage approach: first, adopt existing technologies; then adapt them to local needs; and, over time, move towards frontier AI. It also warns against wasting scarce resources and becoming overly dependent on a single supplier.

Bangladesh is cited as an example of a country that could create substantial value by adopting and adapting existing technologies rather than attempting to build frontier models from scratch.

The problem is that the announcement at the 8 October AI and Innovation Dialogue did not clearly define what “sovereign AI” would mean in practice. Without a definition, there can be no meaningful benchmarks. And without benchmarks, almost anything could be declared a success 12 months from now.

Computing power: Clock is already ticking
The first requirement for sovereign AI is sufficient computing capacity within the country—in other words, high-performance GPU clusters. Training and running a moderately capable Bengali-language model for public services could require hundreds, perhaps around a thousand, GPUs. The cost could run to tens of millions of dollars.

But money alone will not bring the GPUs in. Global demand is soaring, and supply chains are burdened by long waiting lists. Then comes Bangladesh’s public procurement process: tendering, evaluation, objections and retendering.

Experience suggests that this stage alone can easily consume six to nine months. More than half of the 12-month window could therefore disappear before the hardware even arrives.

Then there is electricity. AI data centres require several times as much power as conventional data centres, along with specialised cooling systems. Bangladesh has the National Data Centre at Kaliakoir, but it was not designed for the high-density computing demands of AI.

Keeping equipment worth millions of dollars running safely amid grid-related load-shedding and voltage fluctuations presents a separate challenge.

There is another longstanding problem: we buy infrastructure but fail to budget adequately for operating it. Without annual allocations for electricity, maintenance and upgrades, this cluster, too, could join the ranks of idle equipment within two years. We have seen that fate befall many projects in the sector.

Data is the real foundation of sovereignty
GPUs can be purchased. Data cannot simply be bought. The lifeblood of sovereign AI is our own language and the data generated by our own people. Bengali is spoken by nearly 300 million people, yet it remains underrepresented in global AI models. That gap is also an opportunity.

Seizing it will require a vast, clean corpus of Bengali-language material: government documents, court judgments, textbooks, national archives, radio and television speeches, and recordings that capture regional accents and dialects. But government data in Bangladesh remain fragmented, with much of the material still trapped on paper. The reluctance of one ministry to share data with another is a longstanding problem.

This also raises an important question about the news media. News organisations produce some of the richest and most extensive collections of contemporary Bengali text. If their content is used to train a national AI model, what arrangements will govern copyright and fair compensation? Unless this is settled now, disputes will be inevitable later.

There is no need to start from scratch when developing the model. Fine-tuning an open-weight model for Bengali is achievable within 12 months. The scattered work on Bengali language technology already under way at the country’s universities and start-ups must be brought together under a common framework. Bangladesh also needs its own Bengali-language benchmarks. Without them, there will be no reliable way to test claims that “our model is better”.

Data security is equally important. A legal framework for personal data protection is not enough on paper; it must be enforced in practice. Which categories of data must remain within Bangladesh? Who should have access to which datasets? Unless these questions are answered clearly, the word “sovereign” will ring hollow.

Machines cannot function without people
Hardware can be imported within a year. A skilled workforce cannot be built in the same time. A recent analysis has suggested that Bangladesh, a country of 180 million people, is struggling even to produce 1,000 highly qualified AI engineers. Many of those with the necessary skills leave the country or work from Bangladesh for foreign companies.

There are two priorities for the next 12 months. First, the government should establish a dedicated programme to engage Bangladeshi AI researchers and engineers living abroad, whether on a full-time or part-time basis. Second, universities and start-ups should have straightforward access to the national computing cluster. If researchers must spend months navigating paperwork to obtain permission to use the machines, the country will gain little from the investment.

Institutional responsibility is just as important. The ICT Division, BTRC, the Posts and Telecommunications Division, and the ministries responsible for health, agriculture and education all have a stake in AI.

But who will be in charge? In telecommunications, we have seen for years how decisions stall and responsibility is passed from one authority to another when several institutions share overlapping mandates.

Sovereign AI needs a single, empowered coordinating authority. The 12-month target must also be broken down into monthly milestones aligned with the National AI Policy 2026–2030.

Geopolitical barriers and the limits of connectivity
The hardest questions about sovereignty lie beyond technology itself. High-performance chips will largely come from the US technology ecosystem. A substantial share of open-weight models also originates in the United States or China. The world is increasingly dividing into two competing technology spheres: Pax Silica, led by the United States, and China’s competing technological ecosystem, often referred to as the WAIKO framework.

Which bloc will supply Bangladesh’s chips? Which bloc’s models will underpin its AI infrastructure? These decisions will determine the limits of the country’s technological sovereignty.

The strategy of keeping a foot in both camps has served Bangladesh well in the past. But the room for manoeuvre is shrinking rapidly as competition between technological blocs intensifies. It is better to acknowledge the reality: Bangladesh does not have full sovereignty over its chips or foundational models, and it is unlikely to achieve it. Our objective should be to preserve alternatives and avoid becoming captive to any single supplier.

Consider a domestic example. Puku is now a familiar AI coding platform among developers and students in Bangladesh. It offers a code editor, a command-line interface for working in the terminal, and a browser-based chat interface. According to the “About” page on its website, Puku was developed by Paridhi IO Limited, a Dhaka-based company. Its website lists users from several universities, including BUET, BRAC University and SUST.

By accepting payments through bKash, Nagad and Rocket, the company has made access to world-class AI models easier for users in Bangladesh. It has also developed its own technology for routing each request to an appropriate model based on the task’s purpose and complexity, as well as tools that analyse an entire project’s structure before generating code. Such domestic expertise is an important building block of sovereign AI.

Yet the same “About” page makes it clear that Paridhi’s products use models from multiple AI providers, while the company focuses on model research, fine-tuning and routing. Its packages include closed models such as Anthropic’s Claude Opus 5, whose weights are never hosted on domestic servers.

In other words, when someone uses such a model through Puku, their code and queries are sent to foreign servers for processing. Paridhi has stated that it does not use Puku app users’ data to train its own models and does not sell personal information. But a domestic company has no control over the national laws under which that data may be held or processed on a foreign provider’s servers.

This is not a failing unique to Puku or Paridhi. It reflects the reality of Bangladesh’s AI sector. Even the country’s most promising ventures depend on foreign models. If a foreign provider raises its prices, changes its terms or restricts access for geopolitical reasons, the consequences will be felt directly by these domestic businesses.

Bangladesh is making progress at the application layer, but the underlying “brain” remains in other hands. The promise of sovereign AI will become meaningful only when ventures such as Puku have the option of operating on domestic infrastructure using home-grown or open-weight models.

Connectivity presents another challenge. Bangladesh still depends on a limited number of submarine cables for international bandwidth. Debates over the international internet gateway (IIG) structure and bandwidth costs have been going on for years. The economics of running AI services domestically will depend heavily on the strength and affordability of the country’s own network infrastructure.

The promise will matter if…
I do not oppose this announcement. On the contrary, I would argue that, even if belatedly, the state is now focusing on the right questions. But turning the promise into reality will require several conditions to be met.

First, the government must publicly define what it means by “sovereign AI capability” at the end of 12 months: how much computing power will be available, which models will be deployed, and which services will be delivered.

Second, it needs a fast and transparent special procurement process for computing infrastructure, backed by a guaranteed operating budget for at least five years.

Third, the legal framework for data sharing and protection must be made effective in practice.
Fourth, the single coordinating authority must be given genuine powers to act.

Fifth, progress reports should be published every three months, accompanied by independent evaluations.

Finally, by the end of the year, ordinary citizens should be able to access at least a few government services in Bengali through AI—for health guidance, agricultural information and land-related services, for example. If citizens cannot use it themselves, sovereign AI will remain little more than a presentation in a government ministry.

A solid foundation can be laid in 12 months. But achieving comprehensive sovereign AI capabilities will require at least three to five years of sustained investment and political commitment. The excitement surrounding the announcement must not obscure that reality.

Rased Mehedi is a telecommunications and information technology analyst and editor of Views Bangladesh.