Bangladesh Wants a Bangla AI. Its Own History Is the Warning

Bangladesh has decided to build its own artificial intelligence. The draft National AI Policy 2026-2030, published in February, states that the first project under the policy will be a national Bangla language model, developed as a government-led initiative and managed by a new implementation cell inside the ICT Division. In October the prime minister's adviser on ICT, Rehan Asad, promised a sovereign AI computing facility within twelve months. This month the government asked the Asian Development Bank for an indicative $250 million programme to pay for a Bangla model, cybersecurity and "technology villages", The Daily Star reported.

The ambition is reasonable. The record behind it is not encouraging. For half a century the Bangladeshi state has announced projects of exactly this kind, and a striking number have ended as empty buildings, idle machines and unpaid loans. Before the first taka is borrowed for a Bangla AI, it is worth asking why.

Why a Bangla model matters

A large language model is software trained on enormous quantities of text until it can read, summarise, translate and write. The best-known systems are built by a handful of American and Chinese firms, and they are trained overwhelmingly on English.

Bangla is among the most widely spoken languages on earth, with roughly 237 million native speakers. Yet researchers still class it as a "low-resource" language, because so little high-quality Bangla text is available in a form machines can learn from, as the authors of one recent study note. The practical effect is that AI tools work less well for a farmer in Rangpur than for an office worker in Ohio.

The government's case is strongest inside its own offices. At the October dialogue, Dhaka Forum founder Ashfaq Zaman pointed to about 300 million archived Bangla documents and an electronic filing system running across 17,000 government offices. A model trained on that material and hosted inside the country could help officials draft, search and translate, The Daily Star reported. He also offered the sentence that should be pinned above every desk in the ICT Division: "The real constraint is not the lack of systems but their failure to connect."

What has been decided

Three things are on the table.

The first is the model itself. The draft policy names it as the flagship project and commits the state to prefer domestically hosted, Bangladesh-specific models over foreign services reached through the internet.

The second is computing power. Mr Asad said the government has chosen a small, modular facility over a giant data centre, that sourcing of computing chips is arranged, and that the most advanced processors are still being pursued through government-to-government talks. He acknowledged that electricity and cooling water will be constraints.

The third is money. The finance minister led a delegation to Seoul in early November. Possible partners named include LG AI Research, whose EXAONE model has been cited as a template, the chip maker FuriosaAI and the cloud company NAVER. The ADB's country director confirmed that an indicative $250 million programme is under discussion, with a consolidated request expected by mid-December. Neither the financing structure nor Korean co-financing has been settled.

None of this is unreasonable on paper. Nor were the plans that came before.

Fifty years of unfinished ambition

The factories. In 1972 the new state nationalised jute, textiles, sugar and much else. The symbol of that era was Adamjee, one of the largest jute mills in the world. It was placed under a state corporation in 1972 and closed thirty years later on account of recurring losses and labour unrest, in the words of Banglapedia. The pattern never ended. In the 2020-21 fiscal year, 103 state-owned industrial enterprises were on course for operating losses of Tk 2,387 crore, and finance ministry officials conceded that no policy tried since liberation had fixed them, The Business Standard reported.

The cable. In the early 1990s Bangladesh declined to join the SEA-ME-WE-3 submarine cable, an offer usually described as free of charge. The country did not get its first undersea internet connection until 2006, according to The Daily Star. A later telecom minister said the decision had left the country 14 years behind in information technology.

The laptop. In 2011 the state telephone manufacturer launched Doel, the first laptop assembled in Bangladesh. Within a few years assembly had stopped. The company then imported finished laptops, put the Doel logo on them and sold them as locally made, until demand fell to zero, The Daily Star found.

The satellite. Bangabandhu-1 was launched in 2018 at a cost of about Tk 2,700 crore, with a business plan built on renting capacity to other countries. That market never appeared. By 2024 only 59 percent of its C-band and 25 percent of its Ku-band capacity had been sold, and industry estimates put the eventual loss above Tk 1,500 crore, The Financial Express reported.

The voting machines. Before the 2018 election the Election Commission bought 150,000 electronic voting machines for Tk 3,825 crore. Each cost about Tk 2.35 lakh, eleven times the price of India's machines. By 2024 about 105,000 were out of order, The Daily Star reported, and the Anti-Corruption Commission is investigating.

The parks. Between 2009 and 2023 the government built 39 hi-tech and software parks on the assumption that buildings would produce an industry. Nearly five years after it opened, the Sylhet park had created about 400 jobs against a target of 20,000, Asia News Network reported.

The search engine. Pipilika, a Bangla search engine built at Shahjalal University of Science and Technology with a telecom partner, launched in 2013 and quietly shut in 2021. It was a university project without the funding to become a sustainable service, as one account of its failure explains.

The most relevant failure of all

There is one more case, and it is the closest relative of the new plan.

In 2016 the Bangladesh Computer Council began a project to bring Bangla into modern computing. It had a budget of about Tk 159 crore, a list of 16 tools and a three-year schedule. By early 2022 only one tool had been released, The Business Standard reported.

Ten years on, 12 of the 16 tools exist. The interim government brought several of them together as Kagoj.ai and launched it in December 2025 as the first AI platform for Bangla. A Prothom Alo investigation published this month found that about Tk 80 crore has been spent, that the deadline has been extended three times, most recently to June 2028, and that the platform's text-recognition service was receiving about 50 requests a day. Free tools already on most smartphones do the same job. The ICT secretary's own verdict was candid: "The tools have not reached people in the way they were supposed to."

Not everything failed. The spelling checker tested well and handles about half a million requests a day. The project has also collected real assets: thousands of digitised books and thousands of hours of recorded speech. That is exactly the raw material a Bangla model needs. But this project, run under the same ICT Division that will manage the new model, is the honest baseline for what happens next.

Why the pattern repeats

The failures span several governments and very different technologies, so the explanation cannot be one party or one minister. Five causes recur.

Projects are built for the announcement. A white paper commissioned by the interim government reviewed 52 ICT Division projects and found that at least 12 projects and 65 components had been named after political figures or used for partisan visibility. It described the Digital Bangladesh programme as an "architecture of political slogan", The Daily Star reported. When the purpose of a project is the launch ceremony, nobody is responsible for year five.

Nobody asks who will use it. The satellite had no foreign customers. The hi-tech parks had no tenants. The Digital Sylhet City project delivered Wi-Fi to less than five percent of its intended users, according to The Daily Star's account of the white paper. Demand was assumed, never measured.

Procurement is where the money goes. The same white paper found that some equipment had cost as much as four times world prices, that groups of vendors behaved like cartels and that tenders had been rigged. The voting machines are the clearest example. Expensive hardware is easy to overprice and hard for outsiders to audit, and AI hardware is the most expensive of all.

Buildings come before people. Bangladesh has repeatedly bought the visible part of a technology and neglected the invisible part: engineers, maintenance budgets, working electricity. The country has few graduates with advanced AI skills, and many of the best leave for better pay abroad, as a Financial Express commentary warned last month. The same piece noted that the national grid is too unreliable to run a sensitive data centre on its own.

Nothing is ever closed, and nobody is ever accountable. Deadlines are extended instead of results being examined. Each new government inherits the projects of the last, renames them and carries on. Even the white paper was taken down from the ICT Division's website within days of its publication, a decision Transparency International Bangladesh questioned.

The exception that proves the rule

It would be wrong to conclude that the Bangladeshi state cannot do science. It can, and the proof is in every rice field.

The Bangladesh Rice Research Institute has released more than 120 rice varieties, and rice production tripled in the four decades after independence, The Daily Star reported this year. By 2017 its varieties covered more than 80 percent of the country's rice land.

The differences from the failures above are instructive. The institute has a single measurable goal. It is a permanent research body, not a project with an end date. It works with international partners such as the International Rice Research Institute. And its results are tested in farmers' fields, where failure cannot be hidden.

What a serious Bangla AI would look like

There is also encouraging evidence from the technology itself. In 2025 two researchers published TigerLLM, a family of open Bangla models built on freely available foundations and trained partly on national school textbooks. On standard Bangla benchmarks it outperformed GPT-3.5, an older commercial model. It was the work of a small academic team, not a quarter-billion-dollar programme.

That suggests a cheaper and safer path than the one being discussed.

Start with the data. The scarce resource is clean Bangla text and speech. The state already owns archives, textbooks, court records and the datasets collected under the Bangla computing project. Releasing them openly would do more for Bangla AI than any data centre, because universities and companies could then build on them.

Adapt before building from scratch. Fine-tuning open models on Bangla data costs a small fraction of training a new one. A sovereign model can still be hosted on Bangladeshi soil.

Name the users and the targets first. Which ministry will use the model, for which task, and how will success be measured? If nobody can answer, the project is not ready.

Publish the tests. Independent benchmarks, run by universities and published, are the AI equivalent of the farmer's field.

Keep procurement open. Competitive tenders, published prices and no agreement that locks the state into one supplier's chips or cloud.

Borrow last. A loan for a model that nobody uses still has to be repaid. Grants and pilot funding should come before debt.

Write the ending in advance. Every component should carry a date on which it is independently reviewed and, if it has failed, closed.

The real test

Supporters of the plan can fairly reply that the present government did not build the satellite or buy the voting machines, that it has chosen a modest modular facility over a monument, and that depending forever on foreign AI carries its own risks. All of that is true.

But the previous governments also had good arguments at the beginning. What they lacked was a way of finding out, honestly and early, that something was not working. A Bangla AI is worth building. Whether this one succeeds will depend less on chips from Seoul than on whether Dhaka is finally willing to measure, publish and admit.