When an AI Summit Becomes an API Summit

When an AI Summit Becomes an API Summit

Bangladesh's first national AI business summit, held in Dhaka on 5 November, was billed as a step towards a technology-driven economy. On the evidence so far, it was mostly a trade fair for software that runs on someone else's brain. Call it what it was: an API Summit.

The term needs a word of explanation. An API, or application programming interface, is the socket through which a company plugs its app into a model owned by OpenAI, Google or Anthropic. Plenty of good businesses are built this way. But a country that only plugs in has not built anything it can call its own, and a national summit ought to know the difference.

To be fair to the organisers, they never promised a research conference. The Ministry of Science and Technology and the Dhaka Chamber of Commerce and Industry framed the day around adoption, under the slightly awkward slogan "AI-fication for Tomorrow's Economy". President Mirza Fakhrul Islam Alamgir opened it. Organisers lined up 142 stalls, more than 150 startups, student projects from 20 institutions and around 6,000 expected visitors. Nor did the panels duck the hard questions. According to The Business Standard, speakers discussed job losses, algorithmic bias, deepfakes and the danger of relying on foreign cloud services.

The problem is what came next, which was very little. In the coverage we have seen, no dataset was released, no compute grant offered, no research fund opened and no Bangla model unveiled. The summit diagnosed the dependence and then sent everyone home.

This matters because "AI" now covers several quite different activities. Using ChatGPT at work is adoption. Wiring a model into a bank's customer service is integration. Building or seriously adapting a model, and the data behind it, is development. Sovereignty goes further still. It means owning enough of the chain, from chips to data, that a price rise in San Francisco or a policy change in Washington cannot shut down a public service in Dhaka. Bangladesh is doing plenty of the first two. It is doing very little of the last two, and events like this one make that gap harder to see.

The frustrating thing is that the talent exists. Last year researchers at George Mason University in Virginia published TigerLLM, a family of Bangla language models, at ACL, one of the field's top conferences. They trained it partly on textbooks from Bangladesh's own national curriculum board. Fifty students from Bangladeshi universities helped write its seed tasks. The authors report that it beats every open-source rival on Bangla tests. Yet the work was done in America, and its training examples were generated with GPT-4o and Claude. A national AI strategy should be asking why work like this cannot be done at home.

Other countries have shown what a more serious effort looks like, and not all of them are rich. India's IndiaAI Mission, approved in 2024 with roughly 10,372 crore rupees, now offers more than 38,000 subsidised GPUs to startups and universities. Its summit in February launched home-grown models from firms such as Sarvam AI. Online critics still mocked many of the startups on show as wrappers around OpenAI. The difference is that India had an answer ready. Singapore spent S$70 million on SEA-LION, an open model for Southeast Asian languages, and trained its own engineers along the way. Even SEA-LION now sits on Alibaba's Qwen, a useful reminder that nobody builds every layer alone. The UAE gives its Falcon models away free, precisely so that it is seen as a maker of AI rather than a buyer.

Set against that, Bangladesh's public compute is tiny. The shared GPU cloud the government opened in January runs on just over 20 Nvidia chips of the Volta generation, a design first released in 2017. That is enough for coursework and small experiments. It is not enough to train a serious Bangla model. In October Rehan Asad, the prime minister's ICT adviser, promised a modular sovereign AI facility within a year. If it arrives on time, it will help. It was also a far weightier announcement than anything made at the summit, and it was made somewhere else.

The obvious objection is cost. Bangladesh has power cuts, thin foreign reserves and millions of young people who need jobs. Why not rent the best models in the world and spend the savings elsewhere? For most businesses, that is exactly the right call, and nobody should feel guilty about it. The trouble starts when the state does the same. A government that runs citizen services on a foreign model accepts that model's prices, its rules and its weak grasp of Bangla. Its citizens' data travels abroad and falls under foreign law.

The draft National AI Policy 2026 to 2030 understands this. It calls digital sovereignty a core principle and promises a national Bangla language model. But as bdnews24 reported in January, the policy was still a draft, and a shortage of good Bangla datasets was holding back local work. Good intentions on paper have not yet become labs, budgets or data.

So what should the next summit do? In some ways, less. It needs fewer ribbon-cuttings and more things people can use the next morning. Every exhibitor should say which model its product runs on and where user data goes. That is basic consumer information, not an accusation. Builders working on Bangla speech, Bangla text and local hosting deserve their own stage, not a corner of the hall. And the government should treat the event as a deadline for releasing something useful. Ashfaq Zaman of the Dhaka Forum has noted that the state already holds some 300 million archived Bangla documents. An open dataset drawn from that archive, free hours on National Data Centre GPUs, or a public Bangla benchmark would each do more for the country than another expo.

None of this is an argument against adoption. A garment factory that uses AI to cut waste is doing the economy a favour, whatever model it runs on. But a national summit is not a sales conference. Its job is to tell the country where it stands and what it plans to build. This one mostly showed Bangladesh how much it can buy. If next year's event cannot point to a single model, dataset or lab that did not exist today, the name will stick.

Original reporting: Voice of Startup Bangladesh

← All news