Weekly AI Roundup: What Happened in AI This Past Week (June 13–19, 2026)

Close the window. The monsoon has not properly arrived yet, but the pre-monsoon drizzle is doing its usual work on Dhaka’s streets — the kind of light, almost apologetic rain that makes the city smell new for exactly twenty minutes before the traffic reclaims it. Sit down. This week I have been staring at a single event more than any other in this roundup’s short life: for the first time in history, an American government picked up the phone and told an AI company to turn off its most powerful models for everyone outside the United States. And the company did it. By Friday morning Dhaka time, the AI your nephew was using to write his thesis, the tool your colleague was using to debug a client’s codebase, the model your startup had wired into a customer-service product — gone, with no warning and no timeline for return. I want to sit with that for a moment before we go anywhere else.

As always, I have split this roundup into two halves. Five big global stories first — from the kill switch that remade the global AI conversation overnight to the G7 summit where world leaders and AI lab chiefs broke bread together in the French Alps. Then five Bangladesh stories from the same seven days, where the foreign-dependency question went from theoretical to personal in a matter of hours. I will ask the same two questions on every story: who actually benefits, and who actually pays? And where it matters, I will pull the camera back to ask what this looks like from Dhaka, Sylhet, or Rangpur — rather than from Washington, Seoul, or Évian-les-Bains. Pour something cool. Let us begin.

5 Big International Developments

1. The US government deployed an AI kill switch for the first time — and cut off the whole world from Anthropic’s two most powerful models

The single most consequential event of the week happened on June 13, when US Commerce Secretary Howard Lutnick instructed Anthropic to restrict access to its Fable 5 and Mythos 5 models to US citizens and US persons only, effective immediately. Anthropic, unable to verify the nationality of every user in real time, did what the order effectively required: it disabled Fable 5 and Mythos 5 for all customers worldwide — Canada, Europe, South Korea, Australia, everywhere. For a stretch of hours, the two most capable publicly-accessible AI models on earth were simply off. All other Anthropic models were unaffected; access to Claude Sonnet 4.6, Haiku, and earlier versions continued without interruption. But Fable 5, released barely four days earlier as the most powerful model Anthropic had ever put in public hands, was gone before most of the world had finished evaluating it.

The precipitating cause, as reported by Fortune, Time, Bloomberg, and Tom’s Hardware, was a convergence of two security flags. Amazon researchers had identified vulnerabilities in Fable 5 and escalated them to the White House. Simultaneously, SK Telecom — South Korea’s largest wireless carrier and a $100 million Anthropic investor — was identified by the White House as a company “suspected of having ties to China” with active access to Mythos 5’s restricted cybersecurity capabilities. Anthropic’s AI adviser David Sacks revealed that the administration had offered Anthropic a choice before issuing the directive: fix a discovered jailbreak that could unlock Mythos 5’s most sensitive capabilities, or voluntarily de-deploy the model. Dario Amodei declined both options, on the grounds that the jailbreak was narrow and specific — “not serious,” in Anthropic’s characterisation — and that eliminating all jailbreaks from frontier models is technically impossible with current safety methods. The administration issued the full export-control directive.

Step back from the mechanics and look at what happened structurally. An American government had, for the first time, exercised what can only be described as a sovereign AI off switch — an assertion that a US company’s most advanced AI is a matter of national security, and that the government retains the authority to determine who, globally, gets to use it. This is a capability that governance researchers had theorised for years but never seen deployed. The week it was first deployed happened to coincide with a G7 summit discussing AI governance, with Anthropic opening a new office in Seoul, and with the world still debating what appropriate oversight of frontier AI should even look like. Those are not coincidences. They are the shape of the current moment.

Who benefits, and who pays? The administration’s security position is internally coherent — if Mythos 5 contains capabilities a Chinese-linked entity could exploit, and a known investor holds access, restricting that access is a defensible national-security precaution, however disruptive the method. The ones who pay are, in the immediate term, every non-American user of Fable 5 and Mythos 5 — who found their tools gone through no fault of their own, on the basis of a geopolitical dispute they had no part in. In the medium term, the ones who pay are the AI companies themselves: an AI system that can be turned off by government decree is a fundamentally different commercial product from one that cannot, and the global developer community now knows it. Trust, once damaged this way, does not come back easily.

The Dhaka view is direct and the stakes are real. Bangladesh’s developer and freelancer economy had built substantial workflows on Claude, and Fable 5 was the top of that stack — the model developers reached for when the task was genuinely hard. The overnight cut-off arrived with no warning, no alternative path, and no mechanism for appeal. Every Bangladeshi startup that had integrated Fable 5 or Mythos 5 into a live product discovered in real time that the AI infrastructure they thought they were renting was actually on loan, conditional on geopolitics they cannot control. The event is a practical demonstration of the risk that Bangladesh’s National AI Policy 2026–2030 discusses in abstract terms — “digital sovereignty” — and makes it concrete in a way that no policy document has managed. You cannot build a national AI economy on infrastructure you do not own, and this week proved it in twelve hours flat.

Looking out to 2028–2030, the structural significance is that the US government has demonstrated it can and will use AI access as a national-security instrument, in the same category as semiconductor export controls. That changes the calculus for every country running its AI workloads on American platforms — which, right now, is most of the world. It accelerates the case for sovereign compute infrastructure, for open-weight models that can be run domestically, and for AI governance frameworks that do not assume American goodwill is a permanent structural feature. For Bangladesh, it converts a theoretical argument into a lived experience. Lived experiences change policy faster than theory does.

2. The Anthropic ban triggered a global scramble for sovereign AI — and Mistral just became the most important company in Europe

If June 13 was the event, the days that followed were the reckoning. Across Europe the reaction was swift and visceral. Euronews called it a “wake-up call.” Sifted published a deep analysis titled “What the Anthropic shutdown means for Europe.” Fortune reported that the shutdown “ignites calls for sovereign AI across Europe.” The Conversation ran a piece titled “Sovereign AI? Anthropic shutdown reveals Canada’s weakness” — documenting how the ban landed just as hard in Ottawa as in Paris or Berlin. The word of the week, in every language, was sovereignty: the idea that a country should control the AI systems its citizens and companies depend on, rather than relying on systems that can be switched off by a foreign government without notice or appeal.

The timing landed on top of something the European Commission had already set in motion. On June 3 — ten days before the ban — the Commission had released its European Technological Sovereignty Package, comprising two major legislative proposals: Chips Act 2.0, aimed at rebuilding European semiconductor capacity, and the Cloud and AI Development Act, setting objectives for local cloud providers and home-grown AI. Both were designed to reduce Europe’s dependence on American and Asian technology infrastructure. What had looked like long-term industrial policy became, by the morning of June 14, something that should have been done three years earlier.

The immediate market beneficiary is Mistral AI, the Paris-based open-weight model company. Mistral’s positioning as a European-origin, open-source-friendly alternative to American closed models went from “interesting strategic argument” to “the obvious answer” in roughly 24 hours. Mistral cannot yet match Fable 5 on raw capability benchmarks — no European lab can, today. But in a landscape where the question has become not just “which model is best?” but “which model can I still access next Thursday?”, provenance and reliability of access become competitive attributes that benchmark scores cannot capture. A model you can download and run on your own servers cannot be switched off by Washington.

Who benefits, and who pays? Europe’s AI-independent companies and researchers benefit from renewed political will for sovereign infrastructure investment. Open-weight models — DeepSeek from China, Mistral from France, various academic releases — benefit enormously, because they can be downloaded and run on domestic hardware beyond the reach of any US government directive. The ones who pay are the American AI labs, whose global revenue and goodwill take a durable hit every time this precedent is invoked; and, more broadly, the integrated global AI ecosystem that the past three years had been quietly building, which just took a hard turn toward fragmentation. A world of sovereign AI blocs is a world of slower, more expensive, more duplicative AI development. Whether that cost is worth paying for the security of knowing your AI cannot be switched off is the question every government is now asking.

The Dhaka view is that the sovereign AI argument has just acquired overwhelming empirical force, and Bangladesh needs to act on it. Open-weight models like DeepSeek, which can be downloaded and run on Bangladeshi servers, became strategically more important overnight. The argument for Bangladesh to accelerate its National AI Compute Strategy — to procure GPUs, to establish shared government compute, to build domestic AI infrastructure — is no longer theoretical. The practical demonstration just happened. Every month Bangladesh delays that investment is another month in which the entire country’s AI capability can be eliminated by a single foreign bureaucratic decision.

Looking out to 2028–2030, the structural picture is of AI fragmenting into geopolitical blocs — an American-anchored bloc of closed, government-controllable frontier models; a Chinese-anchored bloc of open-weight alternatives with their own political dependencies; a European-anchored bloc of sovereignty-focused providers; and the vast majority of the world scrambling to figure out which to trust. Bangladesh’s instinct — to diversify, to use multiple providers, to hedge — is exactly right. The tactical move is to run open-weight models domestically for the workloads where capability allows, and to reserve foreign frontier models for tasks only they can do. The strategic move is to build the compute infrastructure that makes that domestication possible. The case for building it is now airtight.

3. SpaceX acquired Cursor for $60 billion — and the AI coding market just reshuffled completely

On June 16, days after SpaceX began trading on Nasdaq as SPCX following its blockbuster IPO, the company confirmed it would acquire Anysphere — the company behind the AI coding tool Cursor — in an all-stock deal valued at $60 billion. The acquisition is expected to close in the third quarter of 2026, pending regulatory approval.

Cursor deserves a moment of explanation for readers who do not follow the software world. Founded in 2022 by four MIT classmates, Cursor is what developers call an “AI-native” code editor — a tool where the AI is built into the writing environment itself, not bolted on as an add-on. It had become one of the most-loved coding tools among professional software developers, with roughly $2.6 billion in annualised revenue and rising enterprise sales. SpaceX had secured an option in April 2026 to either pay $10 billion for a deep partnership or $60 billion for outright acquisition; it exercised the larger option. The strategic logic runs through xAI — Elon Musk’s AI company, which SpaceX merged with earlier this year — which needed a developer-facing product to compete with GitHub Copilot, Claude Code, and OpenAI’s Codex and Windsurf.

The AI coding market, as it now stands, looks like this: GitHub Copilot (Microsoft, increasingly running on Microsoft’s own MAI models), Codex and Windsurf (OpenAI), Claude Code (Anthropic), and now Cursor (SpaceX-xAI). In less than a year, it has gone from “a few startups and one Microsoft product” to “a race between every major tech platform, each with its own model and its own developer distribution.” Cursor’s $60 billion price tag tells you exactly what the big players think control of the developer’s daily workflow is worth.

Who benefits, and who pays? SpaceX-xAI benefits from immediate revenue ($2.6B ARR on day one) and from Cursor’s genuine developer goodwill — which is harder to buy than revenue and usually easier to lose. Cursor’s founders and early investors benefit from an extraordinary return on a three-year-old product. Developers who use Cursor benefit in the short term from SpaceX’s resources: more compute, faster model updates, deeper integration with xAI’s Grok models. The ones who pay are, first, the Cursor users drawn by its independence from big-tech platforms, who now find themselves inside SpaceX’s corporate structure with all the pricing and data-policy uncertainty that entails; and second, the smaller independent AI-coding startups that just lost a category-defining product to consolidation, making it harder to raise money or attract talent in that space. When a $60 billion deal closes, it narrows the field.

The Dhaka view runs through Bangladesh’s enormous freelancing and software-services economy. Cursor is widely used among Bangladeshi developers — particularly those doing contract work for European and American clients who specify the tools they work in. A Cursor owned by SpaceX-xAI is a Cursor whose pricing, data practices, and model choices will be set by Elon Musk’s corporate priorities, not by the independent founders who built it on developer trust. This is the moment for Bangladeshi developers to evaluate tool diversification: to identify which workflows are deeply locked into Cursor, whether Claude Code or Copilot can serve them adequately, and what a switch would cost if Cursor’s post-acquisition pricing shifts. The consolidation of the coding-tool market makes these calculations more important, not less.

Looking out to 2028–2030, the structural pattern is that the AI coding market is becoming rapidly an oligopoly of platform giants — Microsoft, OpenAI, Anthropic, and now SpaceX-xAI, each with owned models and owned distribution. The era of the independent AI coding tool — where a four-person team could build a billion-dollar company on a great product — is not over, but it has compressed significantly. The coding tools of the 2030s will be owned by the same companies that own the underlying models, and the pricing and access decisions will all flow from that ownership. Building skill across multiple tools rather than deep dependency on one is the rational response to a market consolidating at this speed.

4. The G7 summit placed AI lab chiefs at the world leaders’ table — and produced more questions than answers

The 52nd G7 Summit, held from June 15–17 in Évian-les-Bains, France, under the presidency of Emmanuel Macron, was the first in the summit’s 52-year history in which the heads of the world’s leading AI laboratories formally sat alongside heads of state. Sam Altman (OpenAI), Dario Amodei (Anthropic), and Demis Hassabis (Google DeepMind) all attended, and on June 17 they joined US President Donald Trump, Macron, and other G7 leaders for a working dinner in the French Alps that addressed AI as a central strategic agenda item.

The summit’s AI session covered three broad areas: AI infrastructure and energy demands, regulatory alignment across G7 jurisdictions, and the governance of frontier model development — the last of which sat directly in the shadow of Anthropic’s “When AI Builds Itself” paper from two weeks earlier, and, as the week unfolded, the Fable 5/Mythos 5 export-control drama playing out in real time. As ABC News reported, European leaders were pressing for “checks on American dominance” of AI — a theme amplified by the ban on Anthropic’s models, which had just demonstrated to every non-American government exactly how complete that dominance was. As The Hill put it, the summit “focuses on the contentious future of AI and US dominance of the industry.”

No binding agreements or specific regulatory commitments emerged from the AI sessions. The statement language was, by G7 standards, warm but vague — commitments to “responsible AI development,” “safety-by-design,” and “alignment of regulatory frameworks.” The absence of a binding outcome is not unusual for a G7 summit; the point is usually norm-setting and political signalling rather than treaty-making. But the signalling this week was harder to read than usual, because it was occurring simultaneously with the US government imposing export controls on an AI company — an action that looks very different from “regulatory alignment” and rather more like unilateral tech-trade policy.

Who benefits, and who pays? The AI labs benefit from a formal recognition, at the highest level of global governance, that they are geopolitical actors — entities whose decisions carry the same weight as national policy decisions in the domains they touch. That recognition is a form of power, and the labs will use it. G7 governments benefit from having people who actually understand the technology in the room when governance decisions are made, rather than legislating in a fog of technical ignorance. The ones who pay, structurally, are every country not at the G7 table. The governance norms being set in Évian-les-Bains will travel — as rules, as standards, as expectations — to every other country in the world, but they will have been written without those governments present. That is not a new problem with G7 governance, but AI makes it a more consequential one than any previous iteration.

The Dhaka view is one I have made before and will keep making until it changes: Bangladesh is not at the table, and the decisions made there will determine the terms on which Bangladesh accesses, uses, and builds on frontier AI for the next decade. The positive piece of news is that the UN Global Dialogue on AI Governance — with its first session on July 6–7 in Geneva — is exactly the forum where Bangladesh does have a formal voice. The Évian summit reinforced why that July session matters: if the G7 sets the norms among the powerful, and the UN dialogue is where the rest of the world gets to respond, then Bangladesh’s presence, preparation, and clarity of position in Geneva is one of the highest-leverage AI-policy actions the country can take this year. The question remains whether anyone in Dhaka is drafting that position now, or whether July 7 will arrive with Bangladesh showing up as an audience rather than a participant.

Looking out to 2028–2030, the structural significance of this summit is not any particular outcome but the precedent it sets: AI governance is now conducted at the heads-of-government level, with laboratory chiefs as participants. Every future G7, G20, and major multilateral summit will now have an AI agenda and AI-industry representatives in the room. The question for Bangladesh is not how to get into the G7 — that is not available — but how to be present, credible, and vocal in the multilateral forums that are available. Building that presence requires having real positions, real expertise, and real relationships with governments and labs who do sit at the table. That infrastructure does not assemble itself the week before a summit.

5. Anthropic opened its Seoul office — and the Korea story is a study in contradictions

The final major international story of the week is a small but telling portrait of the contradictions at the heart of the current AI moment. On June 17–18, Anthropic formally opened its Seoul office — its third in Asia-Pacific, after Tokyo and Bengaluru — and announced a sweep of enterprise partnerships with some of South Korea’s largest companies. KiYoung Choi, former General Manager of Snowflake Korea, leads the office as Representative Director.

The partnerships are substantial. NAVER — South Korea’s largest internet company, roughly Korea’s equivalent of Google — deployed Claude Code across its entire engineering organisation, with thousands of engineers using it for day-to-day coding work. Nexon, one of Asia’s most successful gaming companies, deployed Claude Code for live-service game development. Samsung SDS launched Claude Cowork and Claude Code across Samsung Electronics. LG CNS deployed Claude across the LG Group. Hanwha Solutions deployed Claude via AWS Bedrock with in-region data controls. And Channel Corp’s Channel Talk platform — used by more than 230,000 businesses — is now powered by Claude. The scale of adoption across the Korean corporate ecosystem is striking even by global standards.

The contradiction is immediate and worth naming clearly. The same US government that had, four days earlier, told Anthropic to cut off all foreign nationals from its most advanced models — including the South Korean citizens who are customers of NAVER, Nexon, Samsung, and LG — was now watching Anthropic open an office in Seoul to deepen exactly those relationships. Anthropic said publicly that the export-control situation “will be resolved soon” and that it was actively working with the US government on a path back to full access. The Seoul office announcement was not postponed or cancelled. If anything, it suggests Anthropic is betting that the June 13 ban is temporary and that its Asia-Pacific expansion is the durable story.

Who benefits, and who pays? South Korean tech giants benefit from Claude Code’s productivity advantages — if NAVER is deploying it to thousands of engineers, the calculus on coding speed and quality is clearly favourable. Anthropic benefits from deepening its revenue base in one of the world’s most advanced technology markets ahead of its IPO. The ones who pay are South Korean enterprises that integrated Fable 5 or Mythos 5 into production systems and had those systems go dark on June 13 without warning. And there is a quieter, more durable cost: every enterprise that watched the June 13 cut-off is now pricing in a “US-government-imposed-outage” risk into every American AI product it evaluates. That risk premium does not disappear when access is restored.

The Dhaka view connects to a practical message for Bangladeshi businesses and developers. The Korea story shows what a serious local AI ecosystem looks like at the enterprise level: major companies deploying frontier AI tools across entire engineering organisations, with local offices, local leaders, and data controls that keep sensitive information within the country’s borders. Bangladesh is not at that level yet — but the pattern Korea is demonstrating is the pattern to build toward. Local representation from the AI labs, in-country data controls, enterprise-scale deployment alongside a functioning regulatory framework, and the domestic technical capacity to supervise and extend what the AI does. That is the destination. The journey requires building institutions and training people who can run that kind of ecosystem, and those are choices the country can make starting now.

Looking out to 2028–2030, the structural question the Seoul story poses is: which countries will be where Korea is today, by 2030? Korea got there through a combination of government commitment, enterprise investment, and a technical talent pipeline that Samsung, Hyundai, and KAIST have been building for decades. Bangladesh’s equivalent pipeline is younger and more fragile, but it exists — in the BdAIO, in AIUB and BUET and NSU, in the freelancing community that has been learning these tools by necessity rather than by formal training programme. The Seoul office shows what the destination looks like. What gets Bangladesh there faster is sustained investment in the pipeline, not in any individual product or partnership.

5 Stories That Hit Closer to Home (Bangladesh)

1. The Anthropic ban hit Bangladesh’s AI economy like a power cut — and the lights may not come back the same way

The most important Bangladesh story of the week did not happen in Dhaka. It happened in Washington, and its consequences landed here within hours. When the US Commerce Department ordered Anthropic to cut off all foreign nationals from Fable 5 and Mythos 5 on June 13, every Bangladeshi developer, freelancer, and startup that had integrated those models into a live product or workflow discovered, without warning, that their tools were gone. No timeline for return, no alternative path, no mechanism for appeal. Just an error message where a working API endpoint used to be.

The scale of the affected community is hard to count precisely, but it is not small. Bangladesh’s freelancing economy of over 650,000 registered freelancers — and a much larger informal ecosystem — had been among the fastest-adopting communities for AI coding and writing tools globally. Claude, in particular, had been widely adopted for high-value tasks: software development contracts, content work for international clients, research assistance, and the back-end automation that powers an increasing share of the country’s digital-services exports. Fable 5 was the top of that stack — the model developers reached for when the task was genuinely difficult.

What makes this week’s event structurally important is not the disruption itself — disruptions happen, and this one may prove temporary. What makes it important is what it proved about the country’s AI infrastructure dependency. Bangladesh’s National AI Policy 2026–2030 commits the country to “digital sovereignty” as a core principle. It states, clearly and admirably, that Bangladesh intends to retain “effective governance and control over critical data, digital infrastructure, and AI systems.” But as of the morning of June 13, every serious AI workload in Bangladesh was running on foreign servers, under foreign terms of service, subject to a foreign government’s export-control authority. The policy says sovereignty. The infrastructure said otherwise. The gap between those two positions closed, hard and fast, on a Friday morning.

Who benefits, and who pays? In the short term, the companies best positioned to benefit from the cut-off are open-weight model providers — DeepSeek, which can be downloaded and run on Bangladeshi hardware, and Mistral, which operates outside US government authority — and Bangladeshi tech firms that had already been building AI products across multiple platforms rather than betting everything on one. The ones who pay are, most directly, the developers and startups that were mid-project with Fable 5 integrations when the switch was flipped. And in the structural sense, every organisation that now has to reprice the risk of foreign AI dependency into its technology planning. That repricing is not cheap, and it is not optional.

The Dhaka view is that this week’s event should be treated as a policy emergency, not an operational inconvenience. The response it warrants is not a workaround — switching to Sonnet 4.6 or finding another provider — but a fundamental recalibration of Bangladesh’s AI strategy. The ICT Division’s National AI Compute Strategy commitment needs to move from a line in a policy draft to a funded, staffed, time-bound programme. The case for sovereign GPU infrastructure, domestic hosting of open-weight models, and a Bangladesh-operated AI layer for government and critical services has never been more obvious or more urgent. The argument used to be “we should do this eventually.” The argument now is “we cannot afford not to have done this already.”

Looking out to 2028–2030, the structural reality is that US export control on AI is not going away. The Anthropic ban is the first use of the mechanism; it will not be the last. As frontier AI models become more capable and more strategically sensitive, the US government’s inclination to control who gets to use them will increase, not decrease. Bangladesh’s path to a resilient AI economy runs through diversification: open-weight models for the workloads they can handle, multiple foreign providers to avoid single-point failure, and domestic compute for the workloads that cannot leave the country’s borders. That is a multi-year programme and an expensive one. The cost of not building it just became much more visible.

2. Bangladesh’s team competed in the Asia-Pacific AI Olympiad from Dhaka

A quieter but genuinely encouraging story to set against the week’s harder news: on June 13 — the same day the Anthropic ban landed — Bangladesh’s eight-member team competed in the first Asia-Pacific Artificial Intelligence Olympiad (APOAI), sitting the online competition from a designated venue in Dhaka. The team was selected by Bangladesh Open Source Network, the organiser of the Bangladesh Artificial Intelligence Olympiad (BdAIO), through a national selection camp held in late May. The squad draws from schools and colleges across the country — not just from Dhaka’s top institutions — continuing a pattern of geographic breadth in the talent pipeline that deserves to be named explicitly rather than taken for granted.

I want to dwell on the timing, because it is worth holding. On the very day the US government proved how completely dependent Bangladesh’s AI economy is on foreign infrastructure, eight young Bangladeshis were sitting an international AI competition that required no foreign infrastructure at all — just trained minds, disciplined mathematical reasoning, and the kind of careful problem-solving that no export control can reach. That is not a coincidence worth ignoring. The most durable form of AI sovereignty is not compute clusters or GPUs. It is the human capital that can build, adapt, and reason about AI systems from first principles. The olympiad pipeline is building that capital, one training camp at a time, and the country has not yet fully understood what it has.

The APOAI is the first of its kind for the region, and Bangladesh’s participation — alongside a four-member team headed to Kazakhstan in August for the International AI Olympiad (IOAI) — reflects a sustained investment in competitive AI talent development that runs several years deep. The BdAIO has completed three national rounds; the selection infrastructure, the volunteer mentors, and the network of schools engaged in the process are a genuinely impressive piece of institutional building, made more impressive by how little external recognition it receives in the national AI conversation about data centres and policy frameworks.

Who benefits, and who pays? The eight students who competed are the most direct beneficiaries — whatever the result, they have been through an experience that sharpens the ability to think carefully about AI under pressure, in ways that a university lecture rarely can. Bangladesh’s reputation as a country building a serious AI-talent pipeline benefits, particularly within the Asia-Pacific region where the country’s technology profile has historically been overshadowed by its garments-and-remittance story. The one honest cost is structural: the olympiad pipeline runs largely on volunteerism, passion, and thin budgets. The Tk 500 crore AI and startup fund in the FY2026–27 budget does not, as far as published accounts show, include specific allocations for the olympiad infrastructure, the regional selection camps, or the mentorship network that makes this achievement possible. A country that is serious about AI talent builds that pipeline on salary lines and recurring budget allocations, not on the goodwill of people who happen to care.

The Dhaka view on this story is one of genuine pride paired with a practical ask. The talent exists and is competing internationally. The question is whether the system around that talent — the training programmes, the career pathways, the domestic AI jobs that give these students a reason to stay in Bangladesh rather than emigrate — is developing at the same pace as the competition results. The olympiad is the easy part. Keeping the competitors in the country, and building the economy that makes staying worth their while, is the work that the startup fund is supposed to enable. The connection between the competition and the fund needs to be made explicit and resourced accordingly.

Looking out to 2028–2030, the structural question is one I have asked before on these competition stories: will the olympiad result be the seed of a domestic AI sector, or a pipeline to emigration? The answer depends not on the students but on the system. A domestic AI venture ecosystem that can absorb olympiad-calibre talent, give it interesting problems to work on, and pay it at rates competitive with a Singaporean job offer — that is what converts a competition result into a national asset. Bangladesh is building the pipeline. The question is whether it is building the destination at the other end.

3. Russia agreed to send an AI delegation to Bangladesh for agriculture and health

A story that arrived just before this seven-day window and continued to resonate through the week: Bangladesh’s Foreign Minister, Dr. Khalilur Rahman, concluded a three-day official visit to Russia in which he proposed — and Russian authorities agreed to — sending an AI delegation to Bangladesh focused specifically on artificial intelligence applications in agriculture and health. Russia also expressed readiness to assist Bangladesh in preparing its national AI policy and in establishing institutions of excellence in AI. The visit’s broader agenda included expanding manpower cooperation (Bangladesh proposed raising the number of Bangladeshi workers in Russia from around 10,000 to 100,000) and expanding bilateral trade in pharmaceuticals and garments.

The AI dimension of the visit is worth unpacking carefully, because it represents something genuinely new in Bangladesh’s AI-diplomacy landscape. Almost all of the country’s AI cooperation has, to date, been structured around Western and East Asian partners — US companies (Anthropic, Google, Microsoft, OpenAI), Korean conglomerates, multilateral bodies anchored in the Geneva-New York axis. A Russia-Bangladesh AI partnership in agriculture and health is a different geopolitical configuration entirely — one that diversifies the country’s AI-cooperation portfolio and, not incidentally, offers an alternative to the American-dominated ecosystem that just demonstrated it can cut Bangladesh off at a moment’s notice.

The specific sectors matter. Bangladesh’s agricultural sector employs nearly 40 per cent of the workforce and is exactly the domain where AI applications offer large, concrete returns at modest compute cost: crop disease identification from smartphone photos, optimised irrigation scheduling, weather and market-price forecasting for smallholder farmers, pest and flood early-warning systems. These are not frontier-model applications. They are AI workloads that can run on relatively modest hardware, use open-source models, and be developed locally with reasonable budgets. Russia’s agricultural AI experience — drawn from large-scale farming technology programmes — is a plausible source of expertise for exactly this kind of application. Similarly in health: AI-assisted diagnostics, medical imaging, and patient-flow management are areas where Bangladesh’s health system has genuine need and the patient data to train on.

Who benefits, and who pays? Bangladeshi farmers and patients are the intended beneficiaries — if the delegation produces concrete AI tools deployed in agriculture and health, the gains in crop yields, disease detection, and healthcare access are real and measurable. The government benefits from a diplomatic partnership that diversifies its AI-cooperation relationships. Russia benefits from a new market for technology exports and the geopolitical signal that it can be a meaningful AI-development partner for countries in the Global South. The cost side is worth examining honestly: Russia’s own AI sector operates under Western sanctions that restrict its access to the chips and cloud infrastructure underpinning frontier AI. Any AI Russia offers to export will be built on alternatives to Nvidia’s best hardware, and its capabilities will reflect that constraint. The delegation’s visit will tell us whether Russia’s agricultural-AI offer is substantive or symbolic.

The Dhaka view is that the instinct behind this partnership is exactly right even if the execution details remain to be seen. The event of June 13 gave the Foreign Minister’s Russia visit a retrospective strategic logic it did not fully have at the time of the trip. Diversifying AI partnerships across geopolitical lines is not a luxury; it is a structural requirement for any country that intends to run its economy on AI it can depend on. Russia may not be the ideal partner for frontier model development. It may be a useful partner for agricultural and health AI where the capability requirements are more modest and the geopolitical independence of the partnership is more valuable than the raw benchmark score of the models involved.

Looking out to 2028–2030, the structural question is whether Bangladesh can build a genuinely diverse AI-partnership portfolio — drawing on Western models for frontier tasks, on open-weight alternatives for domestically deployable workloads, and on diverse bilateral partnerships for sector-specific applications. The Russia agriculture-and-health partnership is one piece of that portfolio. It should not be the whole picture, but it should not be dismissed either. A country that can deploy AI in its fields and hospitals without depending on any single foreign government’s permission is a country with a more resilient AI strategy than one that cannot.

4. Bangladesh launched its “One Citizen, One Digital ID, One Digital Wallet” initiative — and work starts in July

A landmark domestic announcement this week: the government confirmed that work will begin in July 2026 on Bangladesh’s first integrated Digital Public Infrastructure (DPI), structured around a simple but ambitious framework: one digital identity, one digital wallet, and universal access to digital services per citizen. The announcement came from the ICT adviser’s office and was reported by BSS, covered by Biometric Update, and noted by SAMENA Council. The initiative is described as the core of the government’s digital transformation plan — the foundational plumbing through which AI-enabled government services, financial inclusion, and social-benefit delivery will eventually run.

The idea is straightforward even if the execution is not. Bangladesh currently has multiple, non-interoperable identity systems — national ID cards, birth certificates, voter IDs, tax identification numbers — that do not speak to one another cleanly, creating friction everywhere a citizen needs to prove who they are. The DPI proposal replaces that fragmented landscape with a single digital identity that can be verified instantly across both public services and private-sector transactions. The digital wallet sits on top of that identity, providing every citizen — including the roughly 40 per cent who are unbanked or underbanked — with a pathway to digital financial services, allowance disbursement, and market participation. In its most ambitious form, the system would be the national-infrastructure equivalent of the mobile financial services revolution that bKash helped drive a decade ago, but built into the state’s own plumbing rather than sitting on top of it.

AI is woven through every layer of the intended system. Smart-ID verification as collateral for bank lending — already proposed in the FY2026–27 budget for freelancers — becomes practical only when the ID is universal and machine-readable. AI-driven benefit targeting — routing social transfers to the right people without requiring them to navigate a physical bureaucracy — becomes possible when a universal wallet exists. AI fraud detection across government financial transactions becomes scalable when all the transactions run through a common infrastructure layer. The DPI is not itself an AI project, but it is the precondition for most of the AI applications that matter most to ordinary Bangladeshis — not the chatbots and coding assistants, but the crop insurance, the healthcare payment, the school-fee remittance, the government transfer.

Who benefits, and who pays? The most direct beneficiaries are the 40 per cent of Bangladeshis currently excluded from formal financial services — who gain a pathway to a digital wallet that does not require a bank branch, an employer pay-stub, or a credit history. Every citizen benefits from reduced friction in interactions with government: a single identity that works everywhere is a concrete quality-of-life improvement for anyone who has spent an afternoon at a government office re-establishing the same basic facts about themselves from scratch. The ones who pay are, in the near term, taxpayers who fund the infrastructure build; and, in the medium term, anyone whose data-governance concerns are not addressed by the design. A universal identity wallet is a powerful tool in the government’s hands as well as the citizen’s, and the design of its privacy protections matters enormously. A DPI built without robust data-protection provisions is a surveillance infrastructure dressed as a public service.

The Dhaka view is cautious optimism with a specific ask. Bangladesh has announced ambitious digital-infrastructure programmes before, and the historical record is mixed at best. What would make this one different is governance design that puts the citizen, not the state, in control of the identity and wallet — an infrastructure where the citizen can see who accessed her ID record, can revoke that access, and can use her wallet through multiple competing providers rather than a single government-run system. The National AI Policy 2026–2030’s commitment to “citizen dignity” and data privacy is the right language. Whether the DPI’s technical architecture reflects that language is the implementation question that will determine who benefits and who pays in practice.

Looking out to 2028–2030, the structural significance is that Bangladesh’s AI future depends heavily on getting the DPI right. A well-designed, widely adopted universal digital identity is the foundation on which AI-powered public services, financial inclusion, and economic participation can be built at scale. If July 2026 is genuinely the start of that build — and not another announcement that quietly delays — the country could have a functioning DPI by 2028 that positions it well for the decade ahead. The design choices made in the next eighteen months will be very difficult to undo.

5. New research maps the legal response gap in Bangladesh’s AI-driven gender-based violence crisis

The hardest story of the week — and the one I feel most obligated not to soften — is a new academic paper published in the International Journal of Research and Scientific Innovation (IJRSI) under the title “Victims of AI: Addressing the Legal Response Gap to AI-Related Crimes in Bangladesh.” The paper examines, with survey data and legal analysis, the gap between the formal legal protections Bangladesh now has against AI-driven gender-based violence and the practical capacity to enforce them. The gap it describes is not narrow.

The core finding is stark. Bangladesh now has criminal provisions that treat non-consensual deepfake creation, AI-assisted extortion, and image-based abuse as serious offences — progress that is real and should be acknowledged. But 65 per cent of survey respondents believe existing laws are inadequate at preventing new harms from AI, and the paper documents why: prosecuting an AI-generated deepfake case successfully requires law-enforcement officers to have the technical capacity to identify synthetic media, the investigative capacity to trace distribution chains to their origin, the legal framework to compel cooperation from foreign platforms (Meta, TikTok, Google), and the institutional willingness to treat image-based harassment as a serious crime rather than a social dispute. Bangladesh’s law enforcement currently has limited capacity on all four fronts. The number of successful prosecutions for AI-generated image-based violence against women, relative to the documented scale of the harm, remains vanishingly small.

The survey data fills in the picture. 55 per cent of respondents perceive AI misuse as “very common,” with women and children identified as the most vulnerable groups. The pattern of harm is specific: extortionists — usually men, often known to the victim or encountered online — download photos from social media, run them through freely available deepfake tools, generate explicit synthetic content, and contact the victim with threats to distribute it to family, employers, or college administration unless payment is made. This is not a sophisticated operation. It requires a smartphone, a free account on a deepfake platform, and the willingness to cause harm. The technological barrier has essentially disappeared. The legal barrier — prosecution, conviction, deterrence — has not risen to compensate.

Who benefits, and who pays? The ones who pay are, first and most directly, the women targeted — who bear costs in reputation, employment, mental health, safety, and in the worst cases their lives, costs that accumulate whether or not the perpetrator is ever prosecuted. Every woman in Bangladesh who holds or aspires to any form of public life — in journalism, politics, entertainment, activism, or simply with an Instagram account — is paying a preventive tax in the form of restricted self-expression, managed visibility, and constant calculation of what can and cannot be safely posted. That is not a private cost. It is a national one. The beneficiaries of the current legal response gap are the perpetrators and the digital economy of extortion and harassment that has grown around them, knowing that prosecution is unlikely and detection is improbable.

The Dhaka view is that this is the governance test the country cannot afford to fail. The paper uses the term “legal response gap” — the distance between what the law says and what the enforcement system can actually do. Closing that gap is not primarily a technology problem. Bangladesh can acquire the technical capacity to detect synthetic media; it can hire and train officers in cybercrime investigation; it can negotiate platform-cooperation agreements through bilateral diplomacy. What it cannot purchase is the institutional willingness — the political and administrative prioritisation — to treat violence against women in digital spaces as seriously as violence against women in physical spaces. The evidence from this paper, and from every week’s reporting before it, is that the current level of prioritisation is not sufficient. The most useful thing anyone reading this who has access to the ICT Division, the Ministry of Home Affairs, the parliamentary committee overseeing cybercrime law, or the ICT courts can do is to make this the top-line demand of every AI governance conversation until the enforcement infrastructure actually exists.

Looking out to 2028–2030, the trajectory is unforgiving. AI-generated image abuse will get cheaper, faster, and more convincing on a quarterly basis. The generation of tools available to a bad actor today is already beyond what most legal systems can respond to reliably. The generation available in 2028 will be further beyond still. The question is not whether Bangladesh can outpace the technology — it cannot, and neither can any country. The question is whether it can build the enforcement infrastructure, the legal clarity, and the platform-cooperation frameworks that create meaningful deterrence even against threats that are technically very advanced. The IJRSI paper is a clear-eyed description of how much of the available window has already passed. The response it warrants is not more drafting. It is more doing.

Wrapping up

Ten stories in seven days. Globally: the US government deployed an AI kill switch for the first time in history on June 13, ordering Anthropic to cut off all foreign nationals from Fable 5 and Mythos 5 and triggering the most concentrated global conversation about AI sovereignty since the field began — Europe sounded the alarm, Canada acknowledged its weakness, and Mistral quietly positioned itself as the alternative that no foreign government can switch off; SpaceX acquired Cursor for $60 billion on June 16 in an all-stock deal days after its IPO, further consolidating the AI coding market into a small number of giant platform players where independence is the first casualty of scale; the 52nd G7 summit in Évian-les-Bains on June 15–17 brought Sam Altman, Dario Amodei, and Demis Hassabis to the table with world leaders for the first time in the summit’s history, producing no binding agreements but cementing AI as a head-of-government-level governance priority; and Anthropic, even as the export-control storm swirled, opened its Seoul office on June 17–18 and announced enterprise deployments at NAVER, Nexon, Samsung SDS, and LG CNS — demonstrating the gap between the US government’s security concerns and the commercial reality of global AI adoption in a single week.

At home: the Anthropic ban arrived in Bangladesh like a proof-of-concept for every theoretical argument the National AI Policy makes about digital sovereignty — the country’s AI infrastructure is almost entirely foreign-held, runs on foreign terms, and can be switched off by a foreign government decision, and June 13 made that abstract risk concrete and personal for tens of thousands of users overnight; Bangladesh’s eight-member team competed in the first Asia-Pacific AI Olympiad from Dhaka on the same June 13, a reminder that the most durable form of AI sovereignty is trained human minds that no export control can reach; Russia agreed to send an AI delegation to Bangladesh focused on agriculture and health, adding a new and strategically significant dimension to a partnership portfolio that has so far been almost entirely Western-facing; the government announced that work will begin in July 2026 on the country’s first Digital Public Infrastructure — one digital identity, one digital wallet per citizen — the foundational plumbing on which AI-powered public services will eventually run; and a new IJRSI academic paper documented the legal response gap in Bangladesh’s deepfake-driven gender violence crisis, finding that 65 per cent of respondents believe existing laws are inadequate and that successful prosecutions remain vanishingly rare relative to the documented scale of the harm.

If there is one thread running through both halves of this week, it is this: ownership matters more than access. Access to Fable 5 was available to Bangladeshis on June 12. It was not available on June 13. The models did not change; the ownership structure — American company, American government authority — had always been what it was, and this week it was exercised. The lesson is not that foreign tools are bad or that Bangladesh should retreat into technological self-sufficiency it cannot afford. The lesson is that access and ownership are different things, and building an economy on borrowed access is building on ground that can shift. The same lesson runs through the DPI story — a citizen who owns her digital identity is safer than a citizen whose identity is held for her — through the deepfake story — a woman who can enforce her right to be free of image-based violence is safer than a woman who can only hope the law catches up — and through the olympiad story — a country that owns its own trained minds is richer than one that trains them and then watches them emigrate. Ownership — of infrastructure, of rights, of talent — is the thing worth building, slowly and unglamorously, from now until the next time the ground shifts.

If something in this post caught your interest, share it with one person who does not follow tech. The conversation only spreads if we keep starting it.

See you next week. Until then, take care.