Weekly AI Roundup: What Happened in AI This Past Week (May 16–22, 2026)

Come in, sit down, kick the slippers off. The kettle is on and the ceiling fan is doing its best against another warm Dhaka evening. Another seven days of AI news has gone by, and even by the standards of this very loud year it was a heavy week. Two of the largest companies in the world signed a forty-five billion dollar compute contract. The most famous AI educator on the internet changed teams. Google ran its biggest developer conference of the year. The largest pure-play AI company on earth started preparing the paperwork to go public. And eight thousand people at Meta found out their jobs would be ending so the company could spend that money on machines instead.

Like the last few weeks I have split the roundup into two parts. Five big global stories first — the ones that move money, shape policy, and reshape who has leverage over whom. Then five Bangladesh stories that show how the same wave is breaking on our own shore, sometimes with quiet ambition and sometimes with awkward catch-up. I will keep asking the same two questions on every story: who actually benefits, and who actually pays. And then, where it makes sense: what does this look like from Dhaka, Chattogram, or Sylhet rather than from Mountain View or Wall Street. Pour yourself something warm and let us begin.

5 Big International Developments

1. Anthropic agreed to pay SpaceX $45 billion for compute — at $1.25 billion every month

On Wednesday, May 20, SpaceX disclosed in IPO-related filings the terms of what is now the single largest compute contract ever signed by an AI lab. Anthropic will pay SpaceX nearly $45 billion over three years for access to its xAI-merged Colossus data-centre complex, working out to $1.25 billion every month through May 2029, with the meter starting on a reduced rate in May and June 2026 as capacity ramps. The deal gives Anthropic access to more than 300 megawatts of compute and over 220,000 NVIDIA GPUs, spanning the original Colossus 1 and the next-generation Colossus 2 site built around NVIDIA’s GB200 hardware. Either side can walk away with 90 days’ notice, but nobody expects them to. Anthropic confirmed that the new capacity will go directly into Claude Pro and Claude Max, and has already doubled the five-hour rate limits on Claude Code. Almost as a footnote, a separate piece of reporting on May 21 confirmed that Anthropic is also in early talks with Microsoft about renting time on its custom Maia 200 AI chips, looking to diversify beyond its $100 billion AWS Trainium arrangement, its existing Google TPU agreement, and now this xAI relationship.

Pause on the scale for a second. $1.25 billion a month is more than the entire annual revenue of most listed companies in South Asia. It is roughly a third of Anthropic’s projected 2026 annualised revenue, going out the door every year just to one supplier. The winners are obvious if you squint at the cap tables. SpaceX collects 76 per cent of its Q1 2026 capital expenditure back from this single contract, which makes Anthropic effectively the anchor tenant funding Elon Musk’s data-centre build-out ahead of SpaceX’s NASDAQ debut on June 12. NVIDIA wins twice — once on the GPUs it sold to build Colossus, again because every additional dollar Anthropic spends on Claude proves that frontier AI workloads still need NVIDIA silicon. The losers are quieter: Anthropic’s own future margins (every dollar of compute is a dollar that is not profit), the AWS and Google relationships that suddenly look smaller, and the broader ecosystem of mid-tier cloud providers who hoped to capture some of the AI training spend but never had the physical capacity to do so.

There is a darker reading too. Anthropic and xAI are direct competitors in the model layer, and Anthropic is now the largest single customer of xAI’s parent company. That is the AI industry’s first real example of vertical entanglement at the frontier: two labs that disagree about almost everything technically and politically, locked together by the physical reality that one of them owns the only data centre big enough for the other to use. Imagine if Pepsi rented its entire bottling capacity from Coca-Cola, and you get a sense of how strange this is. CEO Dario Amodei has been candid that Anthropic has had “difficulties with compute”. The Microsoft Maia talks are an attempt to keep the supplier list diverse before that entanglement becomes a chokehold.

For Bangladesh, the relevance is less direct but still real. The country’s banking, mobile financial services, and BPO sectors are all increasingly dependent on Claude and ChatGPT for back-office automation. When Anthropic doubles Claude Code’s rate limits in San Francisco, the impact is felt within days in the bKash engineering team in Gulshan, the IT cell at Square Pharmaceuticals, and the dozen software houses in Banani that have quietly standardised on Claude as their pair-programming partner. The bad news is that compute scarcity in California will continue to be priced into the API rates that Bangladeshi developers pay in dollars. The country has no domestic substitute. The good news is that the underlying capability of Claude is going to keep improving at a faster rate than most users can adapt to — and the Bangladeshi developer who learns to wield the latest model well has a productivity multiplier that did not exist twelve months ago.

The longer-term takeaway is that the AI industry is moving past the era when capital was the constraint, and into an era where physical infrastructure — land, power, water, GPUs, fibre — is the constraint. Countries that can build data centres at scale, with reliable power and good network peering, become geopolitically important in a way they were not before. Bangladesh has the population and the bandwidth ambition, but the power grid is the binding constraint, and nobody in the country’s energy policy conversation has yet started talking about AI as the next big load on the system. They will have to soon. The conversation Anthropic is having with SpaceX this month is the conversation Dhaka will be having with somebody in 2029.

2. Google I/O brought Gemini 3.5 Flash, Antigravity 2.0, and a full pivot to agents

Google held its annual I/O developer conference on May 19 and 20, and as expected it was dominated by AI. The headline release was Gemini 3.5 Flash, a faster and dramatically cheaper Flash-tier model priced at $1.50 input / $9.00 output per million tokens — less than half the cost of comparable frontier models. The honest surprise is that Flash now beats Gemini 3.1 Pro on coding and agentic benchmarks (76.2 per cent on Terminal-Bench 2.1 versus Pro’s 70.3 per cent) while running roughly four times faster. Google’s framing was unmistakable: the next phase of useful AI is not better chatbots, it is better agents, and Flash is engineered specifically for the long-horizon, multi-step work that agentic systems need to do at scale.

Alongside the model release, Google launched Antigravity 2.0, a standalone desktop application designed as a control plane for running multiple agents in parallel — you can orchestrate sub-agents, schedule background tasks, and integrate across Google AI Studio, Android, and Firebase. The Gemini API now exposes a Managed Agents service that lets developers run custom agents inside secure Google-hosted environments without managing their own infrastructure. Gemini Omni, the unified multimodal model leaked the week before, also showed up — Google demoed video remixing, in-chat editing, and a single pipeline for cross-modal generation. And in a quietly important consumer move, native Android vibe coding arrived in AI Studio, letting you describe an Android app in plain English and get a working prototype.

The socioeconomic story here is bigger than any single product. Google has been the AI industry’s most patient operator — for two years it watched OpenAI and Anthropic capture the developer mindshare, and it spent that time quietly building the infrastructure (TPUs), the distribution (Search, Android, Chrome, Workspace), and the model line (Gemini 1, 2, 3, 3.5). What I/O 2026 made clear is that Google is now ready to compete on price as well as capability. Flash at $1.50/$9.00 per million tokens is a deliberate undercut of OpenAI’s GPT-5 and Anthropic’s Claude Opus pricing, and because Google owns the TPU stack end to end, it can afford to keep prices low in a way the others cannot. The winners are developers and agentic-app startups, who suddenly get a frontier-class model at a price that lets them run agents continuously rather than in short bursts. The losers are the API resellers and middleware platforms whose business depended on Gemini being a fast follower rather than a price leader.

For Bangladesh, the price cut matters disproportionately. Local software houses and freelance AI builders price their work in dollars to compete on Upwork and Fiverr, but they pay their developers in taka. Cutting the per-token cost of agentic work in half is the difference between an AI-powered automation tool being viable as a side project and being viable as a company. The dozens of small Bangladeshi teams that have spent the last year building Bangla-language summarisers, exam-prep chatbots, and small-business assistants now have the economics to actually charge customers a few hundred taka a month and turn a profit. Until this week, most of them were running margin-negative on inference cost. That changes today.

The deeper trend is that agentic AI — software that takes actions on your behalf, rather than just answering questions — is moving from research demo to production product across the entire industry simultaneously. Google’s Antigravity 2.0, OpenAI’s Deployment Company from two weeks ago, Anthropic’s expanding Claude Skills marketplace, even the Cursor coding agent that just crossed $3 billion in ARR — they are all converging on the same thesis. The next two years will not be about who has the smartest model in a benchmark. They will be about who has the most reliable agentic stack in real businesses. For countries with deep services-export economies, the question is whether the local industry rebuilds itself around agentic delivery before global vendors do it without them. The window is not large.

3. Andrej Karpathy joined Anthropic — and the symbolism cuts deep

On Tuesday, May 19, Andrej Karpathy — co-founder of OpenAI, former director of Tesla Autopilot, founder of Eureka Labs, and the AI world’s single most-followed teacher — announced he is joining Anthropic. He will work on pre-training under team lead Nick Joseph, leading a new effort to use Claude itself to accelerate pre-training research. Karpathy’s personal note on X was characteristically understated: “I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D.” He added that he remains passionate about education and plans to resume his Eureka Labs work in time.

If you are not deep in the AI field, it is hard to convey how much this hire matters as a signal. Karpathy is not a typical research hire. He is the person whose YouTube lectures taught a generation of AI practitioners how transformers actually work, whose nanoGPT and micrograd implementations are the canonical learning materials in the field, and whose career has tracked the actual centre of gravity of frontier AI for fifteen years. He helped found OpenAI in 2015. He ran Tesla AI from 2017 to 2022. He returned to OpenAI in 2023 for a year. He left to build an AI-first education company. And now, in May 2026, he is back inside a frontier lab — and the lab he chose is Anthropic, not OpenAI.

The economic stakes of senior research talent at the frontier are genuinely outrageous. Top-tier AI researchers at Anthropic, OpenAI, Google DeepMind, and xAI are now compensated at levels that previously existed only for hedge fund partners and Premier League footballers — total packages of $10 million to $50 million a year for the very top names, with equity stakes that can run into nine figures if the company exits. The winners of this market are obviously the researchers themselves and the talent agencies (yes, that is now a category) that broker the deals. The quieter winner is Anthropic specifically, because Karpathy joining is the kind of move that signals to every other senior researcher in the field which way the wind is blowing. The loser is OpenAI, whose narrative as the de facto home of AI talent is taking yet another hit in a year of senior departures.

For Bangladesh and the broader global south, Karpathy’s career is one of the most-watched proof points that AI research can be done at the highest level by individuals who built their reputations on open teaching. His online courses have probably introduced more Bangladeshi students to deep learning than every domestic university course combined. The honest news is that the path he walked — undergraduate at Toronto, PhD at Stanford, then frontier industry — is essentially unavailable to a Bangladeshi student today without an extraordinary scholarship and a lucky visa. But the path he opened, where a kid with a laptop and an internet connection can learn the same material the frontier labs use to train their own engineers, is open. The next Karpathy may well be a 19-year-old in Sylhet right now, watching his nanoGPT walkthrough at 2x speed. Whether that kid eventually works in Dhaka or in San Francisco depends on choices the Bangladeshi research and venture ecosystem makes in the next five years. The talent is here. The infrastructure to keep it is not yet.

The broader point about senior researcher mobility is that the AI industry is concentrating. There are perhaps six labs in the world that matter at the frontier — OpenAI, Anthropic, Google DeepMind, xAI, Meta, and a small handful in China — and the senior talent rotates between them with the frequency that movie stars used to rotate between studios. That concentration is, on the one hand, terrifying for anyone who worries about democratic governance of powerful technology. On the other, it is the most efficient way humanity has yet found to push the science forward quickly. Karpathy joining Anthropic does both things at once: it accelerates the frontier, and it concentrates the people who shape the frontier into an even smaller circle.

4. OpenAI is racing to file for IPO — and Musk lost his lawsuit just in time

The most-watched corporate manoeuvre of the week was OpenAI. On Wednesday, May 20, multiple outlets confirmed that OpenAI is preparing to confidentially file a draft IPO prospectus, possibly as soon as Friday, May 22. The company is working with Goldman Sachs and Morgan Stanley on a deal that could value it at up to $1 trillion at debut, against an $852 billion private valuation from its last round. CEO Sam Altman reportedly wants the company ready to list by September. The competitive pressure is precise: SpaceX, which merged with xAI earlier this year and is now valued at roughly $1.25 trillion, is set to disclose its own IPO prospectus on Wednesday, May 21 and debut on the NASDAQ on June 12. If you want to think about it as a story, this is the act break: the two founders who started OpenAI together in 2015, and have spent the past two years insulting each other in court and on social media, are now sprinting to be the first public AI company in history.

The week made that sprint easier for Altman in one important respect. On Monday, May 18, a federal jury in California threw out Elon Musk’s lawsuit against OpenAI and Altman, ruling in less than two hours that Musk had waited too long to sue. Musk had been seeking up to $134 billion in damages and Altman’s removal from OpenAI, arguing that the company had abandoned its founding non-profit mission. The jury found him barred by the statute of limitations — three years for the breach-of-charitable-trust claim, two years for the unjust enrichment claim — and dismissed every charge. Musk announced an immediate appeal on X, calling the verdict a “calendar technicality.” But on paper, OpenAI walked into IPO week without the legal cloud that had been hanging over it for two years.

The socioeconomic implications of OpenAI’s IPO are larger than any other tech listing in living memory. A $1 trillion debut would make OpenAI the third or fourth most valuable company in the world on day one, alongside NVIDIA, Microsoft, and Apple. The winners are the original employees and investors — including Microsoft, which has been the company’s largest backer for six years and stands to realise tens of billions of dollars in paper gains. The structural winners are the global investment banks taking the fees (Goldman and Morgan Stanley split the lead, but the syndicate will include almost every major institution). The losers are subtler. They include every late-stage private investor in the AI ecosystem who paid valuations on the assumption that OpenAI would stay private and disciplined; they will now find OpenAI competing for capital not against Anthropic or Cursor, but against the entire S&P 500. They also include OpenAI’s own product strategy: a public company faces quarterly earnings pressure that a private one does not, and the trade-offs between safety, capability, and revenue look different when you are reporting to retail shareholders.

For Bangladesh, the IPO has two practical consequences worth flagging. First, OpenAI becomes accessible to retail investors here through international brokerage platforms, and the conversations in Gulshan over the next twelve months will inevitably include questions about whether to buy OpenAI stock. The honest advice is: be very careful. AI valuations at IPO are priced for outcomes that may or may not materialise, and the early public-market trajectory of high-profile tech IPOs is rarely kind to retail latecomers. The second, more interesting consequence is that OpenAI’s prospectus will, for the first time, force public disclosure of how much money the company actually makes, how much it loses, and what fraction of its revenue comes from which segments. That disclosure will be the most useful document the AI industry has ever produced, and it will give every CFO, regulator, and policy researcher in the world a baseline for understanding the economics of frontier AI. Bangladeshi institutions that want to plan for the next decade should read it carefully when it lands.

The deeper story is that AI is now a public-markets phenomenon. The era of frontier AI being a small handful of private companies funded by a small handful of VCs is ending. The era of frontier AI being a publicly traded sector, with all the volatility, scrutiny, and short-termism that implies, is beginning. The trillion-dollar question for the next decade is whether public ownership disciplines AI development or distorts it. The answer matters everywhere, including in Dhaka.

5. Meta is laying off 8,000 — and quietly redeploying 7,000 more into four new AI units

On Wednesday, May 20, Meta executed the layoff that has been telegraphed for months. Around 8,000 employees — roughly 10 per cent of the workforce — were let go across Reality Labs, the Facebook social division, recruiting, sales, and global operations. Another 6,000 open requisitions were cancelled, bringing the effective headcount reduction to 14,000 positions. What was newer, and quieter, was the move that preceded the cuts by two days: on Monday, May 18, Meta head of HR Janelle Gale informed about 7,000 existing employees that they were being reassigned into four new AI-focused organisations — Applied AI Engineering, an Agent Transformation Accelerator, Central Analytics, and an enterprise-solutions unit. The new orgs will use what Meta calls “AI native design structures” and run with fewer managers per employee than the rest of the company. The reshuffle is part of Alexandr Wang’s Superintelligence Labs consolidation that has been running since late 2025.

Take a step back and the shape of this is striking. In a single week, Meta has both shrunk by 14,000 positions and redirected 7,000 of its survivors into a brand new organisational geometry that did not exist on Friday afternoon. That is a faster organisational reshape than most large companies attempt in a decade, and it tells you that Mark Zuckerberg and Wang have decided the cost of moving slowly exceeds the cost of moving messily. The winners inside Meta are obvious — the engineers and PMs who landed inside the four new orgs are now sitting in the parts of Meta that get the AI capex dollars. The losers are the 8,000 who got the email. The strategic loser is the broader open-source AI ecosystem that depended on Meta’s Llama series; Llama-related work has been wound down in favour of closed-source Superintelligence Labs models, and the rest of the field is still adjusting.

The international socioeconomic story has not changed much from last week, but the cumulative arithmetic now reads: roughly 20,000 jobs cut at Meta and Microsoft so far in 2026 directly attributable to AI restructuring. That is no longer a forecast. It is one full quarter of layoffs, in two of the largest employers in the technology industry, executed precisely so that the freed payroll budget can be redirected to data centres and chip purchases. The capex math is brutal. Meta’s projected 2026 AI capital expenditure runs between $115 billion and $135 billion. NVIDIA collects perhaps a third of that. The cooling and power industries collect another large chunk. The 14,000 displaced workers collect, in aggregate, perhaps two billion dollars of severance — meaningful for them, a rounding error against the spend.

For Bangladesh, the Meta restructuring matters in two specific ways. First, the continuing pullback from Llama as an open-weights brand is bad news for the local research community that built its early Bangla-language AI work on Llama 3 and 3.1. The pragmatic migration path is to Chinese open models — Qwen, DeepSeek, the GLM series — and to a lesser extent to Mistral. Each of those introduces different latency, hosting, and political-risk trade-offs that local teams will have to navigate. The second, more uncomfortable consequence is that Meta’s reorganisation includes a quiet expansion of trust and safety contracting, much of which historically flowed through India and the Philippines. Bangladesh has tried for years to capture a slice of that contracting pipeline and has been blocked mostly by language coverage and contracting infrastructure. The next 18 months are the window in which Dhaka either builds a serious export-grade trust-and-safety industry on the back of its 165 million Bangla speakers, or accepts that the contracts will go elsewhere.

Looking further out, the deeper pattern across the past three weeks of layoff announcements — Meta this week, Microsoft last month, smaller cuts at Google and Amazon — is that the AI industry has moved firmly into the labour-substitution phase in the West. The argument is no longer whether AI will replace certain kinds of knowledge work; it is which kinds, and how fast. For countries like Bangladesh, where labour is much cheaper, the substitution math is different and the timeline is longer — but the direction is the same. The strategic question is whether the country can build the higher-value AI deployment and orchestration roles fast enough to offset the entry-level coding and data-entry jobs that will get squeezed. The honest answer is that nobody knows yet, and the people who claim to know are usually selling something.

5 Stories That Hit Closer to Home (Bangladesh)

1. The 3rd Bangladesh AI Olympiad wrapped up its national round at BUBT

Last week we previewed it; this week it happened. On Saturday, May 16, Bangladesh University of Business and Technology (BUBT) in Mirpur hosted the national round of the 3rd Bangladesh AI Olympiad 2026, and the country has its champions. In the AI Coding Challenge, Labib Shahriar of Homna Adarsha High School took the “Champion of Champions” trophy, going home with a 3-in-1 Jadu PC alongside the title. Sayed Mohammad Rafan of Cantonment English School and College, Chattogram won “Best of the Best” in the AI Quiz category. Nuzhat Binte Nazim of Ideal School and College picked up the best female participant award in the coding challenge. In total, 14 quiz finalists and 15 coding finalists were declared national winners. The top performers will attend a national camp from May 18 to 20, from which the Bangladesh team for the Asia Pacific AI Olympiad and the International AI Olympiad in Kazakhstan this August will be selected. The Bangladesh Open Source Network ran the event, with BUBT as the platinum sponsor and the Kaggle data-science platform hosting the technical infrastructure.

Step back from the prize-list and consider what the Olympiad really represents. The country has somewhere around 56 million people under the age of 25. The number of those young people who could plausibly hold their own in a Kaggle-graded AI competition is probably in the low six figures — a small fraction of the total, but a meaningfully large absolute number compared to almost any other South Asian country. The Olympiad is one of the very few national infrastructures that exist to find those kids, validate them, and put them in a room with each other. Every other talent-discovery pipeline in Bangladesh — IBA admissions, engineering university competitions, the BCS — selects for skills that have very little to do with AI capability. The AI Olympiad is, for the moment, the only mechanism producing a national leaderboard of teenage technical talent in this specific domain.

The international socioeconomic question is what happens to these kids after they win. The honest pattern over the past decade is that Bangladesh’s most talented technical young people are recruited, with full scholarships, by universities in the US, Canada, Germany, Singapore, and (increasingly) South Korea, by about age 19. Many of them stay abroad. The Bangladeshi state captures very little of the value they produce, beyond the remittances they sometimes send home. The win for Bangladesh is real even when those students leave, but the country is paying for the upbringing and capturing only the residual. The structural fix is to make staying in Bangladesh competitive, which requires the local AI industry to be deep enough by 2030 that staying is plausible rather than a sacrifice. Right now it is not, and that is a problem of capital, of policy, and of patience.

For Bangladeshi families, the more immediate signal is the one being sent to parents. Twenty years ago the aspirational path was medicine, then engineering, then BCS. Five years from now it will be AI, and many middle-class families in Dhanmondi, Bashundhara, and Uttara are already shifting their kids’ tuition load toward maths and coding rather than memorisation. The Olympiad makes that shift visible and legitimate. The serious risk is that the shift remains concentrated — in elite English-medium schools and in a handful of regional cities — while the rural and Bangla-medium school networks get further behind than they were before AI mattered. A serious AI policy treats that widening as one of the country’s most important education questions for the next decade. Otherwise, the demographic dividend everyone keeps forecasting will accrue to a small slice of the population rather than to the country.

If you know a school-age student who lights up when they see code, this is the kind of event that can change a life — not because of the trophy, but because of the realisation that they can compete on the same global platform that working AI researchers use, and they can hold their own. The kids going to Kazakhstan in August are not abstract. They are 15-year-olds from Cumilla and Chattogram and Dhaka. That is a real story, and it is happening this year.

2. Bangladesh’s government laid out a serious AI and semiconductor industrial plan — at last

On Sunday, May 17, at the World Telecommunication and Information Society Day event in Dhaka, Posts, Telecommunications and Information Technology Minister Faqir Mahbub Anam delivered the most ambitious technology-policy speech the country has heard in years. The framing was specific. Bangladesh, he said, “will not only be used as a technology consumer, but will create a visible position in technology production, service delivery and exports through AI, hardware, semiconductors, software, BPO, server security, data centres, startups and digital services.” He invoked the slogans “Made in Bangladesh,” “Assembly in Bangladesh,” and “Serviced in Bangladesh” and tied them to a concrete set of election commitments: AI hardware development hubs, AI-driven data centres, one million new technology jobs, strengthened cyber security infrastructure, and a technology-based cashless economy. Prime Minister’s adviser Rehan Asif Asad added a number to the rhetoric: improved broadband and AI-powered systems could contribute 10 to 15 per cent of GDP growth within five years. The medium-term target is for Bangladesh to be among the top 20 digital economies within five years, and the top 10 within fifteen.

Speeches like this have happened before in Bangladesh and not always landed. The Awami League’s “Digital Bangladesh” framing did real work between 2009 and 2024, but it bent toward connectivity rather than production, and by the end was less about new industry and more about old habits. The interim government’s framing this week is genuinely different in two ways. First, it explicitly names AI, semiconductors, and hardware as production targets — categories that Digital Bangladesh largely avoided because the country’s industrial base was not ready. Second, it ties production to export, which is the discipline that distinguishes a real industrial policy from a domestic-consumption strategy. The reason this matters is that export-discipline forces honesty: you cannot fake your way into a semiconductor packaging contract with TSMC the way you can fake your way into a procurement scheme inside the BCS.

The socioeconomic stakes are unusually direct. Bangladesh’s existing technology export base is roughly $1.9 billion a year in services, dominated by software outsourcing and BPO. The minister’s framing is implicitly proposing to multiply that base — perhaps by ten — over the next decade by adding hardware assembly, chip design services, and data-centre services to the mix. If even half of that materialises, the one million new technology jobs become real, and they would absorb a meaningful slice of the half-million young people who enter the Bangladeshi labour market each year. The losers, if there are any in this picture, are the consumer-imports industry that depends on Bangladesh being a net buyer of foreign electronics, and the small cohort of incumbents who have done well from the existing arrangement of soft-budget software contracts with government.

For the broader Bangladeshi society, the more interesting subtext is the implicit acknowledgment that the country’s current export model — ready-made garments and remittance — is reaching its structural ceiling. Garments are roughly $48 billion of annual exports, against a global apparel market that is no longer growing fast in real terms, and a workforce competing increasingly with automation in Vietnam, Türkiye, and (paradoxically) the US itself. Remittances are roughly $26 billion, sensitive to Gulf migration policy and to AI substitution in the labour-receiving countries. Neither of those numbers is going to do for Bangladesh in 2040 what they did in 2010. The technology bet is the explicit plan B. Whether it works depends on whether the next two budgets actually fund the hardware hubs and the data centres, or whether the announcement gets reduced to a few photo opportunities and a website refresh. The international examples worth watching are Vietnam, which converted its 2010 industrial-policy announcement into real chip-packaging contracts by 2022, and India, which has so far converted its larger ambitions into uneven execution. Bangladesh sits closer to the Vietnamese starting point than the Indian one — smaller, more focused, and with less institutional inertia to fight.

Looking out to 2030, the test is whether the National AI Policy 2026–2030 (currently in late-stage drafting at the ICT Division) and the National AI Centre of Excellence (NAICE) that is meant to coordinate implementation actually get the staffing, the budget, and the political cover they need to operate independently of the political cycle. The speech this week was the easy part. The hard part starts on the next budget day.

3. Bangladesh’s first-ever semiconductor delegation went to South Korea — and signed a real deal

While the minister was speaking in Dhaka, twelve people from the Bangladesh Semiconductor Industry Association (BSIA), led by association president M A Jabbar, were carrying the same message in person to Seoul. The South Korea Roadshow 2026, the first-ever Bangladeshi semiconductor delegation to visit Korea, ran from May 11 to 14. Over three days the delegation met KAIST (Korea Advanced Institute of Science and Technology), SK Hynix, Hana Micron, the Korea Semiconductor Industry Association (KSIA), Koolmicro, Y-Tech, and Growith Associates, with discussions focused on chip packaging, AI hardware, design and testing, and emerging-technology commercialisation. The major outcome was the signing of a Letter of Intent between BSIA, the KAIST Global Commercialization Center, and the Center of Research Excellence on Semiconductor Technology (CREST) — a framework for joint research, talent development, and capacity building in semiconductors. For the Bangladeshi side, this is the first formal research-collaboration arrangement with a major Korean technical institution in this sector. SK Hynix, in particular, is the world’s second-largest memory chip maker and one of the two suppliers (alongside Micron) of the HBM stacks that sit inside every NVIDIA AI GPU. Having a name like that in the room is symbolic, and over time can become material.

The socioeconomic frame for this story is genuinely larger than the trip itself. The global semiconductor market crossed $630 billion in 2024 and is forecast to reach $1 trillion by 2030, driven by AI, electric vehicles, 5G, and industrial automation. Bangladesh’s current share of that market is essentially zero — the BSIA’s stated target is to reach $1 billion in semiconductor exports by 2030, up from about $8 million today, which is a 125x increase and would still leave the country at well under one per cent of the global market. The strategy implicit in the Korean trip is the same one Vietnam, Malaysia, and the Philippines used: do not try to manufacture leading-edge logic chips (which requires tens of billions of dollars of capex and a decade of know-how), but instead specialise in outsourced semiconductor design services, advanced packaging, and AI hardware testing. Those are the parts of the value chain where talent and operational discipline matter more than capital, and where Bangladesh’s labour-cost structure is genuinely competitive.

The winners of a successful semiconductor pivot would be a generation of Bangladeshi electrical and computer engineers, who currently graduate from BUET, CUET, KUET, IUT, and BRAC University into a domestic labour market that does not really value their VLSI training. The losers, in a narrow sense, are India and Vietnam — both of which have publicly stated semiconductor ambitions of their own, and both of which would prefer Bangladesh not to compete. The political risk is that Bangladesh is roughly a decade late to the regional semiconductor conversation. Vietnam already has Intel, Samsung, and Amkor packaging facilities. Malaysia has been doing back-end semiconductor work for fifty years. India just spent $10 billion of subsidies to attract Micron, Tata, and others. Bangladesh has a small handful of design-services startups and almost no fabrication footprint. Catching up requires not just policy ambition but capital — sovereign-fund scale capital — that the country does not currently have.

For the broader Bangladeshi public, the Korea trip is the kind of story that does not show up on the evening news but that quietly determines what the country’s middle-class kids will be doing for work in 2035. If the LOI with KAIST turns into a real research pipeline within three years, the country gets a foothold in one of the most strategically important industries of the century. If it stays a press release, Bangladesh remains a downstream consumer of chips designed and packaged elsewhere, paying full sticker price in foreign exchange every time it imports a phone, a laptop, or a server. The difference between those two outcomes is the seriousness with which the next government treats the Bangladesh Semiconductor Industry Association’s policy asks — specifically the requested 25 per cent cash incentive on semiconductor service exports, the 12-year corporate tax holiday, and the duty-free import of fabrication and testing equipment. Without those, the LOI is just a piece of paper. With them, it is the beginning of an industry.

4. Entrepreneurs are quietly running AI workshops for each other — and 300 turned up this week

On Sunday, May 17, the same day the minister was speaking in Dhaka and the BSIA delegation was wrapping up in Seoul, Entrepreneurs’ Organization (EO) Bangladesh hosted a workshop at Le Méridien Dhaka titled “Learning Presents Leading with AI: Innovation, Strategy & Growth.” Around 300 entrepreneurs, corporate executives, and technology professionals turned up. The keynote ran live demonstrations of the Claude assistant platform for the four functions that small and mid-cap businesses spend most of their time on: finance, operations, marketing, and recruitment. A week earlier, on May 8, the University of Dhaka had hosted a parallel workshop — “AI for Professionals: Future Prospects and Challenges” — jointly organised by Bizz Solutions PLC and the Chartered Technical Training Institute, with keynote speaker Md. Shafiqul Alam (a fellow chartered accountant) arguing that “AI should become a mirror of human intelligence, not a substitute for human wisdom.” Across these two events alone, somewhere north of 500 working professionals received structured, hands-on AI training in the same week. That number is small in absolute terms, but it represents a sharp uptick from anything the previous twelve months produced.

There is something important happening underneath the bare attendance numbers. The Bangladeshi corporate community has gone, in roughly eighteen months, from treating AI as a foreign curiosity to treating it as a practical line item in operating plans. The EO Bangladesh workshop is not a government event, not a university programme, and not a vendor pitch — it is the country’s own entrepreneur community pooling resources to teach itself. That self-organised pattern is how every important shift in Bangladeshi business culture has actually happened, from the early garments factories in the 1980s to the mobile financial services build-out in the 2010s. When 300 owners and CFOs of mid-cap companies show up to learn how to use Claude for invoice processing, you are watching the demand-side ignition of an AI services industry in real time.

The international socioeconomic frame is that the AI services and consulting market — companies that help other companies actually deploy AI — is on track to be a $300 billion industry globally by 2028, according to most analyst forecasts. The winners in that market are the firms that get there first with credible local expertise. In the US and Europe, the major consulting firms (Accenture, Deloitte, McKinsey, BCG) have built dedicated AI deployment practices and are charging $500-plus per hour for the work. Anthropic’s Deployment Company from two weeks ago is a frontier lab going direct into that same market. The losers, globally, are the IT services firms that built their books on traditional software integration; they are being squeezed both from above (by AI-native consultancies) and from below (by tools like Claude for Small Business that let companies skip the consultant entirely).

For Bangladesh, the EO workshop and the Dhaka University event together point to a window of opportunity that is unusual in the country’s economic history. The local technology services export sector is currently dominated by software outsourcing and BPO — work that is steadily being eaten by AI in the very countries that buy from Bangladesh. The pivot the industry needs to make is from “we will write your code for cheaper” to “we will deploy and operate your AI for cheaper.” That pivot is harder, requires senior talent rather than junior, and demands deep industry knowledge of the customer’s business. But the entry-level economics are good: a Bangladeshi AI deployment engineer who can run a Claude or Gemini-based automation project end to end is worth perhaps $60,000 to $100,000 a year to a Western client, versus the $15,000 to $25,000 that a Bangladeshi software developer typically commands. The same workforce moving up the value chain is a national income story, not just a sectoral one. The constraint is talent at the senior end, and the workshop circuit this week is one of the first credible local responses to that constraint.

The risk, as always in Bangladesh, is that the workshop circuit becomes a substitute for industry, rather than a catalyst for it. The country has a long history of producing trained but unemployed graduates — the engineering colleges produce roughly 50,000 a year, against meaningful jobs for perhaps a quarter of them. The AI-services pivot only works if the workshops are tied to actual contracts, real client relationships, and operating businesses. The next twelve months will reveal whether the EO Bangladesh cohort produces a wave of new agentic-AI deployment firms exporting to the West, or whether it produces another wave of LinkedIn certifications and not much else. The signal to watch is whether any of the 300 attendees this week launch credible companies by next spring.

5. A serious essay called Bangladesh’s digital economy “a floor giving way” — and the GPU numbers are brutal

The single most-discussed piece of writing in the Bangladeshi technology community this week was an opinion essay published in The Business Standard with a striking title: “Bangladesh is building a digital economy on a floor that is giving way.” Its central observation was simple and uncomfortable. Stanford University’s 2026 AI Index — the annual audit that everyone serious about AI policy reads — does not mention Bangladesh once. For a country that has staked significant parts of its economic future on digital services, freelancing exports, and a young, tech-capable workforce, the absence is a problem. The essay then walked through the underlying numbers. Between 2018 and 2025, Europe expanded its state-backed AI supercomputing capacity from 3 clusters to 44. North America reached 41. China built 85. South Asia — the entire subcontinent of nearly two billion people — reached a total of 2. Bangladesh, in that picture, has none of its own. The UNESCO–UNDP–EU AI Readiness Assessment published in December 2025 documented “GPU scarcity” in Bangladesh and noted that AI ethics instruction is “nearly absent” from the country’s educational system.

What makes the essay worth dwelling on is that it is not a foreign critic complaining; it is a Bangladeshi commentator writing for a Bangladeshi audience. The argument lands harder because it is internal. And the practical implication is that the country has a yawning gap between the ambition of the National AI Policy 2026–2030 (which explicitly commits to procuring GPUs for shared government use and developing a National AI Compute Strategy) and the operational reality that no Bangladeshi institution today operates a GPU cluster large enough to train a competitive Bangla-language model. The country’s AI researchers either rent compute from AWS, GCP, or Azure — paying full international rates in dollars from limited research budgets — or they downsize their ambitions to fit on a single GPU borrowed from a university lab.

The international socioeconomic frame for compute infrastructure is now uncomfortable for any country that does not have it. The global GPU shortage — driven by limited CoWoS packaging capacity at TSMC and constrained HBM production at SK Hynix — has pushed lead times for data-centre GPUs to 36 to 52 weeks. Hyperscalers with pre-existing relationships (Microsoft, Google, Amazon, Meta, and now xAI/SpaceX) are absorbing essentially all of the available supply. Countries without those relationships — most of the global south — find themselves at the back of an extremely long queue. Even with money in hand, Bangladesh cannot simply place an order for 1,000 H200s and expect them in the data centre this year. The structural lock-out is real. The winners of this dynamic are obvious: NVIDIA, TSMC, SK Hynix, and the handful of countries with sovereign-scale AI compute. The losers, again obvious, are everyone else.

For Bangladesh specifically, the essay’s argument should function as a wake-up call for the next budget cycle. A serious sovereign AI compute commitment — even at a modest scale of, say, 1,000 to 2,000 H200-class GPUs in a shared facility, with electrical and cooling capacity to expand — would cost on the order of $200 to $400 million in capital expenditure, plus a meaningful operating budget. That is a meaningful number against the national budget, but it is small against the cost of not doing it. Without domestic compute, the country cannot fine-tune large models on Bangla data without exporting the data to foreign clouds (a sovereignty problem), cannot train the next generation of local AI researchers without sending them abroad (a talent problem), and cannot offer competitive AI-services exports without paying premium rates for inference (an economics problem). All three problems get worse every year that goes by without action.

The most important practical signal in the essay is the National AI Compute Strategy commitment buried inside the draft AI policy. Whether that strategy gets actual staff, an actual budget, and an actual deadline in the next twelve months will tell us more about Bangladesh’s AI seriousness than any number of ministerial speeches. The honest assessment in May 2026 is that the country has the ambition, the talent, and now the policy text. It does not yet have the infrastructure. Closing that last gap is the work of the next government, the next budget, and the next five years. The clock has been running for a while. It is loud enough now that you can hear it from Dhanmondi.

Wrapping up

Ten stories in seven days. Globally: Anthropic agreed to pay SpaceX nearly $45 billion over three years for compute and started flirting with Microsoft’s Maia chips as a hedge, Google’s I/O kicked off the agentic-AI era in earnest with Gemini 3.5 Flash and Antigravity 2.0, Andrej Karpathy left Eureka Labs to join Anthropic’s pre-training team, OpenAI started racing toward an IPO that could value it at a trillion dollars while a federal jury dismissed Elon Musk’s lawsuit against it in two hours, and Meta cut 8,000 employees on Wednesday while quietly reassigning 7,000 others into four newly minted AI organisations on the same Monday.

At home: the 3rd Bangladesh AI Olympiad crowned its national champions at BUBT and is sending a team to Kazakhstan in August, the interim government’s tech minister laid out the most ambitious AI-and-semiconductor industrial agenda the country has ever heard, a 12-person Bangladeshi delegation signed the country’s first-ever semiconductor research deal with KAIST in Seoul, 300 entrepreneurs at Le Méridien Dhaka spent a Sunday learning how to deploy Claude inside their businesses, and a sharp essay reminded everyone that Bangladesh did not appear in Stanford’s 2026 AI Index even once.

If there is one thread running through both halves of this week, it is that the physical and institutional foundations of AI now matter more than the model layer. The global story is about compute contracts, data centre builds, IPO filings, and 7,000-person reorganisations. The Bangladesh story is about a national AI Olympiad pipeline, a semiconductor research LOI, a workshop circuit teaching working professionals to actually use these tools, and an honest essay arguing that without GPUs the rest of it is theatre. The arguments about which model is smartest have not stopped, but they are no longer the arguments that decide who wins. The arguments that decide who wins are about infrastructure, talent, contracts, and patience. Bangladesh has more of all four than the international AI conversation gives it credit for. Whether the country converts that into a seat at the table over the next five years is the open question.

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. Stay well.