Weekly AI Roundup: What Happened in AI This Past Week (July 11–17, 2026)

Welcome back. The monsoon has been in full voice this week — Dhaka drowned again after an overnight downpour, and Chattogram is doing its usual mid-July arithmetic about which streets will still be passable by evening — and the world’s AI news had a fittingly water-and-silicon character to it. If last week the story was about forums and principles, this week the centre of gravity moved to two very physical things: chips and China. TSMC printed a record quarter on the back of AI demand, the first advanced Nvidia accelerators trickled into China under an American licence, and Shanghai threw open the doors of the world’s largest AI conference alongside a summit on how the technology should be governed. Meanwhile the West’s most anticipated model launch quietly slipped its date, and Europe kept counting down to the day its rules grow teeth.

Two halves as always. Five global, five from home. The same two questions on each — who benefits, who pays — and the same effort to read every story from Dhaka as well as from Shanghai, Hsinchu, or Brussels. And, as ever, a commitment to check each load-bearing fact against its source before repeating it: this week had more than its share of confident rumours dressed up as news, and separating the two is most of the job. Let us begin.

5 Big International Developments

1. The first Nvidia H200 chips reached China under a US export licence

The long, strange saga of American AI chips in China turned another page. On July 14, a senior US Commerce Department official confirmed that a small number of Nvidia’s H200 accelerators had been shipped to customers in China after receiving US export approval. Jeffrey Kessler, the under secretary of commerce for industry and security, said only “very few” had been shipped — “a very small quantity of chips” — and declined to name buyers or quantities; the reporting was carried by CNBC and Bloomberg. (A plain-language note: the H200 is one of Nvidia’s high-end AI training chips — the kind of hardware that does the heavy lifting behind large models. Export controls have for two years governed which versions may be sold to China and in what numbers.) The backdrop is a policy that has zig-zagged repeatedly: exports were halted in April 2025, an earlier China-specific chip (the H20) never actually sold after Beijing discouraged its purchase, and in December 2025 Washington moved to permit H200 shipments — which are the ones now, finally, arriving in small quantities.

The significance is not the volume, which is tiny, but the direction. After years of tightening, the trend line on advanced-chip exports to China has, at least for now, reversed toward a controlled trickle. Whether that trickle becomes a stream is the multi-hundred-billion-dollar question hanging over the entire AI hardware market, and it is a question decided in Washington and Beijing, not in a lab.

Who benefits, and who pays? Nvidia benefits from access, even limited, to the enormous Chinese market it had been largely shut out of; Chinese firms benefit from getting their hands on hardware they cannot yet make at the same level. The ones who “pay” are the hawks on both sides — American security officials who see any export as strategic leakage, and Chinese planners who would rather the country never depend on US silicon at all. The tension between commerce and control is now the defining fault line of the chip world.

The Dhaka view is that this is a story about dependency, watched from a country with almost none of the leverage of either protagonist. Bangladesh does not make advanced chips and cannot buy the top tier at scale; it sits downstream of a supply chain whose taps are turned by two governments. This is precisely why Bangladesh’s “design-first” semiconductor ambition — covered here in recent weeks — is sensible: the country cannot win the fabrication game the giants are playing, so it should build the skills (chip design, testing, packaging) that stay useful no matter whose tap is open. When the elephants negotiate over silicon, the smart small player invests in the one thing nobody can embargo: trained people.

2. TSMC posted a record quarter — and raised its Arizona bet to $265 billion

The money behind the AI boom showed up in a single, staggering earnings report. On July 16, Taiwan Semiconductor Manufacturing Company reported second-quarter revenue of US$40.20 billion (NT$1,270.38 billion) and net income of NT$706.56 billion — a record — with net profit up 77.4 per cent year on year and revenue up 36.0 per cent, on gross margins of 67.7 per cent. (A plain-language note: TSMC is the company that actually manufactures the most advanced chips designed by Nvidia, Apple, AMD and others — the world’s most important contract chipmaker. When AI demand booms, it shows up in TSMC’s numbers first.) CEO C.C. Wei used the results to announce an additional US$100 billion investment in Arizona, bringing TSMC’s total committed US spending there to US$265 billion, and the company lifted its 2026 capital-expenditure guidance to a range of US$60–64 billion, up from US$52–56 billion. These figures come from TSMC’s own quarterly filing with the US Securities and Exchange Commission, not from secondhand summaries.

The significance is that this is the demand story made concrete. There is a lot of talk about an AI bubble; a record quarter with 68 per cent gross margins and a capex guide raised by roughly a fifth is what the non-bubble case looks like on paper — a chipmaker so confident of sustained demand that it is pouring more money into new plants. It does not prove the boom is durable. But it is the single most grounded data point of the week, because it is audited money rather than a valuation or a rumour.

Who benefits, and who pays? TSMC and its shareholders benefit enormously; so do the AI labs whose chips it makes, and the US industrial policy that has courted its Arizona plants. The ones who pay are, potentially, everyone exposed to the concentration risk — an entire industry’s compute needs routed through one company’s fabs — and the Taiwanese public weighing what it means to move ever more of the crown-jewel capacity offshore.

The Dhaka view is one of scale and realism. TSMC’s single-quarter capex increase is larger than Bangladesh’s entire annual development budget for its digital ambitions many times over; this is simply not a game Bangladesh is in, and pretending otherwise would be foolish. The useful takeaway is the same as the Nvidia story’s: the value Bangladesh can actually capture is in the human layer — engineers who can design and verify chips, technicians who can test and package them — not in the fabs that cost a quarter-trillion dollars. TSMC’s numbers are a reminder of which end of the value chain is affordable, and which is not.

3. Shanghai opened the World Artificial Intelligence Conference — and a governance summit alongside it

China staged the week’s largest AI gathering, and paired it, pointedly, with governance. The 2026 World Artificial Intelligence Conference (WAIC) opened in Shanghai on July 17, running through July 20, alongside a High-Level Meeting on Global AI Governance, with organisers describing a record scale — an exhibition floor of more than 100,000 square metres, over 1,100 participating enterprises, and more than 300 products making their global debut. This year’s edition adds a dedicated academic track, WAIC Academic, featuring Turing Award laureates and international researchers. The scale figures come from Chinese state outlets (People’s Daily, CGTN, Global Times), which are reliable for the event’s own logistics even where one reads their framing with care.

The significance is that this is the counterpart to a story told here last week. Earlier this month the United Nations convened its first Global Dialogue on AI Governance in Geneva — the inclusive, universal forum. This week China convened its own governance summit wrapped inside the industry’s biggest trade show — the venue where Beijing sets out its vision of how AI should be built and ruled. Between them lies the central geopolitical fact of AI governance: there is no single table. There is Geneva, there is the G7, and there is Shanghai, and they do not agree on much. A pattern of parallel, competing forums is what “global AI governance” actually looks like in 2026.

Who benefits, and who pays? China benefits from projecting itself as both an AI superpower and a convener of the rules — a soft-power win as much as a technological one. Companies debuting products benefit from the world’s largest AI shop window. The ones who “pay” are those hoping for a single coherent global framework, because every rival summit makes that harder; fragmentation is the price of everyone wanting to host the conversation.

The Dhaka view is that a multipolar governance world is, counter-intuitively, not the worst outcome for a country like Bangladesh — provided it shows up at more than one table. When there is only a G7, Bangladesh has no seat. When there are three or four venues — the UN, Shanghai, Brussels’s rulebook, the G7 — a middle country can play them against each other, take the best ideas from each, and avoid being captured by any single bloc’s agenda. The task is diplomatic bandwidth: sending prepared people to Geneva and to Shanghai and reading Brussels closely, rather than picking one patron. This week’s Shanghai summit is a reminder that the menu of AI norms is now genuinely plural, and that plurality is an opportunity for those nimble enough to use it.

4. Google’s Gemini 3.5 Pro stayed a rumour on its own rumoured launch day

Here is a story about what did not verifiably happen — and why that matters. All week, technology outlets reported that Google would release Gemini 3.5 Pro on July 17, following an unusual decision to scrap the model’s original base version and restart pre-training; leaks described a two-million-token context window and a new “Deep Think” reasoning layer. But as of mid-week there was no model card, no pricing page, and no Gemini 3.5 Pro listing in Google’s public developer documentation, and Google itself had not published an announcement confirming the date or the specifications. (A plain-language note: a “context window” is how much text a model can consider at once; two million tokens would be very large. “Pre-training” is the expensive first phase of building a model — restarting it is a costly, significant decision.) In other words, the single most-anticipated Western model launch of the week was, at the time of writing, a widely-repeated expectation rather than a confirmed release.

I am including this precisely because it is a case study in how modern tech news works, and how easily an expected thing gets reported as a done thing. Dozens of articles stated the July 17 date as fact; almost none of them could point to Google saying so. That is the exact gap this newsletter tries to respect. If Gemini 3.5 Pro ships on schedule, it will be a genuinely important model and I will cover its real, confirmed capabilities next week. What I will not do is describe features and a launch that only exist in leaks.

Who benefits, and who pays? The rumour mill benefits from engagement — a launch-day story writes itself whether or not the launch happens. Google benefits, oddly, from the free anticipation. The ones who pay are readers and developers who plan around a date and a spec sheet that no one has actually confirmed, and who discover the difference only when the API stays empty.

The Dhaka view is a small plea for epistemic discipline, which happens to be a very old value. The Islamic principle of tabayyun — verifying a report before acting on it — maps almost perfectly onto the discipline a technologist needs in a hype cycle. For Bangladesh’s developers and founders, the practical version is simple: build on what is documented and available today, not on what a leak promises for tomorrow. Until the model card exists, Gemini 3.5 Pro is a headline, not a tool.

5. The EU AI Act’s teeth for general-purpose models are now weeks away

Europe’s rulebook moved from theory toward enforcement. The EU AI Act’s penalty provisions for general-purpose AI models are set to become enforceable from August 2, 2026 — now roughly two weeks out — meaning the European Commission will be able to fine providers of foundation models for obligations they have technically been under since August 2025. Those obligations include publishing technical documentation, maintaining a copyright-compliance policy, providing summaries of training data, and, for the most capable models (those trained above a compute threshold of 10^25 floating-point operations), conducting systemic-risk assessments. The Commission also published a July 2026 action plan on cybersecurity and AI. These specifics come from the European Commission’s own AI Act materials rather than from commentary. (A plain-language note: “general-purpose AI” means the big foundation models — the GPTs, Geminis, Claudes and Llamas — as opposed to narrow, single-task systems. The 10^25 figure is a measure of raw training compute used as a rough proxy for “how powerful.”)

The significance is the shift from paper to consequence. A date on which fines become possible — reportedly reaching back to violations since August 2025 — is what turns an aspiration into a compliance department. Whatever one thinks of the Act’s specifics, it is now the world’s most concrete attempt to make frontier-model providers legally accountable, and the rest of the world is watching how it lands.

Who benefits, and who pays? European citizens and downstream businesses benefit from transparency obligations on the models they depend on. The frontier labs “pay” in compliance cost and legal exposure — though the largest of them can absorb it far more easily than smaller providers, which is the perennial risk of heavy regulation entrenching incumbents. The ones who pay most, ironically, may be small European AI firms for whom the paperwork is proportionally heavier.

The Dhaka view connects directly to a Bangladesh story below. The EU is doing at the frontier what Bangladesh is struggling to do at home: attach real enforcement to written rules. Bangladesh’s Cyber Protection Ordinance criminalises harmful AI-generated content on paper, and its draft National AI Policy promises a risk-based framework — but neither yet has the enforcement machinery to match. Europe’s countdown is a useful mirror: the hard part of AI governance everywhere is not writing the rule, it is building the body that can enforce it. Watching how Brussels operationalises its Act is, for Bangladesh’s policymakers, a free education in the gap between a gazette notification and a functioning regulator.

5 Stories That Hit Closer to Home (Bangladesh)

1. Daffodil opened an AI research lab as the government pledged a million ICT jobs

The week’s most tangible home story was a room full of GPUs and a big promise. On Saturday, July 11, Posts, Telecommunications and Information Technology Minister Faqir Mahbub Anam inaugurated an Artificial Intelligence Research Laboratory and the AI Project Competition 2026 at Daffodil International University in Savar, and used the occasion to pledge one million new ICT jobs over the next five years. The lab is equipped with 25 Nvidia DGX Spark workstations and five high-performance machines using RTX 3060 and RTX 4090 graphics cards, staffed by five full-time researchers; the university said its team has already built tools it calls “AI Professor,” “AI Proctor” and “AI Registrar” and is developing a “Smart Classroom” project. The minister presented prizes to 12 teams across six categories, selected from 145 student projects, and said the government would back the most promising ones through its Startup Fund. He also flagged plans for AI training centres nationwide and for expanding satellite-based internet to grassroots communities. The event was reported by The Business Standard and the state news agency BSS.

The significance is that this is exactly the kind of unglamorous, concrete capacity-building that actually moves the needle — not a headline fund or a policy draft, but a physical lab where students can touch real hardware and researchers can work on real problems. Twenty-five DGX-class workstations is a modest amount of compute by frontier standards and a meaningful amount by Bangladeshi ones. (A plain-language note: a DGX Spark is a compact AI development computer; it will not train a frontier model, but it is more than enough for students to learn on and to build serious applied projects.)

Who benefits, and who pays? Students and early researchers benefit most directly — access to hardware is the single biggest bottleneck in Bangladeshi AI education, and this chips at it. The university benefits in prestige; the ministry benefits from a visible win. The risk — the way it could “pay” — is the familiar one of announcement outrunning follow-through: a lab is only as good as the sustained funding, teaching, and industry links around it, and a one-million-jobs pledge is only as good as the training pipeline behind it.

The Dhaka view is warm but watchful. This is the right kind of investment, and it deserves credit: put GPUs in front of students and good things follow. The caution is that a single well-equipped lab at one private university, however welcome, is a pilot, not a system. The million-jobs number is an aspiration that will be met, or not, by whether hundreds of such labs and thousands of trained teachers materialise across the public university system too. The most useful thing the government can do next is make this replicable — fund the compute, train the teachers, and connect the students to real employers — rather than let it stand as a one-off ribbon-cutting.

2. A national roadmap put the digital economy at the centre of a trillion-dollar plan

The jobs pledge turned out to be one line in a much larger blueprint. On July 14, reporting revealed a draft economic strategy prepared by the Planning Commission’s General Economics Division that targets one million ICT jobs as part of ten million jobs overall, with roughly 200,000 direct jobs in high-growth fields — artificial intelligence, cybersecurity, business-process outsourcing, semiconductors, data analytics and Industry 4.0 — and about 800,000 from the continued expansion of freelancing and digital services. The roadmap sets macro targets of 8 per cent GDP growth by FY2029-30, a US$749 billion economy and 5 per cent inflation, on the way to the stated ambition of a one-trillion-dollar economy by 2034. Notably for the country’s huge freelancer workforce, it proposes a national e-wallet with international payment capability, including PayPal, to ease cross-border transactions — a long-standing pain point. The FY2026-27 budget backs the plan with a Tk400 crore entrepreneurship fund and tax relief for startups and freelancers.

The significance is coherence. Individual announcements — a fund here, a lab there — mean more inside a stated destination, and “digital economy as a central pillar of a trillion-dollar economy” is at least a clear one. The PayPal detail matters more than it sounds: Bangladesh has one of the world’s largest freelancer populations and one of its clumsiest cross-border payment situations, and fixing that plumbing would put real money in real pockets faster than almost any grand AI initiative.

Who benefits, and who pays? Freelancers and digital-services workers benefit most, especially from smoother payments; young jobseekers benefit if the training keeps pace. Taxpayers pay if the targets prove to be aspiration without delivery — 8 per cent growth and ten million jobs are very ambitious numbers, and roadmaps have a way of outliving the governments that write them.

The Dhaka view is that the plan is directionally right and should be judged on plumbing, not slogans. The trillion-dollar headline is the least useful part; the e-wallet-with-PayPal line is the most useful, because it is specific, achievable, and would help hundreds of thousands of people this year. The right posture is to press hard on the concrete, near-term deliverables — payments, skills, connectivity — and treat the 2034 macro numbers as the mood music they are. A digital economy is built one solved bottleneck at a time, and this roadmap, for once, names a few of the real ones.

3. The Cyber Protection Ordinance still does not name the harm it is meant to stop

Behind the optimism sits an unfinished protection. Bangladesh’s Cyber Protection Ordinance, 2025 — notable as an early South Asian law to criminalise cybercrime involving artificial intelligence — remains, in practice, an incomplete shield against the deepfake abuse that has been this newsletter’s recurring concern, because deepfakes and AI-generated sexual content are not explicitly named in the statute and enforcement remains limited. The underlying harm is well documented: survey work has found that the large majority of women social-media users have faced online violence, and that most do not report it, out of stigma and a perception that legal recourse is inaccessible or even risky. I will keep this section awareness-oriented and general, without any identifying detail, because the point is the policy gap, not any individual case.

The significance is the mismatch between a law’s ambition and its precision. Criminalising “AI-involved cybercrime” in the abstract is a genuine step — many countries have not done even that — but if the specific, most common harm (non-consensual synthetic intimate imagery) is not named and there is no dedicated capacity to investigate it, the protection stays theoretical for the people who need it most. This is the same lesson as the EU story above, seen from the other side: a rule without enforcement is a promise, not a safeguard.

Who benefits, and who pays? As every week: an unenforced law benefits perpetrators, who face little practical risk, and costs the women pushed out of public and online life. A precisely-drafted, properly-enforced law would reverse that calculus.

The Dhaka view remains what it has been: the missing piece is not more rhetoric but specific investigative-and-prosecution capacity for image-based abuse, plus statutory language that names the harm plainly. Europe’s non-consensual-imagery rules and its enforcement timeline sit right there as templates to adapt. The Ordinance was a start; naming the harm and building the unit that can act on it is the unfinished remainder, and it is cheaper than any AI fund. It simply requires the will to treat this as the priority it is.

4. Connectivity moved up the agenda — with satellites in the mix

One thread ran quietly through the week’s official messaging: getting more Bangladeshis online, faster. At the July 11 Daffodil event the minister explicitly pledged to expand satellite-based internet to grassroots communities, against a backdrop of Bangladesh’s regulator having recently authorised Starlink to route cross-border international bandwidth through the country, and a Starlink partnership aimed at connecting tens of thousands of students in remote schools. (A plain-language note: satellite internet like Starlink beams connectivity from low-orbit satellites directly to a dish, reaching places fibre and mobile towers do not; the cross-border “transborder bandwidth” authorisation lets Bangladesh act as a routing point for neighbours such as north-eastern India, Bhutan and Nepal.) The pieces are early-stage and the economics for ordinary households remain steep, but the direction — treating connectivity as foundational infrastructure for any AI or digital-economy ambition — is the right one.

The significance is that none of the week’s grander plans work without this. A million ICT jobs, a network of AI labs, a freelancer e-wallet — all assume reliable internet beyond the big cities, which large parts of rural Bangladesh still lack. Connectivity is the precondition, not a footnote.

Who benefits, and who pays? Rural students and workers benefit if satellite and terrestrial connectivity genuinely reach them at a price they can afford. The country benefits strategically if it becomes a regional bandwidth router. The ones who pay are households for whom current satellite pricing is out of reach — the digital divide can just as easily be widened by a premium technology as narrowed by it, depending on subsidy and rollout choices.

The Dhaka view is that connectivity deserves to be treated as the first pillar, not the last. It is tempting to chase the shiny end of AI — labs, funds, models — while the unglamorous foundation of “can a student in a haor village get a stable connection?” goes unfinished. The most consequential digital policy Bangladesh can pursue is often the least exciting one: affordable, reliable internet everywhere. Satellites are a genuinely useful new tool in that effort, provided the pricing and public-interest partnerships are built so the benefit reaches the many, not the few.

5. The National AI Policy remained unfinished, again

And so we end where these roundups keep ending. The National AI Policy 2026–2030 remained unadopted this week — its Draft V2.0 finalised after public consultation closed in February, its committee still reviewing, and cabinet approval and gazette notification still pending — leaving its risk-based framework, its bans on mass surveillance and social scoring, its mandatory algorithmic impact assessments, and its promised oversight machinery all short of the force of law. The institutions the policy leans on, including a national governance agency and an independent oversight committee, are not yet operational; the latter would require an Act of Parliament that has not been drafted.

The significance is simply the accumulation, now stretching across many weeks of this newsletter. Bangladesh keeps deploying and announcing — a new lab, a jobs roadmap, a connectivity push, an ordinance — while the single document meant to make all of it coherent sits in draft. Governing AI piece by piece is better than not governing it at all, but it is a second-best that grows more costly the longer the unifying framework stays unfinished, because each new deployment is one more thing the absent policy was supposed to frame.

Who benefits, and who pays? Near-term flexibility accrues to those deploying AI without a binding framework; the cost falls on citizens whose protections remain aspirational and on the coherence of the whole effort. The Dhaka view is the one I will keep making until it is no longer necessary: the draft is good, the delay is unforced, and finalising it is the highest-leverage governance act available to the government this quarter. This week gave the country a lab, a roadmap, and a connectivity pledge — all of them things a finished policy is meant to knit together. Finishing it would make the good news of the week add up to more than the sum of its parts.

Wrapping up

Ten stories in seven days. Globally: the first Nvidia H200 chips reached China in very small numbers under a US licence, nudging the chip-export trend line toward a controlled trickle; TSMC posted a record quarter and raised its Arizona commitment to US$265 billion, the week’s most grounded evidence that AI demand is real money and not just valuation; Shanghai opened the World Artificial Intelligence Conference alongside its own governance summit, confirming that “global AI governance” now means several rival tables rather than one; Google’s Gemini 3.5 Pro stayed an unconfirmed rumour on its own rumoured launch day, a small lesson in verifying before repeating; and the EU AI Act’s penalties for general-purpose models moved to within about two weeks of becoming enforceable, turning Europe’s rulebook from paper into consequence.

At home: Daffodil International University opened an AI research lab as the minister pledged a million ICT jobs, the right kind of concrete, hardware-in-front-of-students investment; a Planning Commission roadmap placed the digital economy at the centre of a trillion-dollar ambition, with a freelancer-friendly PayPal-enabled e-wallet as its most useful specific; the Cyber Protection Ordinance was reaffirmed as an incomplete shield that still does not name the deepfake harm it is meant to stop; connectivity — including satellite internet — rose up the agenda as the precondition everything else depends on; and the National AI Policy 2026–2030 remained unfinished, as it has for weeks.

If there is one thread this week, it is the distance between announcement and enforcement — between a rulebook written and a fine that can be levied, between a lab opened and a system built. The frontier world is learning that governing AI means enforcement machinery, not just communiqués; Bangladesh is learning the same lesson at its own scale, with a good draft policy still waiting on the will to make it law. The chips moved, the money printed, the summits convened — and the harder, quieter work of turning intention into institution is, everywhere, the part still unfinished.

If something here 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.