Weekly AI Roundup: What Happened in AI This Past Week (July 18–24, 2026)

Welcome back. The monsoon kept its side of the bargain this week — Dhaka woke on more than one morning to waterlogged lanes and the familiar negotiation over which rickshaw will brave which flooded corner, and Chattogram had its own drainage arithmetic to do — and the AI world, for its part, spent the week counting money. It was earnings season for the giants, and the quarterly numbers landed one after another like the season’s cloudbursts: Alphabet, Tesla and Intel each opened their books, and each told a different chapter of the same story about how much the AI boom is earning, and how much it is costing. In between the ledgers, Google shipped a fresh batch of models and let slip that the next big one is already training, and Shanghai closed its giant AI conference by trying to build an institution around the technology.

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 Mountain View, Hsinchu or Shanghai. And, as ever, the discipline of checking each load-bearing fact against its source before repeating it. This was a week when the difference between an audited earnings figure and a confident rumour was especially worth respecting. Let us begin.

5 Big International Developments

1. Google shipped a new Gemini Flash family — and confirmed Gemini 4 is already training

The week’s clearest product news came from Google. On July 21, Google released three new models — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and a security-focused Gemini 3.5 Flash Cyber — and, in the same announcement, said it had “started our most ambitious pre-training run yet, for Gemini 4.” The headline model, 3.6 Flash, is priced at $1.50 per million input tokens and $7.50 per million output tokens, and Google says it uses roughly 17 per cent fewer output tokens than the previous 3.5 Flash to reach an answer, with gains on coding and agent benchmarks (for instance 63.9 per cent on MLE Bench, up from 49.7). The smaller 3.5 Flash-Lite is cheaper still, at $0.30 input and $2.50 output per million tokens, and runs at around 350 output tokens per second. The Flash Cyber variant, paired with a code-repair agent called CodeMender, will be, in Google’s words, “exclusively available to governments and trusted partners… as part of a limited-access pilot program.” (A plain-language note: “pre-training” is the long, expensive first phase of building a model; “tokens” are the chunks of text a model reads and writes, and price is quoted per million of them. Fewer output tokens for the same answer means lower cost to run.) All of this comes from Google’s own model blog, not from secondhand summaries.

The significance is less any single benchmark and more the cadence. A cheaper, faster Flash tier is where most real-world applications actually live — chatbots, document tools, coding assistants — because it is the tier businesses can afford to run at scale. And the quiet line about Gemini 4 pre-training beginning is the more strategic signal: the frontier labs are not pausing to catch their breath, they are already spending on the next generation while shipping the current one.

Who benefits, and who pays? Developers and startups benefit most directly: a faster, markedly cheaper model tier lowers the cost of building useful products. Google benefits from keeping enterprise customers inside its ecosystem. The ones who “pay” are rival labs now facing a price-and-speed bar that keeps rising, and, more broadly, anyone worried that the race dynamic — ship one model, start training the next before the ink dries — leaves little room for the slow work of safety evaluation.

The Dhaka view is a practical one: cheaper Flash-tier models are genuinely good news for Bangladeshi builders. The single biggest constraint on local AI products is often the running cost of inference, and every drop in per-token pricing widens what a Dhaka startup can afford to build. The lesson from this newsletter holds: build on what is documented and available today — and 3.6 Flash and Flash-Lite are available today, with published prices — rather than on the next model that is merely rumoured.

2. Alphabet posted a record quarter — and guided AI spending toward $205 billion

The first of the week’s big earnings reports was also the most emphatic. On July 22, Alphabet reported second-quarter revenue up 24 per cent year on year to $119.8 billion, with Google Cloud revenue surging 82 per cent to $24.8 billion and cloud operating income reaching $8.8 billion, against $2.8 billion a year earlier. Chief executive Sundar Pichai said: “Alphabet revenue grew 24 per cent year-over-year. Our AI investments are redefining what’s possible across every part of our business.” The company reported that its cloud order backlog had climbed to $514 billion, that its model interfaces were processing about 22 billion tokens per minute (up from 16 billion a quarter earlier), and that the Gemini app now has 950 million monthly active users. According to CNBC’s coverage of the call, Alphabet raised its 2026 capital-expenditure guidance to as high as roughly $205 billion. The revenue, cloud and usage figures come from Alphabet’s own earnings post; the capex number is from CNBC’s reporting of the call rather than the published blog, and I flag that distinction deliberately. (A plain-language note: “capital expenditure,” or capex, is money spent on physical assets — here, mostly the data centres and chips that run AI. A backlog is contracted future revenue not yet recognised.)

The significance is that this is the demand story made concrete, the same way TSMC’s record quarter was last week. Cloud growing 82 per cent, with operating income more than tripling, is what it looks like when businesses actually pay for AI rather than merely experiment with it. And a capex guide near $205 billion is a statement of conviction: Alphabet is betting an extraordinary sum that this demand is durable. It does not prove the boom will last, but it is audited money, not a valuation.

Who benefits, and who pays? Alphabet and its shareholders benefit; so do the enterprises whose AI workloads its cloud carries. The ones who pay are, potentially, everyone exposed to the risk that $205 billion of spending assumes a demand curve that bends. If AI revenue ever grows slower than the data centres built to serve it, that gap becomes someone’s write-down.

The Dhaka view is about which end of this chain a small country can occupy. Bangladesh will not build $205-billion data centres; it will rent capacity from those who do. That makes cloud pricing and availability a matter of national digital-economy policy, not just procurement. The useful posture is the one this newsletter keeps returning to: invest in the human layer that stays valuable regardless of whose data centre runs the model, and negotiate hard on the cost of the compute the country must import.

3. Intel posted its fastest growth in over fifteen years — while cutting deep into its Foundry unit

The comeback story of the week came with a painful footnote. On July 23, Intel reported second-quarter revenue of $16.1 billion, up 25 per cent year on year — its strongest quarterly growth in more than fifteen years — with adjusted earnings of $0.42 per share against expectations near $0.21, and Data Center and AI revenue up 59 per cent to $6.3 billion; the shares rose about 13 per cent in after-hours trading. Chief executive Lip-Bu Tan pointed to progress on the company’s advanced 18A manufacturing process. The footnote: in the days before the results, multiple outlets reported that Intel planned to cut between 15 and 20 per cent of the staff in Intel Foundry, its contract-manufacturing arm. The earnings figures are consistent across CNBC and Reuters’ reporting; the Foundry-layoff percentages were sourced to industry trackers rather than an Intel press release, so I treat them as reported rather than company-confirmed. (A plain-language note: a “foundry” is a factory that manufactures chips designed by others; Intel is trying to become a major contract manufacturer to rival TSMC, an enormously expensive undertaking.)

The significance is the split screen. Intel’s design-and-products side is finally catching the AI wave — data-centre and AI revenue up 59 per cent is a real recovery — while its manufacturing ambition is bleeding cash and, reportedly, jobs. It is a vivid illustration that in the chip business, designing competitive products and running competitive fabs are two very different, and separately punishing, challenges.

Who benefits, and who pays? Intel’s shareholders benefit from a genuine earnings turnaround. Customers benefit from a healthier second source of AI chips in a market dominated by Nvidia. The ones who pay, plainly, are the Foundry workers facing cuts, and the American industrial-policy bet that Intel can be the domestic manufacturing champion — a goal that gets harder every time the fab side has to shrink to fund itself.

The Dhaka view sharpens a point made here in recent weeks. Intel — a company with decades of expertise and tens of billions to spend — is struggling with the fabrication game. That is the clearest possible argument for Bangladesh’s “design-first” semiconductor instinct: the country should chase the parts of the chip value chain that reward trained people over trillion-taka factories. Intel’s split results are a live case study in which end is affordable. Design, test, packaging and verification skills travel; a leading-edge fab is a bet only a handful of the world’s richest entities can even place.

4. Tesla set a revenue record — and watched its profit collapse under AI spending

The most paradoxical earnings report of the week belonged to Tesla. On July 22, Tesla reported record quarterly revenue of $28.24 billion, up 26 per cent year on year and ahead of forecasts near $26.4 billion — yet operating income fell about 57 per cent to roughly $400 million, cutting operating margin to 1.4 per cent from 4.1 per cent a year earlier, as operating expenses jumped 47 per cent to $4.35 billion on spending for AI, its Optimus humanoid robot and its robotaxi programme. Adjusted earnings of $0.33 per share fell well short of expectations near $0.53. The company said Cybercab production had begun at its Texas plant, that its robotaxi service was now operating in seven metropolitan areas, and that production lines for the first generation of its Optimus robot were being installed. These figures are consistent across CNBC, Electrek and other reputable outlets reporting from Tesla’s own shareholder materials. (A plain-language note: a “robotaxi” is a self-driving car offered as a paid ride service; “Optimus” is Tesla’s humanoid robot. Both are long-horizon bets that cost money now for revenue that may or may not arrive later.)

The significance is that Tesla is the purest example this week of a company sacrificing present profit for a bet on physical AI. Record sales and near-zero profit in the same quarter is a deliberate choice: pour the revenue into robots and self-driving before rivals do. Whether that is visionary or reckless depends entirely on whether the robotaxi and the humanoid become real businesses — and this quarter offers no proof either way, only the size of the wager.

Who benefits, and who pays? If the bet pays off, long-term shareholders and Tesla’s engineers benefit enormously. In the near term, it is those same shareholders who “pay,” in the form of a profit that evaporated into research and factory lines. And there is a broader public question in the robotaxi rollout: the residents of the cities where driverless cars now operate are, in effect, part of a live safety experiment whose terms they did not individually agree to.

The Dhaka view is one of realistic distance. Robotaxis and humanoid robots are not Bangladesh’s near-term story, and pretending otherwise would be a distraction. But the discipline is portable: Tesla’s willingness to burn profit on a long bet is only rational if the bet is genuinely differentiated. For Bangladeshi firms and policymakers, the transferable lesson is to be clear-eyed about which long bets are theirs to make — connectivity, skills, applied software — and which belong to companies with $28 billion quarters to gamble.

5. Shanghai’s AI conference closed by launching a China-led governance body

The week’s governance news came as the world’s largest AI gathering wrapped up. The 2026 World Artificial Intelligence Conference in Shanghai, which ran July 17–20, closed with the formal shape of a new China-proposed body: the World Artificial Intelligence Cooperation Organization, to be headquartered in Shanghai, whose founding agreement was signed on July 16 by representatives of 29 countries — among them Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia and Pakistan. United Nations Secretary-General António Guterres attended the signing, and Chinese Foreign Minister Wang Yi signed on behalf of China; the organisation is described as an independent intergovernmental entity meant to promote international AI cooperation. (I flagged last week that this conference was opening; this week’s news is the institution that emerged from it, and I date the signing precisely to July 16 for accuracy.) The founding details come from Chinese state media and from international outlets including Al Jazeera and The Diplomat; as always with state-convened events, I read the logistics as reliable and the framing with care. (A plain-language note: an “intergovernmental organization” is a body whose members are national governments — like a UN agency — as opposed to a company or a trade show.)

The significance is that this is the counterpart to a pattern this newsletter has been tracking all month. Earlier in July the United Nations held its first Global Dialogue on AI Governance in Geneva; now China has stood up its own standing organisation, headquartered on its own soil, with a coalition drawn heavily from the Global South. There is still no single table for global AI governance — there is Geneva, there is the G7, there is Brussels’s rulebook, and now there is Shanghai’s organisation. The competition to convene the rules is itself becoming the story.

Who benefits, and who pays? China benefits from positioning itself as the convener of a Global-South-inclusive AI order, a diplomatic win as much as a technological one. The founding countries benefit from a seat at a table where, at the UN or the G7, they might have none. The ones who “pay” are those hoping for a single coherent global framework, since every rival institution makes harmonisation harder.

The Dhaka view is that plurality, handled well, is an opportunity rather than a threat for a country like Bangladesh. Bangladesh was not among the 29 founders, but the very existence of multiple venues — Geneva, Shanghai, Brussels, the G7 — means a middle country can take the best ideas from each and avoid capture by any single bloc. The task is diplomatic bandwidth: showing up prepared at more than one table. And as the next story shows, Bangladesh did exactly that this week — its own minister was in Shanghai, making the country’s case.

5 Stories That Hit Closer to Home (Bangladesh)

1. Bangladesh took a four-point case to the Shanghai governance table

Dhaka did not just watch the Shanghai summit; it spoke at it. At the High-Level Meeting on Global AI Governance held alongside the World Artificial Intelligence Conference, Posts, Telecommunications and Information Technology Minister Faqir Mahbub Anam urged the global community to prioritise “safety, equity, adaptability and international cooperation,” with particular emphasis on technology transfer and equitable access for developing countries, and said the future of AI “must also reflect human values, justice and inclusivity.” He told the meeting that the government is formulating an integrated and inclusive national AI policy and plans to set up decentralised AI centres of excellence across the country. The remarks were reported on July 18 by The Business Standard and carried by several other outlets including the state news agency and Dhaka Tribune. (A plain-language note: “technology transfer” is the idea that advanced know-how and tools should be shared with less-developed countries rather than hoarded by those who build them first.)

The significance is that this is Bangladesh doing precisely what the Dhaka view has been urging for weeks: showing up, prepared, at more than one governance table. A minister making a substantive four-point case at the world’s largest AI conference is a small but real exercise of the diplomatic bandwidth a middle country needs to avoid being a rule-taker.

Who benefits, and who pays? Bangladesh benefits from visibility and from aligning itself with the developing-world demand for technology transfer — a framing that plays well in the Global-South-heavy Shanghai coalition. The risk, the way it could “pay,” is the familiar gap between the podium and the gazette: fine words in Shanghai mean little if the promised national AI policy and the “centres of excellence” remain, as they have for months, unfinished at home.

The Dhaka view is warm but insistent. Speaking well abroad is worth something, and credit is due for turning up with a coherent message. But the four priorities the minister named — safety, equity, adaptability, cooperation — are exactly the values a finished, enforceable national policy would encode. The most persuasive thing Bangladesh could say at the next such summit is not another list of principles but a completed law to point to. The podium is easy; the statute is the work.

2. Bangladesh Bank finally eased foreign-exchange rules for freelancers

This was the week’s most materially useful home story, and it lands squarely on a pain point this newsletter has flagged repeatedly. In a circular reported around July 22–23, Bangladesh Bank allowed freelancers and individual service exporters to receive payments on the basis of electronic evidence — platform statements and emails — rather than conventional export documents, permitting inward remittances of up to $20,000 to be credited without formal declaration and payments through online payment gateway service providers of up to $10,000 per transaction. The circular also enables dual-currency cards for freelancers and sets the Exporters’ Retention Quota — the share of earnings a worker may keep in foreign currency — at 50 per cent for ICT-sector freelancers and 30 per cent for other service exporters. The details were reported by The Daily Star and corroborated by the state news agency and other outlets. (A plain-language note: an “online payment gateway service provider,” or OPGSP, is an intermediary like PayPal or Payoneer that moves money between an overseas client and a local worker; a “retention quota” is how much of your foreign earnings you are allowed to hold in foreign currency rather than convert immediately.)

The significance is enormous relative to the modesty of the announcement. Bangladesh has one of the world’s largest freelancer populations and, for years, one of its clumsiest systems for getting paid across borders. Last week this newsletter singled out a proposed freelancer e-wallet with international payment capability as the most useful specific in a trillion-dollar roadmap; this week’s circular is a concrete step in exactly that direction. Fixing the plumbing of cross-border payments puts real money in real pockets faster than almost any grand AI initiative.

Who benefits, and who pays? Freelancers benefit directly and immediately — less paperwork, higher retention of hard-currency earnings, and a legal path for payments they were often receiving through awkward workarounds. The formal financial system benefits as flows that once went grey come onshore. The ones who “pay,” if anyone, are whatever informal intermediaries profited from the old friction.

The Dhaka view is that this is the right kind of policy: unglamorous, specific and immediately helpful to hundreds of thousands of people. It deserves genuine credit. The remaining test is implementation — whether banks actually honour the new flexibility without inventing fresh frictions, and whether the promised dual-currency cards and OPGSP channels work smoothly in practice. A circular is a permission; a working payment is the point. But as a direction of travel, this is the digital economy being built the way it should be — one solved bottleneck at a time.

3. The government promised a special legal panel to chase deepfakes and cyberbullying

The recurring concern of this newsletter got a promised response this week. At the weekly Secretariat press briefing on July 21, the Prime Minister’s Information and Broadcasting Adviser, Dr Zahed Ur Rahman, said all administrative procedures for a special panel of lawyers to pursue deepfakes, disinformation and cyberbullying were complete, adding: “Only the formal government notification remains to be issued, which is expected within the next one or two days as the government has taken zero-tolerance policy against these offences.” He said the panel, under the Ministry of Information and Broadcasting, would be able to initiate legal action on its own initiative rather than only on complaints, noted that women are among the most frequent victims of cyberbullying, and warned that deepfakes create “social instability.” The remarks were reported by the state news agency. I will keep this awareness-oriented and general, as always, because the point is the policy, not any individual case. (A plain-language note: a “deepfake” is synthetic audio, image or video generated by AI to make someone appear to say or do something they did not; the harm most often falls on women through non-consensual fabricated imagery.)

The significance is that, for weeks, this section has argued that Bangladesh’s Cyber Protection Ordinance is an incomplete shield because it does not name the specific harm and lacks dedicated capacity to act. A panel empowered to initiate action on its own — rather than waiting for victims to come forward, which most do not — is, on paper, a direct answer to the second half of that critique. It is exactly the investigative-and-prosecution capacity this newsletter has repeatedly said was missing.

Who benefits, and who pays? If it materialises with real teeth, the beneficiaries are the women and others driven out of online life by image-based abuse, who would no longer bear the entire burden of reporting. The ones who “pay” would be perpetrators who have so far faced little practical risk. The whole benefit, though, is conditional on the panel actually being constituted and resourced.

The Dhaka view is cautious optimism, with the emphasis on cautious. “Only the formal notification remains” is a sentence this newsletter has learned to hold lightly; the gap between a briefing-room promise and a functioning enforcement body is precisely where Bangladesh’s AI governance has repeatedly stalled. This is genuinely the right idea — proactive enforcement of image-based abuse is the missing piece — and if the notification issues and the panel is staffed with people who understand digital forensics, it will be one of the more consequential steps of the year. The test, as ever, is whether the gazette follows the podium. Next week may tell us.

4. A National Semiconductor Symposium was announced — along with a planned National AI Institute

The country’s chip and AI ambitions got an institutional push this week. At a press conference in Dhaka on July 23, Science and Technology Secretary Md. Anwar Hossain announced the National Semiconductor Symposium and BEAR Summit 2026, to be held on July 25–26, jointly organised by the Ministry of Science and Technology, the Bangladesh Semiconductor Industry Association and the Silicon River Bangladesh programme — and said a National AI Institute is being established through government-industry-academia collaboration to advance AI research and support policy formulation. A BEAR Innovation Expo with 35 exhibition stalls was to run alongside the summit. The announcement was reported by the state news agency and The Business Standard. (A plain-language note: a “semiconductor” is the class of chips at the heart of all modern electronics; “design-first” — the strategy this newsletter has praised — means focusing on designing chips, which rewards trained engineers, rather than fabricating them, which requires factories costing many billions.)

The significance is that two of this newsletter’s running threads — the sensible design-first semiconductor instinct and the long-promised AI institutional machinery — advanced in a single announcement. A symposium that convenes industry, academia and government around chip design is exactly the coalition Bangladesh needs, and a National AI Institute tasked partly with supporting policy formulation could, in principle, help finish the very framework that keeps stalling.

Who benefits, and who pays? Students, researchers and the small community of Bangladeshi chip-design firms benefit from a national convening point and, potentially, from a research institute to anchor the field. The ministry benefits from a visible initiative. The risk of “paying” is the one this newsletter names most often: another body announced, another summit held, while the substance — funded research, trained people, a finished policy — lags behind the ribbon-cutting.

The Dhaka view is encouraging, with the usual caveat. Pairing a symposium with an institute that feeds policy is smart architecture. But an institute is only as good as its budget, its staffing and its independence, and “being established” is not the same as operating. The most useful outcome would be for the promised National AI Institute to become the body that finally drives the unfinished National AI Policy across the line, rather than one more entry on a list of things announced. Institutions knit announcements together — but only once they actually exist.

5. BTRC tightened how it tests mobile-network quality

The connectivity thread ran on this week too, in the unglamorous but important form of measurement. On July 23, the Bangladesh Telecommunication Regulatory Commission approved a stricter drive-test methodology for assessing mobile-network quality of service, replacing its 2021 framework: the minimum sample size for key performance indicators rises to 500 from 300, the download-test file grows to 50MB from 30MB, the required test-server bandwidth increases to at least 1Gbps from 300Mbps, and highway testing speed rises to 80 kilometres per hour from 50. Commissioner Iqbal Ahmed said: “The new methodology clearly sets out how QoS measurements should be carried out and how the results should be verified. These changes will benefit mobile users across the country.” For context, the regulator counted roughly 188.6 million mobile subscribers as of May. The changes were reported by The Daily Star. (A plain-language note: a “drive-test” is exactly what it sounds like — engineers drive around with measuring equipment to check real-world signal, speed and call quality; “quality of service,” or QoS, is the umbrella term for how well the network actually performs, as opposed to what is advertised.)

The significance is that measurement is the quiet foundation of accountability. Every grand plan in this newsletter — a million ICT jobs, a network of AI labs, freelancers paid smoothly from abroad — assumes reliable connectivity, and you cannot hold operators to a standard you do not rigorously measure. Tightening the test is how a regulator turns “improve the network” from a slogan into a checkable obligation.

Who benefits, and who pays? Ordinary mobile users benefit if tougher testing translates into real pressure on operators to improve coverage and speed. The regulator benefits from sharper enforcement tools. The ones who “pay” are the mobile operators, who now face a higher bar and, potentially, the cost of upgrades to meet it.

The Dhaka view is that this is connectivity policy doing its least glamorous and most necessary job. This newsletter has argued that affordable, reliable internet everywhere is the first pillar of any AI ambition, not a footnote; a credible QoS regime is part of what makes that pillar load-bearing. A regulator that measures honestly and holds operators to the result is worth more to a haor-village student’s actual connection than any number of conferences. The best digital-economy work is often the work nobody applauds.

Wrapping up

Ten stories in seven days. Globally, it was Big Tech’s earnings week, and the numbers read like AI’s report card: Google shipped a cheaper, faster Gemini Flash family and confirmed that Gemini 4 is already pre-training; Alphabet posted a record quarter with Google Cloud up 82 per cent and guided AI spending toward roughly $205 billion; Intel logged its fastest growth in more than fifteen years even as it reportedly cut deep into its Foundry unit; Tesla set a revenue record while its profit collapsed under the weight of AI and robotics spending; and the World Artificial Intelligence Conference in Shanghai closed by launching a China-led cooperation organisation with 29 founding countries, adding one more table to an already crowded governance landscape.

At home, Bangladesh took a four-point case for equity and technology transfer to that Shanghai table; Bangladesh Bank finally eased foreign-exchange rules for freelancers, a concrete fix for a long-standing payment headache; the government promised a special legal panel empowered to chase deepfakes and cyberbullying on its own initiative; a National Semiconductor Symposium was announced alongside a planned National AI Institute; and BTRC tightened the way it tests mobile-network quality, the quiet infrastructure of accountability.

If there is one thread this week, it is the price of ambition — measured, for once, in audited money. The giants showed exactly what the AI boom is earning and what it is costing: record revenues sitting beside collapsing margins and $205-billion spending plans, conviction and risk in the same ledger. Bangladesh’s version of the same lesson is smaller but sharper: the country made real, useful moves this week — a payments fix that helps today, a promised enforcement body, an institutional push on chips and AI — and every one of them now waits on the same thing, the follow-through that turns a circular into a working payment, a briefing into a gazette, an institute into an operation. The money moved, the models shipped, the summit convened — and the harder, quieter work of delivery is, everywhere, the part still ahead.

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.