When AI models become interchangeable, Indonesia's child-safety filing, and the HBM chain through Korea, Taiwan and Japan

This Week's Analysis

When AI models become interchangeable, the work moves to routing

Short Reads

1. Indonesia's child-safety rule puts risk filing alongside under-16 deactivation

   Indonesia has moved from setting out a child-safety framework in 2025 to putting operators through a real compliance process in 2026. They have to assess the products children may use, document the risks and file those assessments with the regulator. Services classified as network and social-media services must also deactivate under-16 accounts. The public conversation naturally focuses on how many accounts are being removed. I found the filing requirement more interesting. It creates a record of how operators assess the risks their products pose to children, even if the public cannot yet see what those assessments conclude.

   Read on Relay Asia →
2. Korea makes the memory. Taiwan packages it. Japan enables the yield.

   Extra high-bandwidth memory does not become AI compute on its own. Korea makes the memory, Taiwan packages it, and Japan supplies the tools and materials that help make the process work at scale. TSMC is adding packaging capacity quickly, while C.C. Wei continues to describe packaging as a constraint on customers' growth. This Short Read looks at why the bottleneck sits at the join.

   Read on Relay Asia →

Suggested Reads

• Thailand's data-centre boom is running into a clean-power squeeze. Thailand wants the data-centre investment, though finding enough clean power and water to support it without squeezing everyone else could be a stumbling block. A 15 September 2026 roundtable reported by The Nation described the risk as a "green squeeze": data centres could consume renewable electricity that manufacturers, SMEs and other businesses need for their own decarbonisation. Attracting an AI infrastructure boom is only one part of the equation; the infrastructure underneath it has to grow too.
• Thai social-media users put job security first as AI use rises. This one caught my attention because it moves the AI conversation away from investment and into people's working lives. RealWatch Lab found that job loss accounted for 33.35% of AI-related concerns in the Thai social media discussions it analysed from January to August. That comes as AI adoption is rising, with the share of Thailand's working-age population using AI increasing from 9.1% in the first half of 2025 to 12.4% in the first quarter of 2026, according to Microsoft. The technology is spreading and the anxiety is spreading with it.
• Southeast Asia is still growing. The paths are diverging. Bain, DBS and Vriens & Partners' latest outlook projects the six largest Southeast Asian economies to grow by an average of 4.8% a year from 2026 to 2035. What I found more interesting is the divergence underneath it. Vietnam remains the growth leader while Thailand lags, and the report sees Indonesia, the Philippines and Thailand as more exposed in a downside scenario. The region is still connected by trade, investment, energy, technology and supply chains. But the next decade will not look the same from every capital.
• China's CXMT says a new DRAM platform has entered mass production. China's CXMT, the country's leading producer of DRAM (Dynamic Random-Access Memory) chips, says its fifth-generation DRAM platform has entered mass production, with new 24-gigabit LPDDR5X chips and a claimed 50% increase in dies per wafer. This is not the same part of the memory stack as the HBM story in this week's Short Read. DRAM is the working memory inside devices, while HBM is designed for high-bandwidth computing. Still, it is worth watching alongside the HBM story. China's memory industry is moving further into territory dominated by Samsung, SK Hynix and Micron.
• Australia's National AI Centre wants the whole team at the table. Australia's National AI Centre published a set of scenario exercises that ask organisations to bring people from different parts of the business together and work through a fictional AI incident before a real one happens. The idea is simple: responsible AI is not just a policy document. It is something a team gets better at by practising. That feels much closer to how AI governance actually works inside an organisation.
• A woman travelled to Chiang Mai to meet a pilot who was never there. The Bangkok Post reported that a 66-year-old woman from Lampang travelled to Chiang Mai International Airport to meet a pilot she had met online. Airport officers concluded that the photograph she had been sent was AI-generated. Tourist police eventually found her son and put her on a bus home. It is an unsettling story, and it made me wonder whether enough is being done to improve fraud literacy and protect people from technology-enabled scams as AI becomes more accessible.