Prepared with AI assistance from the linked primary sources. Company and government claims are attributed to their issuers. How we use AI.

This watchlist covers October 2026 through September 2027, using sources checked on 2 October 2026. It is a set of questions about implementation, not a prediction that particular projects will finish within that window. The most informative developments will show whether announced policy, infrastructure and technical capabilities become usable in practice.

Will policy plans turn into operating guidance?

Hong Kong’s September Policy Address says society-wide guidance for AI-agent safety is planned for 2027. The concrete event to watch is publication: its scope, intended users, implementation arrangements and relationship with other requirements. The announcement itself does not establish that future guidance is already effective. The official Policy Address.

Japan’s government GenAI guideline v2.0 took effect in September 2026 and includes procurement checks. Watch how agencies apply them, beyond publication of the document. The official overview.

For suppliers, the practical questions are specific: do procurement documents state evaluation criteria, data handling terms and responsibilities for incidents? Those signals would help distinguish an institutional process from a policy statement that has not yet influenced purchases.

Which infrastructure milestones become verifiable?

Ubitus says its Maizuru project entered construction in September. TCS says HyperVault secured Hyderabad land for a campus planned at up to 1 GW. These are different stages, and neither release establishes an operating facility. Ubitus’s update; TCS’s announcement.

For both, watch for documented progress appropriate to the stage: completed works, commissioned equipment, service terms and customer availability. No opening date should be inferred where the cited announcement does not provide one. A phased project’s ultimate target also needs to remain separate from whatever first becomes operational.

Will regional model gains survive deployment tests?

The September 2026 SEA-LION-v4.8 technical report describes adaptations of NVIDIA Nemotron models and reports improved SEA-HELM results across Southeast Asian language evaluations. Those are the authors’ benchmark findings, not measurements of every potential business application. The technical report.

The next useful evidence would test the same capabilities on concrete tasks: mixed-language customer enquiries, specialist documents and cases where the correct response is to ask for clarification. Readers should look for reproducible methods, model versions and failure examples alongside average scores.

What would count as a meaningful year?

Our assessment is that a convincing year of progress would connect technical and institutional changes to repeated use. That could mean a public body documenting how it buys an AI service, a facility offering contracted capacity, or an application publishing results on representative local work.

A quiet implementation update may therefore matter more than a prominent launch. The key takeaway for the next 12 months is to follow transitions: promise to document, document to process, construction to service, and benchmark to useful task. Each transition requires its own evidence.

Sources

See something we should correct? Contact the newsroom.