Region Focus | Southeast Asia

Region: Southeast Asia

This region last appeared in Region Focus on 9/18 — a 6-week gap. Two clear threads converged this week: governments are trying to keep AI capability in-house through a three-pronged push, while platforms are simultaneously buying more AI capability from outside and cutting their own R&D spend. Read together, they say more about where this region stands than either thread alone.

Key Developments This Week

Singapore's SWITCH 2026: Champions of AI, MARVEL-AI and SME workshops, all at once

Singapore's Deputy Prime Minister and Minister for Trade and Industry, Gan Kim Yong, announced a new package of AI policies at Singapore Week of Innovation and Technology (SWITCH 2026) on 10/27, building on the four national AI missions — manufacturing, finance, healthcare and connectivity — announced earlier this year in Budget 2026. The first initiative is Champions of AI, led by Enterprise Singapore, with an initial cohort of 24 local companies including DBS, OCBC, UOB, Singapore Airlines, Singtel, Keppel and Sembcorp; the government will help these companies build long-term AI transformation roadmaps, and it's officially billed as the "national flagship" program. The second is MARVEL-AI, an A*STAR-led materials-science AI accelerator backed by roughly S$150 million (about US$116 million) in government funding. The third is an expanded run of agentic AI workshops for small and medium enterprises — officials say the pilot phase already drew more than 280 micro and small businesses, who used tools like ChatGPT, Claude and Manus to build customized prototypes. (e27)

Taken together, the three programs map onto three different company sizes with three different adoption paths: large enterprises get government-assisted roadmapping, SMEs get hands-on workshops, and research institutions get a government-funded shared accelerator. Singapore's government is explicitly positioning itself as the gap-filler, rather than waiting for the market to grow these capabilities on its own.

MAS finalizes AI risk guidelines for financial institutions; agent governance enters a two-year countdown

The Monetary Authority of Singapore (MAS) formally released the final version of its AI Risk Management Guidelines (AIRG) on 10/7, requiring financial institutions to assess every AI use case on impact, complexity and reliance, with lower-risk cases eligible for a lighter governance track — but still required to document each agent's identity, tool access and safeguards. The guidelines specifically call out generative AI and agents, requiring their safeguards to be tested for effectiveness, while easing one earlier expectation: institutions no longer need to stand up a dedicated AI committee and can instead fold AI governance into their existing structures. The full guidelines carry a phased timeline, giving institutions roughly two years to fully comply. (MAS official release)

What stands out about MAS's approach is the sequencing — voluntary framework first, binding supervisory expectations second. The Infocomm Media Development Authority's (IMDA) Model AI Governance Framework for Agentic AI, published as voluntary guidance last year, now has part of its logic formally written into binding supervisory expectations for financial institutions through MAS.

Sea and Grab deepen outside AI partnerships even as R&D spend shrinks

Sea Group, Shopee's parent company, announced on 10/15 a deepened partnership with the Google AI Futures Fund, under which Sea Labs will gain access to select frontier models and Google expertise for applied AI product development. The deal builds on an agreement the two companies signed in February 2026 covering gaming, agentic commerce and agentic payments — but neither company disclosed investment figures, product milestones, or progress on that earlier partnership. Notably, Sea's fourth-quarter R&D spend fell 7.8% year over year to $277.6 million even as group revenue grew 38.4% over the same period — deepening an AI partnership while cutting the company's own R&D budget is an odd pairing to hold side by side. (e27)

Grab took a different path: on 10/14 it announced expanding its OpenAI partnership (running since 2024) to Thailand, training local drivers and merchants to use agentic AI — delegating multi-step tasks, building simple websites, setting up repeatable workflows. Grab's Driver AI Assistant now reaches more than 500,000 drivers. (CNN) The same week, Grab signed a framework agreement with SoftBank and Petroleum Sarawak Berhad (PETROS) to explore building an AI and digital infrastructure platform in Sarawak, Malaysia, aligned with the state's goal of reaching high-income status by 2030. (SoftBank official release) Southeast Asia's two largest consumer platforms are both choosing to buy AI capability from American or Japanese partners rather than build it themselves.

Deep Dive

I'd read this week's Southeast Asia story through Resource Dependence Theory, because the three developments this week sit on opposite ends of the same dependency relationship — one side is deepening it, the other is trying to buffer against it.

Platforms are deepening outside dependence while shrinking their own resource base. Neither Sea nor Grab has a roadmap for building frontier models themselves; both chose to buy AI capability from Google, OpenAI and SoftBank instead. Resource Dependence Theory holds that when an organization hands the acquisition of a critical capability to an outside partner, it gives up some of its own strategic flexibility. Sea's 7.8% year-over-year drop in R&D spend is that trade-off made concrete — not a temporary pullback, but resources that might have gone toward in-house building redirected into an outside-procurement budget instead.

Government is running the opposite play, trying to buffer with national-level resources. Singapore's three initiatives — Champions of AI, MARVEL-AI, SME workshops — are all, at bottom, aimed at reducing enterprises' dependence on any single outside AI vendor: helping large companies build their own transformation roadmaps (rather than simply adopting a vendor's template), funding a materials-science AI platform research institutions can call their own, and teaching SMEs to prototype things themselves. MAS's risk guidelines extend the same logic into regulation: singling out third-party AI for extra scrutiny is a way of telling financial institutions, in policy terms, that depending on an outside AI vendor is itself a risk to be governed, not a neutral technical choice.

The power asymmetry in a dependency relationship decides who bears the risk. Resource Dependence Theory's core claim is that whoever controls the scarce resource holds the power. Google, OpenAI and SoftBank control the scarce resource here — frontier models and infrastructure — so regional platforms like Sea and Grab, however large their market caps and user bases, remain the resource-dependent party in this relationship. Singapore's government clearly sees this, which is why it's choosing to backfill "resource independence" at the national level rather than letting every company in the market drift toward the same kind of outside dependence.

Takeaways for Taiwanese Founders

• If you advise on or build tools for enterprise AI transformation: Singapore's three-layer approach — flagship companies, SME workshops, sector-level accelerator — is worth studying less for the funding amounts than for the logic of tailoring the capability-building path to company size. If Taiwan designs a similar program, copy that tiering logic first, rather than offering every company the same kind of subsidy
• If your customers are in financial services: MAS's tighter scrutiny of third-party AI, and its two-year phased compliance clock, look like a template other Asian financial regulators — Taiwan's FSC included — may well reference. Teams serving financial-services clients should start preparing agent-identity records and tool-access inventories now, rather than waiting for Taiwan's own guidelines to land
• If you run a platform business and are weighing whether to build AI capability in-house: Sea and Grab's case is a warning sign. "Ship features fast on a big lab's model" looks efficient in the short run, but the cost is handing over control of a critical capability — and that dependence rarely shows up on a balance sheet until R&D spend starts quietly declining. Before signing an outside AI partnership, be clear about which capabilities you absolutely cannot outsource
• If you're assessing infrastructure positioning in Southeast Asia: the Grab-SoftBank-PETROS AI infrastructure project in Sarawak, Malaysia, is part of a broader data-center and compute buildout accelerating across Southeast Asia. That's a positive signal for Taiwan's hardware supply chain — servers, thermal, power management — regardless of who eventually wins at the application layer

Cognitive Delta

I used to think this week's Southeast Asia story was "platforms racing to ship agentic features." Looking closely at Singapore's three SWITCH 2026 programs alongside Sea's earnings numbers, the real divide turned out to be different: not who has the strongest agent features, but who's actually building their own AI capability versus who's just renting it. Sea and Grab chose to buy from outside — and cut their own R&D spend while deepening those partnerships — while Singapore's government is the one backfilling, through MARVEL-AI and Champions of AI, trying to keep a slice of national-level AI capability in-house rather than ceding the whole ecosystem's AI autonomy to a handful of American or Japanese giants.

References

• e27 — Inside Singapore's Southeast Asia AI Push: Lighthouses, Rice Importers and Robot Labs
• MAS — Singapore MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial Institutions
• e27 — Sea (Southeast Asia) and Google Open Chapter Two of Their AI Pact Before Chapter One Ships
• CNN — Grab (Southeast Asia) Announces New Thailand Initiative With OpenAI
• SoftBank — SoftBank Group, Grab and PETROS Sign Framework Agreement to Explore AI Infrastructure in Sarawak, Southeast Asia