ASEAN Sovereign AI in 2026: Models, Compute, and the Regulatory Patchwork

Singapore: NAIS 2.0 as the regional anchor #

Singapore's National AI Strategy 2.0, released in December 2023 and operationalized through 2024 and 2025, set fifteen actions across activity drivers, people and skills, and the enabling environment. The headline allocation is more than S$1 billion over five years for AI compute, talent, and industry deployment, with the Strategic AI Compute fund anchored on a Google Cloud agreement that gives AISG and approved national projects priority access to TPU v5p and Trillium pods inside the Singapore region. The Infocomm Media Development Authority (IMDA) runs the demand side through the AI Verify governance toolkit and the AI Trailblazers program, which by early 2026 had moved more than 150 enterprise GenAI use cases from sandbox to production.

AI Singapore, hosted at the National University of Singapore, is the technical execution arm. The SEA-LION family (Southeast Asian Languages In One Network) is now on its third generation, with SEA-LION v3 8B and 70B variants trained on roughly 1 trillion tokens of Southeast Asian text spanning English, Bahasa Indonesia, Bahasa Melayu, Thai, Vietnamese, Tagalog, Tamil, and Burmese. The model is permissively licensed and serves as the multilingual base layer that Indonesian, Thai, and Vietnamese teams fine-tune against rather than rebuilding tokenizers from scratch. Singapore concedes the chip layer to the United States hyperscaler stack, retains the model and data layer through SEA-LION and AISG, and uses regulation through IMDA, the Personal Data Protection Act (PDPA), and the AI Verify Foundation to keep its policy autonomy.

The national programs side by side #

Read across the six largest ASEAN economies, the AI funding picture in 2026 looks coherent but uneven. Singapore is spending the most per capita and has the deepest model bench. Indonesia and Thailand have credible national models with first-language training data advantages. Malaysia is leaning on the semiconductor and data center side rather than on model development. Vietnam has the most active private model lab in VinAI plus the FPT AI Factory infrastructure play. The Philippines is the latest entrant, anchored on its January 2026 AI Act.

The funding numbers below are the publicly disclosed envelopes through 2026 to 2028, converted to United States dollars at year-end 2025 rates. They include only sovereign or sovereign-adjacent commitments, not private capital expenditure on data centers, which is tracked separately in the Malaysia Johor and broader ASEAN data center brief. Sisyphus treats these as floors rather than ceilings, since most programs are layered on top of existing innovation agency budgets that are not always cleanly disclosed.

Indonesia and Thailand: Sahabat-AI, Typhoon, and the multilingual stretch #

Indonesia's Sahabat-AI launched in late 2024 as a public good co-built by GoTo Group, KOMINFO, and Indosat Ooredoo Hutchison, with academic partners including Tokopedia AI Lab and ITB. The launch covered five languages: Bahasa Indonesia, Javanese, Sundanese, Batak, and Balinese. By early 2026 the family includes 8B and 70B parameter variants, with the larger model trained on roughly 600 billion tokens dominated by Indonesian-language web, news, and government corpora. The model is open-licensed and is being plugged into government service portals, BPJS health navigation, and electoral information bots through the Komdigi platform.

The economic logic for both Jakarta and Bangkok is the same. Domestic enterprises pay roughly 30 to 50 percent more per token of useful local-language generation when routing to GPT-4o or Claude than to a properly tuned local model, because tokenizers built on English-heavy corpora fragment Indonesian and Thai morphology. Governments also want citizen-facing AI services to run in country, both for latency and for compliance with Undang-Undang Pelindungan Data Pribadi (UU PDP), which entered full enforcement in October 2024 with fines of up to 2 percent of annual revenue, and the Thai PDP, which has run since 2022.

Thailand's Typhoon model, developed inside SCB 10X with public funding flow-through from the Thailand AI Council, is the highest-quality Thai-language open model in 2026. Typhoon 2 70B and the smaller Typhoon-Audio variants are trained on roughly 500 billion tokens with a Thai-heavy mix and benchmark above Llama 3.1 70B on Thai cultural and legal evaluation suites. The compute side is the Thai-Apex AI initiative, which envisages a national pool of 2,000 to 4,000 H100-class GPUs accessible through allocation to universities, the Thailand AI Council, and small and medium enterprise partners under the AI Industry Promotion program, concentrated in NSTDA's data center facility outside Bangkok with a secondary node planned in the Eastern Economic Corridor.

Malaysia and Vietnam: chip strategy and private champions #

Malaysia's MyDigital AI funding sits under the National Semiconductor Strategy launched in May 2024, with a stated industry envelope of more than RM 25 billion and a stated public funding envelope of around RM 5 billion through 2030. The AI piece is the smaller of these envelopes but more concentrated. MyDigital coordinates AI Sandbox, the National AI Office (NAIO) established in late 2024, and a planned national large language model, branded MyAi-Sertu, that has so far been more aspiration than artifact. The bigger story for Malaysia is data center cluster build-out in Johor and Selangor, covered separately, and the country's Tier 1 status under the United States Diffusion Rule, which lets it absorb GPU import volumes that Indonesia and Thailand cannot.

Vietnam's sovereign AI is led from the private sector. VinAI Research, founded in 2019 inside Vingroup, ships PhoGPT, the leading open Vietnamese language model, with PhoGPT 4B and 7.5B variants trained on roughly 480 billion tokens of curated Vietnamese text. FPT Corporation operates the FPT AI Factory in Quang Ninh and Ho Chi Minh City, anchored on Nvidia DGX systems and reportedly covering more than 4,000 GPU equivalents at full ramp through 2026 and 2027. The Ministry of Science and Technology updated the National AI Strategy in 2025 with a 2030 target of placing Vietnam among the top 50 countries in the Tortoise Global AI Index. Both Malaysia and Vietnam are using AI policy as an FDI magnet rather than as a digital sovereignty doctrine, a credible bet only if talent retention can be solved, since FPT and VinAI together still lose roughly one in six senior researchers per year to Singapore, the United States, or Chinese labs.

Models, languages, and the low-resource problem #

The shared technical wall facing every ASEAN sovereign program is scaling-law economics for low-resource languages. The Chinchilla-style optimum points to roughly 20 tokens of training data per parameter, but high-quality public Bahasa Indonesia text is in the low hundreds of billions of tokens, Vietnamese sits around 200 to 300 billion, Thai is closer to 150 billion, and Tagalog is meaningfully smaller. Pure single-language training above 20B parameters runs out of clean data well before it runs out of compute, which is why every credible regional model now uses multilingual pretraining anchored on English plus regional languages.

The practical consequence is that SEA-LION, Sahabat-AI, Typhoon, and PhoGPT are all variants on the same recipe: take an open-weights English-heavy base (Llama 3.1, Qwen 2.5, or Mistral), continue pretraining on a curated regional mix, then fine-tune on instruction and preference data sourced through local partnerships. Genuinely from-scratch sovereign training runs above 70B parameters remain economically irrational anywhere in ASEAN through the 2026 to 2028 horizon. The exception is Singapore, which has the compute and the budget but has chosen not to attempt it because the marginal benefit over a continued pretrain on Llama-class weights is small.

Compute, the Diffusion Rule, and the regulatory patchwork #

The United States Bureau of Industry and Security framework on advanced computing, the so-called Diffusion Rule, classifies countries into three tiers. Singapore sits in Tier 1, with effectively unrestricted access to advanced GPUs subject to standard end-use checks. Malaysia is also in Tier 1 after the January 2025 designation, anchored on its hyperscaler buildout. Indonesia, Thailand, the Philippines, and Vietnam sit in Tier 2, where individual customers face an aggregate cap of roughly 50,000 H100-equivalent units without a Validated End User license, and country-level allocations through 2027 are constrained to a stated cap that the Bureau publishes in tranches. For the Tier 2 ASEAN states, the binding question is whether they can secure enough VEU licenses to scale national models and enterprise deployment without hitting the cap.

On data sovereignty, ASEAN runs a five-jurisdiction patchwork. Singapore's PDPA, in its 2020 amended form, is the most enterprise-friendly. Indonesia's UU PDP and Thailand's PDP are stricter and now actively enforced. Malaysia's PDPA, amended in 2024, added breach notification and a Data Protection Officer requirement. The Philippines runs the 2012 Data Privacy Act, with the new AI Act layered on top in 2026. The ASEAN AI Governance and Ethics Guide, adopted in June 2024, is a non-binding floor rather than a unifying instrument. Argus expects the next two years to be defined by country-specific enforcement rather than by ASEAN-wide harmonization, with cross-border data flow agreements doing more useful work than treaty-level convergence.

What to watch through 2026 and 2027 #

Three variables determine whether ASEAN sovereign AI compounds or stalls. The first is talent. Singapore continues to import researchers through Tech.Pass and the AISG residency program, but the wage gap to Silicon Valley is still wide enough that mid-career attrition runs at roughly 12 to 18 percent annually for senior model engineers across the region. Indonesia and Vietnam, where the gap is wider and salary disclosure norms are weaker, leak proportionally more. Sisyphus tracks the McKinsey ASEAN Digital and Tortoise Global AI Index updates as the proxy series.

The second variable is the business model. Sovereign LLMs in ASEAN are overwhelmingly business to government first: tax authorities, citizen service portals, healthcare navigation, legal research for the judiciary. The conversion to private-sector enterprise revenue is real but slower than the marketing implies, since enterprise buyers in banking, telco, and oil and gas still prefer Anthropic Claude or OpenAI GPT-4o for English-heavy workloads and use the local model only for regional language coverage. Argus expects the B2G to B2B transition to take three to five more years.

The third variable is private-sector adoption under regulatory uncertainty. Where the AI Act and PDP frameworks are clear, as in Singapore and increasingly Thailand, enterprise GenAI deployment is moving faster. Where rules are still being written, as in the Philippines through 2026 and parts of Indonesia, deployment is dominated by pilots that struggle to escape proof-of-concept purgatory. Sisyphus assigns a roughly 60 percent probability that at least three ASEAN states reach 25 percent enterprise GenAI penetration in their largest 200 firms by year-end 2027, a base case that depends more on regulation finalization than on model quality.

Sources #

• AI Singapore (AISG) and SEA-LION
• IMDA Singapore National AI Strategy 2.0
• KOMINFO Indonesia and Sahabat-AI
• Thailand AI Council and NSTDA
• MyDigital Malaysia and National Semiconductor Strategy
• MOST Vietnam National AI Strategy
• IEEE Spectrum coverage of ASEAN sovereign LLMs
• Tortoise Global AI Index
• McKinsey ASEAN Digital and AI report series
• Tech in Asia ASEAN AI tracker