Southeast Asia’s Mixed AI Stacks: Pragmatic but Not Risk-Free

Southeast Asian governments are often portrayed as facing a binary choice in artificial intelligence (AI): align with the American ecosystem or adopt Chinese technology. That pressure might be becoming more explicit, as the US is preparing to warn partners that deeper participation in a US-led AI coalition may be incompatible with joining Beijing’s competing initiatives.

Yet Southeast Asia’s emerging AI landscape suggests a more complicated reality. Governments, technology firms and research institutions are building mixed AI stacks. These stacks have two key elements: foreign foundational models, chips and cloud infrastructure coupled to software with local-language data, evaluation capacity and regionally developed applications.

A single mixed stack system might use an American-designed chip, a Chinese open-weight model (where users have access to downloadable model parameters), a regional cloud provider and a locally curated dataset. Rather than reproduce the entire AI value chain, regional actors are investing selectively in the layers that matter most for local relevance and practical control.

This mixing of technologies creates a growing tension: Southeast Asian actors have practical reasons to combine the best available technologies from multiple sources, even as geopolitical competition may increasingly make such arrangements harder to sustain.

Singapore’s AI programmes provide one of the clearest examples of selective capability building. Southeast Asian Languages in One Network (SEA-LION) and MERaLiON focus on Southeast Asian languages, multilingual processing, code-switching and multimodal capabilities. Yet different versions of SEA-LION have been built on foundation models developed by Meta, Google, and Alibaba before being further trained for regional contexts.

Similar approaches are emerging elsewhere. Indonesia’s Sahabat-AI is being developed by Indosat, GoTo and AI Singapore. It adapts models for Bahasa Indonesia, Javanese and Sundanese. Its local contribution lies in language data, adaptation, evaluation and application development, while parts of the underlying stack rely on foreign foundation models and NVIDIA hardware.

Thailand’s Typhoon family builds on Meta’s Llama and Alibaba’s Qwen models before applying further training for Thai language and cultural contexts. Vietnam’s FPT is expanding domestic compute and deployment capacity through its AI Factory, using NVIDIA chips and software.

These initiatives operate at different layers. SEA-LION, Sahabat-AI and Typhoon concentrate on model and language adaptation, while FPT’s AI Factory focuses more heavily on infrastructure and deployment. None reproduces the entire stack from either China or the US.

Taken together, these initiatives do not fit neatly into US- or China-led technology blocs. Their components come from firms and institutions across multiple countries.

Mixed stacks offer Southeast Asia three main advantages.

First, they improve local relevance. Global models remain uneven in their treatment of Southeast Asian languages, multiple scripts, local idioms and frequent code-switching. Local adaptation can improve performance in vital areas such as healthcare, education, public services and translation without requiring governments to train an expensive foundation model from scratch.

Second, they can improve efficiency. Smaller or more specialised models may be cheaper to operate and easier to customise. Open-weight models can also give developers greater freedom to adapt systems and, where local hosting is feasible, reduce dependence on continuous access to a foreign application programming interface (API).

Third, mixed-stack development can strengthen evaluation and governance capacity. Building regional datasets and benchmarks allows governments and firms to compare competing systems, test their performance in local contexts and identify weaknesses before deployment. Singapore’s exploration of “nutrition labels” for AI applications reflects this broader emphasis on disclosing systems’ capabilities and limitations.

The strategic value of a local AI programme may therefore lie less in owning a model than in being able to evaluate, modify and replace it.

The benefits of mixed stacks are real, but the language used to describe them requires precision.

A model can be locally adapted without being locally hosted. It can be locally hosted without being locally owned. It can be locally owned while remaining dependent on foreign chips, development tools, or cloud software. Describing a system as “national” or “sovereign” can therefore exaggerate the degree of control that governments or domestic firms actually possess.

Open-weight availability does not automatically resolve this problem. An organisation may have access to a model’s weights but lack the specialised hardware, engineering expertise, energy or financing required to operate it independently.

Mixed sourcing can also conceal concentration. An agency might use three different models, while hosting all of them on the same cloud or running them on the same chip architecture. If the shared infrastructure fails or access conditions change, diversity at the model layer offers little protection.

The real test of a mixed stack is not simply how many countries or companies supply its components. It is whether critical models, infrastructure, data and tools can be replaced within an acceptable time and budget, and whether governments retain the political freedom to make those substitutions.

Local hosting does not necessarily mean local control either. Foreign investment can expand Southeast Asia’s data centre capacity, but ownership, technical operations, pricing and decisions about pricing or access may remain outside the host country. A facility located in Southeast Asia can therefore strengthen domestic capacity while also creating a new point of external dependence.

Mixed does not mean portable. AI systems become embedded in proprietary data formats, cloud credits, security tools, fine-tuning processes, application interfaces and organisational routines. A stack may appear modular on paper but can be costly and disruptive to dismantle in practice.

There may also be a political limit to diversification. Mixed stacks assume that governments and firms will remain free to combine technologies and partnerships from competing ecosystems. Recent US pressure on its AI partners raises the possibility that this freedom could be narrowed. Export controls, procurement rules, security requirements or conditions attached to investment could eventually make some combinations more difficult.

Southeast Asia’s mixed-stack approach is a rational strategy for a region that cannot — and need not — replicate every part of the global AI value chain.

The real test of a mixed stack is not simply how many countries or companies supply its components. It is whether critical models, infrastructure, data and tools can be replaced within an acceptable time and budget, and whether governments retain the political freedom to make those substitutions.

Southeast Asia’s mixed stacks preserve room for manoeuvre only when their dependencies are visible, components are genuinely portable and governments retain the freedom to move between suppliers. Otherwise, they may replace one obvious dependency with several less visible ones.

2026/265

---

This is an adapted version of Trends in Southeast Asia 2026/24, published on 14 September 2026. The paper and its references can be accessed at this link.