Vietnam’s sovereign AI relies heavily on open-weight models developed in the United States. However, as Washington and Beijing are moving to restrict access to these technologies, Hanoi’s bamboo diplomacy offers little protection to its AI ambitions.
In mid-July, Vietnam approved its National Digital Transformation Strategy for 2026-2030 under Decision 1266/QD-TTg, and the National Data Strategy under Decision 1308/QD-TTg. Both strategies aim for national digital sovereignty, domestic self-reliance and state-level data governance on the assumption that artificial intelligence (AI) models would remain a public good. Within days, the United States (U.S.) and China signalled their readiness to restrict access to those models.
Made in America
Vietnam’s current AI systems are modified versions of foreign tech. On the ground, Viettel, the state military telecom giant spearheading Vietnam’s AI goals, announced its VT-Super-120B-A12B Vietnamese language model had matched the accuracy of major global models of similar size. It was built by adapting Nemotron, Nvidia’s freely downloadable model family, to Vietnamese data. Viettel’s earlier model was also built on Meta’s Llama 3, trained with Nvidia tooling, and run on a cluster of 22 DGX B200 supercomputers at its Hoa Lac centre.
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
Moreover, VNG’s GreenNode subsidiary introduced GreenMind-Medium-14B-R1, the first open-source Vietnamese reasoning model built to run on Nvidia’s software and a single Nvidia H100 graphics chip. Meanwhile, FPT Corporation committed $200 million to build its AI factory powered by Nvidia hardware.
At the state level, Vietnam’s Ministry of Science and Technology met with senior executives from Qualcomm in Hanoi on July 17. Deputy Minister of Science and Technology Hoang Minh and Qualcomm’s Executive Vice President Durga Malladi discussed expanding cooperation in AI semiconductors, research and development, and workforce training.
While Vietnam has made its bets on the U.S. for its AI ambitions, other Southeast Asian countries are leaning towards China. Indonesia’s Indosat Ooredoo Hutchison partnered with AIonOS on DeepSeek-powered services, and Malaysia’s Communications Ministry launched a sovereign full-stack AI ecosystem running on Huawei hardware.
Outside Two Blocs
AI governance is dividing into two blocs. One bloc is the U.S.-led Pax Silica initiative, which coordinates trusted supply chains for semiconductors, critical minerals and AI infrastructure with 24 signatories. Singapore, the Philippines, India, Japan and South Korea are among them. The Philippines converted membership into commitments by agreeing in April 2026 to a 4,000-acre economic security zone in the Luzon economic corridor designated as the initiative’s first AI-native industrial hub.
The other is the Chinese-backed World Artificial Intelligence Cooperation Organization, signed into existence in Shanghai on July 16 by 29 governments. The membership also includes Russia, Belarus, Cuba, Venezuela and most of Central Asia, alongside Vietnam’s neighbours Laos, Cambodia and Myanmar.
Membership in either bloc could offer access to supply chains, technical assistance and software distribution during a diplomatic crisis.
However, Vietnam belongs to neither group because of its long-standing bamboo diplomacy, a policy of balancing relations between Washington and Beijing without taking sides. This leaves Hanoi in an awkward position. Vietnam depends heavily on American technology, but enjoys none of the guarantees or protections of one.
Weaponising Access
Export controls on chips work because processors are physical goods, subject to customs enforcement. On May 31, the Bureau of Industry and Security extended licensing requirements to any China-parented buyer worldwide, closing loopholes in Singapore and Malaysia.
AI models do not behave the same way. Access to a closed system can be revoked instantly by flipping an application programming interface (API) key. For example, on June 12, Anthropic suspended access to its Fable and Mythos models to comply with U.S. Commerce Department export controls, restoring access only on July 1 after those controls were lifted. Such events largely explain why governments prefer AI models they can host locally.
In contrast, an open-weight model, once downloaded, cannot be recalled by any foreign regulator. Instead, global superpowers exert control by forcing major tech companies and code-sharing platforms to block downloads from specific regions or countries. They can also pressure developers to restrict future model updates to dodge penalties from Washington or Beijing.
This fight over AI access is now an open battle. On July 16, Chinese startup Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model that independent evaluators say matches top American models at a fraction of their operating cost. On July 21, Treasury Secretary Scott Bessent signalled that Washington could sanction Chinese tech firms, citing American-model watermarks found inside Chinese ones. Days later, China’s Ministry of Commerce called the investigations groundless, threatened countermeasures, and began consulting Alibaba, ByteDance and Z.ai on export controls covering model weights, training data and chip designs.
Why It Matters
For years, nations have built digital capacity cheaply and quickly by customising open-weight AI models. Kimi K3 seemed to promise that era would continue. Instead, the geopolitical fallout exposed the fragility of relying on superpower goodwill.
For Vietnam, the real threat is getting left behind. Washington or Beijing cannot delete the AI models already sitting on Vietnamese servers. What they can block is future releases. If both superpowers restrict open-weight models, Vietnam’s AI ecosystem gets stuck using today’s tools while the rest of the world moves forward. A national tech stack built on frozen updates decays one generation at a time. States that have not localised model weights face an even harsher reality. Their access relies on live connections and downloads that can vanish overnight with a new policy.
At its core, this is a problem of time. Vietnam’s bamboo diplomacy relies on having time to adapt. Trade deals and defence agreements move slowly, giving Hanoi room to bend without breaking. AI access, however, moves instantly as access disappears with a revoked key or a blocked download link.
Diplomatically, Vietnam tries to stay neutral at all costs. However, its technology does not try to do so. All of its major AI models are built on American weights, run on American chips, and improve when American companies release new ones. Vietnam acts as if it can delay picking a side, but with every new AI update or blocked release, the cost of delay becomes more expensive. In the past, diplomatic pressure moved slowly through international summits. Today, that pressure speeds up with every new model release.
