AI, Software, Coding, Internet Security Thread

[Indos]

 

[Indos]

Is MIT Still Worth It in the Age of AI? I Asked Students​

 

[HERO]

Artificial intelligence is only an aid. Mathematics, physics, and chemistry are the underlying foundation. Without strength in these disciplines, nothing else can even be discussed!

 

[Indos]

 

[Indos]

 

[Indos]

 

[Indos]

Asia’s middle powers shouldn’t waste money on an AI race they won’t win​

William B Smith The Australian National University
Published: 29 September 2026

In Brief​

Asia’s middle powers are investing in domestic foundation models to reduce dependence on US and Chinese providers. But staying near the frontier means financing successive generations of increasingly expensive models, alongside talent, experimentation and failed training runs — and the resulting models depreciate fast. Governments may gain more durable leverage by investing in established strengths such as semiconductors and robotics, which competitors find far harder to replicate.

As US–China competition over artificial intelligence (AI) heats up, Asia’s middle powers are seeking a place in the emerging AI value chain. Several governments have responded by backing domestic foundation models — machine learning models trained on vast datasets — to reduce their dependence on foreign providers. But this approach risks directing scarce resources towards one of the hardest parts of the AI economy in which to catch up.


Foundation models sit near the headwaters of the AI economy. Expectations about their future capabilities are driving the buildout of compute infrastructure worldwide.


Cutting-edge development of these models remains concentrated on either side of the Pacific, with US and Chinese models exchanging the lead on various benchmarks since early 2025. Composite benchmark results suggest that the strongest foundation models developed elsewhere reliably lag behind the frontier. Even Naver’s CLOVA X, one of South Korea’s leading challengers, was named as the most-used generative AI service by only 2 per cent of South Korean users in a 2025 government survey.


The concentration of frontier development has prompted concerns about foreign dependence in several of Asia’s advanced and emerging economies. Such anxieties have been intensified by unpredictable US export decisions under President Donald Trump, whose administration imposed export controls on Anthropic’s leading models in June 2026. Relying on China’s open-weight models offers no guarantee of durable access either, with Beijing reportedly considering similar export controls to combat foreign misuse of its systems.


In August 2025, South Korea selected five companies to compete in developing a homegrown model, supplying 530 billion won (US$374.6 million) to build models rivalling those of the leading US and Chinese firms. India has committed Rs 10,372 crore (US$1.1 billion) to a similar project through its IndiaAI mission, much of it directed towards developing indigenous models trained on Indian languages and datasets. Japan’s GENIAC program likewise supplies compute resources and support to domestic developers.


This approach draws on Asia’s tradition of industrial policy, which is widely credited with driving much of the region’s catch-up growth since the 1970s.


But foundation models are a particularly difficult target for industrial catch-up. The capital required to remain near the frontier is high and rising rapidly. Training Grok 4, a leading model, cost an estimated US$490 million. Epoch AI estimates that training costs for frontier language models have risen roughly 3.5 times annually since 2020.


The next generation of foundation models will cost several times more to train — quickly exceeding the modest subsidies of Asia’s catch-up players. Yet even these estimates understate the cost of reaching the frontier. Firms must also compete for talent, fund extensive experimentation and absorb the costs of failed training runs.


Frontier firms’ advantages compound over time. They can convert superior capabilities into large global customer bases, generating revenue to reinvest in more capable models. Token- and subscription-based business models also make consumers highly responsive to even small differences in model capabilities. So lagging firms have significantly less revenue to invest in improving their models.


This is magnified by the unusually research-intensive nature of foundation-model development. Anthropic spent roughly 1.5 times its revenue in 2025 on research and development, while many of its Chinese rivals spent several times their revenue. For Asia’s middle powers, staying within reach of the frontier would require financing successive generations of increasingly expensive models at a scale far beyond current programs.


The resulting asset also depreciates quickly. In Asia’s early catch-up years, an automobile factory could remain productive for decades and hold market share even when foreign competitors released a better product because product-development cycles were slower and switching was costly. Foundation models are different. A slightly inferior model has weak differentiation unless it offers lower inference costs, better local-language performance, enhanced privacy or superior performance in specialised domains. None of these require governments to train a frontier model from scratch.


Singapore is pursuing this narrower objective. Its SEA-LION v4.5 model, designed for Southeast Asian languages and cultural contexts, was developed by fine-tuning existing open-source models from Alibaba and Google DeepMind. With a budget of only S$70 million (US$55.1 million), the program captures many of the localisation benefits at a far lower cost.


But what fine-tuning can achieve is constrained by the existing model’s capabilities. The most advanced of these, in areas of strategic significance such as cyber operations and scientific research, will remain closely tied to access to increasingly capable foundation models.


Countries seeking insurance against restrictions on frontier systems may be better placed to direct scarce public funds towards parts of the AI value chain where they possess durable comparative advantages. These include South Korea’s memory-chip economy, Taiwan’s semiconductor industry and Japan’s strength in robotics.


These positions are already protected by a deep moat — large upfront investment, significant technical expertise and supplier ecosystems that have taken decades to develop — making them costly to replicate at scale. Public investment can deepen this moat, raising the costs for competitors seeking to displace them while giving middle powers greater leverage to secure reliable access to top foundation models.


Asia’s middle powers are right to worry about dependence on foreign AI systems. But attempting to recreate foundation models at home is a costly endeavour that will likely fail to secure access to the most strategically significant foundation models. For most middle powers, doubling down on existing advantages offers a more reliable route to security.


William B Smith is Project Assistant at the East Asian Bureau of Economic Research, Crawford School of Public Policy, The Australian National University.

https://eastasiaforum.org/2026/09/29/asias-middle-powers-shouldnt-waste-money-on-an-ai-race-they-wont-win/