If Asia spends money chasing basic AI models, it may choose the wrong battlefield from the beginning. (Reuters photo)
[Financial Channel/Comprehensive Report] Many middle-power Asian countries are investing in developing local AI basic models, hoping to reduce their dependence on US and Chinese suppliers. However, foreign media believe that the training cost of basic AI models increases by about 3.5 times every year, and the training cost of the next generation of basic models will be several times higher than now. It will soon exceed the scale of subsidies that Asian catch-up countries can currently provide, warning that middle powers in Asia may fall into a money-burning race.
Foreign media pointed out that looking at the cost of model training alone still underestimates the cost of truly approaching the technological frontier. Companies must also compete for talents, pay a lot of experimental fees, and bear the losses caused by training failures. The basic model is particularly difficult to replicate Asia’s past successful industrial catch-up models because the capital investment required to stay at the technological frontier is not only huge but also rising rapidly.
Please read on...
The article pointed out that training Grok 4, one of the current leading models, is estimated to cost US$490 million (approximately NT$15.53 billion); the research institution Epoch AI estimates that the training cost of cutting-edge language models has increased approximately 3.5 times every year since 2020. This means that to continue to remain competitive, governments must invest higher amounts from generation to generation.
Despite rapidly rising costs, many middle powers in Asia are still investing public resources in the hope of establishing local basic models and reducing their dependence on US and Chinese suppliers. South Korea selected five companies to compete for the domestic model project in August 2025, providing 530 billion won (approximately NT$12.4 billion) in funding, hoping to develop models that can compete with leading companies in the United States and China.
India has invested approximately US$1.1 billion (approximately NT$34.8 billion) through the IndiaAI project, a considerable part of which is used to develop models trained in Indian languages and local data; Japan's GENIAC project also provides computing resources and support to local developers.
This approach continues Asia's tradition of relying on industrial policies to catch up with advanced countries since the 1970s, but the basic model is significantly different from traditional industries such as automobiles and electronics. Car factories can last for decades, and even if competitors come up with better products, incumbents may still be able to maintain the market due to long product cycles and high switching costs. However, the basic model is updated very quickly. If a slightly backward model does not have lower inference cost, better local language capabilities, higher privacy protection or special field advantages, it is difficult to make a significant difference.
And the advantages of leading companies accumulate over time. More capable models can attract larger global customers and generate revenue before investing in next-generation models. The business model based on token pricing and subscription also makes consumers very sensitive to small differences in model capabilities. Therefore, lagging companies can earn less revenue and it is more difficult for them to continue investing in research and development.
The basic AI model itself is highly dependent on research expenditures. Anthropic's R&D expenditure in 2025 will be approximately 1.5 times the company's revenue, and many Chinese competitors' R&D investment will even reach several times revenue. For Asia's middle powers, if they hope to stay close to the frontier model in the long term, the investment required will far exceed the scale of current policy plans.
Singapore has chosen a different route. Its SEA-LION v4.5 model is specifically targeted at Southeast Asian language and cultural contexts. It does not start training from scratch, but fine-tunes it based on the existing open source models of Alibaba and Google DeepMind. The entire project budget is only SGD 70 million (approximately NT$1.73 billion), and it can still achieve a considerable degree of localization effect.
However, what fine-tuning can do is still limited by the capabilities of the original model, and areas of strategic significance such as cyber operations and scientific research are still highly dependent on more advanced basic models. If Asian countries are truly worried about being restricted by cutting-edge AI systems in the future, it may be more sustainable to invest limited public resources in areas where they already have long-term comparative advantages, including South Korea's memory chips, Taiwan's semiconductors, and Japan's robotics industry. These industries require huge upfront investment, technical capabilities, and supply chains that have been established over many years. The cost for competitors to replicate on a large scale is much higher than retraining a model.
The article pointed out that it is not unreasonable for Asia's middle powers to worry about relying on foreign AI systems. However, if they try to re-create a basic model in their own country that can compete with the US and China's cutting-edge models, it will not only be costly, but also may not truly ensure the acquisition of the most strategically valuable AI capabilities. In comparison, deepening existing industrial advantages may be a more sustainable choice.
Grasp the economic pulse with one hand. Click here to subscribe to Free Finance Youtube channel
