Several Asian countries are pouring public money into building their own AI models, and the cost curve says they will lose that race — bad news for national AI programs, better news for chip and robot makers.

• Training costs for top AI models rise about 3.5 times every year, so each new generation quickly outgrows what government budgets can cover.
• South Korea, India, and Japan are all funding home-grown models to cut their dependence on American and Chinese suppliers.
• The real bill is bigger than training alone — talent wars, failed runs, and endless experiments. Anthropic spent more on research than it earned in revenue, and Chinese rivals spend several times theirs.
• Singapore took a cheaper path: fine-tuning existing open-source models from Alibaba and Google DeepMind for Southeast Asian languages, at a fraction of the cost.
• The suggested alternative is to double down on what these countries already dominate — Korean memory chips, Taiwanese semiconductors, Japanese robots — because rivals cannot copy those cheaply.

Outlook: Expect Asian governments to keep funding local models for now, but budget pressure will likely push more of them toward Singapore-style fine-tuning and hardware bets.