Why full AI self-sufficiency would hold Asia Pacific back

There is a growing consensus that artificial intelligence (AI) has the potential to raise productivity, improve competitiveness, and underpin long-term economic growth. Yet at the same time, governments are placing more weight on AI sovereignty, seeking greater control and oversight over the data, compute infrastructure, AI models, tooling, and applications that increasingly underpin economic and public-sector activity.

This creates a tension between valid concerns about economic security and societal interests on the one hand, and the economic gains AI can offer on the other—raising important questions about how much it will cost, how widely it can be rolled out, and how quickly it will happen.

Trade-offs in AI sovereignty policy design

The push for AI sovereignty has grown quickly across Asia Pacific. Governments want more control over the data, infrastructure, and digital systems that support their economies. As AI becomes more deeply built into essential services and business activity, policymakers are putting greater focus on making sure sensitive work can be managed within trusted frameworks and stays under domestic oversight.

But this creates trade-offs. Targeted domestic investment in AI infrastructure can help advance sovereignty, resilience, industrial policy, and capability-building goals. But when restrictions are applied more widely across sectors in the Asia-Pacific region, our research shows costs increase sharply, AI adoption falls meaningfully, and the broader economy faces growing opportunity costs on top of the direct fiscal burden.

These trade-offs can translate into billions of dollars in additional direct costs, foregone productivity gains, slower business adoption of AI, and greater demands on energy and water resources.

Our analysis looked at five broad policy approaches, ranging from open, assurance-led frameworks through to highly restrictive, ownership-focused models. These approaches vary in how much restriction they apply across the AI stack, which allows us to compare their economic effects on a consistent basis.

A domestically owned full AI stack comes at a high price

The impetus to create fully domestic AI infrastructure will involve buying large volumes of advanced chips or significant investment in domestic compute infrastructure well ahead of revenues arriving. In the most restrictive scenarios, in which governments require a domestically owned full AI stack, large economies such as Japan and India could incur an estimated additional US$149.7 billion and US$102.5 billion, respectively, in direct costs between 2025 and 2035 relative to an assurance-led approach, equivalent to around 0.3% and 0.2% of GDP.

Smaller and more open economies may face lower total costs, but the burden can still be significant in relative terms. Singapore’s direct costs under the most restrictive scenario are estimated at US$29.1 billion, which amounts to 0.4% of 2025–2035 GDP, or double that of India.

Opportunity cost of missed adoption growth

At higher levels of restrictiveness, AI adoption by firms could be delayed by three to five years, reflecting the time needed to build domestic infrastructure, tools, and skills. When adoption picks up again, it remains on a lower long-term path because firms face higher costs, less choice, and weaker incentives to innovate.

These slower adoption pathways translate into substantial opportunity costs. The impacts across businesses and public sector organisations mount up, reducing the long-term economic gains normally expected from this kind of general-purpose technology. These effects on adoption lead to substantial opportunity costs, measured as lost GDP gains compared with an unrestricted adoption path. By 2035, Japan’s opportunity cost under the most restrictive scenario is estimated at US$58.2 billion by 2035, while India’s is estimated to be $54.7 billion.

In our first article in this series, we explained how this phenomenon is akin to dysregulation found in biology. Problems arise when individually sensible interventions begin to pull in different directions, gradually reducing the coherence of the wider policy environment. That can make it more difficult for firms to invest, innovate, and scale new technologies.

Hidden environmental price of fragmentation

But the danger created by the dysregulation at the heart of this discussion goes beyond the financial penalties of lower GDP and productivity into an often overlooked impact of strict AI sovereignty: the environmental costs.

Hyperscale cloud providers deliver strong energy efficiency through purpose-built infrastructure, advanced cooling, and economies of scale. Requiring AI workloads to run in smaller, fragmented domestic data centres would mean giving up those efficiency gains.

In more restrictive scenarios, duplicating infrastructure can lead to materially higher carbon emissions and water use, particularly where workloads move into smaller, less efficient, or lower-utilisation domestic facilities. This is especially important in APAC, where many economies still depend heavily on carbon-intensive power grids and face significant water stress.

Getting the balance right

These findings are very significant for the countries of the Asia Pacific region that are looking to oversee the development of an AI stack within their economy. The policy question is therefore not whether APAC economies should develop domestic AI capability. It is how they can strengthen local control, resilience, and skills without cutting themselves off from the global technology frontier.

For most APAC economies, sovereignty should not mean owning the entire AI stack but having the ability to govern it effectively. By combining global capabilities with local control, governments can protect national interests without giving up the economic gains promised by AI. Getting the balance wrong could leave countries facing higher costs, slower adoption, and weaker access to the technologies needed to capture AI’s productivity gains; an example of the dysregulation effects that we are tracking in this series of articles.