For government policymakers and regulators across Asia, 2026 presents a high-stakes dilemma: how to assert digital sovereignty over Artificial Intelligence without throttling the very innovation and productivity gains that AI promises to deliver.
This tension is not abstract. According to the Oxford Economics report, Economics of Sovereign AI, the region stands at a crossroads.
On one path lies the promise of AI as a general-purpose technology capable of reversing a decade of sluggish productivity growth. On the other lies the rising tide of "AI sovereignty"—a push for domestic control over data, compute infrastructure, and models driven by legitimate security and economic concerns.
The predicament for governments is that while sovereignty sounds strategic, overly restrictive implementation carries a heavy price tag: higher costs, delays of up to five years in adoption, and potential GDP losses equivalent to 1.4% in leading economies like Japan.
For enterprise leaders, this unfolding policy patchwork across Asia is no longer a distant regulatory concern—it is an immediate operational and financial reality.
The adoption gap: Why speed still matters
The urgency to get policy right is underscored by a stark regional lag. While the economic potential of AI is immense—with the IMF calling it "our best chance" to boost supply-side constraints—Asia-Pacific and Japan (APJ) are trailing global leaders.
The report notes that AI user share across APJ is only 13.5% of the working-age population, significantly behind North America (33.5%) and Europe & Central Asia (21.7%) [source: Microsoft/Oxford Economics, Page 18].
This gap represents a massive, missed opportunity. Faster adoption, underpinned by balanced policies, could deliver substantial increments to regional GDP. However, if governments prioritise restrictive sovereignty measures, they risk cementing this lag rather than closing it.
As Henry Worthington,managing director for Oxford Economics, observes, the risk for enterprises is not simply higher compliance costs but a deeper structural problem.
He notes that differing AI sovereignty approaches across Asia “are likely to increase the complexity of deploying AI applications consistently across borders… The risk is not simply higher compliance burden, but greater fragmentation in how AI systems are built and scaled.”
The spectrum of restriction: From assurance to isolation
Understanding the policy landscape is critical for business leaders. The Oxford Economics report breaks down AI sovereignty into five distinct levels of restrictiveness, ranging from Level 1 (Assurance-led) to Level 5 (Domestically owned full stack).
Most countries do not sit neatly in one box, but the direction of travel is towards hybrid models that mix global capability with local control.
- Assurance-led (e.g., Japan, Singapore): Focuses on governance, transparency, and risk-based controls. These nations maintain access to global cloud providers while applying data residency only to the most sensitive public workloads.
- Hybrid (e.g., India, Malaysia): Public sector restrictions are tightening (local data hosting, certified providers), while the private sector remains largely open to global innovation.
- Restrictive (e.g., elements in South Korea): Stricter controls on cloud provision under programmes like CSAP, and economy-wide rules under the AI Basic Act, signal a move towards greater domestic ownership.
For multinational enterprises, this means the rulebook changes every time you cross a border. Worthington summarises the operational challenge succinctly: “Enterprises may increasingly need to adapt AI systems market by market, rather than deploy a single scalable regional architecture.”
This leads to fragmented infrastructure, duplicated processes, and slower time-to-value.
The hard numbers: Costs, delays, and lost growth
The report provides quantifiable evidence of the trade-offs. When governments move towards highly restrictive, ownership-centric models (Level 5), the economic consequences for the region are profound.
1. Direct costs skyrocket
Building domestic AI capacity from scratch is capital-intensive. By 2035, the additional direct costs (compared to an open model) are staggering:
- Japan: US$149.7 billion
- India: US$102.5 billion
- South Korea: US$88.4 billion
- Singapore: US$29.1 billion
[Source: Oxford Economics, Page 10-11]
2. Adoption is delayed and diminished
Under the highest restrictiveness levels, AI adoption by firms is delayed by three to five years. Once resumed, adoption follows a permanently lower trajectory. By 2035, firm-level adoption rates under open models reach as high as 44% (Singapore), but under restrictive models, they collapse to single digits—just 3.4% in Japan and 2.3% in Indonesia [Source: Oxford Economics, Page 11].
3. Opportunity costs (foregone GDP)
The "opportunity cost" of sovereignty—the GDP lost because AI is not being used productively—is enormous. By 2035, Japan faces cumulative losses exceeding US$58 billion (1.4% of its GDP) , while Singapore could lose 3.2% of its GDP [Source: Oxford Economics, Page 12].
For CFOs, the message is clear. Worthington highlights that restrictive policies directly impact the bottom line:
“CFOs should prepare for higher infrastructure, governance, and operational costs in markets moving towards localisation… Additional investment costs [could reach] up to US$150 billion in Japan and US$29 billion in Singapore between 2025 and 2035.”
He further notes that “sovereign cloud regions cost an estimated 15–30% more” and that restrictive policies “could delay enterprise AI adoption by three to five years.”
The hidden burden: SMEs and the environment
The predicament is not uniform across the economy. Two overlooked casualties of restrictive sovereignty are small businesses and the environment.
The SME burden: Smaller firms depend on low-cost, embedded AI services from global hyperscalers. Restrictive policies that limit access to global tooling or mandate domestic-only hosting raise barriers to entry. Because public-sector procurement often sets the pace for adoption (firms are 70% more likely to adopt when the government leads), slowing down government AI use under restrictive rules also dampens private sector confidence and investment.
The environmental paradox: Restrictive sovereignty can worsen a country's carbon footprint. Hyperscale global providers operate data centres with advanced cooling and higher utilisation rates (PUE ~1.1). Forcing workloads onto smaller, domestic facilities (PUE 1.6-2.0) increases electricity and water consumption per unit of AI output. In Asia, where power grids remain carbon-intensive, this is a significant risk.
A way forward: control-and-choice
So, what is a government to do? The report points away from isolation and towards what it calls "control-and-choice" models—or what the Tony Blair Institute describes as "strategic agency" rather than technological autarky.
This balanced approach rests on three pillars:
- Risk-based data classification: Only the most sensitive workloads (approx. 10% of government data) need full localisation.
- Assurance through transparency: Using verifiable safeguards (e.g., encryption key management, auditability) rather than blanket bans on global providers.
- Partnership models: Blending global technology with local oversight. Notably, the CNAS Sovereign AI Index cited in the report finds that foreign companies are involved in roughly 70% of national AI projects [Source: CNAS, Page 73].
How enterprises should respond
For CIOs and COOs in Asia, the volatile regulatory environment demands a new playbook for 2026.
- Adopt a "portable" architecture: Avoid deep lock-in with any single cloud provider. Design systems that can move workloads between global, regional, and sovereign cloud environments as rules change.
- Prioritise optionality: As Worthington advises, “CIOs are likely to place greater emphasis on optionality, portability, and resilience as AI regulations become more fragmented… Technology strategies will need to remain flexible enough to comply with local requirements while striving to achieve deployment and productivity gains at scale.”
- Engage with regulators: The rules are not final. Proactive engagement on what "assurance" looks like (e.g., ISO 42001 standards) can help shape a business-friendly environment.
The bottom line is this: In the race for AI leadership, the countries that win will be those that learn to balance the legitimate need for sovereignty with the undeniable economics of innovation. For enterprises, navigating this fragmented landscape will require as much strategic foresight as technical skill.
