When Digitization Becomes Dependency: How AI Infrastructure Is Redrawing the Sovereign Risk Map

On April 15, 2026, Mihnea Constantinescu, Deputy Governor of the National Bank of Moldova, sat on a panel at the IMF Spring Meetings in Washington and said something that should have alarmed every finance minister in the room.

He was not describing a currency crisis or a debt spiral. He was describing something newer and harder to hedge: the moment Moldova realized it was running critical sovereign functions through AI infrastructure it did not own, could not audit, and could not guarantee access to. Macroeconomic forecasting. Financial fraud detection. Public sector digitization. All of it running on systems controlled by companies in California.

Moldova is not an outlier. It is a preview.

The IMF’s “New Economy Forum: AI and the Resilience Gap,” which convened that afternoon, crystallized what policymakers have sensed but few have quantified. The same technology promising to accelerate development is, concentrated in a handful of US firms, constructing a new architecture of dependency. It does not look like debt. It does not look like capital flight. But it behaves exactly like both.

We ran this dynamic through our Institutional DNA Test (IDT) a framework that measures whether a country’s institutions can actually absorb and sustain a given technology across five country groupings. The results define a new fault line in global economic competition. Emerging market policymakers have a narrowing window to act on it.

The Intelligence Utility: AI Is Critical Infrastructure Now

Peter McCrory, Head of Economics at Anthropic, framed the stakes plainly at the New Economy Forum. AI today can compress the iteration cycle on complex analytical tasks downloading data, running regressions, modeling scenarios. But it still makes mistakes. The system produces high-quality output only when a skilled human operator with deep domain knowledge guides the process.

That observation matters more than it sounds. AI is not a plug-and-play productivity tool. It is infrastructure like electricity or broadband that amplifies the capacity of skilled institutions and produces negligible or negative returns in institutions that lack the human capital to direct it.

Neil Thompson, Director of MIT FutureTech, reinforced this from the labor economics side. His April 2026 working paper, “Crashing Waves vs. Rising Tides,” argues that AI’s labor market impact will not be the sudden, concentrated obsolescence that headline coverage suggests. It will be a broad, gradual lowering of barriers to complex task completion across many sectors. A rising tide.

The catch: that tide only rises where the boats are seaworthy. Countries with the digital infrastructure, the skilled workforce, and the institutional capacity to absorb AI tooling will capture the productivity gains. Countries without that substrate will watch the tide rise under someone else’s boats.

IMF Deputy Managing Director Bo Li described the required policy response: stress-test macroeconomic frameworks for AI disruption scenarios, shift from single-point forecasting to probabilistic scenario planning, and prioritize policies with high “option value” adaptable labor markets, portable skills credentials, robust social insurance that preserve flexibility as the technology evolves.

That is the right prescription. The problem is that most emerging markets are making none of those three adjustments while simultaneously building their sovereign functions on AI infrastructure controlled by firms in Menlo Park and Seattle.

The Data Is Not Ambiguous

The IMF Working Paper “The Global Impact of AI: Mind the Gap” (WP/25/076, Cerutti et al., April 2025) runs a multi-sector dynamic general equilibrium model and arrives at a conclusion that is both precise and brutal: the estimated AI-driven growth impact in advanced economies could be more than double that in low-income countries.

In the IMF’s high-productivity baseline, global Total Factor Productivity (TFP) the portion of economic output not explained by capital or labor inputs, a measure of efficiency rises 1.8% over five years and 2.4% over ten. Those gains are overwhelmingly concentrated in advanced economies with large shares of white-collar service employment and the digital infrastructure to integrate AI tooling. In the low-productivity scenario, global TFP gains compress to 0.8% over ten years. The asymmetry between rich and poor countries does not compress with it.

There is a structural reason for this divergence that goes beyond skills and infrastructure. The IMF paper flags what it calls an inverse Balassa-Samuelson effect: AI-driven productivity gains in advanced economies will hit the non-tradable sector services that cannot be imported or exported, like local retail and domestic government functions hardest. This matters for EM central banks because traditional currency adjustment loses its power. You cannot devalue your way back to competitiveness when the productivity gap is embedded in AI systems you do not own.

The BIS Bulletin on “Global Giants in the AI Supply Chain” (February 2026) maps the supply chain concentration driving this divergence. US firms account for roughly 70% of all deals in the global AI model market. The top seven global AI firms by market capitalization are all US-listed and together are worth more than twice the next thirteen combined. Stanford HAI’s 2025 AI Index confirms the innovation asymmetry: US institutions produced 40 notable AI models in 2024. China produced 15. Europe produced 3. All of Sub-Saharan Africa, Latin America, and most of South and Southeast Asia combined produced approximately zero to one.

These are not market share statistics. They are sovereignty statistics.

The Access Risk Is Not Theoretical

Constantinescu’s concern was not hypothetical. The vulnerabilities he described are already being realized not as dramatic outages, but as quiet regulatory seizures.

The most acute disruptions in EM AI access have come from US providers unilaterally updating Terms of Service to restrict API access the programming interface through which external software connects to a model in unsupported or sanctioned territories. A startup in Nairobi or Chisinau that built its software stack on a US foundation model wakes up one morning to find the service blocked. Not because of anything the startup did. Because a compliance department in San Francisco made a judgment call about geopolitical risk.

This is the defining vulnerability of the current moment: access to what is now effectively an “intelligence utility” can be severed instantaneously by foreign corporate policy or Washington-directed export mandates. US Bureau of Industry and Security export controls have already restricted NVIDIA H100 and H200 GPU access to specific jurisdictions. The next control regime could restrict inference access meaning the ability to run queries through a model directly.

The CSIS “Advancing Human Security Through AI” session (April 2026) sharpened what is at stake. When healthcare systems, financial oversight mechanisms, and public sector forecasting run on foreign-hosted AI, the host nation loses the ability to audit the models for local biases, ensure operational continuity during geopolitical disputes, or adapt the system to local conditions. Sovereign AI architecture, CSIS concluded, is now a baseline national security requirement.

Moldova did not choose dependency. It chose digitization, which in the current environment means the same thing.

IDT Applied: Scoring the AI Sovereignty Spectrum

We applied our Institutional DNA Test (https://juncturepolicy.org/frameworks/institutional-dna-test) to five country groupings to quantify where emerging markets sit on the AI sovereignty spectrum. The IDT measures compatibility between the productivity promise of a technology and the domestic institutional capacity to realize it. Scores run from 1 (minimal capacity) to 10 (full capability).

In the AI sovereignty context, we score five dimensions:

• Compute Infrastructure: domestic GPU and data center capacity as a share of need
• Human Capital Readiness: AI-skilled workforce depth and engineering talent density
• Data Governance Maturity: legal frameworks, data sovereignty laws, audit capacity
• Financial Capacity: ability to fund sovereign AI infrastructure at meaningful scale
• Institutional Adaptability: government execution capacity for multi-year technology strategy

The IDT gap between the top and bottom of this table 7.2 vs. 1.6 is the quantified version of the resilience gap the IMF is warning about. It is also a roadmap: each dimension is a policy lever.

Moldova: The Anatomy of Maximum Exposure

Moldova scores 3.1 on the AI IDT. That number reflects structural realities that Constantinescu understands better than anyone in that IMF session room.

Moldova has near-zero domestic compute capacity. Its AI talent base is real the country has produced capable software engineers but they are in Vienna, Bucharest, and Berlin, not Chisinau. Emigration has hollowed out exactly the skilled workforce that would need to staff a sovereign AI program.

What Moldova does have is EU accession ambition and the digital connectivity that comes with it. The country is rapidly digitizing its public sector to meet European integration standards. That digitization is happening almost entirely through foreign-hosted cloud and AI infrastructure. Every step toward EU integration is, paradoxically, a step toward deeper dependency on infrastructure that could be withheld under a geopolitical shock. An EU-aspiring neighbor of Ukraine cannot dismiss that scenario as theoretical.

The World Bank‘s 2025 Moldova Economic Update identifies digitalization of procurement and EU integration as the key reform channels. AI capacity-building does not appear as a standalone priority. That gap between the digitization agenda and the sovereignty agenda is precisely what the IDT is designed to diagnose: Moldova is importing productivity tools without building the institutional substrate to own them.

Constantinescu’s proposed answer organizing government data as public infrastructure and rethinking how sovereign functions are architected is the right diagnosis. The question is whether Moldova can execute it without the financial capacity or compute base the response requires. Our IDT score suggests the institutional will is present. The resource constraint is the binding limitation.

What Separates the Winners

Three emerging markets have moved from dependency to strategy. The pattern across India, UAE, and Saudi Arabia is consistent enough to constitute a playbook.

India has the most replicable model. The IndiaAI Mission, approved in 2024 with an initial outlay of roughly $1.24 billion, targets procurement of more than 10,000 GPUs to build a sovereign compute grid. The goal is explicit: ensure Indian startups and researchers do not have to route through US hyperscalers for foundational inference capacity. India’s IDT score of 6.2 reflects its genuine advantage the world’s largest concentration of AI engineering talent alongside its still-developing data governance framework and the execution challenges of a continental-scale federal system.

The India AI Summit (February 2026, New Delhi) marked a strategic shift. India is no longer framing its AI agenda around safety first. It is framing it around implementation first. India is positioning itself as the leader of a Global South AI agenda that prioritizes development outcomes over alignment debates. The CSIS Futures Summit confirmed this framing, noting ongoing coordination and governance challenges while recognizing India as the most credible EM voice in global AI architecture discussions.

UAE scores 7.2 the highest in our EM set because it solved the financial constraint completely and moved directly to scale. Stargate UAE, OpenAI’s first international deployment, envisions a 10-square-mile AI campus in Abu Dhabi with up to 5 gigawatts of planned power capacity and 200 megawatts launching in 2026. The Falcon model series, developed by the Technology Innovation Institute and released open-source, establishes UAE as the only emerging market with a credible sovereign foundation model outside the US-China duopoly. The UAE’s score reflects a core IDT insight: when financial capacity is effectively unlimited through sovereign wealth fund backing, the binding constraints shift entirely to human capital and governance. Both are addressable with targeted investment.

Saudi Arabia scores 6.6 and is executing the most capital-intensive program. HUMAIN, launched in May 2025 under the Saudi Public Investment Fund, plans 11 data centers with 2,200 megawatts of combined capacity and several hundred thousand NVIDIA GPUs. The Tuwaiq Academy targets 100,000 AI and machine learning trained programmers by 2030. AWS is investing $5.3 billion in a Saudi cloud region launching in 2026. Saudi Arabia’s limiting factor reflected in its 4.5 human capital score is the time required to develop domestic talent depth at the scale its infrastructure investment demands. The infrastructure is ahead of the human capital, which is the inverse of India’s situation.

The common thread: all three countries made a deliberate decision to treat AI infrastructure as sovereign infrastructure, not as a vendor relationship. They committed national balance sheets before the dependency became irreversible.

The Countries That Cannot Do What the Gulf Did

The Gulf model has one requirement that most emerging markets cannot meet: sovereign wealth. India is the partial exception, using state capacity rather than resource wealth to anchor its program.

Sub-Saharan Africa scores 1.6 on our IDT. Africa holds less than 1% of global data center capacity while hosting roughly 18% of the world’s population. Only about 5% of Africa’s AI talent has access to adequate computing resources. The continent’s infrastructure financing gap exceeds $100 billion per year (AfDB). No Sub-Saharan government has launched a sovereign foundation model program. All AI capacity on the continent is foreign-hosted.

Microsoft has pledged $300 million for AI infrastructure in South Africa and $1 billion for a geothermal-powered data center in Kenya. These investments are real. They are also, structurally, the opposite of sovereign AI. They extend the hyperscaler model into new geographies while deepening dependency on foreign-controlled platforms.

The IMF SDN “Bridging Skill Gaps” (January 2026) documents the second-order problem: demand for IT and AI skills is recomposing labor markets in advanced economies at twice the rate it is doing so in emerging markets. The skill gap is widening at the same time the compute gap is widening. Both compound.

The BPO and call center sectors the number-one employer in the Philippines, Moldova, Ethiopia, and dozens of other EMs face what our research classifies as extreme disruption risk. Large language models can already automate 80 to 90 percent of the tasks that define these employment categories. The labor cost arbitrage advantage that drove foreign direct investment into these sectors for three decades is being erased by the same AI systems these countries cannot afford to build themselves.

For Moldova specifically: its services sector is exposed to exactly this disruption, it has zero domestic AI R&D capacity, and its best AI-capable workers have emigrated to the EU. Three converging pressures productivity gap, access risk, labor displacement make Moldova the most acute case study for what happens when the digitization agenda runs ahead of the sovereignty agenda.

Scenarios

“Sovereignty Sprint, Divided Results” 30% probability

India, UAE, Saudi Arabia, and a small cohort of middle-income EMs successfully establish sovereign AI infrastructure over the 2026 to 2030 period. The resilience gap bifurcates. Countries that moved early build genuine strategic autonomy. The rest Moldova-tier economies and most of Sub-Saharan Africa fall into a permanent dependency relationship with US hyperscalers and become the AI equivalent of commodity exporters: providing data and labor while extracting minimal value.

For investors: Significant differentiation within EM fixed income and equity. Country-specific AI sovereignty scores become a legitimate factor in sovereign credit analysis. India and Gulf tech ecosystems attract sustained foreign direct investment.

For corporates: Companies operating in sovereign-AI countries gain access to locally auditable, adaptable AI tooling. Companies relying on BPO services in dependency-tier countries face supply chain disruption as labor cost arbitrage disappears.

For EM officials: The window to join the sovereignty cohort is open but closing. Countries with IDT scores above 5.0 have a credible path. Below 4.0, the binding constraint is financial capacity that requires multilateral intervention.

“Platform Lock-In” 50% probability

The status quo deepens. US hyperscalers extend their infrastructure into EMs under investment pledges that nominally address the compute gap but structurally entrench dependency. Most EM governments lack the political will or financial capacity to fund sovereign alternatives. The 2x productivity divergence documented by the IMF widens to 3x by 2030 as developed-market firms compound their AI advantage through continuous model iteration that EM enterprises cannot match or audit.

For investors: US hyperscaler equity remains structurally advantaged. EM growth stories become increasingly dependent on technology adoption narratives that rest on foreign-controlled infrastructure.

For corporates: Cost advantages in tradeable services compound as AI widens the productivity gap. EM manufacturing and BPO cost advantages erode faster than projected.

For EM officials: The cost of building sovereign alternatives rises with each year of delay as hyperscaler infrastructure becomes embedded in procurement, financial, and health systems. Dependency management negotiating data rights, ensuring operational continuity guarantees, building open-source fallback capacity becomes the primary AI policy task.

“Bifurcation Shock” 20% probability

Escalating US-China technology competition forces a hard bifurcation in global AI infrastructure. Export controls extend from hardware to software and inference access. EMs face a binary choice between the US and Chinese AI ecosystems with no credible non-aligned option. Countries that have not built domestic capacity face simultaneous access risk from both sides. Countries that chose one side lose access to the other’s models.

For investors: Highest tail risk scenario for EM sovereign debt in countries with significant technology sector exposure. Geopolitical alignment premiums re-emerge across asset classes.

For corporates: Supply chain restructuring under this scenario is more disruptive than the friend-shoring dynamic of 2022 to 2024. AI stack decisions become geopolitical decisions.

For EM officials: Countries with the highest option value are those that have invested in open-source model capacity and domestic compute. The UAE’s Falcon model and India’s compute grid provide the most protection against bifurcation shock.

Why This Matters

The AI productivity gap is not a digital divide. It is a sovereignty divide. When access to reasoning infrastructure can be revoked by a foreign compliance department, digital dependency becomes geopolitical vulnerability.

The IMF put a number on what is at stake: advanced economies will capture more than twice the AI-driven productivity gains of low-income countries. The BIS mapped the concentration that makes this outcome structural: 70% of AI model market activity runs through US firms.

What neither institution has fully articulated is the compounding mechanism. The productivity gap drives the financial gap. The financial gap drives the compute gap. The compute gap drives the human capital gap. The human capital gap widens the productivity gap further. The IMF’s inverse Balassa-Samuelson warning that currency adjustment cannot restore AI competitiveness means there is no traditional macroeconomic correction available to EMs that fall behind. There is no exchange rate lever for this problem.

Constantinescu had it right in Washington: sovereign functions running through offshore black-box models represent unacceptable operational and geopolitical risk. The question is which EMs have the IDT score to act on it.

Recommendations by Audience

For EM central bankers and finance ministers:

Conduct an AI sovereignty audit of your critical sovereign functions. Map every process running on foreign-hosted AI infrastructure. Quantify the disruption cost if access were severed for 30, 60, and 90 days. Treat this as you would a single-counterparty concentration risk in foreign reserves management. Bo Li’s three-part prescription from the IMF session is the right starting framework: stress-test your macroeconomic models, shift to scenario-based planning, and prioritize high-option-value policies.

For countries with IDT scores below 4.0, unilateral sovereign AI programs are not financially viable in the near term. The more productive path is coalition-based compute pooling a model the IMF and World Bank have not yet operationalized but which the research clearly supports.

For fixed income and equity investors:

Integrate AI sovereignty scores into country analysis alongside traditional institutional quality metrics. The correlation between IDT AI scores and long-run productivity trajectories will tighten as AI integration deepens in advanced economies. Countries with IDT scores above 6.0 India, UAE, Saudi Arabia warrant a structural productivity premium in growth forecasts. Countries below 3.0 face a compounding drag that traditional metrics underweight.

Watch the BPO sector in the Philippines, India Tier-2 cities, Moldova, and Ethiopia for the first measurable evidence of AI labor substitution. This is the leading indicator for EM current account deterioration in service-dependent economies.

For corporate strategists:

The hyperscaler infrastructure expansion into the Global South is not a de-risking event. It is a concentration event. Companies building operational infrastructure in EMs on the basis of foreign-hosted AI should model API access loss with the same rigor applied to supply chain single-source risks.

Companies operating in the sovereign-AI cohort India, UAE, Saudi Arabia gain access to locally auditable systems and domestically enforceable data rights. Factor AI sovereignty status into market entry decisions with the same weight given to intellectual property protection and contract enforcement.

Bottom Line

The AI resilience gap is real, quantified, and widening. The IMF has documented the 2x productivity divergence. The BIS has mapped the 70% concentration that makes it structural. Moldova has shown what sovereign dependency looks like in practice.

The window for EMs to build sovereign AI capacity before dependency becomes irreversible is measured in years, not decades. The IDT scores tell us which countries have the institutional substrate to sprint and which will need multilateral intervention to avoid permanent relegation to the bottom of the AI productivity curve.

The intelligence utility is the new utility. Access to it is no longer a technology question. It is a sovereignty question.

For the full Institutional DNA Test methodology, see: https://juncturepolicy.org/frameworks/institutional-dna-test

For the Policy Transplant Rejection Index framework referenced in related analysis: https://juncturepolicy.org/frameworks/policy-transplant-rejection-index

For our India AI Mission implementation analysis: https://juncturepolicy.org/analysis/india-ai-mission-implementation

Sources:

IMF New Economy Forum: AI and the Resilience Gap April 15, 2026. https://meetings.imf.org/en/2026/spring/schedule/new-economy-forum

IMF Working Paper: “The Global Impact of AI: Mind the Gap” (WP/25/076) April 2025. https://www.imf.org/en/publications/wp/issues/2025/04/11/the-global-impact-of-ai-mind-the-gap-566129

IMF Staff Discussion Note: “Bridging Skill Gaps” (SDN/2026/001) January 2026. https://www.imf.org/-/media/files/publications/sdn/2026/english/sdnea2026001.pdf

IMF Note: “Global Economic and Financial Implications of Artificial Intelligence” April 2026. https://www.imf.org/-/media/files/publications/imf-notes/2026/english/insea2026002.pdf

BIS Bulletin: “Global Giants in the AI Supply Chain” February 2026. https://www.bis.org/publ/bisbull122.pdf

Stanford HAI AI Index 2025. https://hai.stanford.edu/ai-index/2025-ai-index-report

CSIS: “Advancing Human Security Through AI” April 10, 2026. https://www.csis.org/events/advancing-human-security-through-ai

CSIS: “Outcomes from the India AI Summit” April 16, 2026. https://www.csis.org/events/outcomes-india-ai-summit-futures-summit

IMF Article IV Consultation: Republic of Moldova March 5, 2026. https://www.imf.org/en/publications/cr/issues/2026/03/05/republic-of-moldova-2025-article-iv-consultation-press-release-staff-report-561374

UNDP: AI Divergence Report December 2025. https://www.undp.org/asia-pacific/press-releases/ai-risks-sparking-new-era-divergence-development-gaps-between-countries-widen-undp-report-finds

S&P Global EM Outlook Q1 2026. https://www.spglobal.com/ratings/en/regulatory/article/global-fixed-income-economic-outlook-emerging-markets-q1-2026-ai-will-drive-trade-divergence-in-2026-s101657542