The AI Paradox: Why Emerging Economies May Win by Not Racing
Lacking chips and cash, poor countries may leapfrog rich ones by training small models on local knowledge
Emerging economies appear catastrophically disadvantaged in the AI revolution. The United States and China control over 90% of private AI investment. Hyperscale computational infrastructure concentrates in a handful of wealthy nations. Advanced semiconductor manufacturing resides almost entirely in Taiwan, South Korea, and the United States. By conventional metrics, the gap widens rather than closes.
Yet this analysis may miss the fundamental dynamic. Emerging economies face a strategic choice that could transform apparent disadvantage into unexpected advantage: whether to compete in a race toward artificial general intelligence that demands resources they lack, or to pursue a different path entirely - one built on open-source models, on-premises deployment, and specialized applications solving local problems extraordinarily well.
Three realities converge to create this possibility. The third wave of the information revolution is built on open-source foundations that democratize capability in unprecedented ways. Intensifying geopolitical competition under the Trump administration is accelerating the weaponization of digital infrastructure, making sovereignty concerns paramount. And the race toward ever-larger general-purpose models may be orthogonal to where development impact actually resides - in specialized systems trained on local data.
Understanding how these dynamics intersect reveals why choosing not to race may constitute the winning strategy.
The Third Wave: Democratized Production, Oligopolistic Execution
The third wave of the information revolution differs fundamentally from its predecessors: it is built on open-source architectures.
Previous waves operated through proprietary platforms. Microsoft Windows, Oracle databases, Cisco routers remained closed, licensed, controlled. The second wave’s cloud infrastructure followed similar patterns through AWS, Azure, and Google Cloud.
The third wave emerged differently. Most foundational AI models are open in the meaningful sense - downloadable, modifiable, deployable independently. Meta’s Llama 3.1 (July 2024), Mistral’s models (Apache 2.0), Alibaba’s Qwen 2.5 series (Apache 2.0), and DeepSeek’s R1 (January 2025, MIT license) exist as accessible weights that any developer can adapt. According to Hugging Face, over one million models were publicly available by early 2025.
This openness emerged not from idealism but from industry structure. Early AI companies needed vast web-scraped training data without explicit permission. Publishing model weights and encouraging collaboration created legal cover by framing research as open science rather than commercial appropriation. The architecture persists despite intensifying commercial competition.
Yet this democratization exists in structural tension with radical centralization. As of Q4 2024, Amazon Web Services holds approximately 31% of global cloud infrastructure market share, Microsoft Azure approximately 20%, and Google Cloud approximately 11%, according to Synergy Research Group. Combined, these three U.S. providers control roughly 62% of the global market. In Europe specifically, U.S. cloud providers account for approximately 69% of spending.
The paradox is stark: the means of production are democratized through open models, but the means of large-scale execution remain oligopolistic. Anyone can download Llama 3.1’s 405 billion parameters, but training such models from scratch or running inference at massive scale requires infrastructure concentrated in remarkably few hands.
The paradox is stark: the means of production are democratized through open models, but the means of large-scale execution remain oligopolistic. Everyone can build; only a handful control where things run.
For emerging economies, this creates both vulnerability to infrastructure denial and opportunity to deploy models locally in ways that bypass dependency entirely.
Digital Infrastructure as Geopolitical Weapon
The weaponization of digital infrastructure is not hypothetical. Between 2005-2010, Microsoft’s Windows Genuine Advantage demonstrated that foreign corporations could remotely disable or degrade operating systems across sovereign borders. China’s October 2008 “black screen” incident - where unlicensed Windows installations received persistent warnings affecting millions of users - sparked governmental alarm about foreign control over critical infrastructure. These incidents seeded digital sovereignty doctrines subsequently codified in China’s Cybersecurity Law (2017), Data Security Law (2021), and Personal Information Protection Law (2021).
When digital infrastructure becomes an instrument of state power, it becomes subject to the same sovereignty concerns governing energy security or defense industrial capacity.
Contemporary dynamics operate at vastly greater scale. The October 17, 2023 U.S. Commerce Department rule expanded semiconductor export restrictions, targeting NVIDIA’s A100 and H100 GPUs and establishing performance thresholds that effectively restricted chips capable of training large AI models. By early 2025, NVIDIA’s latest H200 and forthcoming B100/B200 architectures face similar restrictions.
The Trump administration’s return to power in January 2025 accelerated this trajectory. In March 2025, the administration announced substantial reductions to foreign development assistance, with USAID’s budget facing 100% cuts . Development programs focused on digital infrastructure, technology transfer, and AI capacity building were all affected. The administration’s “America First” framework explicitly prioritizes domestic AI competitiveness over international capacity building.
Simultaneously, restrictions on AI model exports remain under active consideration. While no formal rules had been implemented by early 2025, policy discussions in Washington center on whether advanced AI capabilities constitute dual-use technologies requiring export licensing - potentially restricting even open-source model weights if they exceed certain capability thresholds.
The investment asymmetry reinforces concentration. According to the Stanford AI Index 2024 Report, the United States attracted $67.22 billion in private AI investment in 2023, while China attracted $7.76 billion - roughly a 9:1 ratio that has widened substantially since 2020. Europe attracted $15.97 billion in 2023, placing it distant third. Preliminary data for 2024 suggests these trends continued, with the U.S. attracting approximately $75-80 billion.
Even wealthy nations recognize vulnerability. The EU’s Important Projects of Common European Interest in Cloud Infrastructure and Services (IPCEI-CIS), approved December 2023, allocates up to €1.2 billion across eight member states for European cloud capabilities. The Gaia-X initiative, launched in 2020 by France and Germany, aims to create federated data infrastructure under European governance. France’s Mistral AI, despite raising €600 million in 2024, remains orders of magnitude smaller than OpenAI or Anthropic.
For emerging economies, the structural reality is considerably more severe. Sub-Saharan Africa hosts less than 1% of global data center capacity despite accounting for approximately 15-18% of world population. According to the International Finance Corporation’s analysis, Africa’s data center market remains significantly underdeveloped relative to population and economic needs. Latin America’s data center capacity concentrates heavily in Brazil and Mexico, leaving smaller economies dependent on international connectivity. South and Southeast Asia outside China and India similarly lack indigenous hyperscale infrastructure.
The strategic implication is unambiguous: dependency on externally-controlled computational infrastructure represents not merely commercial risk but strategic vulnerability to geopolitical pressure. When digital infrastructure becomes an instrument of state power - which the Trump administration’s policies explicitly confirm - it becomes subject to the same sovereignty concerns governing energy security or defense industrial capacity.
Yet the open-source foundation creates crucial asymmetry. While semiconductor fabrication and hyperscale cloud infrastructure can be restricted through export controls, the models themselves - the intelligence encoded in weights and parameters - can be copied, modified, and deployed independently. China’s DeepSeek R1, released January 2025, reportedly achieved performance competitive with OpenAI’s models using substantially less computational resources and training on hardware not subject to the most stringent U.S. export controls. This demonstrates that access to cutting-edge chips, while advantageous, is not absolutely determinative.
For emerging economies, this creates strategic options fundamentally different from previous technological competitions where core technologies remained proprietary and inaccessible.
Two Architectures: Cloud Dependency Versus Sovereign Deployment
Geopolitical competition has crystallized two distinct deployment models.
The U.S. model operates through centralized hyperscalers delivering cloud services. Computation occurs in large data centers accessed via subscription and APIs. This assumes robust internet connectivity, permissive cross-border data flows, and trust that foreign infrastructure operators will maintain service availability regardless of geopolitical circumstances. It optimizes for performance and scale but embeds dependency.
China’s model emphasizes hybrid and on-premises deployment. This developed from market dynamics and regulatory requirements. Software-as-a-service never achieved the dominance in China that characterizes Western markets. China’s regulatory framework - the Cybersecurity Law (2017), Data Security Law (2021), and Personal Information Protection Law (2021) - created strong incentives for on-premises deployment of sensitive workloads. According to IDC’s analysis of China’s cloud market, China’s public cloud market reached approximately $33 billion in 2024, smaller as percentage of IT spending than comparable U.S. figures despite China’s large economy. Enterprise spending on on-premises infrastructure remained proportionally higher.
For emerging economies lacking hyperscale data centers and increasingly unlikely to receive development assistance for building them under current U.S. policy, China’s approach may constitute unexpected advantage.
On-premises deployment offers tangible benefits. Data and processing remain within national jurisdiction, insulated from external political pressure. A minimum viable deployment requires one or two servers with 64-128GB VRAM - approximately $15,000-35,000 per server in early 2025 - within procurement capacity of universities, ministries, and larger domestic enterprises. This proves far more feasible than constructing hyperscale data centers requiring hundreds of millions in capital, specialized workforce, and reliable electrical infrastructure at 10+ megawatt scale.
An ‘AI in a box’ running within national borders may prove strategically superior to cloud dependency - even sacrificing some performance or scale.
Capital expenditure on hardware compares favorably to perpetual cloud subscriptions when foreign exchange constraints bind. According to Epoch AI’s 2024 analysis, inference costs for language models decreased dramatically - some tasks showing 10x+ annual cost reductions between 2020-2024 driven by hardware improvements, software optimization, and model efficiency. Under foreign exchange pressure, one-time CapEx models denominated in domestic currency outperform perpetual OpEx subscriptions denominated in dollars.
With U.S. development assistance declining sharply under the Trump administration, emerging economies must pursue strategies requiring minimal external support. On-premises deployment using open-source models requires initial hardware investment but no ongoing foreign currency expenditure for cloud subscriptions or licensing.
The Chinese deployment model, developed for regulatory reasons, inadvertently addresses sovereignty concerns and infrastructure constraints binding emerging economies. An “AI in a box” running within national borders may prove strategically superior to cloud dependency - even sacrificing some performance or scale.
Local Knowledge, Local Impact: The Hidden Advantage
The most consequential insight may be that the race toward artificial general intelligence is irrelevant to development imperatives - and that emerging economies possess a structural advantage precisely where it matters most.
Frontier model development consumes extraordinary resources. OpenAI CEO Sam Altman stated GPT-4’s training exceeded $100 million. Industry estimates for next-generation models range from $500 million to over $1 billion. This trajectory pursues general capability across all domains - answering questions about global history, writing code in multiple languages, analyzing images from anywhere, generating creative content on any topic.
But the most economically valuable data is rarely global. It is local, domain-specific, and rooted in conditions that vary dramatically by geography, industry, and context.
In mining, geological formations, ore composition, equipment performance characteristics, and optimal extraction techniques vary by specific site. A model trained on South African platinum mining data provides limited value for copper extraction in Zambian conditions. But a model trained specifically on data from Zambian copper mines - equipment sensor readings, geological surveys, historical yield data, local labor practices, regulatory requirements - can optimize operations in ways a general model never could. The knowledge that matters is hyper-local: this mine, this ore body, these specific conditions. For operations where a 5% improvement in ore recovery translates to millions in additional revenue, the value is tangible and substantial.
In agriculture, soil composition, microclimate patterns, pest pressures, crop varieties, and farmer practices differ radically between regions. A model trained on U.S. Midwest corn farming provides little guidance for smallholder cassava cultivation in Nigeria. But a model trained on data from Nigerian agricultural extension services, local soil maps, regional weather patterns, indigenous crop varieties, and verified yield outcomes from similar farms can provide guidance calibrated to specific Nigerian conditions. A 2020 peer-reviewed synthesis in Remote Sensing found that precision agriculture techniques using site-specific, data-driven management can increase crop yields by up to 30% under optimal implementation conditions . Additional research from FAO on precision agriculture in developing countries documents similar potential benefits. For smallholder farmers operating on thin margins, a 20% yield improvement represents transformative income increase - the difference between subsistence and surplus.
In healthcare, disease prevalence, genetic susceptibilities, environmental risk factors, infrastructure capabilities, and treatment protocols vary significantly by geography. A diagnostic model trained primarily on North American or European patient data may perform poorly on African or Southeast Asian populations with different genetic backgrounds and disease profiles. The WHO’s 2021 guidelines on tuberculosis screening conditionally recommend computer-aided detection software for tuberculosis screening via chest X-ray as replacement for human readers in individuals aged 15+. Published studies show these systems achieve 82-95% sensitivity depending on settings - comparable to human radiologists. A systematic review in The Lancet Digital Health analyzing CAD4TB performance found pooled sensitivity of 89% and specificity of 78% across multiple settings. Research published in PLOS ONE on AI for TB diagnosis further validates performance in resource-limited settings. Critical for resource-constrained environments: these systems operate entirely offline, require no internet connectivity, and process no patient data externally. They deliver diagnostic capability to clinics that would otherwise have none.
The most economically valuable data is rarely global. It is local, domain-specific, and rooted in conditions that vary dramatically by geography, industry, and context.
In infrastructure and utilities, power grid characteristics, water distribution system topologies, road network configurations, and maintenance requirements are fundamentally local. A model optimizing power distribution must understand this specific grid’s topology, generation capacity, consumption patterns, and failure modes. For cities struggling with unreliable electricity or water supply, even modest improvements in infrastructure reliability deliver significant quality of life improvements and economic benefits.
In governance and administration, tax codes, business registration procedures, import regulations, and administrative precedents are jurisdiction-specific. Document analysis systems trained on jurisdiction-specific regulatory frameworks can assist business registration, import/export procedures, tax compliance, and license applications. Research by the World Bank’s Doing Business project has documented how bureaucratic complexity creates significant barriers to formal business operation in many emerging economies. The World Bank’s Ease of Doing Business Index (archived) quantified these barriers across multiple dimensions. Reducing time and cost of compliance increases formalization rates and economic activity.
This represents a fundamental inversion of conventional advantage. Advanced economies excel at accumulating massive datasets across global populations - billions of internet users, millions of digitized medical records, extensive satellite coverage, comprehensive sensor networks. They have more data.
But emerging economies possess something potentially more valuable: the undigitized domain expertise of millions of practitioners working in local conditions. The miner who knows from experience how ore behaves in specific geological formations. The farmer who understands through generations of practice how crops respond to local microclimate variations. The healthcare worker who recognizes disease presentation patterns specific to local populations. The civil engineer who knows which road segments flood during specific weather conditions.
This knowledge exists but remains largely undigitized, unstructured, inaccessible to conventional AI approaches. Small language models trained on carefully curated local datasets - combining whatever digital records exist with structured interviews capturing practitioner expertise, with sensor data from local conditions, with verified outcome data - can encode this knowledge in deployable form.
Emerging economies possess something potentially more valuable than massive datasets: the undigitized domain expertise of millions of practitioners working in local conditions.
Several open model families support this approach with computational requirements accessible to emerging economy institutions. Meta’s Llama 3.1 (8B parameters), Mistral 7B (Apache 2.0 licensed), Microsoft Phi-3 (3.8B parameters), Alibaba Qwen 2.5 (0.5B to 72B parameters), and DeepSeek R1 (MIT licensed) all permit local deployment and fine-tuning. Models in the 7-13 billion parameter range require 64-128GB VRAM for full-precision inference, though 4-bit and 8-bit quantization reduces requirements to 16-32GB with modest performance degradation, as documented in research on quantization techniques for large language models and efficient LLM inference methods. For sufficiently narrow domains, CPU-only inference becomes viable, eliminating GPU requirements entirely.
These applications share defining characteristics: narrow functional scope optimized for specific domains, training data sourced from local conditions and verified for accuracy, computational requirements compatible with accessible hardware, and deployment architectures maintaining data sovereignty and operational independence.
They represent fundamentally different value proposition than frontier models. GPT-4 can answer questions about almost anything but with accuracy limitations and no guarantee of relevance to local contexts. A small model trained on Zambian mining data cannot discuss poetry or solve calculus problems - but it can optimize copper extraction in Zambian geological conditions with accuracy grounded in local empirical reality rather than global averages that may not apply.
The economic logic is compelling. A general-purpose model accessed through cloud subscription creates recurring foreign exchange outflows while delivering generic capabilities. A specialized model trained on local data and deployed on local infrastructure creates one-time capital expenditure in domestic or regional currency while delivering capabilities specifically optimized for local economic activities that generate domestic revenue.
This is where development impact materializes. Not in accessing ChatGPT through cloud subscriptions embedding foreign exchange exposure and geopolitical vulnerability, but in deploying specialized models solving local problems extraordinarily well using local knowledge on local infrastructure under sovereign control.
Strategic Divergence: The Leapfrog Opportunity
For four decades, emerging economies participated in globalization as rule-takers. Technology platforms, infrastructure standards, and service architectures originated in advanced economies; emerging economies adapted these with varying success. Development models assumed convergence - that emerging economies would progressively adopt solutions pioneered elsewhere, gradually closing capability gaps.
The AI revolution presents something different: the possibility of divergent paths optimized for distinct constraints rather than convergent paths toward a singular model.
Mobile telephony provides precedent. In the 1990s, conventional wisdom held countries needed robust landline infrastructure before mobile networks could succeed. Emerging economies proved this wrong by leapfrogging landlines entirely when mobile offered superior economics and better alignment with local conditions. Kenya’s M-Pesa, launched 2007, demonstrated mobile-first financial services could outperform branch-based banking. According to Central Bank of Kenya data, by 2024, over 80% of Kenya’s adult population used mobile money regularly - far higher than mobile banking penetration in most advanced economies. Research from MIT on M-Pesa’s economic impact documented how mobile money lifted approximately 194,000 Kenyan households out of poverty.
In mobile telephony, emerging economies won by leapfrogging. In AI, they may win by not racing at all.
AI may present comparable inflection point. Three converging realities create this possibility.
Open foundations enable capability without dependency. Unlike proprietary platforms of previous waves, foundational AI models are accessible, modifiable, deployable independently. You could not download and modify Windows or Oracle. But you can download Llama 3.1 or DeepSeek R1, fine-tune on local data, and deploy on your own hardware.
Geopolitical pressure makes sovereignty imperative. The Trump administration’s policies - semiconductor export controls, reduced development assistance, potential AI model export restrictions - explicitly confirm digital infrastructure as instrument of state power. This makes dependency on foreign infrastructure a strategic liability but validates alternative architectures. On-premises deployment optimized for local needs offers sovereignty, economic sustainability under foreign exchange constraints, and resilience to external pressure.
The AGI race may be irrelevant to development. Frontier models optimize for breadth - doing ten thousand things adequately. Development impact may reside in depth - doing specific things extraordinarily well using local knowledge that global models cannot access. The most economically valuable knowledge in emerging economy contexts is domain-specific, geographically specific, and rooted in local conditions that vary dramatically from global averages.
The question is not whether emerging economies can compete in the race toward artificial general intelligence - a competition requiring resources they lack and which U.S. policy now actively restricts them from accessing. The question is whether policymakers will recognize that not competing in that race may constitute the winning strategy.
The conventional framing says emerging economies are losing because they lack computational power, investment capital, and technical talent concentrated in the U.S. and China. This assumes a single race with a single finish line.
The real victory may lie in solving specific problems extraordinarily well rather than attempting everything adequately, in building sovereign capability rather than accepting dependency.
But if the relevant metric is development impact - progress toward Sustainable Development Goals, improvements in health outcomes, agricultural productivity, governance capacity, economic inclusion - then the race looks entirely different. Emerging economies may win precisely by not racing toward AGI, by not pursuing cloud-dependent architectures, by not attempting to replicate Silicon Valley or Shenzhen.
They may win by building something different: specialized models trained on local data, deployed on local infrastructure, solving local problems extraordinarily well, under sovereign control, immune to geopolitical pressure, economically sustainable without perpetual foreign exchange outflows or development assistance.
The apparent disadvantage - lack of hyperscale infrastructure, limited investment capital, distance from AI superpowers - may constitute hidden advantage if it forces different choices better aligned with development imperatives. Emerging economies possess what matters most for local economic value creation: deep domain expertise in local conditions, accumulated practitioner knowledge, and acute understanding of which problems most urgently need solving.
The apparent disadvantage -lack of hyperscale infrastructure, limited investment capital, distance from AI superpowers - may constitute hidden advantage if it forces different choices better aligned with development imperatives.
In mobile telephony, emerging economies won by leapfrogging. In AI, they may win by not racing at all - by recognizing that the competition commanding global attention is not the competition that matters for their populations. The real victory may lie in solving specific problems extraordinarily well rather than attempting everything adequately, in building sovereign capability rather than accepting dependency, in optimizing for local impact using local knowledge rather than competing on global benchmarks using global data.
Whether this potential is realized depends less on technology than on strategic choice - recognizing that apparent disadvantage may contain hidden advantage, and that the race everyone else is running may not be the race worth winning.
Rafal Rohozinski is the founder and CEO of Secdev Group, a senior fellow at the Centre for International Governance Innovation (CIGI), and co-chair of the Canadian AI Sovereignty and Innovation Cluster.