VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Super Intelligence | Small States and the Global South | Access, Capability and the Risk of an Intelligence Divide

Three students work together at a classroom table, reviewing open textbooks and notes.

Super Intelligence (SI) will not arrive into a flat world. Countries begin the AI transition with very different levels of compute, connectivity, capital, skills, language coverage, research capacity, public-sector capability and bargaining power. That starting inequality matters. The same model can be transformative in one country and marginal in another because the surrounding system determines whether capability can actually be absorbed, adapted and governed.

For small states and countries across the Global South, the central question is therefore not simply whether frontier AI becomes available. It is whether societies can turn access into local capability: the ability to select useful systems, adapt them to local needs, evaluate them, protect public interests, build businesses around them and retain enough institutional knowledge to avoid permanent dependency.

Super Intelligence Could Create an Intelligence Divide

The digital divide historically concerned access to computers, networks and information. Advanced AI adds a deeper layer. A society may have internet access yet lack the compute, local-language data, engineering talent, procurement capacity or research institutions needed to use powerful AI well. The divide is therefore not only between connected and unconnected populations. It can become a divide between countries that shape intelligent systems and countries that mostly consume them.

UNDP’s 2026 work describes this risk as a potential Next Great Divergence: AI can widen development gaps when countries experience disruption before they have the infrastructure, skills and governance capacity to capture the corresponding productivity gains.

The Adoption Gap Is Already Measurable

UNDP’s 2026 diffusion work reports a substantial difference in generative-AI use between the Global North and Global South. Its AI Diffusion Jobs and Tools Programme cites 27.5% usage among working-age people in the Global North versus 15.4% in the Global South, with the gap widening rather than automatically closing.

That matters because adoption compounds. Organisations that begin using AI earlier build internal knowledge about prompts, workflows, data architecture, verification and process redesign. The capability gap can therefore grow even when everyone eventually receives access to similar models.

Compute Is Highly Concentrated

ITU leadership highlighted in 2026 that only a small minority of countries host inference-grade or training-grade AI compute within their borders, while the Global South holds far less data-centre capacity than its share of global internet users. Stanford’s 2026 AI Index similarly shows extreme concentration in data centres, high-end chips and frontier-model development.

This does not mean every country needs a frontier training cluster. Most organisations do not. But lack of domestic or regionally accessible compute can increase cost, latency, dependency and vulnerability to external commercial or geopolitical decisions.

Small States Do Not Need to Win the Frontier-Model Race to Benefit

A small country cannot always match the capital expenditure of the largest technology powers. That does not make SI strategically irrelevant. Small states can specialise in adoption, domain integration, public-service transformation, safety evaluation, multilingual systems, trusted data, regulation, education and narrow high-value applications.

The strategic mistake would be treating frontier training as the only meaningful form of AI capability. Value can also be created through model adaptation, orchestration, domain data, evaluation, software integration and institutional redesign.

Singapore Shows One Small-State Path: Orchestrate Rather Than Imitate Scale

Singapore provides a useful small-state case because its AI strategy relies on dense infrastructure, strong institutions, enterprise adoption, public-sector capacity and international connectivity rather than attempting to reproduce every layer of the global AI supply chain domestically. Its 2026 Enterprise Compute Initiative, for example, supports access to cloud compute and implementation support for local companies rather than requiring each company to build its own data centre.

The broader principle is transferable: small states can build orchestration capability—the ability to connect external frontier intelligence with local data, policy, industries and public systems.

Local Language Is Infrastructure

AI systems perform best in languages and cultural contexts that are well represented in training and evaluation. ITU’s 2026 WSIS discussions on local-language AI note that billions of people speak languages underrepresented in digital systems. Weak language coverage can reduce accuracy, cultural fit and safety even when a model appears strong in English.

UNDP’s Every Language Matters initiative treats local-language capability as a full value chain involving data, institutions, funding, community participation and governance. This is the right frame: translation alone is not enough if concepts, local services and cultural references remain poorly represented.

Language Gaps Become Service Gaps

A citizen who cannot reliably use AI in their strongest language may receive poorer education, health information, agricultural advice, legal guidance or government support. When AI becomes embedded in services, language inequality can become functional inequality.

For SI, local-language evaluation should therefore include not only translation quality but task accuracy, cultural context, numerical reasoning, safety, dialect variation and performance on local institutions and documents.

Local Data Can Be More Valuable Than More Global Data

Frontier models know a great deal about the world, but many high-value tasks depend on current local data: laws, prices, school systems, government forms, transport routes, agricultural conditions, disease patterns and business records. Countries that build trustworthy data infrastructure can improve AI usefulness without training giant models.

Data quality, provenance, permission and update frequency matter more than merely accumulating documents. A local knowledge system becomes valuable when it can answer: Which source is authoritative? Which version is current? Who owns the data? Who can access it? How is an error corrected?

Public Digital Infrastructure Can Multiply AI Value

Digital identity, payments, registries, interoperable data systems and secure cloud platforms can make AI easier to integrate into public services. Countries without such foundations may have excellent chatbots while still being unable to automate real administrative workflows safely.

This is why SI readiness is partly an institutional infrastructure question. Intelligence must connect to systems of record before it can improve state capacity reliably.

SMEs Matter More in Many Global South Economies

UNDP notes that small and medium-sized enterprises account for more than 90% of businesses in many African and Latin American economies. That changes the diffusion problem. A strategy centred only on large enterprises misses where most firms and jobs actually sit.

SMEs often need low-cost, easy-to-deploy tools, local-language interfaces, trusted intermediaries and practical training. The bottleneck may be workflow redesign rather than model capability.

The Real Unit of Diffusion Is the Workflow

AI adoption statistics count users, but economic value appears when work changes. A farmer uses weather and market intelligence differently. A small exporter uses translation, documentation and customer research. A clinic uses AI for triage support or records. A school uses tutoring and planning. Each use case requires local integration.

SI diffusion should therefore be measured in completed workflows, productivity gains, error reduction and improved access—not only accounts created or prompts sent.

Education Determines Whether Countries Become Users or Builders

Broad AI literacy helps citizens use tools safely. Advanced technical education creates engineers, researchers, evaluators and entrepreneurs who can adapt systems. Domain education matters too because AI without local subject expertise can produce generic solutions disconnected from real needs.

Countries that only teach prompt use risk creating consumers. Countries that combine AI literacy with mathematics, computing, language, science, design and institutional knowledge can create local builders.

Research Capacity Is More Than Publishing Papers

Local research institutions can evaluate whether frontier systems work for local populations, languages and industries. They can identify failures that foreign benchmark suites miss. They also create a pipeline of people capable of negotiating with global technology providers from a position of technical understanding.

For small states, strong universities and applied research institutes can therefore be strategic leverage even when frontier-model training occurs elsewhere.

Open-Weight Models Can Reduce Some Access Barriers

Open-weight models allow countries, universities and firms to host, inspect and adapt systems without depending entirely on one provider’s API. That can support local languages, sensitive data and sovereign deployment.

The trade-off is infrastructure. Hosting open models still requires compute, engineering and operational security. Openness shifts the dependency rather than eliminating it.

Closed Frontier Models Can Offer Capability Without Infrastructure Ownership

Cloud APIs can give small organisations access to powerful models they could never train or host themselves. This can accelerate adoption dramatically. It also creates dependence on pricing, terms of service, data policies, model availability and network connectivity.

A resilient strategy can combine external frontier models with portable workflows and local alternatives rather than relying on a single access model.

Procurement Is a Strategic Capability

Governments and large institutions in small states may not build frontier models, but they can influence markets through procurement. Requirements around interoperability, data residency, auditability, local-language performance, accessibility and exit rights can shape how vendors design services.

Poor procurement can create lock-in. Strong procurement converts purchasing power into institutional capability.

Competition Policy Matters When Intelligence Becomes Infrastructure

If a few global firms control models, cloud infrastructure and distribution, smaller economies may face weak bargaining power. Competition policy, portability and interoperability can reduce dependency without requiring domestic replication of every layer.

The relevant question is contestability: can a user, business or government switch providers without rebuilding the entire workflow?

Local AI Companies Can Build Value Above Frontier Models

Most economic value may not come from training the base model. It can come from building trusted applications, domain datasets, workflow software, evaluation systems, compliance layers, customer support and industry-specific tools on top of frontier intelligence.

This creates space for companies in smaller markets to compete through fit rather than raw scale.

AI Diffusion Can Create New Jobs Before It Eliminates Old Ones

UNDP’s diffusion programme emphasises work around model adaptation, integration, evaluation, training and local service delivery. These are not necessarily frontier-research jobs. They are the jobs required to make AI useful inside real organisations.

For the Global South, capturing these diffusion-layer jobs may be as important as competing for a small number of frontier-model roles.

Labour Exposure Differs From Labour Benefit

A worker can be exposed to AI-driven task change without receiving access to the tools or training needed to benefit from it. That creates a dangerous asymmetry: disruption diffuses faster than capability.

Policy should therefore distinguish exposure, adoption and augmentation. The goal is not merely to estimate which jobs change, but whether workers gain the tools and skills to move upward in the new workflow.

Women, Rural Communities and Informal Workers Can Face Different Barriers

AI access is shaped by device ownership, connectivity, education, language, formal employment and payment access. National averages can therefore hide large internal gaps. A country may appear digitally connected while specific communities remain excluded from high-value AI use.

SI inclusion should be measured by who can obtain useful outcomes, not merely who can technically reach a chatbot.

Public-Service AI Can Produce Outsized Benefits in Capacity-Constrained States

Where administrative capacity is scarce, AI can help summarise case files, translate communications, draft documents, triage requests and support front-line staff. These uses can increase state capacity without requiring every citizen to become an AI expert.

The risk is that weak institutions may also have less ability to audit vendors, correct biased systems or maintain data quality. Capability and governance need to grow together.

Health Systems Can Benefit From Knowledge Access Before Full Automation

In regions with specialist shortages, SI can support clinicians through translation, literature synthesis, decision support and patient education. These benefits do not require autonomous diagnosis. Augmentation can be valuable even when professional authority remains human.

Local validation is essential because disease prevalence, treatment availability and health-system constraints differ across countries.

Education Can Become a Major Equaliser—If Language and Curriculum Fit

AI tutors can provide individualised practice at far lower marginal cost than one-to-one human tutoring. For countries with teacher shortages or large class sizes, this can be transformative. But the system must understand the local curriculum, examination structure, language and cultural context.

Generic global tutoring can widen gaps if affluent students receive locally adapted tools while others receive weaker, less contextual systems.

Agriculture Is a High-Value Localisation Test

Farmers need local weather, soil, crop, pest, market and input information. A globally capable model can provide general agronomy yet still fail if it lacks current local conditions. Agricultural SI therefore depends on data integration and local extension knowledge.

This demonstrates a larger principle: the last mile of intelligence is often local data and trusted delivery.

AI Sovereignty Does Not Require Complete Technological Autarky

“Sovereign AI” can mean many things: domestic compute, local-language models, protected public data, secure hosting, regulatory authority or the ability to switch providers. A small state may not control the entire AI stack and can still preserve meaningful strategic autonomy.

The more practical objective is selective sovereignty: identify which layers are critical enough to require domestic or trusted control and which can remain globally sourced.

Regional Cooperation Can Create Scale Without Erasing National Control

Neighbouring countries can share compute infrastructure, language resources, evaluation datasets, standards and research capacity. Regional cooperation can make projects economically viable that would be too expensive for one small state.

The design challenge is governance: who pays, who gets priority access, how data is protected and how benefits are distributed.

Data Centres Are Not Automatically Development

A country can host large data centres while capturing limited local value if ownership, engineering, research and downstream businesses remain elsewhere. Infrastructure can still create jobs, tax revenue and connectivity, but the development outcome depends on local linkages.

SI strategy should therefore ask not only “How much compute is located here?” but “What capabilities, firms and institutions grow around that compute?”

Foreign Investment Can Build Capacity or Create Dependency

Foreign AI investment can bring capital, cloud infrastructure, training and access to global markets. It can also create reliance on external providers if local firms and institutions never develop the skills to negotiate, adapt or replace the system.

The strongest development outcome usually combines external investment with local skill transfer, local suppliers, research partnerships and portable institutional knowledge.

Governance Capacity Is a Competitive Asset

Clear rules, efficient public administration, strong data protection, credible courts and technically capable regulators can make a small state attractive for trustworthy AI deployment. Governance is therefore not merely a brake on innovation. It can become part of the value proposition.

Small states often have one advantage here: institutions can sometimes coordinate more quickly because the number of major actors is smaller.

Speed of Coordination Can Partly Offset Lack of Scale

Large countries possess capital and research depth but can face complex coordination across agencies and regions. Small states may be able to align education, procurement, regulation, research and industry policy more quickly. The advantage is not automatic; it depends on institutional trust and execution capability.

In a fast-moving SI transition, response time can become strategic.

Small States Can Become Evaluation Hubs

Frontier developers need independent testing across languages, sectors and regulatory environments. A country with strong technical institutions and trusted governance can specialise in evaluation, certification, safety research and regulated testbeds.

This creates a high-value role between model creation and mass deployment.

Local Benchmarks Prevent Imported Blind Spots

A global benchmark may not test local laws, dialects, school curricula, social-service forms, medical pathways or business conditions. Countries that build their own evaluations can detect failures before imported models become embedded in high-consequence systems.

Evaluation capacity is therefore part of digital sovereignty.

Local Content Producers Matter to the Future Training Ecosystem

When local languages are poorly represented online, future models receive less training signal about those communities. Journalism, educational content, cultural archives, public records and research publications therefore become part of AI infrastructure.

Countries can strengthen future language and cultural representation by investing in digitisation, licensing frameworks and high-quality public knowledge resources.

Open Public Data Can Create Local AI Markets

Well-governed open datasets can help startups build applications in transport, environment, education, trade and urban planning. The value comes from structured, current and documented data rather than dumping raw records without maintenance.

Public data policy can therefore shape the local application layer even when the base model is foreign.

AI Development Finance Is Becoming a Distinct Policy Question

Infrastructure, language data, skills and public-sector modernisation all require investment. Yet many lower-income countries face tighter fiscal constraints precisely when the AI transition demands new capability. Development finance institutions may therefore increasingly fund compute access, digital public infrastructure, skills and responsible AI systems.

The quality of financing matters. Debt-heavy infrastructure without local value creation can worsen dependence rather than close the intelligence divide.

Small States Need Exit Options

A country can become operationally dependent on a provider whose prices, policy or availability change. Exit options include open standards, data portability, model abstraction layers, multi-cloud architectures and contracts that preserve access to records.

Resilience is easier to build before dependency becomes acute.

Cybersecurity Capacity Must Scale With AI Adoption

AI adoption expands the digital attack surface because more workflows, credentials and data move through automated systems. Smaller states may face shortages of cybersecurity professionals even as AI dependence grows.

SI can strengthen defence by helping with detection and analysis, but institutions still need identity, access control, incident response and secure infrastructure.

AI Can Strengthen or Weaken Administrative Legibility

A strong SI deployment can make public processes more legible: cases are logged, evidence is linked, delays are visible and decisions can be audited. A weak deployment can do the opposite by hiding administrative judgement behind a model output no one understands.

Small states should treat legibility as a design requirement, not an accidental by-product.

The Best Small-State Strategy Is Portfolio-Based

No single investment creates SI readiness. A resilient portfolio includes connectivity, cloud access, public data, cybersecurity, local-language resources, education, research, procurement, evaluation and industry adoption. Some elements may be domestic; others may be regional or internationally sourced.

The portfolio should reflect the country’s economic structure rather than copy the strategy of a much larger power.

Worked Example: A Small Island State

A small island state may lack the scale for frontier compute but possess strong tourism, shipping and public administration. Its SI strategy could prioritise resilient cloud links, local-language and local-law assistants, disaster response, port logistics, education and remote healthcare. Regional compute agreements can supply infrastructure that is inefficient to build domestically.

The success metric is not number of domestic foundation models. It is improved national capability and reduced dependency in critical functions.

Worked Example: A Fast-Growing African Economy

A country with a young workforce and rapidly expanding digital services may prioritise mobile-first AI, agriculture, payments, education, SME productivity and local-language systems. The key constraints may be electricity reliability, cloud cost, data availability and skills rather than demand.

Policy can focus on removing those bottlenecks while avoiding the assumption that every AI initiative should begin with a national large-language model.

Worked Example: Singapore as a Networked Small State

Singapore can combine global cloud and semiconductor networks with domestic regulation, public-sector digital systems, enterprise support, research and multilingual capability. Its advantage is not raw population or natural resources. It is the density of coordination across government, industry, universities and infrastructure.

That model cannot be copied mechanically, but it demonstrates that small size does not preclude strategic influence when institutions are strong.

What Should Countries Measure?

Useful indicators include AI adoption by income and region; compute access and cost; local-language performance; number of organisations with internal AI capability; availability of evaluators and cybersecurity professionals; domestic AI startups; public-sector deployment quality; research output; procurement maturity; and the ability to switch providers.

These indicators reveal whether the country is building capability or merely consuming services.

What Would a Closing Intelligence Divide Look Like?

It would not require every country to own equal compute. It would mean useful AI adoption spreading beyond wealthy firms; local languages and institutions performing well; domestic organisations building adaptation and evaluation skills; public services using AI responsibly; and countries having enough strategic autonomy to negotiate access rather than accept whatever is offered.

Convergence is about capability to benefit, not identical infrastructure.

What Would Worsening Divergence Look Like?

Warning signs include widening adoption gaps, declining bargaining power, local-language exclusion, AI-driven productivity concentrated in foreign firms, loss of skilled workers, dependence on inaccessible cloud systems, public institutions deploying tools they cannot audit and domestic firms becoming permanent consumers of imported intelligence.

These are measurable structural outcomes, not abstract fears.

Super Intelligence Changes the Development Question

If SI becomes dramatically more capable, the returns to local integration, data and institutional capacity rise rather than disappear. A very powerful model still needs to understand the country’s laws, systems, languages and priorities. The intelligence may be global; the problem it solves remains local.

This is why the Global South is not merely a late adopter in the SI story. It contains billions of people, languages, markets and problems that cannot be represented well without local participation.

RFE Closure: Diffusion Must Become Local Capability

The problem is unequal starting conditions that allow AI benefits, jobs and institutional capability to concentrate faster than the technology diffuses. The function of an inclusive SI strategy is to turn access into local capability across infrastructure, skills, language, data, organisations and governance. The receiver is the society trying to improve human development rather than merely consume advanced tools.

The exit condition is to change strategy when adoption rises but local capability does not: when businesses remain dependent, local languages remain weak, public institutions cannot audit the tools they use or citizens experience disruption without corresponding gains. SI diffusion is successful only when intelligence becomes usable, governable and productive where people actually live.

Continue the Super Intelligence (SI) Series

Continue to Super Intelligence | Scientific Discovery, then explore mathematics, software engineering, healthcare, energy and the other systems through which advanced intelligence could change civilisation.


Why Frontier Access Is Only the First Layer of SI Readiness

A country can purchase access to the best model in the world tomorrow and still fail to obtain meaningful economic value. The model may not connect to local systems, employees may not know how to verify outputs, data may be too fragmented to retrieve reliably, procurement rules may prevent integration and managers may redesign no workflows. This is the difference between model access and absorptive capacity.

Development economics has long recognised that technology transfer produces unequal outcomes because institutions, skills and complementary capital differ. SI intensifies the same principle. The intelligence layer can be globally available while the complementary systems required to turn intelligence into outcomes remain deeply local.

The Capability Stack for a Small State

A useful national SI stack can be pictured as nine layers: connectivity; compute access; secure cloud infrastructure; trusted data; language coverage; skilled people; organisations capable of redesigning work; evaluation and cybersecurity; and governance able to set rules and procure responsibly. Weakness at any layer can limit the value of stronger layers above it.

This stack prevents a common strategic mistake: spending heavily on visible AI applications while ignoring the quieter institutional plumbing that makes them reliable.

Connectivity Still Matters in the Age of Intelligence

Frontier AI often depends on reliable cloud connectivity. High latency, unstable networks or expensive data plans can make advanced systems unusable outside major cities. Edge models can reduce this dependency, but they require local hardware and engineering support.

For rural regions and island communities, AI inclusion can therefore remain an old-fashioned infrastructure problem. Better models do not remove the need for fibre, mobile networks, power and device access.

Electricity Reliability Is an AI Policy

AI discussions often begin with models and end with models. Yet every layer depends on electricity. Data centres, school devices, clinics, mobile networks and local servers all require reliable power. In countries with frequent outages, the effective availability of AI can be far lower than national connectivity statistics imply.

This makes energy planning part of SI readiness even for countries that never host frontier training clusters.

Cloud Credits Do Not Automatically Build Local Engineering Capability

Subsidised compute and cloud credits can accelerate experimentation. The deeper development outcome depends on what teams learn while using them. If the subsidy ends and organisations cannot maintain, evaluate or migrate the systems, the programme created temporary access rather than durable capability.

Well-designed support should therefore pair compute with architecture guidance, cybersecurity, training, evaluation and a pathway to sustainable operating cost.

The Last Mile of AI Is Organisational

Most organisations do not fail at AI because the model cannot write text. They fail because nobody owns the source data, the workflow is undocumented, exception cases are ignored, permissions are unclear and staff do not know which outputs must be reviewed.

This last mile can be a comparative advantage for countries with strong management, professional standards and coordinated institutions even if they do not train frontier models.

AI Literacy Should Be Role-Specific

A minister, teacher, software engineer, nurse, small-business owner and regulator do not need the same AI literacy. Generic awareness campaigns can produce shallow familiarity without usable competence. Role-specific training connects the technology to decisions people actually make.

For example, a teacher needs to understand verification, learning design and student agency; a regulator needs evidence standards and audit concepts; an SME owner needs workflow and data discipline. National capability emerges from these specialised literacies working together.

Universities Can Become National AI Translation Layers

Universities do more than conduct frontier research. They can test imported models, train public servants, support SMEs, build local-language datasets and convene industry with government. In a small state, one strong university or research institute can influence a large share of the national ecosystem.

This makes applied evaluation and extension work strategically important even when it produces fewer high-profile publications than frontier-model research.

Brain Drain Can Become Intelligence Drain

High-skill AI workers are globally mobile. Small states and lower-income countries can train talented people only to lose them to larger technology hubs. Remote work creates one countervailing possibility: skilled workers can participate in global AI markets while remaining locally based.

Retention therefore depends on research opportunity, compensation, infrastructure, quality of institutions and whether local organisations offer intellectually meaningful work.

Diaspora Networks Can Extend National Capability

Countries do not need every expert physically resident to benefit from their knowledge. Diaspora researchers and engineers can mentor teams, review policy, connect local firms to global markets and create research collaborations. Small states have long used networks to compensate for limited domestic scale.

SI makes those networks more important because frontier knowledge changes quickly and informal knowledge transfer can move faster than formal curricula.

Government Should Know When Not to Build

Public-sector enthusiasm can produce national platforms that duplicate mature commercial services at higher cost. Strategic capability includes deciding which layers genuinely require public ownership and which are better purchased competitively.

The build-versus-buy decision should consider sovereignty, security, market maturity, switching cost, local capability development and whether the function is truly strategic.

Government Should Also Know When Buying Is Not Enough

Some capabilities are too foundational to outsource completely. Identity, authoritative registries, public records, critical infrastructure controls and sensitive national data may require strong public stewardship even when commercial AI is used around them.

SI can sit above public systems of record without becoming the record itself.

Local Evaluation Markets Can Become New Industries

As AI regulation and procurement become more demanding, organisations will need testing for accuracy, safety, bias, cybersecurity, privacy, language quality and domain suitability. Independent evaluation can therefore become a local professional service sector.

Countries that develop credible evaluation institutions can export trust services even if they do not export frontier models.

Certification Can Fail If It Becomes a Checkbox Industry

Evaluation and certification create value only when they discriminate between stronger and weaker systems. If every vendor can obtain a badge through superficial documentation, certification becomes theatre.

Small states considering an evaluation-hub strategy therefore need technical depth, independence and mechanisms to update tests as models improve.

Local Language AI Needs Community Governance

Building datasets in underrepresented languages can involve sensitive cultural material, oral traditions or community-owned knowledge. Treating such data as a free raw material for model training can create legitimate concerns about consent, attribution and benefit sharing.

Local-language AI programmes should therefore involve speakers and communities in decisions about data collection, representation and acceptable use.

Translation Is Not Cultural Alignment

A sentence can be translated correctly while the system still misunderstands politeness, legal context, educational norms or local institutions. Cultural alignment requires more than multilingual tokens. It requires examples, evaluators and domain knowledge drawn from the society itself.

This is another area where local participation can create capability that global scale alone cannot supply.

AI Safety Capacity Should Not Be Concentrated Only in Frontier Countries

Countries that import AI systems still need to evaluate them. A model can be safe under one language and workflow while behaving differently elsewhere. Safety research therefore needs geographic and linguistic diversity.

International funding for evaluation labs, incident reporting and public-sector expertise can help prevent a world where only model-producing countries understand the systems everyone else uses.

Incident Reporting Should Capture Local Failure Modes

A globally significant AI incident database can miss failures that occur in local languages, informal economies or under-resourced public services. National and regional incident reporting can reveal problems that never reach international headlines.

Sharing anonymised patterns internationally can then improve global safety without requiring every incident detail to become public.

Informal Economies Need Different AI Interfaces

Many workers in developing economies operate outside large formal organisations. They may not have structured enterprise data or paid software suites. Mobile messaging, voice interfaces and low-cost assistants can therefore matter more than complex enterprise platforms.

SI diffusion strategies should follow how people actually work rather than assuming every economy looks like a multinational corporation.

Voice Can Be More Important Than Text

In communities where literacy levels vary or local languages have limited written digital content, speech interfaces can widen access. Voice systems also create challenges around accent recognition, privacy and the cost of high-quality speech data.

Multimodal SI may therefore reduce some literacy barriers while creating new evaluation requirements.

Public Trust Can Become a Scarce Resource

Countries with limited institutional trust may struggle to deploy AI in public services even when the technology is strong. Citizens can reasonably worry about surveillance, bias, data misuse or opaque decisions. Trust cannot be generated by a communications campaign after deployment.

It is built through transparent purpose, credible safeguards, visible correction and institutions that respond when systems fail.

AI Procurement Should Include an Exit Plan

Every major public or enterprise SI procurement should ask what happens if the vendor changes price, stops service, changes the model, is acquired or fails. Data export, documentation, interoperability and migration support should be planned before the contract creates deep dependency.

Exit planning turns resilience from an emergency response into an architectural property.

National AI Strategies Should Include Replacement Cycles

Models improve rapidly. A country that builds policy around one vendor or model generation can become obsolete quickly. Procurement and regulation should assume that systems will be replaced repeatedly.

Versioned architecture, benchmark suites and modular integration make replacement cheaper and reduce lock-in.

Digital Public Goods Can Lower Entry Costs

Open standards, open datasets, open-source tools and reusable public software can help smaller countries avoid rebuilding common components. International collaboration can therefore create shared infrastructure that leaves room for local adaptation.

The best shared resource is one that reduces duplication without forcing every country into the same policy or culture.

Regional Language Models May Have a Distinct Role

Not every language or region needs an independent frontier model. But regional models can be useful where data sovereignty, local-language performance or offline operation matters. They can also serve as specialists within a larger system that uses frontier models for general reasoning.

Hybrid architecture can therefore provide local control without requiring full frontier replication.

AI Capability Can Support Better Policy Analysis

Governments in small states often have limited specialist staff. SI can help compare policy options, summarise evidence, simulate scenarios and draft implementation plans. This can increase analytical capacity dramatically.

The risk is epistemic dependency: officials may accept model output without enough internal expertise to challenge assumptions. Public-sector augmentation should therefore strengthen evidence and staff capability rather than replace institutional judgement.

Strategic Foresight Becomes More Accessible

Advanced models can help smaller governments monitor global technological, economic and environmental trends that once required large analytical teams. This can improve preparedness for shocks.

Forecasts remain uncertain, so SI should support scenario planning rather than convert uncertain futures into apparently precise predictions.

AI Can Reduce Some Expertise Scarcity

A small country may have few specialists in rare diseases, tax law, industrial engineering or advanced software. SI can make expert-level information more accessible and help local professionals cover wider domains.

This does not remove the need for human experts in high-stakes cases. It can change how scarce expertise is distributed and when escalation is needed.

The Global South Is Not One Market

The phrase “Global South” covers countries with radically different incomes, languages, institutions, industries and digital readiness. Policy recommendations that fit India may not fit Laos, Peru, Kenya or Pacific island states.

The useful common theme is unequal starting conditions, not identical development needs.

Small States Are Not One Category Either

Some small states are wealthy financial or logistics hubs. Others are remote island economies with limited fiscal space. Some have strong universities; others rely heavily on imported expertise. Size creates certain coordination opportunities and scale constraints, but starting conditions still dominate strategy.

SI planning should therefore begin with a national capability map rather than a generic “small state playbook.”

A National Capability Map Should Find the Binding Constraint

For one country the bottleneck may be compute. For another it may be local-language data. For another it may be public procurement, broadband, cybersecurity or management. The highest-value intervention is the one that releases the current binding constraint rather than the most fashionable AI programme.

This is an operating-system view of development: fix the weak link that prevents the rest of the stack from delivering value.

AI Policy Needs Time Horizons

Short-term policy can focus on adoption, training and safe procurement. Medium-term policy can build data infrastructure, research and evaluation capacity. Long-term policy can address compute, education pipelines and strategic industries.

Trying to solve every layer at once can diffuse resources. Sequencing matters.

Metrics Should Track Capability Retention

After a pilot ends, does the organisation still know how to operate the system? After a foreign consultant leaves, can local staff maintain it? After a grant ends, can the service pay for inference? These questions measure whether capability was retained rather than rented temporarily.

Development success should therefore include institutional memory and operating sustainability.

The Risk of a Two-Speed Intelligence Economy

Within one country, large firms and elite schools may adopt frontier SI quickly while small firms, public schools and rural services lag. National averages can then improve even as internal inequality grows.

Diffusion policy should watch the distribution of capability, not only the frontier users.

Public Libraries and Community Institutions Can Become AI Access Points

Where households cannot afford premium tools, trusted public institutions can provide access, training and guidance. Libraries, schools, community centres and workforce agencies can play a role similar to earlier phases of internet diffusion.

Access paired with human support can be more valuable than access alone.

Benchmarking Against Peers Can Improve Strategy

Small states can compare themselves with countries that share similar scale, income or economic structure rather than only with the largest AI powers. Peer benchmarking can identify realistic improvements in cloud access, digital government, startup formation and skills.

The goal is to improve national capability relative to plausible peers, not to win an irrelevant global model-size contest.

Global AI Governance Should Include Diffusion Metrics

International AI debates often focus on frontier safety, which is important. They should also measure whether safe access, evaluation capacity and economic benefits are diffusing. A world in which a handful of countries govern frontier models while the rest remain dependent can create its own instability.

Safety and inclusion should therefore be measured in parallel rather than treated as competing agendas.

Local Agency Is the Core Development Outcome

The deepest objective is not national ownership of every technology. It is the ability to make informed choices: select providers, protect data, adapt systems, train people, negotiate contracts, regulate high-risk use and leave a vendor when necessary.

That is what turns imported intelligence into national capability.

Reality Check: SI Could Narrow Some Gaps While Widening Others

Low-cost intelligence can make high-quality tutoring, translation, legal information and business assistance available where expert labour is scarce. That is a powerful equalising force. At the same time, countries controlling models, compute and capital can capture outsized economic rents.

Both dynamics can occur at once. The correct policy response depends on which layer of the value chain is widening and which is narrowing.

Thresholds for a Healthy SI Diffusion Strategy

A country should ask whether AI adoption is accompanied by rising local productivity, stronger worker capability, better service access, improved language coverage, more domestic firms, stronger evaluation capacity and lower switching cost. If usage rises while these indicators stagnate, the country may be becoming a larger market without becoming more capable.

The strategy should change before dependency hardens.

Final Principle: Intelligence Must Reach the Institution That Can Use It

Super Intelligence can compress the cost of cognition, but societies still need institutions capable of absorbing that cognition. A village clinic, a ministry, an SME and a university laboratory each require different interfaces, data and governance.

The last mile of SI is therefore not the last mile of broadband. It is the last mile from model capability to a person or institution that can act on the result correctly.


Financing the Intelligence Layer Without Creating Fiscal Fragility

AI infrastructure can be capital-intensive, while many developing economies face limited fiscal space and competing needs in health, education, transport and energy. A national SI strategy therefore needs capital discipline. The question is not whether AI is important; it is whether each investment produces enough durable capability to justify what the country gives up elsewhere.

Public investment makes the most sense where the market underprovides shared infrastructure, research, language resources, evaluation or skills. Commercial services may be more appropriate where competition already delivers usable capability at lower cost.

Development Finance Should Fund Capability Chains, Not Isolated Pilots

An AI pilot can look impressive while depending on foreign consultants, temporary cloud credits and manually cleaned data that no agency maintains. Once the grant ends, the system disappears. Development finance creates more durable value when it funds the entire operating chain: data ownership, staff training, security, maintenance, evaluation and recurring inference cost.

The measure of success is whether the local institution can still operate and improve the system after external support declines.

Domestic Compute Can Be Small and Still Strategic

A small state may never train the world’s largest models, yet modest domestic compute can remain strategically useful for sensitive workloads, local model fine-tuning, research, emergency continuity and evaluation. The optimal scale should reflect actual workloads rather than prestige.

This creates a layered model: hyperscale frontier capability may be imported, while selected sensitive or latency-critical workloads run locally.

Inference Sovereignty May Matter More Than Training Sovereignty

Most organisations consume intelligence through inference rather than training. For many states, the strategic question is therefore whether important services can continue to run, whether data can remain protected and whether pricing or access can be changed unilaterally by an external provider.

Resilient inference can involve multi-provider contracts, local fallback models, cached knowledge, regional cloud capacity and emergency procedures.

Regional Compute Pools Can Convert Fragmented Demand Into Scale

Several countries may each have too little demand to justify a large facility but collectively support one. A regional compute pool can provide research and public-sector access while spreading capital cost. It can also support shared language or climate models relevant across borders.

Such systems require clear rules for allocation during peak demand, funding, cybersecurity, data residency and dispute resolution. Shared hardware without shared governance simply moves the bottleneck.

Public Institutions Need Their Own AI Product Management Capability

Governments often procure technology as fixed projects. SI systems behave more like continuously changing services. Models update, prompts drift, data changes, evaluations saturate and vendors alter interfaces. Public institutions therefore need product-management capacity: ownership of objectives, metrics, versions, incident response and user feedback over time.

This capability can be more important than writing the underlying model code.

Local Government Can Be an Important Diffusion Layer

National strategies set direction, but many citizen-facing services operate at municipal or regional level. Local governments understand transport, permits, waste, housing, community services and language needs differently from central ministries.

Small, well-scoped SI deployments can therefore emerge locally before becoming national infrastructure. Successful pilots should be designed for reuse rather than remain isolated demonstrations.

AI Sandboxes Can Reduce the Cost of Learning

Regulatory and operational sandboxes let institutions test systems with limited scope, real users and explicit safeguards before mass deployment. This is especially useful in small states where a national rollout can quickly affect a large share of the population.

A good sandbox has entry criteria, measurable hypotheses, incident reporting and a clear decision at the end: stop, repair, expand or redesign.

National AI Capability Needs a Repair System

Every deployed system will eventually fail somewhere. The strategic question is how quickly the country can detect the failure, identify ownership, correct data, update the model or vendor configuration and notify affected users. A country with fewer AI systems but faster repair can be more resilient than one with widespread deployment and weak recovery.

Repair capacity should therefore be measured alongside adoption.

Public Feedback Is a Sensor for Imported AI

Citizens encounter local edge cases that benchmark designers elsewhere will never see. Complaint channels, public feedback and professional reporting can reveal language failures, inaccessible interfaces, incorrect local law or culturally inappropriate recommendations.

Feedback becomes useful when it routes to someone who can actually change the system rather than disappearing into a generic support inbox.

National Champions Are One Option, Not a Requirement

Some countries may choose to support domestic AI firms as strategic companies. Others may rely on competition among global providers while investing in local applications and infrastructure. The best choice depends on market size, capital, talent and strategic objectives.

National capability should not be confused with the existence of one national AI brand. An ecosystem can be strong even when the frontier model is imported.

Local Ownership Can Exist at the Application Layer

A hospital may own the patient workflow, data architecture, evaluation and clinical governance while using an external model for reasoning. A logistics firm may own the optimisation data and customer relationship while renting inference. Strategic value can therefore sit above the base model.

This layered view helps smaller economies identify where ownership matters most.

National Data Assets Need Stewardship, Not Extraction

Public records, cultural archives, language corpora and scientific datasets can improve AI capability. Treating them as a one-time resource to be handed to external model developers may create little long-term value. Stewardship asks how access, licensing, attribution and public benefit will be managed over time.

Strategic data policy should preserve the ability to support multiple future models rather than lock the asset into one partnership.

Small States Can Move Faster on Shared Standards

Where government, universities and major industries have strong communication channels, a small state can standardise model evaluation, procurement clauses, incident reporting and public-sector architecture relatively quickly. These common standards reduce duplication across agencies.

The advantage disappears if coordination becomes centralised micromanagement. Shared infrastructure should make local teams faster, not less able to adapt.

AI Readiness Should Be Tested Through Stress Scenarios

What happens if the primary cloud provider goes offline? If inference prices triple? If a widely used model is withdrawn? If a local-language update degrades accuracy? If a major prompt-injection vulnerability appears? Stress scenarios reveal dependency before a crisis does.

National SI readiness is therefore partly the ability to maintain essential functions under degraded conditions.

Strategic Reserves Can Include Models, Data and Skills

Countries maintain reserves of fuel, medicine or financial liquidity because some resources are too important to leave entirely to normal market conditions. The AI equivalent need not be a giant sovereign model. It might include secure copies of open-weight models, critical datasets, documentation, fallback cloud capacity and trained personnel.

The appropriate reserve depends on which services would fail if normal access disappeared.

The Most Important Import Is Sometimes the Method

A country can import a model and remain dependent, or import the methods needed to evaluate, integrate and improve models. The second creates compounding capability because local teams can apply those methods to future systems.

Technology-transfer agreements should therefore value knowledge transfer, documentation and training alongside access to the immediate product.

AI Can Strengthen South–South Knowledge Networks

Countries facing similar agricultural, linguistic, climate or public-service challenges can exchange datasets, evaluations and workflow designs. An effective solution developed in one lower-resource environment may be more transferable to another than a system designed for a wealthy market.

SI can accelerate this exchange by lowering translation and adaptation costs.

Do Not Measure Success by Model Size

A smaller country can waste enormous resources chasing a frontier-model benchmark that has little connection to its development priorities. Model size is an input, not a national outcome.

Better metrics are service quality, business productivity, research capacity, language inclusion, cybersecurity, resilience and the share of the AI value chain captured locally.

Do Not Measure Success by Adoption Alone

High adoption can reflect useful diffusion or shallow dependence. If everyone uses foreign AI tools but domestic firms do not improve productivity, workers do not gain skill and public institutions cannot evaluate the systems, adoption may not translate into durable development.

The direction of capability accumulation matters more than the raw number of users.

The Strategic Question Is Who Learns From Every Deployment

When a ministry, school or company deploys SI, someone should capture what worked, what failed, which data was missing and which skills were required. If that knowledge remains only with an external vendor, the country pays for the deployment but loses the learning.

A mature ecosystem turns every project into institutional learning that improves the next one.

Final Development Test: Can the Country Say No?

The strongest test of strategic agency is whether a country can refuse a poor vendor, reject an unsafe deployment, switch providers, protect its data and choose a slower path when the evidence does not justify the risk. Dependence becomes visible when those choices disappear.

Super Intelligence should expand national option space rather than shrink it.

Closing Threshold: From AI Consumer to SI-Capable Society

A society crosses an important threshold when it no longer treats advanced AI as a mysterious external service. Public institutions know how to procure and audit it. Businesses know how to redesign workflows around it. Schools teach students to use and challenge it. Engineers can connect it to local systems. Researchers can test it on local problems. Citizens can understand where it is used and how to contest consequential decisions.

At that point, the country may still import frontier models, chips and cloud infrastructure, but it has stopped importing the entire decision architecture. It possesses enough knowledge to decide where SI belongs, where it does not, what evidence is required and how to recover when a provider or model fails.

The Strategic End State Is Capability With Choice

The deepest development goal is not technological independence in every component. It is capability with choice: enough infrastructure to participate, enough skill to adapt, enough evidence to evaluate, enough competition to switch, enough local knowledge to contextualise and enough institutional authority to refuse a deployment that does not serve the public interest.

If Super Intelligence eventually compresses the cost of cognitive work as dramatically as many expect, these complementary capabilities will determine whether that compression broadens human development or simply deepens dependence. Small states and the Global South do not need to reproduce the largest AI powers. They need to make sure that intelligence arriving from elsewhere can be converted into capacity that remains at home.

The final test is practical: when the next model generation arrives, can local institutions assess it, integrate it, compare alternatives, preserve their data, protect users and keep the gains they learn from deployment? If yes, the country is accumulating SI capability rather than merely renting intelligence.