AI Swing States: How Latin America Will Shape US-China Competition and Global AI Safety

Latin America will not decide the future of artificial intelligence by training the largest frontier model. At least not yet.

It will decide it in a different way: through the ecosystem choices it makes now.

Which cloud providers host public-sector AI? Which chips power national compute plans? Which open models are adapted for Spanish, Portuguese, and Indigenous languages? Which safety evaluations become mandatory in government procurement? Which data centers are approved, where they are located, how much water and energy they consume, and whether their contracts create local capacity or long-term dependence?

These choices may look technical. They are strategic.

The United States wants to export a full American AI technology stack: chips, servers, cloud systems, models, cybersecurity, data pipelines, and sectoral applications. China is building its own global AI offer around "AI Plus," open cooperation, digital infrastructure, open-source ecosystems, and Global South diplomacy. Europe is exporting regulation. Frontier labs are exporting model access. Cloud companies are exporting dependency -- unless governments negotiate otherwise.

Latin America should not respond by choosing a flag. It should respond by choosing the rules of the ecosystem.

"Latin America's leverage is not neutrality. It is conditional alignment: accept capital, infrastructure, models, and partnerships from any side only when they strengthen interoperability, auditability, local capacity, democratic control, environmental responsibility, and AI safety."

That is the central proposal of this article: Latin America should organize itself as a bloc of AI swing states.

An AI swing state is not a country that sits passively between Washington and Beijing. It is a country whose procurement, infrastructure, data, talent, and safety choices can swing the diffusion of AI ecosystems. Latin American governments may not control frontier labs, but they can control the conditions under which frontier AI enters schools, hospitals, courts, tax agencies, farms, banks, and critical infrastructure.

This argument extends a thread I've been pulling on for the last year -- in Governance as Advantage on open-weight AI and compute diplomacy, and in Human First, AI Frontier on Europe's institutional moment. The same logic applies south of the Rio Grande -- with different stakes, different leverage, and different languages.

Five Key Judgments

The Evidence: Adoption Without Leverage

Latin America is already an AI market. It is not yet an AI power.

The region ranks highly in mobile-app adoption: ECLAC's ILIA 2025 index, drawing on Sensor Tower data, reports that Latin America and the Caribbean hold a stable 15-20% global share of generative-AI mobile-app downloads [1]. But the same index shows deep structural gaps. Roughly 90% of regional supercomputing capacity is in Brazil; Brazil and Mexico account for about 68% of researchers; five countries produce around 90% of publications; and most national AI strategies lack budgets, implementation mechanisms, or evaluation systems [1].

This is the central paradox: Latin America is adopting AI faster than it is shaping AI.

WEF and McKinsey estimate that broader AI adoption could lift Latin America's productivity growth by 1.9-2.3% annually and generate $1.1-$1.7 trillion in additional yearly economic value -- modelled potential, not currently realized gains [2]. In their 2025 regional AI capabilities survey of approximately 129 senior respondents, only 23% reported any positive AI impact on EBIT, and only 6% reported more than a 5% improvement [2].

That gap is not mainly about enthusiasm. It is about institutions.

AI productivity does not appear automatically because a ministry buys a chatbot, a university licenses a model, or a bank deploys copilots. It appears when organizations redesign workflows, train operators, connect data, manage risk, measure impact, and prevent systems from becoming opaque infrastructure controlled by external vendors. This is the same pattern I described in Intelligence Is No Longer Scarce. Trust in It Is. -- the bottleneck is verification capacity, not model access.

This is where Latin America's choices become geopolitical. A weak state buys tools. A strategic state shapes markets.

The US Offer: Export the Full Stack

The United States has made its AI strategy explicit.

The U.S. AI Action Plan calls for driving global adoption of American AI systems, hardware, and standards; exporting American AI to allies and partners; countering Chinese influence in international governance bodies; strengthening compute export controls; and evaluating national security risks in frontier models [3].

A July 2025 Executive Order goes further. It directs the federal government to promote "full-stack American AI technology packages" abroad -- chips, servers, accelerators, storage, cloud systems, networking, data pipelines, labeling systems, models, cybersecurity, and applications in education, health, agriculture, transportation, and other sectors [4].

This is not simply technology export. It is institutional export.

A country that adopts a full-stack AI package may receive useful infrastructure, cloud access, security tools, and productivity gains. It may also -- and this is an analytical inference, not a claim made by the U.S. documents themselves -- inherit dependency on U.S. vendors, U.S. chips, U.S. licensing terms, U.S. export-control politics [9], U.S. standards, and U.S. assumptions about acceptable AI risk. The strategy documents define the offer; the dependency outcome depends on the conditions under which the offer is accepted.

The China Offer: AI Plus, Infrastructure, and Global South Diplomacy

China's AI strategy is also explicit.

In August 2025, China's State Council issued guidance to accelerate the "AI Plus" initiative across science and technology, industrial development, consumption, public well-being, governance, and global cooperation. The guideline calls for strengthening AI models, data supply, intelligent computing power, open-source ecosystems, and talent teams. It targets more than 70% penetration of new-generation intelligent terminals and AI agents by 2027, and more than 90% by 2030 [5].

China's Global AI Governance Action Plan, published at the 2025 World Artificial Intelligence Conference, frames AI as an international public good and calls for digital infrastructure, global adoption, open cooperation, safety, controllability, fairness, and respect for national sovereignty [6].

This is a powerful message for the Global South. China can offer infrastructure, financing, industrial policy experience, open-source models, telecom integration, and a diplomatic narrative that criticizes AI monopolies by a small number of countries and firms.

But China's offer is also strategic. A country that accepts Chinese AI infrastructure may gain access and affordability. It may also -- again, as an analytical inference rather than a claim made by the official Chinese documents -- inherit dependencies in telecoms, cloud, surveillance architectures, data governance, industrial standards, and geopolitical exposure to U.S. export controls [9]. The dependency risk in either direction is conditional, not automatic; it materializes when full-stack offers are adopted without portability, auditability, and local-capacity safeguards.

Again, the answer is not rejection. The answer is conditionality. Latin America should make both Washington and Beijing compete to satisfy regional rules: interoperability, safety testing, public-interest access, environmental reporting, and local technical transfer.

The Third Force: Regional Models and Cultural Infrastructure

The most important AI infrastructure for Latin America may not be a hyperscale data center. It may be a corpus.

Latam-GPT is an early signal. Launched by Chile's National Center for Artificial Intelligence (CENIA), with support from more than 30 institutions across eight Latin American countries, it was designed as an open-source language model trained on regional cultures and data [10]. Its initial release focuses on Spanish and Portuguese; Indigenous-language coverage is planned for later phases, not yet supported at launch.

The lesson is not that Latam-GPT will compete with the largest global frontier models. It probably will not. The lesson is that regional AI infrastructure creates bargaining power.

A country that understands how models are built is better positioned to regulate them. A region that develops its own datasets can evaluate foreign systems more effectively. A public sector that can run smaller open models for specific tasks is less dependent on opaque APIs. A school system, court system, or health system that can test models in Spanish, Portuguese, Quechua, Guarani, Aymara, Rapa Nui, or Caribbean contexts can detect failures that English-centric benchmarks miss.

"Cultural infrastructure is safety infrastructure."

This is the same logic I applied to Spain's reference personas: a country -- or in this case, a region -- that builds its own evaluation substrate is no longer a price-taker on what "good performance" means. The benchmark is the leverage.

The Data Center Question: Sovereignty, Water, Energy, and Lock-In

AI is turning data centers into strategic assets.

Latin America is becoming more attractive for digital infrastructure because of cloud demand, AI growth, nearshoring, renewable energy potential, and proximity to U.S. markets. Brazil and Mexico are emerging as anchor markets: White & Case reports Brazil accounts for more than 40% of announced data-center capital investment in the region, with Querétaro emerging as a major Mexican hub [15]. CBRE's regional tracker lists São Paulo, Santiago, Bogotá, and Querétaro as the four largest Latin American data-center markets by inventory, with São Paulo leading at roughly 493 MW in Q1 2025 [16].

But data centers are not automatically sovereign infrastructure. They can become extractive infrastructure.

A country can host data centers and still lack compute access for universities, startups, public-interest research, local safety testing, and national development priorities. It can supply electricity and water while foreign firms capture most of the strategic value. It can approve "AI infrastructure" without knowing whether the facility serves domestic innovation or merely offshore demand.

The same principle applies to critical minerals. Latin America supplies roughly 35% of the world's lithium and 40% of its copper, making it essential to the energy transition and to AI infrastructure supply chains [13]. But mineral wealth becomes AI leverage only if it is linked to responsible supply chains, traceability, industrial upgrading, and digital infrastructure strategy.

The region should not export minerals, import cloud, and call that digital transformation.

Why This Matters for Global AI Safety

Most AI safety debates still focus on frontier labs, model weights, evaluation protocols, and catastrophic misuse scenarios. That focus is necessary but incomplete.

AI safety will also be decided in deployment regions.

A model that performs well in English may fail in Spanish-language legal reasoning, Indigenous-language public services, local medical triage, agricultural extension, disaster response, or social-benefit eligibility. A public agency may deploy an AI system without version control, audit logs, appeal rights, or incident reporting. A police force may procure a predictive system without bias testing. A hospital may use a clinical assistant without local validation. A school system may deploy tutors without child-safety evaluation.

These are not abstract concerns. The 2026 International AI Safety Report -- a broad expert review led by Yoshua Bengio with more than 100 contributors -- notes that general-purpose AI systems are becoming more capable in coding, mathematics, and autonomous tasks, while risks include malicious use, malfunction, systemic effects, cyber capabilities, biological and chemical misuse, and open-weight models whose safeguards can be removed and whose weights cannot be recalled once released [7]. Those are global risk categories. Latin America does not yet have a comparable regional incident-evidence base; building one is part of the proposal in this article.

Latin America's contribution to global AI safety should therefore be practical: build the institutions that test, monitor, and govern AI where it touches people. Latin American AI governance is still nascent, and many regional bills remain vague or lack actionable measures. Regional initiatives should evolve from declarations of intent into concrete commitments, accountability, and context-specific safety institutions [8].

Proposal: The Latin American AI Swing States Compact

Latin America should launch a Latin American AI Swing States Compact: a voluntary but operational agreement among governments, universities, development banks, AI labs, civil society, and private-sector partners.

Its purpose would not be to create another declaration of principles. The region has enough principles. Its purpose would be to define the minimum technical, institutional, and procurement conditions under which AI ecosystems enter Latin American public life.

To move beyond slogan, each pillar of the Compact needs a lead institution, a funding source, a first-year deliverable, and an honest implementation risk. A first sketch:

The point of the table is not to fix every parameter. It is to remind that each pillar fails without naming who leads, who pays, and what ships in year one.

1. Create a Regional AI Safety and Evaluation Network

The region should establish a Latin American AI Safety and Evaluation Network -- a Red Latinoamericana de Evaluación y Seguridad de IA.

Not one centralized institute. A distributed system of specialized nodes:

• Brazil: compute-intensive evaluation, industrial AI, cybersecurity, frontier-model risk.
• Chile: language evaluation, cultural datasets, public-interest model infrastructure.
• Mexico: cross-border supply chains, nearshoring, U.S.-linked AI infrastructure standards.
• Colombia and Uruguay: public-sector deployment, civic participation, rights-based evaluation.
• Caribbean and Central America: disaster response, climate risk, public services, small-state capacity.

The network should publish an annual Latin American AI Safety Report, aligned with international AI safety work but focused on regional risks: Spanish and Portuguese model failures, Indigenous-language exclusion, public-sector automation, data protection gaps, cyber vulnerability, election integrity, environmental costs, and AI use in security systems.

2. Adopt a Public AI Procurement Standard

The fastest way to shape AI markets is procurement. Every public agency buying or deploying AI above a defined risk threshold should require:

• A system card or model card.
• Local-language performance evaluation.
• Human-rights and public-impact assessment.
• Audit logs accessible to authorized oversight bodies.
• Incident reporting within a defined period.
• Version-control and rollback mechanisms.
• Clear rules on data use, retention, transfer, and model training.
• Exit clauses that allow the agency to switch providers.
• Interoperability with alternative cloud and model providers.
• Disclosure of energy and water footprint for compute-intensive services.
• Human authority over high-impact decisions.

This is not anti-innovation. It is pro-competition. Bad procurement locks governments into opaque systems. Good procurement forces vendors to compete on reliability, transparency, portability, and safety.

3. Build a Regional Model and Dataset Commons

Latam-GPT should be treated as a beginning, not a one-off project. Latin America needs a regional model and dataset commons that supports Spanish, Portuguese, Caribbean linguistic diversity, and Indigenous languages. It should include public-domain corpora, licensed educational material, government service datasets where appropriate, cultural datasets, safety benchmarks, and domain-specific evaluation sets for health, education, justice, agriculture, and climate.

The goal is not to replace global models. The goal is to make global models accountable.

A regional evaluation set allows governments to ask: does this system work here, for our people, in our languages, under our laws, with our failure modes?

4. Finance Compute With Public-Interest Conditions

Development banks, national banks, and regional institutions should create a compute-financing platform for AI infrastructure -- but the financing should be conditional.

If public incentives, tax benefits, concessional loans, or public land support AI data centers, the resulting infrastructure should reserve capacity for public-interest uses: universities, startups, hospitals, public agencies, safety testing, and regional model development.

No public subsidy should create private compute enclaves with no domestic capability spillover.

5. Use "Two-Door Diplomacy" With the U.S. and China

Latin America should avoid binary alignment. Instead, it should use two-door diplomacy.

Through one door, the region accepts U.S., European, and allied investment in cloud, chips, standards, cybersecurity, and sectoral applications. Through the other door, it accepts Chinese and other Global South partnerships in infrastructure, open-source systems, industrial AI, and applied deployment.

But both doors should lead into the same room: Latin American rules.

If a U.S. vendor passes those tests, buy from the U.S. If a Chinese vendor passes them, buy from China. If a European, Brazilian, Chilean, Mexican, Indian, Japanese, or open-source provider passes them, buy from them. The strategic objective is not neutrality. It is leverage.

6. Link AI Safety to AI Adoption

Latin America's AI opportunity is not only in building models. It is in applying AI safely to sectors where the region has urgent needs: health access, education quality, public administration, tax compliance, climate adaptation, agriculture, logistics, judicial backlogs, disaster response, and small-business productivity.

Brazil's "AI for the Good of All" plan proposes around USD 4 billion across roughly 54 measures covering infrastructure, dissemination, training, public services, business innovation, and regulatory and governance processes [11]. The figure is a plan, not yet disbursed spending; tracking actual execution will matter as much as the announcement. Chile has updated its national AI policy and proposed a risk-based AI bill after applying UNESCO's Readiness Assessment Methodology [12]. Colombia approved a national AI policy through CONPES 4144 in February 2025 [14].

These are promising moves. But national strategies will matter only if they become operational systems: budgets, procurement standards, evaluation labs, training programs, implementation dashboards, and public accountability.

"A strategy without execution is a PDF. A procurement rule changes the market."

7. Build an AI Operator Corps

The bottleneck is not only AI researchers. It is operators.

Latin America needs people who can translate AI capability into institutional performance: civil servants who understand evaluation, teachers who can use AI tutors safely, doctors who understand clinical decision support, judges who understand algorithmic evidence, municipal officials who can procure systems responsibly, and engineers who can maintain AI infrastructure.

A regional AI Operator Corps could train thousands of professionals each year in practical AI deployment, safety evaluation, procurement, and workflow redesign. This connects directly to the work I've described in Democratizing AI: Scaling Education Across 12 Countries -- the Saturdays.AI experience showed that distributed, applied AI training works at scale across Latin America when institutions commit to it.

This matters because AI failure is often organizational, not technical. The model may be good, but the institution may be unprepared.

Country Archetypes: Where the Swing States Are

Not every country has the same role. The region's strength will come from combining them.

The honest premise is that Latin America is heterogeneous. ILIA's own categories distinguish pioneers, adopters, and explorers; countries differ sharply in compute capacity, institutional quality, fiscal space, procurement law, cloud-market maturity, data-protection regimes, foreign-policy alignment, and electricity and water constraints [1]. A "swing states" Compact does not require uniform adoption; it requires interoperable rules among willing first movers, with on-ramps for others.

• Brazil is the anchor. It has scale, research capacity, industrial policy tools, data center gravity, and an ambitious national AI plan. Its choices will influence the region's compute, regulation, and industrial AI trajectory.
• Chile is the governance and cultural-infrastructure node. Its updated AI policy, UNESCO-linked readiness work, CENIA, and Latam-GPT position it to lead on evaluation, language, and responsible public-sector deployment.
• Mexico is the nearshoring and North American interface. Its AI infrastructure choices will be shaped by U.S. proximity, Chinese supply chains, data centers, manufacturing, and trade politics.
• Colombia can become a public-sector AI governance laboratory, especially if its national AI policy translates into procurement rules, civic participation, and measurable implementation.
• Uruguay, Argentina, Costa Rica, Panama, the Dominican Republic, Ecuador and others can lead in agile deployment, digital government, education, and regional interoperability -- if they avoid becoming passive buyers of foreign systems.
• The Caribbean and Central America should not be treated as afterthoughts. Smaller states face distinctive AI safety needs: climate risk, disaster response, public-sector capacity, tourism dependence, language diversity, cybersecurity vulnerability, and cross-border data governance.

The Strategic Mistake to Avoid

The biggest mistake would be to treat AI policy as a race to attract vendors.

That approach leads to weak bargaining. Governments offer tax incentives, expedited permits, data access, public contracts, and political legitimacy. In return, they receive cloud credits, pilot projects, memoranda of understanding, and impressive launch events. Then the dependency begins.

A better approach is to treat AI policy as ecosystem design. Latin America should ask every vendor, lab, and government partner:

• Will your system make us more capable after the contract ends?
• Can our universities evaluate it?
• Can our regulators inspect it?
• Can our startups build on it?
• Can our citizens challenge it?
• Can our public agencies switch away from it?
• Can our languages and cultures be represented?
• Can our energy and water systems sustain it?
• Can our safety institutions learn from it?

If the answer is no, the deal is not strategic.

The Bottom Line

Latin America's AI future will not be determined only in Silicon Valley, Washington, Beijing, Brussels, or Shenzhen. It will be determined in procurement offices, data protection agencies, university labs, development banks, ministries of education, energy regulators, data center permits, public datasets, and regional safety networks.

The region should not ask: "Should we align with the United States or China?"

It should ask: "What conditions must any AI ecosystem satisfy to operate in our societies?"

Latin America can be a market for other people's AI systems.

Or it can become a bloc of AI swing states.

The difference will be decided now.