Why the US AI export offer misses Latin American demand, and how China and the US each meet it
Note: This essay was originally written in Russian and translated into English with the help of Claude Opus 5.5, as English is my second language. The analysis, arguments and conclusions are my own.
The United States has built its AI export strategy around the frontier and the full technology stack: the goal is to make American technology the standard that other countries build on and stay tied to. This essay looks at that strategy from the buyer's side – 20 Latin American and Caribbean countries that do not build frontier AI themselves – and compares how the US and China meet their demand.
The main argument is that the American offer does not contradict the region's demand as a whole, but it diverges from it exactly where demand is most acute. In capital, compute infrastructure and the developer ecosystem, the US remains the strongest supplier. But regional demand has two layers. A broad mass layer – public services, business, agriculture – looks for low cost, sufficient quality and control, and here Chinese open-weight models and bundled packages fit better. A narrow critical layer – cyber defense and national security – genuinely needs the frontier, yet this is precisely where American access is least reliable, as the June 2026 suspension of Anthropic's models showed. The financing promised by the American AI Exports Program has so far barely reached the region.
Countries respond to the rivalry not by choosing a side but by mixed sourcing across layers – the same way they vote with the West on values and with China on economics at the UN. Political anchoring to the US shapes this choice less than one might expect: strongly and weakly anchored countries mix suppliers in similar ways. Mixing, however, only protects against being switched off. It does not settle the main question: whether a country becomes a co-owner of AI infrastructure or remains a host site collecting rent for land and energy.
In July 2025 the US administration set out its AI strategy in America's AI Action Plan and Executive Order 14320 of 23 July 2025, "Promoting the Export of the American AI Technology Stack". The essence of the strategy is to lock in American leadership in AI and convert it into a lasting economic and political advantage. The main method is to make the American AI stack the global standard: the more countries build their AI infrastructure on it, the larger the market for American companies and the less room there is for the technologies of strategic rivals, above all China. In this logic exports are not merely trade but a way of extending the American ecosystem beyond the country's borders.
The strategy's main instrument is the Commerce Department's American AI Exports Program. It works from the demand side, promoting "full-stack American AI technology packages" abroad. Proposals are submitted by industry consortia – groups of companies led by a single "anchor" firm. Foreign firms can join only with restrictions, and each proposal must explain how it advances US national interests. Partner countries are not members of the consortia; they are the buyers. The first application window, open from 1 April to 30 June 2026, was for "pre-set" consortia required to offer the whole five-layer stack: AI-optimized hardware, data pipelines and labeling, models, cybersecurity tools and sector applications (DLA Piper). The next stage envisages "on-demand" consortia assembled for a specific deal and offering only the layers it needs. Approved packages receive government-to-government advocacy, priority export-licence review and help accessing federal financing through EXIM and DFC.
The program's logic is ecosystem integration. A package promises the buyer compatibility across layers, a single accountable supplier and lower integration risk. The flip side is the same thing: the deeper a country is embedded in the American ecosystem, the more expensive it is to switch. What the package does not promise matters too: it does not promise low cost. Its arguments are integration and security, not the price of components. As the following sections show, this is where the seller's logic starts to diverge from the buyer's.
The second element is Pax Silica, a State Department framework on AI and semiconductor supply chains adopted on 12 December 2025 (State Department fact sheet). Formally it is a non-binding declaration without enforcement mechanisms. Its stated purpose is to reduce allies' "coercive dependencies", above all on Chinese suppliers, in critical links: minerals and processing, energy, semiconductors, and compute and digital infrastructure. Concrete content is still thin: supply-chain mapping, joint investment and a US pledge, announced in March 2026, to put $250 million into an investment consortium for energy and critical-mineral supply chains. At the second summit on 26 June 2026, Argentina, Chile, Costa Rica, El Salvador and Panama joined the declaration (State Department). Pax Silica is therefore better understood as a political framework and a coalition signal than as an instrument with its own levers: Washington can apply pressure not through the declaration itself but through what stands behind it – export licences, access to the US market and financing.
The two elements assign Latin America different roles. In the export program's logic the region is a market for the American stack. In Pax Silica's logic it is part of the stack's resource base. Among the Latin American signatories are Chile, the world's largest copper producer and one of the leading holders of lithium reserves, and Argentina, the second pillar of the so-called lithium triangle (USGS). For Washington the region matters not only as a buyer but also as a source of raw materials for the very stack it is being invited to buy.
This combination is familiar to the region from long before AI. Raúl Prebisch, the first executive secretary of CEPAL – the organization that today publishes the ILIA index – formulated with Hans Singer the thesis that a periphery exporting raw materials and importing the core's manufactured goods loses out in its terms of trade over time. Immanuel Wallerstein generalized this structure into the core–periphery model of the world-system. It would be wrong to claim that the American strategy deliberately reproduces this pattern: this is an interpretation of the architecture of two instruments, not their stated aim. Pax Silica formally includes cooperation on mineral processing, leaving room for value added on the region's side. But if processing stays on paper, the combination of "buy our stack and supply the raw materials for it" will reproduce exactly the structure of exchange that Latin American structuralism has argued against since the mid-twentieth century.
The American focus on the frontier and the full stack is not a misperception. It follows from a consistent chain of reasoning.
The US invested earlier and more heavily than anyone else in AI. American companies financed research for decades, built the largest compute clusters first and now spend hundreds of billions of dollars a year on infrastructure. Consequently the US has the best closed models: by most assessments American labs hold the frontier. A caveat is needed straight away: Chinese open-weight models lag that frontier by only two to five months, and the buyer's logic builds on this.
If the American advantage lies in quality, it is reasonable to build the strategy on superior quality. But quality does not matter equally for every task. It matters most where outcomes depend on capabilities relative to an adversary – in cybersecurity, intelligence and defense. If the attacker has a stronger model than the defender, the defense loses, and "good enough" stops being good enough. The frontier is therefore not only a commercial product but a dual-use technology: capabilities useful for cyber defense are also usable for attack. The next step follows: the frontier can only be distributed under control, choosing who gets access and on what terms. The full stack is convenient precisely for this: American clouds, models and security tools in one package leave the supplier able to exercise control.
Industry economics add to the security logic. Hundreds of billions of dollars of investment only pay off if there is paying demand outside the US for expensive frontier services and customers stay inside the ecosystem. Exporting the full stack is a way to create and retain that demand.
If the export policy works, it should set off a compounding lead: revenue and data from foreign customers feed new compute and new models, the lead grows, and countries that settle for "good enough" today will find it ever harder to catch up. It is widely believed in Washington that this self-reinforcing dynamic is what distinguishes AI from earlier technologies.
The same logic shapes the American understanding of sovereignty. On 20 February 2026 in New Delhi, Michael Kratsios, director of the White House Office of Science and Technology Policy, said that "real AI sovereignty means owning and using best-in-class technology for the benefit of your people" and "charting your own national destiny" (White House, a US government source). He urged countries to pursue strategic autonomy alongside rapid AI adoption rather than full self-sufficiency. The White House release added that running on the American stack lets countries keep sensitive data at home.
In this definition sovereignty is defined through owning the best technology and through autonomy without self-sufficiency. Since the best technology is, by the American assessment, American, the definition is not merely compatible with dependence on the US – in practice it presupposes it. This is sovereignty compatible with dependence.
There is an asymmetry here. For itself, the US understands technological sovereignty primarily as control: export restrictions on advanced chips, restrictions on Chinese technology in its own infrastructure and, as the Fable episode would show, export controls on its own models. Partners, by contrast, are invited to understand sovereignty as access to the best technology. In the same speech Kratsios said that AI adoption would not bring a bright future if it were "subject to bureaucracies and centralized control". Four months later the Commerce Department switched off a frontier model for all foreign users with a single letter.
A country that does not build the frontier solves a different problem from the country that sells it. Model quality matters, but it is not the first priority. What matters is the balance of three parameters: cost, sufficiency for the task at hand, and control.
This triad is not speculative: sources of quite different kinds arrive at it independently. The Latin American AI Index ILIA 2026, produced by CEPAL and Chile's CENIA, defines technological sovereignty not as self-sufficiency but as preserving choices. The Brazilian government, splitting its AI infrastructure package in August 2026 between Huawei and, probably, Nvidia, explained that its strategy was "not to depend on a single company, technology or country" (China-Global South Project). The US-based Stimson Center (Jesse Marks and Mathew Burrows, 2026) describes buyers in the market for ordinary AI adoption through four requirements: adequate performance, affordable price, local control and confidence that access will not be revoked. A regional international organization, the government of the region's largest country and a Washington think tank arrive at the same picture of demand.
Each of the three parameters needs explanation.
Sufficiency means that most applied tasks of governments and businesses do not require the best model in the world. Document processing, citizen advice on public services, agronomic recommendations and translation into local languages are all handled well by models a few months behind the frontier. Quality requirements here are threshold requirements: a model either handles the task or it does not, and further gains in quality add almost nothing.
Under these conditions cost becomes decisive. If two models both handle a task and one is several times cheaper, the choice is obvious.
There is an important exception to this rule. Threshold logic works where a model solves a task on its own and stops working where the outcome depends on the balance of forces with an adversary – above all in cybersecurity and defense. Protecting the information systems of a state, its banks or its power grids is measured not against an abstract threshold but against the attacker's capabilities: if the attacker's model is stronger than the defender's, a "good enough" defense loses. In adversarial domains there is no "good enough". Here the priority of quality returns, and here the buyer genuinely needs the frontier. This is the same dual-use logic on which the US builds its strategy, seen from the buyer's side: for the buyer the frontier matters not everywhere but in a narrow yet critical area.
Control is the most complex of the three parameters because it is not binary but a spectrum. At the bottom, data are stored in the country but on a foreign company's cloud, and the company continues to run both the infrastructure and the model. Higher up is a government cloud, where the state controls the infrastructure but still uses someone else's model through an external service. At the top, an open-weight model runs on the country's own hardware: the country can adapt and audit it and keep using it even if relations with the developer break down. Different countries, and different agencies within one country, sit at different points on this spectrum, and, as shown below, this largely determines their choice of supplier.
Put the three parameters together and the logic of choice becomes clear, but it has two layers. For most applied tasks a model that is sufficient, several times cheaper and runnable on one's own hardware satisfies all three requirements at once, while a closed frontier model wins only on quality, which is not the priority here. In cybersecurity and defense the balance is reversed: quality becomes the first requirement, and the buyer has to accept a higher price and less control. Regional demand for AI is therefore dual-layer. A broad mass layer is oriented toward price and control; a narrow critical layer toward the frontier. The American offer is built around the second layer, the Chinese around the first. It also matters that the risk of being switched off, discussed in section 1.4, bites hardest in the critical layer: where the frontier is irreplaceable, losing it costs the most.
Usage data show that such choices are actually being made. The OpenRouter and a16z study "State of AI: An Empirical 100 Trillion Token Study" analysed more than 100 trillion tokens processed through OpenRouter between November 2024 and November 2025. Over that period the share of Chinese open-weight models rose from 1.2% of weekly volume at the end of 2024 to almost 30% in some weeks by late 2025, averaging about 13%. The shift began in late 2024, before the reliability of American suppliers became a political issue.
These data have an important limitation. OpenRouter users are mostly developers and start-ups who pick a model per request on price and quality and can switch in minutes. States procure differently: through tenders, with requirements on data storage, support and supplier liability, and for years ahead. The OpenRouter data therefore describe the developer market, not state demand. But the developer market matters for this essay too: it shows the lower, spontaneous layer of demand that forms without political decisions. A CSIS report on AI partnerships in the Global South describes the same layer: developers in Kenya, Nigeria and Uganda build products on Chinese open models, citing price and quality in local languages. How this spontaneous layer relates to state procurement is discussed in section 5.2.
Analysts who follow the American strategy see the gap between the seller's and the buyer's logic from three angles.
The first concerns where competition actually takes place. Marianne Lu and Sam Winter-Levy (Carnegie, July 2025) show that for the Global South the decisive race is for diffusion and adoption, not for the frontier: it will be won by whoever offers reliable infrastructure, applied solutions and financing. In this race American leadership is far less secure than in the race for model quality. Marks and Burrows specify the mechanism: full-stack lock-in works in the frontier market, where alternatives are few, but not in the market for ordinary adoption, where buyers assemble hybrid systems from different sources and keep the option to switch.
The second is a dispute over the concept of sovereignty itself. Kratsios's formula, which defines sovereignty through owning the best technology, is directly opposed by Ruby Scanlon and Vivek Chilukuri of CNAS (April 2026): countries "are not looking to become passive consumers of American AI, or worse, vassals of America's AI 'dominance.' They want to be active participants in the AI transition." For the buyer, sovereignty means first of all control over data, models and the ability to choose, while the quality of the purchased technology comes second. CNAS therefore proposes including vetted open-weight models in American export packages, giving the buyer the top rung of the control spectrum.
The third concerns the US's ability to implement its own strategy. Winter-Levy and Teddy Tawil (Just Security, June 2026) note that the export program names neither priority markets nor clear selection criteria, and that its first promoted project was a data centre in South Korea – an ally where American companies already operate. This is not an accident but a consequence of the program's design. It is industry-led: companies submit the proposals, and they propose deals where they see safe business – ready infrastructure, established contracts and institutional trust. American financing agencies such as EXIM are traditionally risk-averse. As a result the program gravitates toward safe markets and bypasses exactly those where it competes with China. A Stanford and Carnegie memo (August 2026) adds that financing instruments like EXIM and DFC are inherently hard to deploy effectively. According to GPPi, as of August 2026 the program was still "in its initial phase", with 78 applications received by July.
Until the summer of 2026 the gap between the seller's and the buyer's logic concerned mainly price and control. On 12 June 2026 a third dimension was added: reliability of access.
The risk that a technology supplier might switch its product off had long been discussed, but almost always in relation to China. Washington used the argument for years: in the campaign against Huawei in 5G networks, including in Latin America, and in May 2025, when undocumented communication devices were found in Chinese solar inverters and American media began to speak of built-in "kill switches" (New York Sun). The logic was clear: dependence on a strategic rival's technology is dangerous because the rival can exploit it in a crisis.
On 12 June 2026 US Commerce Secretary Howard Lutnick sent Anthropic an export directive: suspend access to the newly released Claude Fable 5 and Claude Mythos 5 for any foreign national anywhere in the world (Nextgov). These were among the most capable frontier models on the market; Mythos was designed specifically for cybersecurity and was distributed mainly to cyber-defense organizations through Project Glasswing. Anthropic is not a peripheral player: in early 2026 it was valued at $380 billion with annualized revenue of at least $18 billion. According to Axios, the directive was triggered by another company's claim to have broken Mythos's safeguards. CSIS adds that Amazon researchers had found ways around restrictions meant to stop the model answering questions about finding cyber vulnerabilities. The grounds for the shutdown were thus exactly the dual-use risk on which the American logic of controlling the frontier rests.
Anthropic could not reliably distinguish foreigners from US citizens and switched both models off for all customers. The requirement was withdrawn on 30 June, and from 1 July Fable 5 was again available to users worldwide after the Center for AI Standards and Innovation tested an improved safety classifier (BankInfoSecurity). Mythos, however, according to Bloomberg Government, remained available only to a limited number of US organizations; talks on wider access, including international access, through Project Glasswing continued without a clear timeline.
The shutdown lasted nineteen days, but its significance for buyers lies not in its duration but in the precedent. For the first time an export control was applied not to a chip or software but to a specific model, and on the basis of the user's nationality. The fear that the US had for years directed at Chinese technology – that what you buy can be switched off by a foreign government's decision – was for the first time realized by the US itself: the kill-switch risk. Argentine lawyer Andrés Gil Domínguez, writing in Infobae, pointed to another detail: GPT-5.6 underwent government review before release, whereas Fable was switched off after release. The American state now controls access to the frontier at both ends – before and after a model reaches the market.
This precedent directly contradicts the full-stack logic described in section 1.1. The package promises compatibility and lower risk, but the deeper a country is embedded in the American ecosystem, the more it loses if access is cut. The risk the full stack was supposed to protect against turned out to be built into the stack itself.
Reactions differed by region. In Europe the shutdown was read as a signal about technological sovereignty, and the European Commission officially stated that such measures should not discriminate against partners. In Latin America no official reaction from governments, regulators or central banks could be found. The media covered the episode, and Gil Domínguez observed that "in Latin America there was no comparable regional political reaction"; instead, "a deafening silence prevails".
No broad conclusions should be drawn from this silence. Fable 5 was released only days before the shutdown, and few had built it into their workflows. Mythos was available to a narrow circle of organizations anyway. Users had alternatives – earlier models in the same line and competitors' models. Governments' silence shows only that the episode did not touch their critical infrastructure. It says nothing about how important the frontier is to regional demand as a whole; that question is answered by the criterion-by-criterion analysis in section 4.
The work discussed above analyses the American strategy from the seller's side – its goals, instruments and capacity to deliver them. This is a necessary perspective, but an incomplete one. This essay looks at the same strategy from the buyer's side: it takes the American strategy in its own terms and compares it with what 20 Latin American and Caribbean countries need, and with how China responds to the same demand.
The analysis is organized around criteria that countries themselves set out in their strategies and act on in procurement, with particular attention to sensitive areas – government data and cybersecurity. Section 2 describes how the region chooses between the two powers and how the countries differ. Section 3 sets out what they need. Section 4 compares the Chinese and American responses criterion by criterion. Section 5 examines how countries combine both offers in practice. Section 6 concludes.
To understand how Latin American countries choose AI suppliers, one first has to understand how they choose between the US and China in general. AI does not create relationships from scratch: it is laid over an existing structure of trade, investment, credit and military-political ties.
Historically Latin America has been a sphere of predominant US influence. That influence rests primarily on security and politics: military cooperation, arms supplies, participation in American coalitions and diplomatic coordination. Its formal expression is Major Non-NATO Ally status, granted to Argentina (1998), Brazil (2019), Colombia (2022) and, in January 2026, Peru (Federal Register).
Over the past twenty years a second centre of gravity has emerged, and it works through the economy. According to AS/COA, China has overtaken the US as South America's main trading partner; it has been Brazil's top partner since 2009 and Peru's for most of the past fifteen years. By the mid-2010s Chinese state banks had lent the region more than $125 billion, about half of it to Venezuela (Inter-American Dialogue). Twenty-two countries of the region now participate in the Belt and Road Initiative. At the 2025 China–CELAC forum Beijing announced a credit line of 66 billion yuan (about $9.2 billion).
Hence the region's defining feature: the two forces act in different planes. The economy pulls increasingly toward China; security and politics still anchor the region to the US. The rivalry therefore rarely looks like a straight choice of one side. More often it is an American effort to prevent economic dependence on China from turning into political dependence. The clearest recent episodes are Panama's exit from the Belt and Road Initiative in February 2025 under pressure from Washington (Diálogo Américas, a US Southern Command publication) and, in its extreme form, the US military intervention in Venezuela on 3 January 2026, which ended with the capture of Nicolás Maduro (AP via WDET).
If the two forces act in different planes, it is rational to expect countries to follow each force in its own plane rather than choose one side wholesale. That is what happens.
Ernesto Talvi and Gabriel Leiva of the Elcano Royal Institute analysed 31 UN General Assembly resolutions from 2001–2023 on which the US and the EU voted against China. They divided them into questions of values (human rights, sovereignty and territorial integrity) and questions of economic interest (trade and development, sanctions). Latin America's democracies – excluding Cuba, Nicaragua and Venezuela – voted with the West on values in roughly four cases out of five, and with China on economic questions in more than nine cases out of ten. The largest countries, Mexico and Brazil, behave the same way. Importantly, left-wing and right-wing governments differ only slightly in this behaviour.
Talvi and Leiva's conclusion is key for this essay: the region is neither neutral nor balancing; it chooses a side depending on the issue – alignment by issue, not by camp. Because this behaviour persists across changes of government, it cannot be explained by ideology. It follows from the structure of interests described above: where security and political values are at stake, countries stay with the US; where trade and development are at stake, with China.
This is the structural basis of what the literature calls hedging. Individual countries show it clearly. Peru is an almost ideal example: China is its main trading partner, its Pacific trade runs through the Chinese-built port of Chancay, and yet in 2026 Peru became a Major Non-NATO Ally and signed an AI cooperation memorandum with the US embassy. Colombia has been a Major Non-NATO Ally since 2022, yet joined the Belt and Road in May 2025 (China-Global South Project). Brazil trades with China more than with anyone else, but the US remains its largest foreign investor.
Two conclusions follow. First, it is pointless to expect monolithic pro-Americanism from the region, even where ties to the US are especially dense. Second, if countries choose sides by issue, a split by layer is to be expected in AI as well: a pull toward the US where AI touches security, and toward China where it touches price and economic development. This matches the dual-layer demand of section 1.3. How this split looks in practice, and where it breaks down, is shown in section 5.
The split sets the general background, but the region's countries are too different to describe with a single formula. To compare demand and supplier responses, they need to be grouped.
What mainly determines a country's AI needs is how developed its AI ecosystem is: a country without compute or a strategy needs something different from one building its own supercomputers. This is the basis of the typology.
In a region with such a strong US presence, however, a second question arises immediately: does political anchoring to Washington change what a country buys and from whom? The denser the ties to the US, the costlier it is to diverge from Washington, so one might expect strongly anchored countries to have a narrower choice of suppliers. We do not turn this into a separate axis of the typology but treat it as a question in its own right: section 2.5 shows how anchored the countries are, and sections 3.4 and 5 show what follows.
We deliberately do not use government ideology as a criterion. Talvi and Leiva's data show it explains the region's foreign-policy behaviour poorly, and Latin American governments change often. The typology should rest on structure, not on the current course. For both level of development and political anchoring there are established measures, so the groups are built on them rather than on the author's judgment.
The first measure is the Latin American Artificial Intelligence Index, ILIA, produced annually by the UN Economic Commission for Latin America and the Caribbean (CEPAL) and Chile's National Center for Artificial Intelligence (CENIA). The 2026 edition, the fourth, was presented on 20 October 2026 and covers 20 countries (CENIA). The region has more countries: Nicaragua, Haiti and most small Caribbean states are not covered, so the essay's conclusions apply to these 20.
The index assesses three dimensions. Enabling factors are what an AI ecosystem cannot develop without: digital infrastructure (connectivity, compute, cloud services and data centres), data (availability and governance) and talent (from basic AI literacy to master's and doctoral programs). R&D and adoption show how far a country produces AI knowledge and turns it into practice: publications and research centres, start-ups and investment, software development, and use of AI by firms, citizens and the state. Governance assesses the framework a country has built: a national strategy and responsible institutions, international engagement, data and cybersecurity regulation, and ethical standards.
The overall score is a composite from 0 to 100. On this basis CEPAL divides countries into three groups with thresholds at 60 and 35 points. These groups form the basis of our typology.
Pioneers – countries scoring above 60: Chile, Brazil and Uruguay. By CEPAL's definition they have strong digital infrastructure, developed talent and an innovation environment, and AI is embedded in national strategies. Brazil holds most of the region's compute, Chile leads in research, Uruguay in business adoption. These countries can not only buy AI but produce some of its components themselves. Many countries in the region want to own part of the stack, but for pioneers this is the next realistic step, and their freedom to choose suppliers is the greatest.
Adopters – countries scoring 35 to 60: Colombia, Costa Rica, Argentina, Peru, Mexico, Ecuador, the Dominican Republic, Panama, El Salvador, and Trinidad and Tobago. They are beginning to embed AI in the economy and public administration; research is moderate and AI policy is taking shape. The group's common weakness is moving from strategy to actual deployment: many have better-developed strategies and institutions than research and practical AI use. This is the largest and most heterogeneous group: demand for AI already exists, but the resources to meet it independently do not. Freedom to choose suppliers therefore matters most for them, and they show most clearly how political anchoring to the US works.
Explorers – countries scoring below 35: Paraguay, Jamaica, Cuba, Guatemala, Bolivia, Honduras and Venezuela. They are taking first steps: basic capabilities are limited, AI policy is at a preliminary stage, and some have no national strategy at all. Their main problem is not choosing a supplier but building the foundation on which AI can be used at all, and their choice of supplier is largely determined by who arrives with financing.
The second measure is the Formal Bilateral Influence Capacity index (FBIC), maintained by the Pardee Institute for International Futures at the University of Denver. It measures how far one country is able to influence another on the basis of their relationship. It has two components. Bandwidth captures the breadth of ties: trade and trade agreements, diplomatic representation and shared international organizations, military alliances and arms transfers. Dependence captures how much one country relies on the other: what share of its trade, aid received and arms imports the partner provides. The index runs from 0 to 1. We use the October 2025 release with data to 2024. FBIC matters because it measures not governments' declarations but the structure of ties: the greater US influence over a country, the more it stands to lose in a conflict with Washington.
The first step is to establish who is the main external centre of influence for each country. For each of the 20 countries we ranked every country in the world able to influence it. The result is unambiguous. The US ranks first for 15 of the 20 countries, including all ten adopters. China ranks first only for Cuba and Venezuela – for Venezuela this is the picture of the Maduro era. For three more countries – Uruguay, Paraguay and Bolivia – the main external centre of influence is Brazil, a reminder of its role as a regional power. For most countries China is not even in the top three: it ranks second only for Chile and Peru, and elsewhere it is behind not only the US but also European countries and regional neighbours.
[ВСТАВИТЬ РИСУНОК 1]
Figure 1. US and Chinese influence capacity (FBIC) over each country in 2024, grouped by ILIA 2026 tier. Within each tier countries are sorted by US influence. Source: Pardee Institute FBIC v. October 2025; author's calculations.
At first sight this contradicts Talvi and Leiva's finding that the region votes with China on economic questions. But there is no contradiction, and its absence confirms the picture in section 2.1. FBIC combines trade, diplomacy, international organizations, alliances and arms into one measure. In every one of these planes except trade, the US has a huge accumulated advantage and China has none. China's weight is concentrated in a single plane – the economy – so in the composite index it trails not only the US but also European countries with their dense diplomatic and institutional ties. Structurally the region is anchored to the US, while economically it is pulled increasingly toward China. This combination is what creates room for hedging.
This determines how to measure anchoring. Dividing countries into "pro-American" and "pro-Chinese" makes no sense: for almost all of them the main centre of influence is the US. The meaningful question is how tightly a country is anchored to the US. The measure is therefore absolute US influence by FBIC, not the US share of the US–China pair. The share can mislead: Argentina's is high only because Chinese influence is small, while absolute US influence over Argentina is moderate and its room for independent choice is real.
Among adopters this measure varies considerably. Mexico is the most strongly anchored – more than any other country in the sample – followed by El Salvador, Colombia, the Dominican Republic and Costa Rica. For Peru, Ecuador, Argentina, Panama, and Trinidad and Tobago, US influence is markedly lower, and the sharpest break in the whole series falls between Costa Rica and Peru. If anchoring determined the choice of supplier, the first five should have a noticeably narrower choice than the second five. We therefore refer to them below as more and less strongly anchored adopters, but do not treat them as separate groups: as sections 3.4 and 5 show, anchoring changes channels and the strictness of restrictions rather than demand itself or the general pattern of choice. Argentina is also a reminder that structural freedom does not predetermine political course: under Milei it deliberately turned toward Washington.
Explorers are also heterogeneous in their anchoring: Guatemala, Honduras and Jamaica are closely tied to the US, Paraguay and Bolivia look first to Brazil, and Cuba and Venezuela to China. But their capacities are so limited that differences in anchoring affect their demand less than adopters'. Full values are given in the appendix.
Two caveats concern specific countries and are needed up front.
Venezuela. The FBIC data end in 2024 and describe Maduro-era Venezuela, when China was its main external centre of influence. After the US intervention of January 2026 this picture has most likely changed radically, so all conclusions about Venezuela refer to the period before 3 January 2026.
Cuba. Long isolation and sanctions make data on Cuba's AI ecosystem and external ties less reliable than for other countries.
To compare the Chinese and American offers, one must first understand the demand they answer. The central claim of this section is that countries' AI needs are determined by the development of their AI ecosystems, not by their political orientation. The list of needs is almost the same everywhere because all face the same structural deficit. Groups differ in which need comes first. Political anchoring to the US changes not the needs themselves but how freely a country can meet them from different suppliers.
The claim rests on two complementary kinds of source. The ILIA index shows countries' deficits from outside, using a single methodology. National strategies and projects show how governments themselves see these deficits and what they want to obtain.
The region lacks not frontier models but what precedes any use of them. ILIA 2026 identifies four such deficits, and each generates a need.
The first and most acute is compute. According to ILIA, the five benchmark countries (Germany, Spain, Estonia, Portugal and Singapore) have on average about 29 times more GPUs per million inhabitants than Latin America. Even this average flatters most countries, because 87.7% of all GPUs in the region are in Brazil alone; the other nineteen share an eighth of an already modest capacity (TeleSemana). For most of them the question is not which model to train but what to run even someone else's model on. Hence the first need: compute infrastructure.
The second deficit is a legal basis for data. ILIA notes that many countries lack clear rules on text-and-data mining. Data may exist, but cannot lawfully be assembled into a corpus for training or fine-tuning. The problem has a reverse side: if the rules on data use are undefined, data cannot be protected either – from a foreign supplier or from one's own contractors. This deficit cannot be solved by buying equipment, only by legislation. Hence the second need: legal certainty and control over data.
The third deficit is adoption. AI is used by 27.6% of large firms in the region, 12.8% of medium-sized and 6.8% of small ones, against 66%, 41% and 24% in the benchmark countries. The gap widens as firms get smaller, which makes sense: small businesses cannot afford expensive models and lack staff to deploy them. Generative AI is used by 16.1% of people aged 15–64. Hence the third need: affordable models and people to deploy them.
The fourth deficit is money. Seventeen of the 20 countries have national AI strategies and 15 have institutions responsible for them, but only two – Brazil and Colombia – have a dedicated budget. Strategies thus exist mostly as intentions without funding. Hence the fourth need: financing.
Together these four needs – compute, control over data, affordable models and talent, financing – form a common regional list. No country can close all the gaps alone, which is why ILIA frames the region's goal not as technological self-sufficiency but as preserving choices: if external suppliers cannot be avoided, the sensible aim is not to depend wholly on any one of them.
If the list is shared, the groups differ in which need comes first – where their binding constraint lies. The logic is that of any system requiring several resources at once: development is held back by the scarcest resource, and until it is supplied, investment in the others yields little. A country without data rules gains little from a supercomputer, and a country without compute gains little from its own language model. It is worth distinguishing what is wanted from what is attainable: many countries want to own part of the stack, as the case of Mexico will show, but the binding constraint determines what a country can get now.
For pioneers the binding constraint is scale. Basic conditions are broadly in place. Chile leads the region in research, Uruguay in the share of firms using AI in every size class, Brazil in infrastructure and above all compute. Their demand is therefore aimed not at building a foundation but at owning part of the stack and expanding it.
Brazil shows what this looks like in practice. In August 2026 the government allocated about R$2.3 billion (about $440 million; here and below at roughly R$5.2 per dollar) from the national science and technology fund to AI infrastructure (China-Global South Project). About R$1.28 billion (about $245 million) will go to a supercomputer in Rio de Janeiro that the state research network RNP is building with Huawei, responsible for compute, and iFlytek, responsible for Portuguese-language models; training of 3,000 specialists is built into the project, and the cooperation agreement takes effect in July 2027. About R$1 billion (about $190 million) is reserved for a tender for a second supercomputer in Rio Grande do Norte; the science minister expects Nvidia to be the supplier, but the tender has not yet been held. A government cloud, RISC-V chip development with Spain and an algorithmic transparency centre remain plans. Chile released Latam-GPT in February 2026 – an open 70-billion-parameter language model for Latin America.
Yet even for pioneers scale is limited by money, and absolute figures show this more clearly than any percentages. Brazil's entire AI Plan for 2024–2028 amounts to R$23 billion – about $4.4 billion over five years – of which about R$5.8 billion (about $1.1 billion) goes to infrastructure and development (Convergência Digital). For comparison, according to Bruegel, private AI investment in 2024 alone was $4.5 billion in the UK and $109 billion in the US. The comparison is imperfect – a public plan against private investment – but it conveys the scale correctly: a five-year public program of the region's largest country is comparable to one year of private AI investment in one European country. Chile, the index leader, operates on an even smaller scale: the supercomputer at the University of Tarapacá in Arica cost, by various estimates, $5–10 million, and a chip-purchase tender by the state agency Corfo about $7 million (Americas Quarterly). What pioneers need, then, is not ready-made technology but whatever lets them expand their own part of the stack on limited budgets: access to compute, joint development and partners with whom to build.
For adopters the binding constraint is moving from intention to practice. At the level of strategies and institutions many look respectable, sometimes on a par with pioneers. Costa Rica's governance score exceeds 80 points and El Salvador's is about 61, but their R&D and adoption scores are 45 and about 25 (Revista E&N). Strategy on paper is well ahead of AI in practice.
Colombia shows that this gap is hard to close even with money. It is one of only two countries in the region with a dedicated AI strategy budget: policy CONPES 4144, adopted in February 2025, provides about 479 billion Colombian pesos (about $115 million) through 2030 (Impacto TIC). Its most visible projects are two AI centres of excellence in Bogotá and Zipaquirá, at 50 billion pesos each (together about $25 million), and a network of about a hundred small access centres. But as Nearshore Americas reports, project details remain vague and public access will be limited. Former ICT minister Sandra Urrutia points to a more basic problem: about 40% of the population has no internet access, about 70% in rural areas. What adopters need, then, is less new strategies than what turns strategies into practice: investment in infrastructure, talent and applied solutions for government and business.
For explorers the binding constraint is hardest to identify because the group is heterogeneous. But Paraguay shows that the usual image of such countries as unconnected and therefore unready for AI is inaccurate. According to DataReportal, 83.1% of Paraguayans used the internet as of October 2025, and median fixed broadband speed was about 100 Mbit/s at year-end. Paraguay is connected. Its weakness lies elsewhere: ILIA rates its research and adoption as weak, and its group neighbours Guatemala and Honduras are weak in governance too, at around 20 points.
Paraguay does, however, have an asset pioneers have to buy: surplus cheap hydropower from the Itaipú dam, which it has exported for decades. Paraguayan AI is now being built around this asset. On 22 October 2026 the US company Crusoe and Paraguay's Araico announced an AI data centre next to the Itaipú substation: a first phase is planned for mid-2027, then up to 106 MW of compute capacity, and about 280 MW in the long run (Crusoe). US Deputy Secretary of State Christopher Landau called it possibly the largest private foreign investment in Paraguay's history. In parallel Araico is promoting the "Iguazú AI City" with announced investment of $12 billion over three years; it names Founders Fund, IDB Invest and the US International Development Finance Corporation (DFC) among participants in its financing structure (Intelligent CIO). Confirmation of DFC's participation from the corporation itself could not be found.
These projects show what explorers' demand really looks like. What to build: data centres around cheap energy. Who pays: foreign private capital, possibly with American and multilateral finance. On what terms: this is where the main question lies. It closely resembles the model that Kim, Tadepalli and Levine describe for foreign data centres in developing countries: the host country receives fixed payments for land and energy, while the margin on compute goes to the owners of chips and models – a host site instead of a stake. The main need of explorers is therefore not access to models but the ability to turn a natural asset into a stake in the AI economy rather than remain a site for other people's compute. For that they need above all institutions and legal frameworks that let them bargain over terms – exactly where the group is weakest.
The groups thus differ not in what they want but in what stage they are at. Pioneers want to own part of the stack, adopters to turn strategies into practice, explorers to turn their resources into a stake in the AI economy. The list of needs is the same, but the order differs.
If needs are set by development level, political anchoring to the US should not change them much. This is clearest for the country anchored most strongly: if even its demand resembles that of countries far from Washington, anchoring must be acting on something else.
That country is Mexico. If anchoring shaped needs, one would expect Mexico to want integration into the American stack. The opposite is the case. The National AI Plan, prepared by the Agency for Digital Transformation and Telecommunications and presented in 2026, aims for the state to be able to create, adapt, control and operate AI systems itself rather than rely on others', and explicitly names reducing dependence on private and foreign suppliers as a priority (Xataka México). The plan has three pillars: a legal and ethical framework, talent and public-sector software, and strategic infrastructure for technological sovereignty. Under the third, the state supercomputer Coatlicue is being built with about 14,500 GPUs at a cost of more than 6 billion pesos (about $330 million), along with a government cloud and an "AI Factory" for in-house development of public services (Infobae). In August 2026 President Sheinbaum added a resource dimension: Mexico needs data centres for its own AI, but should not become a data centre for other countries, because that takes too much water and energy (Expansión).
In substance this is the same sovereignty agenda as Brazil's – a sovereignty agenda inside the US orbit. Anchoring to the US does not make Mexico's demand "American". It acts on something else: the channels through which Mexico can meet its demand.
Trade shows how. In 2025 the Trump administration pressed Mexico to raise tariffs on Chinese goods, arguing that they were reaching the US through Mexico under USMCA preferences (El CEO). In December 2025 Mexico adopted tariffs of up to 50% on goods from China and several other Asian countries, and Beijing publicly demanded their withdrawal (Canal 26). Sheinbaum justified the decision by the priority of USMCA, which gives the country "enormous advantages". The BYD story follows a similar logic: in 2025 the company shelved plans for a Mexican plant for 150,000 vehicles and 10,000 jobs, citing geopolitical instability and Washington's trade stance (bne IntelliNews). The example is not clear-cut: according to the Financial Times, China's Commerce Ministry was also holding up approval for fear of technology leaking to the Americans (Mexico News Daily).
In AI no direct case could be found of Mexico rejecting an attractive Chinese offer under pressure from Washington: Huawei's cloud region in Mexico keeps operating. Anchoring shows itself differently – in what is absent. Brazil has a state supercomputer being built with Huawei. Mexico, with the same sovereignty agenda, has no major state AI projects with Chinese companies. The freedom Brazil uses openly is, for Mexico, limited by the cost of diverging from its main trading partner.
A caveat is needed: Coatlicue is not yet running – construction began in June 2026 and is scheduled for 24 months. This is exactly what makes Mexico an adopter despite pioneer-level ambitions: its binding constraint is the group's – moving from plan to working system. But for understanding demand the intention is what matters: it shows what the country wants, whether or not it manages to get it.
Section 1.3 divided demand into two layers: a broad mass layer where price and control decide, and a narrow critical layer – cybersecurity and defense – where quality decides and the frontier is needed. The layers are distributed unevenly across groups.
The mass layer exists everywhere and dominates everywhere. Public services, document workflows, agriculture and education make up most of what countries in all groups want from AI.
The critical layer also exists everywhere, because cyberattacks do not pick countries by development level. But the value of frontier cyber defense depends on whether a country has the people and institutions to use it. Pioneers and some adopters do, so their demand for the frontier in the critical layer is real. Explorers need more basic things first: cybersecurity laws, protection of critical infrastructure, minimum staff qualifications. A frontier model adds little where no one can use it.
Between the two layers lies an intermediate zone: public services and citizens' data. Control matters more here than in the rest of the mass layer, because a leak or shutdown hits the state itself. But the frontier is not needed: tasks remain threshold tasks and a "good enough" model handles them. This is therefore not part of the critical layer but a zone where the demand for control is strongest and the choice of supplier most contested – state data as a contested zone. Sections 4.3 and 5.2 show what it looks like in practice.
This has an important consequence for comparing suppliers. The American offer, built around the frontier, finds demand mainly in the critical layer and mainly among countries able to absorb it. The Chinese offer, built around price and open models, lands in the mass layer, which exists everywhere.
These findings yield the criteria on which section 4 compares the Chinese and American responses. Price comes first, because it answers strategies without budgets and the adopters' binding constraint and shapes every other choice. Compute, infrastructure and capital answer the region's main deficit, the missing financing and the pioneers' binding constraint. Control over data and models answers the missing legal basis for data and the sovereignty agenda shared by Brazil and Mexico. Talent and adoption answer the gap in AI use between large and small firms. Reliability of access answers the risk of being switched off, which became a precedent after the Fable episode, and the demand to preserve choices that ILIA puts at the centre of its understanding of sovereignty.
These criteria weigh differently for different groups. For pioneers compute and control matter more; for adopters price, financing and adoption; for explorers the terms on which their resources are built into other people's infrastructure. In section 4, the description of demand under each criterion therefore shows how it looks for each group.
The five criteria derived in section 3 are discussed below in the same way: what countries mean by the criterion and how it differs by group, how China responds, how the US responds, and whose offer fits demand better. We start with price, because it decides at every level – from choosing a model for a public-services chatbot to who builds a data centre and who owns it.
What countries need. Price is the common denominator of all the other criteria. For adopters, who have strategies but no budgets for them, it decides directly. For pioneers it determines how much of their own infrastructure they can build on limited money. For explorers it decides whether a deal happens at all. Price has two components: what it costs to query a model, and what the compute it runs on costs.
How China responds. On model prices the Chinese advantage is most visible. In November 2026 GLM 5.1 from China's Zhipu costs about $0.90 per million input tokens and $2.83 per million output tokens on OpenRouter. OpenAI's flagship GPT-5.5 costs $5 and $30 respectively (Morph) – roughly six times more on input and ten times more on output. According to Rest of World, Chinese open models perform some tasks 50–90% cheaper than leading American ones overall. There is a caveat: according to AlphaSense data for August 2026, on some complex tasks American models are cheaper per solved task because they use fewer tokens. In the mass layer, where tasks are simple, the gap persists.
The reason for the gap lies in market structure, and it matters more than the figures. A closed model is sold by a single supplier, and every request carries its margin, which recovers the billions spent on training. An open model can be run by anyone: on OpenRouter itself, one open model is served by several independent providers at once, and competition among them pushes the price toward the cost of compute. A country that runs an open model on its own hardware, or fine-tunes it for its own tasks, pays only for compute and electricity, and there is no developer margin in that price at all. Over the long run, therefore, a country's own fine-tuned model is cheaper, even if the initial investment in hardware is higher.
How the US responds. American suppliers can compete on price mainly with their smaller lines: GPT-5.4-mini costs $0.75 per million input tokens and $4.50 per million output tokens, comparable to Chinese models. But it is a scaled-down model far from the frontier, and it is still closed: the supplier's margin stays in the price, and the model cannot be moved to one's own hardware. American open models – OpenAI's gpt-oss, Google's Gemma, Nvidia's Nemotron – offer the same economics as Chinese ones but, as section 4.3 shows, lag further behind in quality.
Verdict. On price China wins in the mass layer almost without qualification, and it wins through openness, which removes the developer margin from the price. The American offer can match it on price only with smaller closed models that give neither the frontier nor control. For adopters and explorers this is a decisive argument; for pioneers it is an argument for their own infrastructure built around open models.
What countries need. Compute is the region's main deficit, but demand for it differs from what the American frontier logic assumes. Clusters for training frontier models cost tens of billions of dollars, are beyond even Brazil, and Brazil does not need them. Countries need capacity to fine-tune existing models, run them in public services and business, and support scientific computing. Pioneers want world-class machines that let them own part of the stack. Adopters want their first serious public compute capacity, like Mexico's Coatlicue. Explorers act here more as suppliers: their asset is cheap energy on which others build data centres.
Compute is the one criterion where the main question becomes capital: who pays for construction, who owns the result and who earns the income. Chinese and American deals are therefore examined below along exactly these lines.
How Chinese deals are structured. There are three formats in the region.
The first is a state partnership with technology transfer. In August 2026 the Brazilian government allocated about R$1.28 billion (about $245 million; here and below at roughly R$5.2 per dollar) over five years from the national science and technology fund for a supercomputer in Rio de Janeiro. The Brazilian state pays, and the operator is the state research network RNP. Chinese companies take part as technology partners. Huawei is China's largest maker of telecommunications and server equipment. iFlytek is a Chinese company specializing in speech recognition and language technology; since 2019 it has been on the US Entity List (TechNode). The agreement covers technology transfer, joint research, a "sovereign software stack" and a national Portuguese-language model, plus training for 3,000 Brazilian specialists over five years; launch is planned for July 2027 (Convergência Digital). Open sources do not say which chips the machine will use. The capital here is Brazilian; China contributes technology, know-how and training; and the machine and the model stay with the Brazilian side.
The second format is a private data centre for a Chinese company's own needs. ByteDance, the owner of TikTok, is building its largest data centre outside China at the Pecém industrial complex in Ceará. ByteDance is the anchor tenant, and the capacity is intended primarily for training its own models (Kiara Capital). The site is being built and developed by Omnia, the data-centre platform of the Brazilian investment firm Pátria Investments (DCD). Power is supplied under a 20-year, $2 billion contract by Brazil's Casa dos Ventos, 34% of which is owned by France's TotalEnergies (TNW). Announced investment is about R$200 billion (about $38 billion) for a target capacity of about 1 GW; the first phase is valued at R$50 billion (about $9.6 billion) (InvestNews). The government, for its part, signed a provisional measure allowing data-centre equipment to be imported tax-free (Bloomberg Línea). How stakes and financing are divided among the participants has not been disclosed, and neither has the chip supplier. There is an indirect guide: outside China, ByteDance builds compute on American chips – in Malaysia, through its partner Aolani Cloud, it is deploying Nvidia Blackwell systems (Wall Street Journal, via Seeking Alpha). If Pecém works the same way, Chinese private capital in Brazil will run on American hardware. The economics for Brazil are clear: land, energy sales, taxes net of incentives and construction jobs – but no stake in the compute and no stake in the models.
The third format is Chinese corporations' own clouds. Huawei cloud regions operate in Santiago, São Paulo, Buenos Aires and Mexico City. In June 2026 Huawei Cloud's Latin America president Mark Chen said the company plans to bring its own Ascend AI processors into its cloud services in the region – that is, to offer customers compute on Chinese chips instead of Nvidia (South China Morning Post; Descubre). For Huawei this is a way around the shortage of American chips; for customers it is a new risk: in May 2025 the US Commerce Department's Bureau of Industry and Security warned that using Ascend chips "anywhere in the world" violates US export controls (Bloomberg).
How American deals are structured. There are three formats here too.
The first is the hyperscalers' private clouds. In January 2025 AWS announced a $5 billion, 15-year investment in a cloud region in Querétaro (Mexico News Daily); Microsoft and Google are also building there. Capital, ownership and income here are American; the country gets taxes, jobs and the ability to keep data within its borders. These clouds are built for the region's commercial customers and are already running, and this is where the American advantage in private capital shows.
A second advantage is tied to the first: hardware. In advanced AI chips American companies are close to a monopoly, and this is visible even in other people's deals. Brazil's science minister expects the second Brazilian supercomputer in Rio Grande do Norte (about R$1 billion, about $190 million) to come from Nvidia (China-Global South Project). In October 2025 the Dominican Republic signed a memorandum on sovereign AI directly with Nvidia (El Nuevo Diario). Even ByteDance builds compute on Nvidia outside China. Wherever a data centre stands and whoever owns it, it is likely to contain American chips.
The second format is private projects built on local resources. In Paraguay the American company Crusoe and local partner Araico are building a data centre at the Itaipú dam (section 3.3; Crusoe): the capital and compute are private, and Paraguay provides the site and cheap energy. The deal's terms – taxes, equity, the price of energy – have not been made public.
The third format is megaprojects for future demand. Stargate Argentina, announced on 10 October 2025, two weeks before midterm elections, is a letter of intent for a data centre of up to 500 MW costing up to $25 billion. In this scheme OpenAI is only a possible future buyer of capacity and invests nothing; the money must be raised by Argentina's Sur Energy through a joint venture with a cloud developer (DPL News). In April 2026 the economy ministry reported that no application for tax incentives had been filed (Buenos Aires Times), and as of 23 October 2026, according to Bloomberg Línea, there was still no final contract.
Why American capital does not add up. American capital is present in the region, but it is private and goes where a private investor sees a return: large markets with existing demand. That is why the hyperscalers build in Mexico, Brazil and Chile, where commercial customers exist, and do not build where a project would have to create its own demand. Stargate Argentina stalled on exactly this: OpenAI is willing to buy capacity, but $25 billion in a country with Argentina's macroeconomic risks has to be found by a local developer.
Public financial institutions were supposed to close this gap, but they are weak by design. EXIM, the Export-Import Bank of the United States, is the government export credit agency that lends to foreign buyers of American goods (China has its own counterpart, the Export-Import Bank of China). DFC, the US International Development Finance Corporation, provides loans and guarantees and takes equity in projects abroad. Until the end of 2025 DFC could operate freely only in low- and lower-middle-income countries. Most Latin American countries are upper-middle-income, so each required a separate national-security justification. Deals above $10 million required notification to Congress, and profits went to the Treasury rather than being reinvested. DFC's active portfolio in the region since 2020 was estimated at about $4.4 billion, while Chinese infrastructure investment over two decades exceeded $285 billion (IE University; see also CSIS). In December 2025 Congress reauthorized DFC, raised its exposure cap to $205 billion and allowed it to work in almost all countries except the twenty richest (DFC, a US government source). The effect of these changes is not yet visible in the region: no confirmed DFC loans for data centres or clouds in Latin America could be found. International financing for the region's data centres has flowed through other channels, for example through IFC, the World Bank's private-sector arm, which invested in the Brazilian operator ODATA (DCD). American public money also carries a built-in condition: Section 889 of the 2019 National Defense Authorization Act bars federal agencies from using loan or grant funds to procure Huawei and ZTE equipment. Finally, the AI Exports Program itself is industry-led and, as section 1.3 showed, gravitates toward markets where American companies already operate.
Does China face similar difficulties? It does. Chinese state lending to the region has fallen sharply: in 2010 the China Development Bank (CDB) alone lent the region almost $25 billion, while in 2023 the Inter-American Dialogue and Boston University database recorded just two loans, both to Brazil, totalling $1.3 billion (Business & Human Rights Resource Centre). The instruments remain. CDB reports a $22.5 billion special loan fund for China–Latin America and China–Caribbean infrastructure projects and participation in a $10 billion production-cooperation fund, with telecommunications among the sectors listed (CDB, a Chinese state bank). At the China–CELAC forum in May 2025 Beijing announced a 66 billion yuan (about $9.2 billion) credit line naming the digital economy and AI as priorities (Dialogue Earth). But no direct loan from Chinese state banks to a specific AI project or data centre in the region could be found. Commercial loans and direct investment by Chinese companies have taken the place of the policy banks. In other words, cheap Chinese state money for AI is also nearly absent from the region today. China's advantage rests on its companies' willingness to bring technology, training and know-how into projects that the country itself pays for.
How the US shapes supplier choice. American influence over procurement is also written into documents. On 8 September 2026 Peru's prime minister Luis Galarreta and US ambassador Bernie Navarro signed a memorandum under which Peru undertakes to "import and adopt US technologies" (Infobae; El Comercio); Navarro called AI "developed and controlled by extra-hemispheric suppliers" a threat to national security. Argentina, Chile, Costa Rica, El Salvador and Panama signed the Pax Silica declaration (State Department). In Panama the US supported replacing Huawei telecom towers at 13 sites (CSIS). On their own these documents bind no one: the memorandum and Pax Silica carry no penalties for non-compliance. The real levers lie elsewhere. The first is chips: the US decides who may buy advanced processors, and the Ascend warning extends American export controls to anyone using the Chinese alternative. The second is money: by law, American public financing is unavailable to projects built on Huawei and ZTE equipment. The third is market access: Mexico's tariffs of up to 50% on Chinese goods came under pressure from Washington and to preserve access to the American market (section 3.4). The declarations merely formalize pressure that works through chips, finance and trade.
Verdict. Behind the many projects stand two different economic models. The American one is private capital that sells a service and keeps ownership and income for itself. It works well in large markets with existing demand, and together with a near-monopoly in hardware this is a real advantage: American clouds are already built, and American chips sit even in other people's data centres. But the American system offers almost no state partnerships in which the country receives the machine and the model as its own property: private capital will not do it, and until the end of 2025 public financial institutions were constrained by country-income rules, small limits and caution. At the state level the Chinese model works differently: the country pays itself, Chinese companies bring technology, training and know-how, and the machine stays with the country. At the private level Chinese capital behaves just like American capital: ByteDance builds in Brazil for its own needs and, probably, on American chips.
Hence the main conclusion of this criterion: what matters for countries' demand is who owns the result, and the supplier's nationality is secondary. Brazil's RNP supercomputer belongs to Brazil. The Querétaro clouds belong to American companies, and Mexico hosts them. The data centres at Pecém and Itaipú follow the model that Kim, Tadepalli and Levine describe as fixed rent instead of a share of AI revenue – a host site instead of a stake. For pioneers, who need to own part of the stack, the Chinese partnership model is closer to their demand. For adopters, who need capacity without large investment of their own, American clouds are a workable option. For explorers the risk of remaining a host site is the same whoever the supplier is.
What countries need. Control, as section 1.3 showed, is a spectrum: from data stored in the country on someone else's cloud, through a government cloud, to an open-weight model on one's own hardware. Pioneers aim for the top – their own or adapted models. Adopters want above all a government cloud and control over citizens' data. For explorers even the bottom rung – keeping data in the country – matters. Honduras's deal is telling: in November 2025 the state telecom operator Hondutel and the US company MeetKai signed a memorandum on hosting AI platforms "entirely within Honduras" (PR Newswire). For a country at this level even that is real progress up the spectrum.
How China responds. China's main instrument here is open weights. DeepSeek, Qwen, Kimi and GLM models are released under permissive MIT and Apache 2.0 licences: they can be downloaded, run on one's own server and fine-tuned without sending data to China. The buyer gets the top rung of the control spectrum without developing a model of its own. The scale of diffusion shows in derivative models: on the Qwen family alone, about 150,000 derivatives had been created on Hugging Face by August 2026. At the level of state infrastructure, China's answer to the language question is iFlytek's partnership with Brazil's RNP and the state IT company Serpro to build Portuguese-language models.
How the US responds. The American offer is strong on the lower rungs. AWS, Microsoft, Google and Oracle cloud regions allow data to be stored in-country, and this is what the White House means when it says the American stack lets countries keep sensitive data within their borders. Higher up the offer is weaker. Leading American models are closed. American open models exist, but according to Nathan Lambert they lag the closed American frontier by roughly six to nine months, against two to five for Chinese open models, and in the Artificial Analysis ranking 15 Chinese models outperform the best American open model (Rest of World).
Chile's Latam-GPT is a telling case, and its structure is worth spelling out. The project is led by Chile's National Center for Artificial Intelligence, CENIA – a private corporation with public funding – together with more than 200 contributors from 15 countries and 33 institutional partners (Tech Policy Press). The 70-billion-parameter model is built on Llama 3.1, an open model from the American company Meta. The first version was trained on AWS with a budget of about $550,000: roughly $300,000 from CENIA and $250,000 from the Latin American development bank CAF (Euronews). Future versions are to be trained on the University of Tarapacá supercomputer in Arica, Chile (investment estimates range from $5 to $10 million). The training corpus comprises more than 8 terabytes of text from 20 countries of the region and Spain; where the data are stored has not been publicly specified. The model is distributed openly and can be adapted by developers, companies and governments. The region's most prominent sovereign model is thus built on an American open model in an American cloud. American infrastructure can serve a sovereign project perfectly well – provided the American offer includes open weights.
Verdict. On control over models China wins: open weights are a technical answer to the demand for sovereignty. On control over data the two are roughly at parity: both offer in-country hosting. The Chinese advantage has one limit that should be stated calmly. Open models develop in the wake of the frontier: they use shared research methods and, in part, distillation – training on the outputs of stronger models; Lambert estimates that distillation accounts for roughly one to two months of the lead Chinese labs have over American open models. The frontier is thus upstream of the supply the region relies on: if frontier progress slows, the ceiling of "good enough" will also rise more slowly.
Sensitive area: public services and citizens' data. The stakes here are higher than in the rest of the mass layer: a leak or a shutdown hits the state itself. Countries therefore aim for the upper rungs of the spectrum, but reach them by three different paths. The first path is their own infrastructure: Mexico is building a government cloud and an "AI Factory" so that public services do not depend on contractors. The second is a foreign supplier that hosts everything inside the country: this is how Honduras's deal with the American MeetKai works. The third is a foreign technology partner that helps the country build its own: Brazil is building a Portuguese-language model with iFlytek and Serpro, and Huawei, according to its own materials, is building e-government clouds with Chilean state bodies (Data Centre Magazine). In all three cases what decides is the rung of control the supplier allows; its origin moves to second place. This is where the American offer finds it hardest to compete: a closed model by definition leaves control with the supplier. But the US does have tools, and, as CNAS suggests, they rest on trust: confidential computing, in which the provider cannot see the data being processed, and independent audits.
What countries need. According to ILIA, no country in the region reaches a high level of human capital, and small businesses barely use AI. Talent is a deficit for every group, but it means different things. Pioneers need researchers and engineers able to develop their own models. Adopters need specialists able to deploy ready-made solutions in government and business. Explorers need basic qualifications without which they can neither deploy AI nor bargain with foreign investors. Mexico shows the scale of demand: its state AI training centre, supported by more than a dozen technology companies, enrolled more than 10,000 participants in its first intake and aims for 25,000 certificates a year.
Outflow compounds the deficit. According to ILIA 2025, based on LinkedIn's methodology, almost every country in the region is a net exporter of AI specialists: people leave for more mature ecosystems, most visibly from Argentina and Venezuela (Derecho Artificial). Training a specialist is not enough; the specialist also has to stay.
How China responds. With scale and ties to its own projects: 10,000 training places under the CELAC plan and 3,000 specialists being trained in the Brazilian project with Huawei and iFlytek (Convergência Digital). The second form matters more. Specialists trained on Chinese hardware and models themselves become an element of lock-in to the Chinese stack.
How the US responds. Direct training programs in the region come mainly from private companies, universities and embassies. Google announced 15,000 scholarships in Colombia and 7 million educational licences (DPL News). In Peru, alongside the memorandum with the US embassy, Miami Dade College and the cultural centre ICPNA signed an AI training agreement. In scale these are comparable to Chinese programs, but they are fragmented and unconnected to infrastructure procurement.
The main American advantage under this criterion is historical. CUDA, the platform through which Nvidia GPUs are programmed, has existed since 2007, and PyTorch, created at Meta, since 2016; almost all of the world's AI development grew up on this pairing. Cloud skills are certified by AWS, Azure and Google Cloud. The region's engineers learn these tools by default, simply because the market runs on them. How strong this lock-in is shows in the fact that even Google, according to Reuters, is working with Meta to make PyTorch run fully on its TPU chips so that developers need not stick to CUDA.
Chinese strategy aims precisely at weakening this historical trend. Huawei is building its own pairing – the CANN software environment for Ascend chips and the MindSpore framework – and in August 2025 it open-sourced CANN (TrendForce) and released an adapter for running PyTorch on Ascend. Analysts at ChinaTalk note that developers complain these tools are awkward and unstable, and that Huawei is betting on compatibility with PyTorch more than on fully replacing CUDA. Training inside Chinese packages fits the same strategy: if the Brazilian supercomputer runs on Huawei hardware (not yet confirmed), its 3,000 specialists will learn on Chinese tools, and the historical lead of the American environment will begin to erode exactly where China arrives with a package.
This has a less obvious consequence for talent outflow. Skills in American tools are portable: an engineer who has mastered CUDA, PyTorch and AWS is in demand in any country, above all in the US. American training programs raise exactly the qualifications the global market buys, which means part of the investment in people may leave with the people. Skills in the Chinese environment are less portable: they keep a specialist in local projects but tie them to the Chinese stack. This is a hypothesis that still needs testing, but it explains why neither side solves the region's talent problem on its own.
Verdict. Neither side meets the demand for talent, and lock-in through talent works differently for each. For the US it arose historically and is sustained by the market: American tools are what everyone learns. China builds its lock-in deliberately, through packages and an open alternative to CUDA. For the US this means that under this criterion it wins through its ecosystem, and consolidating the advantage here is cheaper than anywhere else: training needs no export licences, does not run into the question of model control and costs little. For now private companies use this advantage; government strategy hardly does. For countries the lesson is different: people must be trained together with projects that give them work at home, otherwise training becomes an export of specialists.
What countries need. A guarantee that access to a technology will not be revoked by a third country's government. The criterion became tangible after the Fable episode, but for some countries it was key before. For pioneers, building their own systems for years ahead, reliability is a matter of long-term planning: a revoked component breaks the whole structure. For adopters, whose public services run on ready-made cloud services, it is a matter of the continuity of government. For explorers, who take whatever arrives, it is a question of whether their only supplier might be cut off.
How China responds. The difference is clearest when two shutdown scenarios are compared. If Washington restricts access, an American model running as an external service stops at once: in June 2026 Fable was switched off for all customers in a single day (Nextgov). If Beijing decides to restrict access, an already downloaded Chinese open model keeps running on the buyer's servers. Beijing can stop releasing new open versions, halt updates and support, or impose legal restrictions on use. This creates legal friction and leaves the buyer without further development of the model, but the downloaded weights physically remain with the buyer. At the model level, therefore, the Chinese offer is more reliable.
China's weak point lies in infrastructure. If an open model runs on a Huawei cloud or on Ascend chips, the leverage returns to the supplier: Huawei controls hardware supply, maintenance, firmware and the CANN software environment without which Ascend does not work. Leaving such a cloud is as hard as leaving an American one, and replacing the chips means moving the whole software stack. In addition, as section 4.2 showed, using Ascend exposes the customer to American export controls. Open weights protect against the recall of a model, but a country that has built its compute on Chinese hardware depends on China no less than a country on an American cloud depends on the US.
How the US responds. Here the American offer loses because of its own precedent. The Fable episode showed that access to the American frontier can be revoked with a single Commerce Department letter, and the terms of its restoration confirmed the government's right to do it again. Mythos, the cyber-defense model, remained available after restoration only to a limited number of American organizations (Bloomberg Government). Latin American governments did not react officially, but, as section 1.4 argued, this says more about the episode not touching their critical infrastructure. In addition, Washington ties access to a demand to exclude alternatives, as in the Peruvian memorandum (section 4.2), asking partners to give up their fallback at exactly the moment it has shown that its own access can be revoked.
Verdict. At the model level China wins on reliability: open weights cannot be recalled. At the infrastructure level both sides hold the buyer equally tightly, through different links: the US through clouds and access to the frontier, China through hardware and Huawei's software environment. The US loses additionally because it demands that alternatives be excluded, which directly contradicts the region's understanding of sovereignty as preserving choices.
Critical layer: cybersecurity. Here, as section 1.3 showed, "good enough" stops being enough and the demand for the frontier is real. And here the main American advantage and the main American vulnerability meet. Mythos was distributed narrowly from the start – through Project Glasswing, mainly to cyber-defense organizations – and after the Fable episode it remained available only to a limited circle of American organizations. The restriction here is structural: frontier cyber defense was never widely available to the region, and the Fable episode only cemented that. In the one area where the region genuinely needs the American frontier, access to it is least reliable. This is perhaps the sharpest expression of the mismatch between the seller's and the buyer's logic: the US is strongest where it restricts most, and it restricts where the buyer needs it most.
Table 1 brings the five criteria together: who is stronger under each, for which groups the criterion matters most and why.
| Criterion | Stronger offer | Matters most for | Why |
|---|---|---|---|
| Price | China | All groups, above all adopters and explorers | Open models remove the developer margin: competing providers serve them, or the country runs them on its own hardware; the US matches on price only with smaller closed models. |
| Compute, infrastructure and capital | US in private capital and hardware; China in state partnerships | Pioneers (ownership), explorers (terms of hosting) | American capital is private and keeps ownership; US public instruments were constrained until late 2025; China brings technology and training into projects the country pays for; state money for AI is scarce on both sides. |
| Control over data and models | China (models); parity (data) | Pioneers and adopters, especially for government data | Open weights give the top rung of control; American open models lag further; Latam-GPT shows a sovereign project can be built on an American base. |
| Talent and adoption | US (historically, by default) | All groups | The global toolchain is American; China erodes it through packages and open-sourced CANN; the portability of American skills feeds talent outflow. |
| Reliability of access | China (models); neither (infrastructure) | Pioneers and adopters; the critical layer above all | Downloaded weights cannot be recalled; the US holds buyers through clouds and frontier access, China through hardware and Huawei's software environment; the US also demands that alternatives be excluded. |
Table 1. Summary of the comparison by criterion. Source: sections 4.1–4.5.
The table shows a mismatch between two logics, with no simple lead for either side. The seller offers what it is strong in: quality, private capital, hardware and integration into its ecosystem. The buyer is looking for what it lacks: low cost, ownership and control. The US is ahead where its advantage arises by itself, through the market – in private capital, chips and the developer environment. It falls behind where it would have to adapt its offer to the buyer: cut prices through openness, finance state projects that transfer ownership, give up the right to switch off. The reason lies in how the American offer is built: private companies assemble it, selling a service and keeping ownership, while the state takes part mainly through restrictions. China wins where its companies are willing to give away what American ones keep – model weights and technology in state projects – but retains its leverage in hardware and the software environment. The sharpest case is cybersecurity: the US is strongest where it restricts most, and restricts where the buyer needs it most. How countries in the region combine both offers in practice is examined in the next section.
Section 4 compared the Chinese and American offers criterion by criterion. Countries, however, choose a set of suppliers for their entire AI infrastructure at once, and when that choice is viewed as a whole, the main empirical feature of the countries examined becomes visible: a country's political orientation and its technological choices diverge. Countries publicly oriented toward the US use Chinese technology, and countries moving closer to China keep relying on American technology. Mixed sourcing by layer is becoming the prevailing practice: a country takes its cloud, chips, models and training programs from different suppliers, although the depth of mixing and its distribution across layers differ from country to country.
First, it should be made clear what counts as mixing here. A supplier may have a physical presence in a country (a cloud region, an office); its services may be available on the market; companies may actually use them; the state may procure them; and, finally, the country may depend on the supplier strategically. Data on cloud companies' actual customers are almost entirely unavailable. Mixing is therefore recorded below at the two levels that can be observed: suppliers' presence on the market, and state projects and agreements. It does not follow that every supplier is actually used in every national stack.
A second caveat concerns the word "orientation". A country's ties to a power may be diplomatic (ally status, declarations signed), commercial (share of trade), infrastructural (who builds the ports, cables and networks) or security-related. These ties need not coincide, and it is precisely their divergence that the cases below show.
The divergence is visible in all three groups.
Pioneers mix openly. Brazil is paying from its own budget for a supercomputer being built with China's Huawei and iFlytek, and at the same time expects Nvidia equipment for a second one (section 4.2). In July 2026 it became a founding member of the World AI Cooperation Organization (WAICO), a forum China created as an alternative to Western AI governance forums. Diplomatically, Brazil has remained a Major Non-NATO Ally of the US since 2019, and the US is still its largest foreign investor. TNW summed up the point of this split: it costs Brazil efficiency but buys it insulation from pressure by either side. Chile signed the American Pax Silica declaration in June 2026, and its flagship AI project, Latam-GPT, is built on the American open model Llama and trained on AWS (section 4.3). Alongside the American hyperscalers, Santiago hosts a Huawei cloud region, and according to the company's own materials Huawei is building e-government clouds together with Chilean state bodies (Data Centre Magazine).
Adopters mix regardless of how strongly they are anchored to the US. Mexico, the most strongly anchored, consistently follows Washington on trade – up to 50% tariffs on Chinese goods – and its cloud market is built around the American hyperscalers in Querétaro. Huawei's cloud region in Mexico City nevertheless keeps operating. In the public sector the Chinese presence is noticeably weaker: Mexico has no major state AI projects with Chinese companies so far (section 3.4).
Among less strongly anchored adopters the picture is essentially the same. Argentina under Milei made an unambiguous diplomatic choice: it signed Pax Silica, has long been a Major Non-NATO Ally and is promoting Stargate Argentina with OpenAI. On the market, however, Huawei has a cloud region in Buenos Aires, and in July 2026 the company signed an agreement on connectivity, data and the use of AI in agriculture with the cooperative Agricultores Federados Argentinos, which brings together about 36,000 producers (Los Andes; Agrofy News). Peru moved closer to the US diplomatically and on security: in 2026 it became a Major Non-NATO Ally and signed a memorandum committing it to "import and adopt US technologies" (section 4.2). Its commercial and infrastructural ties, however, are Chinese: its main trading partner is China, and its Pacific trade runs through the port of Chancay, built and operated by China's COSCO Shipping.
Mexico's strong anchoring and the weaker anchoring of Argentina and Peru thus produce a similar result: an American core in clouds and chips, a Chinese presence in market layers and no major state AI projects with China. All the adopters examined mix, and the degree of structural anchoring does not change this. What changes the picture more is direct pressure on a specific strategic asset. Panama, weakly anchored by FBIC, left the Belt and Road under pressure from Washington in 2025, signed Pax Silica and, with American support, is replacing Huawei telecom towers (CSIS), moving from mixing to an unambiguous choice. The reason is the canal: strategic infrastructure whose control matters disproportionately to the US and therefore narrows Panama's room for foreign-policy manoeuvre. This yields a methodological conclusion for the whole essay. A general anchoring index such as FBIC may poorly reflect a country's vulnerability in particular sectors: a country with weak overall anchoring may have a single asset on which pressure is at its maximum.
Explorers mostly take whoever arrives, but pure cases are rare here too. Honduras broke with Taiwan in favour of China in 2023, then in November 2025 signed a memorandum on sovereign AI infrastructure with the American MeetKai; the government of Nasry Asfura, which took office in early 2026, is reviewing deals with China (Bloomberg), although it formally maintains recognition of Beijing. Paraguay, one of the few countries in the world that keeps relations with Taiwan, mixes differently. The American Crusoe is building a commercial data centre at the Itaipú dam (section 4.2), and in May 2026 presidents Santiago Peña and Lai Ching-te signed a memorandum on a sovereign AI computing centre to be run by the binational entity Yguazú Digital; China's foreign ministry called Peña's visit a provocation (EconoJournal). Taiwan acts here as a third technology partner with a special political status, and the case shows that choice in the region need not reduce to the US–China pair.
Mixed sourcing has three causes, and all three are rational.
The first is political. The region's countries have long separated their relations with the great powers by functional area: they support one side on security and cooperate with the other on the economy. Talvi and Leiva showed this logic in UN voting: the region chooses sides issue by issue. AI procurement becomes one more such area, divided along the layers of the technology stack. Voting and procurement are different kinds of decisions: the first is short-term and symbolic, the second binds a country for years. But both are driven by the same desire not to make all relations depend on a single power, which is why political multi-alignment shows up in the stack.
The second is economic. Different suppliers win in different layers, and in each layer the buyer takes the component that is optimal in terms of price, quality and availability. In advanced chips American companies have almost no competitors on quality, and even China's ByteDance builds compute on Nvidia outside China. In models, Chinese open weights win on price and control, because their price carries no developer margin (section 4.1), and their quality is sufficient for most applied tasks. Rest of World describes the outcome of this calculation: many Global South countries buy American chips and clouds and run Chinese open models on them. This is how a buyer behaves in any market where suppliers are strong in different things.
The third concerns insurance and bargaining. A supplier that holds every layer can cut the buyer off entirely and, over time, raise prices, because leaving is too costly. Several suppliers weaken both threats: none of them can switch the country off completely, and each knows the country has a fallback. This insurance comes with a condition without which it does not work. Mixing reduces the risk of being switched off only if the alternative components are genuinely interchangeable and the country is able to move critical functions between them. If a Chinese open model runs only on American chips, and an American cloud is the only place those chips are available, then there are formally several suppliers but a single critical dependency. As section 4.5 showed, open weights protect against the recall of a model but leave dependence on hardware and the software environment. Supplier diversification is not dependency diversification, and a country's resilience is determined by the number of genuinely workable alternatives; the number of suppliers by itself says little.
Where, then, does the line between layers run? Section 2.2 derived an expectation from issue-based alignment: on security countries should lean toward the US, on the economy toward China. The cases examined show that the expectation holds at the two ends of the spectrum and fails in the middle.
At one end is the mass layer: models for applied tasks, clouds for non-sensitive services, sector solutions in agriculture and industry. Here Chinese suppliers are present freely – up to Huawei clouds in Mexico City and Buenos Aires and Huawei's agreement with Argentine farmers.
At the other end is the security core: cyber defense, defense and the core of telecommunications networks. There are as yet no major AI procurements for cyber defense in the region, so the closest analogues come from neighbouring areas. In 2021 Brazil admitted Huawei to commercial 5G networks, but the rules for a separate government network effectively excluded the company (Teletime; Poder360). In 2023, after President Chaves met President Biden, Costa Rica introduced a 5G cybersecurity regulation that excludes equipment from countries that have not joined the Budapest Convention on Cybercrime, which means above all China (Nearshore Americas); as of 2025 the courts had kept it in force (TeleSemana). After the Conti ransomware attack on its government systems in 2022, Costa Rica received $25 million in cybersecurity assistance from the US (CyberScoop). Two cases do not establish a norm for the whole region, especially since Brazil's and Costa Rica's grounds differ: the first applied a technical rule to one network, the second adopted a general regulation after talks with Washington. But they show something important: even countries with different political orientations draw a line around the core, including Brazil, which works with Huawei in other layers. Political anchoring appears to shape the strictness and form of such restrictions more than the underlying effort to separate the core.
In the middle lie government data and services, and here the expectation of section 2.2 does not hold. As section 4.3 showed, some countries build their own infrastructure, like Mexico; others host government platforms with an American supplier inside the country, like Honduras; still others build their own with Chinese partners, like Brazil and Chile. What decides here is the rung of control the supplier allows. The supplier's nationality moves to second place, but only where guarantees are acceptable: it still determines whose jurisdiction and export controls the system falls under, as the American warning about Ascend chips makes clear. Government data therefore remain a contested zone, where open weights and integrated Chinese offers give China an opening and the American offer has to compete on trust.
Mixing proceeds at different speeds at different levels. The spontaneous developer layer discussed in section 1.3 mixes earliest and most freely: a developer switches models in minutes and passes no political filters. State procurement binds for years and passes through those filters, so mixing there is slower and more cautious. The direction, judging by the cases examined, is the same: American hardware and clouds, with more and more open models on top.
Mixed sourcing has three limits, all already visible in the region.
The first is the cost of interoperability. In itself it depends on how closed suppliers' stacks are. Where open standards and compatible interfaces exist, components from different companies fit together without great expense: an open model can run in any cloud. In the hardware layer the stacks are closed. Nvidia chips are programmed through CUDA and Ascend chips through CANN (section 4.4), and moving a workload from one to the other means moving the software stack. A country that splits its compute between Nvidia and Ascend has to keep specialists for both, although how much this raises costs depends on how workloads are distributed. Julie George of CSET points out that fragmentation of stacks will aggravate interoperability problems, especially in security. The more incompatible stacks a country combines, the higher its costs of integration, maintenance and training.
The second is shrinking room for manoeuvre. Mixing is possible as long as both powers tolerate their partners working with the rival. Signs of narrowing are already visible, and they run through the levers described in section 4.2. The American warning about Ascend chips makes mixing risky precisely in the hardware layer, where a Chinese alternative is most needed for independence from Nvidia. The ban on directing American public money to Huawei and ZTE equipment cuts projects with mixed equipment off from American financing. Mexico's tariffs, Panama's exit from the Belt and Road and Peru's memorandum show that pressure is already moving from documents into practice. The sharper the rivalry, the costlier mixing becomes and the likelier it is that countries will be required to choose.
The third limit matters most, because it does not depend on the number of suppliers. Mixing protects against being switched off but leaves the question of ownership open. A country can take a data centre from an American company, a model from a Chinese one and a cloud from a third – and in all three cases remain a host site collecting rent for land and energy, while the income from compute and models flows to the owners. The risk is common to Pecém and Itaipú (section 4.2), although the deals are structured differently: at Pecém the anchor tenant is a Chinese company building for its own needs, at Itaipú an American company is building a commercial data centre with a local partner, and the distribution of income may differ between them.
Two different notions of sovereignty follow, and it is important not to conflate them. Sovereignty of choice is the ability to change suppliers and avoid depending entirely on a single external actor; this is what ILIA puts at the centre, and this is what mixing provides, provided the components are interchangeable. Economic sovereignty is the ability to own part of the technology chain and earn income from the value created. One can have the first and almost none of the second. The first requires diversified suppliers, open models and one's own compute. The second requires different instruments: negotiating the terms of investment, revenue sharing, developing local companies, technology transfer and accumulating national capabilities. In essence this is the question Prebisch and Singer argued over (section 1.1): a periphery that sells resources and buys finished technology risks losing on its terms of trade, however many suppliers it has.
Three observations follow. Each rests on a small number of cases, so they are directions for further analysis rather than established regularities.
First, China's advantage lies in an integrated technology offer. The American offer arrives in pieces: a cloud from one company, chips from another, public financing that has not yet reached the region. Chinese companies arrive with a turnkey technological solution: compute, a local-language model, specialist training and technology transfer in one deal. This is about technology specifically: cheap Chinese state money for AI is almost absent from the region, and Brazil is paying for its own supercomputer (section 4.2). For countries with limited staff such integration is a decisive convenience. It does not cancel the American advantage in the ecosystem: a country can buy a Chinese solution for a specific task and keep building its general skills on American tools. But integration carries a risk. The more layers a country takes from a single supplier, the deeper its dependence may become, and in China's case that dependence is concentrated in Huawei's hardware and software environment (section 4.5). How deep it turns out to be depends on the terms of technology transfer, the openness of the software, the ability to replace the hardware and the training of national specialists. The Chinese offer may solve today's demand at the cost of narrowing tomorrow's freedom of choice – the whole stack, first, at the risk of deeper dependence.
Second, countries' leverage is temporary. A niche gives a country bargaining power only if it is continuously needed and hard to replicate. Paraguay's cheap hydropower is a competitive advantage, but not a resource unique in the region. Besides the price of energy, a data centre needs reliable supply, grid and backbone connectivity, legal predictability, the possibility of a long-term contract and a suitable climate, so different countries' offers are only partly interchangeable. But if neighbours – Brazil, Uruguay, Costa Rica, Colombia – can offer comparable terms on price, reliability and infrastructure, Paraguay's bargaining power will weaken. Competition between host countries can also run through tax incentives: Brazil has already exempted imported data-centre equipment from taxes (section 4.2). Such an incentive is not in itself a loss: future tax revenue, jobs and infrastructure may justify it. It has to be assessed by comparing its cost with the expected benefit, and if neighbours start competing on incentives, that comparison will increasingly favour the investor. A country therefore needs to lock in favourable terms while its asset is still needed.
Third, large deals are often announced before the terms needed to assess their benefit to the host country become public. The three projects examined are at different stages. Stargate Argentina is a letter of intent announced two weeks before an election; a year later, according to Darío Clemente of an Argentine observatory monitoring the RIGI investment regime, there is no reliable updated information on it. Crusoe's Paraguayan data centre has been announced as a concrete project with a timetable, but its terms – taxes, equity, the price of energy – have not been made public. Pecém is under construction, but the division of stakes and financing among its participants has not been disclosed either. This does not prove that governments failed to assess the deals: they may hold information that is not public. But it does mean that the public and independent analysts cannot assess them, and that the investor understands the project's economics better than the country providing land, energy and tax terms. Under such conditions it is hard to tell whether a country is gaining a stake in the AI economy or only rent for a site.
The American AI offer in Latin America diverges from the region's demand where that demand is most acute, even though it matches it on many points. In private capital, advanced chips and the developer environment, the US remains the strongest supplier. American clouds in the region are already built, American chips sit even in Chinese companies' data centres, and the region's most prominent sovereign model, Latam-GPT, is built on an American open model and trained in an American cloud. But on price, control over models and reliability of access, an offer assembled around the frontier and ecosystem integration fits poorly with countries that do not build the frontier themselves. The American system offers almost no state partnerships in which a country receives compute and a model as its own property. The public financing promised by the AI Exports Program has not yet reached the region.
Behind this lies a mismatch between two logics. The seller offers what it is strong in: quality, private capital, hardware and lock-in to its ecosystem. Its offer is assembled by private companies that sell a service and keep ownership, while the state takes part mainly through restrictions. The buyer is looking for what it lacks: low cost, ownership and control. China wins where its companies are willing to give away what American ones keep: model weights and technology in state projects. Its advantage lies in an integrated technology offer. Cheap Chinese state money for AI is almost absent from the region, and the most telling Chinese project, Brazil's supercomputer, is paid for by Brazil itself.
Countries respond to the rivalry between the two powers with mixed sourcing across layers, just as at the UN they vote issue by issue rather than choosing a side wholesale. The line runs between zones of demand. The mass layer takes what is cheap and sufficient, countries separate the security core from Chinese suppliers, and government data remain a contested zone where what decides is the rung of control the supplier allows. Political anchoring to the US changes this picture less than one might expect. Strongly and weakly anchored countries mix in similar ways, and anchoring shows mainly in the strictness of restrictions around the security core and in the channels available. Pressure on a specific strategic asset matters more than structural anchoring, as Panama shows.
The essay's main analytical result concerns what mixing does and does not provide. It provides sovereignty of choice – the ability to change suppliers and avoid depending wholly on a single power. But it does so only on one condition: the alternative components must be genuinely interchangeable. A country's resilience is determined by the number of workable alternatives in each critical layer, and the number of suppliers by itself says little about it. Mixing provides no economic sovereignty at all. A country can keep five suppliers and remain a host site collecting rent for land and energy, while the income from compute and models flows to the owners. Supplier diversification preserves room for political manoeuvre. Technological autonomy and a share in value creation come only from building national productive and technological capabilities – a move from choosing between external offers to building one's own capabilities.
If the aim of the American strategy is to make its stack the region's standard, it has to sell control and guarantees alongside quality.
That implies several concrete steps. Include vetted American open-weight models in export packages, as CNAS proposes: Latam-GPT shows that the region readily builds sovereign projects on an American base when that base is open. Make export-control procedure predictable: after Fable, partners need clear rules on when and how access may be restricted, and in cybersecurity they need guaranteed access to frontier defensive tools, because this is where the region needs the American frontier most and where it is currently least available. Use DFC's new authorities: in December 2025 Congress raised its exposure cap to $205 billion and lifted the country-income restriction that excluded most of the region (DFC); for this to change the picture, the money has to go to markets where competition with China is taking place and to projects that private capital will not finance on its own. Accept state partnerships in which a country receives compute and a model as its own property – the format in which China currently has no American competitor. And build on the cheapest American advantage, the developer environment, by investing in training the region's specialists and linking that training to projects inside the countries; otherwise American training programs will partly feed talent outflow to the US itself.
The integrated technology offer works, and open weights are China's strongest card, because they answer the demand for sovereignty technically. But China's position has three weaknesses.
The first is money. State lending to the region has fallen sharply, and the CELAC credit line has been announced but has not yet turned into specific AI projects. The Chinese offer rests on technology, and countries pay for the projects themselves, which narrows the circle of those able to take it up.
The second is hardware. Open weights remove dependence at the model level, but China's leverage returns at the infrastructure level: Huawei clouds, Ascend chips and the CANN software environment hold the buyer just as tightly as American clouds do. The harder Huawei pushes Ascend in the region, the more the Chinese offer resembles what countries want to insure against. In addition, the American warning about Ascend chips makes their use legally risky for customers.
The third is the political price. Every major Huawei project becomes a pretext for American pressure, and the sharper the rivalry, the more a Chinese choice costs countries. The main argument in China's favour depends on open weights staying open. If Beijing decides that models are a strategic asset and closes them, that argument disappears.
For the countries themselves, the conclusion is that mixing suppliers is useful but not sufficient. It protects against a complete shutdown if the components are interchangeable, and does nothing about a country's structural position in the AI economy.
Several practical principles follow. Test interchangeability: in every critical layer a country should have a workable alternative to which functions can be moved without rebuilding the whole system, and where possible it should prefer open standards and portable solutions – an open model that can run in any cloud, and clouds that can be left. Write into contracts what Americas Quarterly calls "sovereignty by design": data portability, interoperability, audit rights and clear exit terms. Use national assets – energy, data, location – as leverage while they are needed, and seek in return a share of revenue, ownership of part of the infrastructure or guaranteed access to compute; assess tax incentives for investors by weighing their cost against the expected benefit, so that neighbours' competition for investors does not turn into a contest of incentives. Disclose the terms of the largest deals: while they remain closed, the public and independent analysts cannot tell whether a country is gaining a stake in the AI economy or only rent for a site. Train people together with projects that give them work at home; otherwise training becomes an export of specialists. And pool resources regionally, as ILIA recommends and as Latam-GPT, built with support from the Latin American development bank CAF, demonstrates: what is beyond one country may be within reach of the region.
The essay's conclusions have limitations, and they should be stated plainly.
The indices underlying the typology measure capacity and the structure of ties, not behaviour. National strategies are declarations, and many of the projects mentioned, from Coatlicue to Stargate Argentina, have not been built. The FBIC data end in 2024 and the Talvi and Leiva data in 2023. The sample is limited to the 20 countries covered by ILIA. Mixing is recorded at the level of suppliers' presence and state projects: data on cloud companies' actual customers are almost entirely unavailable. The absence of American deals was established by searching open sources, not through access to financing agencies' closed data, and DFC's claimed participation in the Paraguayan project could not be confirmed. For the security core there are no direct AI procurements in the region, and conclusions about it rest on two cases from telecommunications and cyber assistance, so they remain preliminary.
A general anchoring index such as FBIC captures pressure on individual strategic assets poorly, and the case of Panama shows that a country's vulnerability in a specific sector can differ greatly from its overall anchoring.
The essay's framework is bipolar. The European Union and individual European countries – for example Spain in Brazil's RISC-V chip project – as well as the Gulf states, Korea and Japan are left out of the analysis. Yet FBIC shows that European countries outrank China in overall influence over many countries in the region, and the Paraguay–Taiwan AI centre shows that third partners are already appearing. How they change the logic of mixing is a separate topic.
Politics in the region changes fast. Lula was re-elected in Brazil in October 2026, Colombia has had a new government since August 2026, Chile and Honduras changed governments in 2026, and Argentina votes in 2027. The typology describes the space of choice, not the choice a particular government makes. Separate caveats, set out in section 2.6, concern Venezuela and Cuba.
Finally, the technology itself changes fast. The quality gap between Chinese open models and the American frontier, on which much of the buyer's logic rests, is measured in months, and any new model release from either side can shift it.
| Country | ILIA 2026 group | US influence | China influence | Top three external influencers |
|---|---|---|---|---|
| Chile | Pioneer | 0.316 | 0.189 | US 0.316; China 0.189; Brazil 0.174 |
| Brazil | Pioneer | 0.300 | 0.103 | US 0.300; France 0.190; Germany 0.112 |
| Uruguay | Pioneer | 0.181 | 0.111 | Brazil 0.300; US 0.181; Argentina 0.141 |
| Mexico | Adopter | 0.681 | 0.078 | US 0.681; France 0.133; Netherlands 0.118 |
| El Salvador | Adopter | 0.480 | 0.056 | US 0.480; Guatemala 0.210; Honduras 0.155 |
| Colombia | Adopter | 0.470 | 0.068 | US 0.470; Germany 0.156; South Korea 0.081 |
| Dominican Republic | Adopter | 0.451 | 0.060 | US 0.451; Spain 0.089; Italy 0.061 |
| Costa Rica | Adopter | 0.433 | 0.090 | US 0.433; Guatemala 0.100; China 0.090 |
| Peru | Adopter | 0.328 | 0.160 | US 0.328; China 0.160; South Korea 0.137 |
| Ecuador | Adopter | 0.282 | 0.089 | US 0.282; Spain 0.146; Germany 0.137 |
| Argentina | Adopter | 0.274 | 0.059 | US 0.274; Brazil 0.183; France 0.135 |
| Panama | Adopter | 0.215 | 0.096 | US 0.215; Ecuador 0.110; Colombia 0.103 |
| Trinidad and Tobago | Adopter | 0.179 | 0.126 | US 0.179; Guyana 0.147; China 0.126 |
| Guatemala | Explorer | 0.470 | 0.049 | US 0.470; El Salvador 0.142; Spain 0.114 |
| Honduras | Explorer | 0.426 | 0.052 | US 0.426; Guatemala 0.175; El Salvador 0.154 |
| Jamaica | Explorer | 0.409 | 0.094 | US 0.409; Netherlands 0.115; China 0.094 |
| Paraguay | Explorer | 0.234 | 0.042 | Brazil 0.240; US 0.234; Argentina 0.207 |
| Venezuela | Explorer | 0.220 | 0.264 | China 0.264; US 0.220; Spain 0.160 |
| Bolivia | Explorer | 0.082 | 0.123 | Brazil 0.125; France 0.124; China 0.123 |
| Cuba | Explorer | 0.029 | 0.073 | China 0.073; Spain 0.051; Mexico 0.030 |
Table 2. FBIC scores (0–1) for 2024. Source: Pardee Institute, FBIC v. October 2025; ILIA 2026 groups from CEPAL and CENIA; author's calculations.
Sources tied to the US government are marked where relevant in the text (White House, State Department, DFC, and Diálogo Américas, a US Southern Command publication); China Development Bank is a Chinese state source; Chinese state-media figures are cross-checked against independent outlets. Sources are grouped by the section in which they are first cited.