The AI Money Trail in Latin America, Explained

Our July piece framed the regional AI question precisely: political conditioning baked into training data, embedded censorship, and the underlying truth that whoever controls the compute holds the leverage regardless of which flag is on the model card.

That argument now has a balance sheet attached. Every major AI lab, American and Chinese, has capital committed here — some of it as far back as December 2024, some of it landing three weeks ago. The clearest single frame arrived on August 27, 2026, when OpenAI and Alibaba Cloud both opened in São Paulo. Same city, same day. A week earlier, Brazil had already decided it wasn’t choosing between them.

The pattern in the spending says more about how this plays out than the ideological framing did.

OpenAI launched commercial operations in Brazil on August 27, its first office in the Americas outside the United States, and the numbers behind that decision are worth sitting with. Brazil is one of ChatGPT’s three largest markets by weekly active users, with users nearly doubling over the past year and roughly 215 million messages sent each day. By number of developers using the OpenAI API, Brazil ranks second globally, and it is Codex’s largest Latin American market — weekly Codex users up more than elevenfold since the start of 2026, daily interactions nearly thirtyfold. ChatGPT Enterprise seats in Brazil have grown fivefold year over year. In June 2026, 35 percent of classified messages from individual accounts in Brazil were work-related against 30 percent globally: the difference between a consumer market and a commercial one.

OpenAI switched on ChatGPT advertising in Brazil and Mexico on August 11, in the same wave as the United Kingdom, Japan and South Korea, taking the ads pilot to nine markets after an initial US test in February and a first expansion to Canada, Australia and New Zealand. Free and Go tiers see ads; paid tiers don’t. Latin America made the first monetization cohort rather than a later one.

Anthropic moved differently, and earlier. Chile was named the first Latin American home for a Claude Impact Lab in May 2026, aimed narrowly at using AI to translate Chilean financial regulation into plain language for ordinary citizens — a small, specific, credibility-building bet rather than a land grab. Six weeks later, on June 30, Anthropic signed a multi-year alliance with Globant, making the Latin American-born software firm one of the first Preferred Services Partners in the Claude Partner Network and giving 28,500 Globers access to Claude. It’s a distribution play that runs through a company the region already trusts.

Google has run a Querétaro cloud region since December 4, 2024 — its 41st worldwide, third in Latin America, part of a five-year US$1.2 billion regional commitment. Microsoft has pledged US$50 billion to the Global South by 2030. And Amazon — not itself a model developer, but the infrastructure everything else runs on — committed more than US$4 billion in May 2025 to a three-availability-zone AWS region in Chile, its largest single infrastructure bet in the region to date.

Meta took a different route years ago and is still cashing in on it. Its July 2024 expansion carried Meta AI, running on Llama, into Latin America for the first time — Argentina, Chile, Colombia, Ecuador, Mexico and Peru, part of a rollout that reached 22 countries worldwide — live inside Facebook, Instagram and WhatsApp with no new data center required. In a region where WhatsApp is closer to a utility than an app, distribution acquired that way is a bigger reach story than any of this year’s office openings, even though it predates the current capital wave entirely.

Then there’s xAI, whose most notable footprint isn’t a facility or a partnership — it’s a regulatory precedent. Indonesia blocked Grok on January 10, 2026 and Malaysia followed the next day, the first two countries to do so, after regulators found its image tools producing sexualized imagery of women and minors at scale. No Latin American regulator followed, but Colombia is legislating hard in the same territory on its own account: an AI bill running through Congress on a risk-based model, and Decree 0769 of 2026 obliging digital platforms to design against risks to minors. Not every US lab’s presence here is a positive one, and the difference between arriving with capital and arriving as a headline matters.

The Chinese side of the ledger moved just as fast. Alibaba Cloud, which opened its first Latin American data center in Mexico last year, launched its first Brazilian cloud region on August 27 — two new data centers in São Paulo, plus a suite of agentic AI services aimed at local enterprises, the same day as OpenAI’s office. It draws on a roughly US$56 billion AI infrastructure programme.

Huawei is the more complicated case, because it’s playing on both the infrastructure and the model layer. It launched Xinghe AI Fabric 2.0 for the regional market at its Network Summit in Medellín on August 5. Its cloud chief told the South China Morning Post in June that deploying Huawei’s own Ascend AI chips in Latin America is under consideration — not announced, not scheduled, but the thing to watch, because if it lands the compute-dependency question changes shape entirely. And most concretely, Huawei is now a formal partner alongside iFlytek on a Brazilian government supercomputer in Rio de Janeiro aimed at large language model development.

DeepSeek hasn’t needed a single announced facility to matter here. Its pitch — a capable, cheap, open-weight model available to countries that can’t afford closed API bills at scale — is exactly the appeal the July piece flagged, and it has driven a genuine regional debate about whether Latin America should build on Chinese open weights at all. What hasn’t happened yet is the conversion: as of early 2026, Latin America and the Caribbean sat at the low end of global regions for DeepSeek adoption in financial services. The interest is real; the production deployments largely aren’t there yet. Which is precisely when the political-conditioning findings in that piece stop being academic — they matter at the procurement decision, not after it.

If you want the clearest evidence that Latin American governments have already rejected the “pick a side” framing, look at what Brazil did with its money on August 20. The government committed roughly R$2.3 billion (US$444.2 million) to AI supercomputing and split it.

• Just over half — R$1.3 billion (US$251 million) funds a Rio de Janeiro supercomputing project developed with Huawei and iFlytek, to be used primarily to develop large language models for general and sector-specific applications.
• About R$1 billion (US$193.1 million) goes through a tender for a machine in Rio Grande do Norte, chosen for its energy potential, that Brazil expects to rank among the world’s ten most powerful AI processing systems. Reuters reported officials expect Nvidia to win it; Science and Technology Minister Luciana Santos told Folha de S.Paulo she anticipated the same.

The government’s own framing was explicit: the strategy is not to depend on a single company, technology or country, and the investments are aimed at strengthening national sovereignty over data. That’s not indecision. It’s a bargaining position — and it’s the single most useful data point in this whole story for anyone trying to understand how the region intends to play this: not by choosing a bloc, but by making both blocs compete for access.

Here’s the detail that best captures why “closed US versus open China” was always too clean a frame. Chile’s Latam-GPT — launched February 10, 2026, the region’s first open large language model, coordinated by CENIA with more than 60 institutions across 15 countries and more than 100 professionals, trained on over 300 billion tokens of licensed regional text on a Llama 3.1 70-billion-parameter base, built expressly to correct for the fact that Spanish and Portuguese make up roughly 4 percent and 2 percent of the data behind most models — was trained with Amazon Web Services.

The project built to reduce dependence on foreign infrastructure needed foreign infrastructure to exist at all — and Chile knows it. A National AI Computing Centre is being built at the Universidad de Tarapacá in Arica under a 2023 agreement with CENIA: a roughly 10-billion-peso project (about US$10 million) over three years, with a first tranche of some 4.5 billion pesos earmarked to buy and install the hardware, funded from the state universities’ Fortalecimiento 2030 fund plus the university’s own money. Separately, Corfo and the science ministry awarded US$14 million — the largest supercomputing investment the Chilean state has made — to two other AI centres, the Universidad de Chile’s SCAI-LAB and a Tecnoera-run facility in Viña del Mar, financed out of Corfo’s Salar de Atacama lithium lease revenues. Science minister Aisén Etcheverry put the purpose plainly: to develop models like Latam-GPT “with our own technological infrastructure, prioritising our needs and without depending on foreign providers.”

Sovereignty at the model layer turns out to be achievable on a modest budget. Sovereignty at the infrastructure layer is a different order of problem — Chile is spending lithium money on it — and almost nobody in the region currently owns enough of that layer to avoid renting it.

None of this is as clean as “AI companies are betting big on Latin America and it’s working.” Two things complicate it.

First, announcements aren’t capital. In October 2025, two weeks before Argentina’s midterms, OpenAI and Sur Energy signed a letter of intent for Stargate Argentina — up to US$25 billion in Patagonia, to be structured under Milei’s RIGI incentive regime. Seven months later, on May 28, 2026, Buenos Aires Times found the project had never been presented to the government: not among RIGI’s 17 approved projects, not among the 23 under evaluation. Both companies declined to comment, and the Economy Ministry said it doesn’t even count Stargate among announcements, “because there has really been no mention of timescale, sums or localities.” Nothing further has been made public since. It may still happen. Right now it’s a press release, not a data center — a useful reminder to treat every regional AI announcement as a claim to verify.

Second, the physical footprint is generating real friction. Chile’s data center count has grown from a handful a decade ago to roughly 65 facilities today, and a community in Cerrillos, outside Santiago, got an environmental court to partially reverse a Google permit in 2024 over use of Santiago’s strained aquifer. A Google data center in Uruguay was held up amid the country’s worst drought in 74 years after estimates put its cooling demand at 7.6 million litres of potable water a day. Both projects survived by redesigning to air cooling rather than by winning the argument — and in January 2026 Chile’s Servicio de Evaluación Ambiental published its first formal evaluation criterion for data center projects, under a National Data Centers Plan, precisely to standardize how these questions get asked. Meanwhile Brazil is planning an “AI City” in Eldorado do Sul, a Rio Grande do Sul district devastated by the 2024 floods; the developer, Scala, has 1.8GW of grid connection available to it near-term against a guaranteed 5GW over time and a processing target near 4.75GW, and says the site was chosen for ground high enough to have stayed dry — a claim residents who lived through 2024 aren’t taking on faith. None of that stops the capital from flowing. It does mean siting, permitting and water access are becoming as decisive as the geopolitics, and they vary enormously by country.

That last point is the practical takeaway. Brazil is moving toward binding domestic AI regulation while also signing a Digital Partnership with the EU in Brasília on June 12, 2026, covering data governance, AI, digital infrastructure and online platforms. Argentina is doing close to the opposite, though so far only on paper: Milei and Sturzenegger’s June 2026 framework proposes keeping AI free from prior state regulation and creating a “non-human corporation” category for AI-run companies, but no AI statute is in force. Mexico is moving toward screening: a bill Sheinbaum sent to Congress on August 30, 2026 would require National Foreign Investment Commission approval for foreign acquisitions above 49 percent in named sectors including artificial intelligence, semiconductors and data storage — with no affirmative silence, so a missed deadline counts as a rejection.

“Latin America” is not one AI market. It’s at least three different regulatory environments, each negotiating with both Washington and Beijing on different terms — which is exactly why the companies that get this right treat market entry here as a country-by-country problem, not a regional rollout, and the ones that get it wrong call a letter of intent a data center.