Latambusiness.org ·Short Read·July 30, 2026

The AI Battle in Latin America, Explained

Latin America is quietly becoming a decisive front in US–China AI competition — and the contest isn't over chips. It's over model architecture: who ships open weights, who sells closed access, and who ends up dependent on someone else's compute either way.

The two ecosystems are diverging — though not as cleanly as the "closed US vs. open China" framing suggests. US frontier labs mostly sell access to closed models through paid APIs, but Meta's Llama and OpenAI's open-weight gpt-oss release (its first open weights since GPT-2) complicate that binary. Meanwhile, Chinese labs — DeepSeek, Alibaba's Qwen, Moonshot, and Z.ai — are releasing capable open-weight models on aggressive schedules that anyone can download, self-host, and fine-tune.

For capital-constrained developers and governments in the region, the cheap, adaptable option is increasingly the Chinese one. Deutsche Bank analyst Adrian Cox has framed the broader split as Android versus Apple — open and customizable versus closed and proprietary — and on that axis, Chinese labs are currently the more aggressive open-weight shippers. A 2026 RAND analysis goes further, arguing both Washington and Beijing now treat open models as instruments of soft power — while cautioning that hard data on who's actually adopting what remains limited.

The risks here are real. But they're often overstated into "sleeper cell" rhetoric that doesn't survive scrutiny.

🇨🇳
Politically Conditioned Behavior
27.2% flaw rate
CrowdStrike tested DeepSeek-R1 across 30,250 coding prompts. Baseline code quality matched Western peers — but prompts mentioning Tibet, Uyghurs, or Falun Gong saw severe-vulnerability rates climb to 27.2%, an increase of almost 50% over baseline, and the model sometimes refused outright. CrowdStrike's own hypothesis: emergent misalignment from censorship-aligned training, not a deliberately planted backdoor — a distinction that matters for how you mitigate it (rigorous local testing, not blanket bans).
🌐
Embedded Censorship
In the weights
Chinese models demonstrably filter or reframe politically sensitive topics, and some of that behavior lives in the weights themselves, not just hosted deployments — though the effect isn't uniform; some models answer abroad what they'd block inside China. Noema's Nathan Gardels counters that Western models carry their own embedded cultural and policy behaviors — open weights at least let you inspect and adapt what you're running.
🔌
Dependency, in Both Directions
Compute is leverage
Americas Quarterly's analysis of DeepSeek's regional impact flagged single-source dependency and algorithms designed elsewhere as the core risk of any foreign stack — American or Chinese. Free weights still require expensive compute, and whoever supplies the data centers holds the real leverage.

In February 2026, Chile launched Latam-GPT — an open-source model coordinated by CENIA with more than 60 partner institutions across 15 countries, trained primarily in Spanish and Portuguese, with Indigenous-language inclusion planned as a future phase. The model draws on an 8-plus-terabyte regional text corpus and roughly 300 billion tokens (about 230 billion words) compiled from official regional sources, aimed at cutting the cultural bias baked into models trained mostly on English-language web text.

The initial build ran lean: about $550,000 in core funding, mainly from development bank CAF, with a separate ~$5 million supercomputer at Universidad de Tarapacá slated to host future training runs. Then-President Gabriel Boric personally led the launch, framing it as a way to make the region "un actor activo y soberano" — an active and sovereign actor — in the AI economy. Tellingly, version 1 was trained on AWS: sovereignty here is being built incrementally, not declared.

An open question worth watching: Latam-GPT was a flagship of the Boric government, which handed power to José Antonio Kast's right-wing administration in March 2026. Whether a project premised on regional integration and shared public data repositories survives that transition intact will show whether "digital sovereignty" in Latin America is a durable state policy or a temporary political trend.

The honest lesson: the region's answer to foreign open weights is its own open weights, paired with local compute infrastructure, rigorous auditing standards, and diversified cloud suppliers. Non-alignment isn't about picking the "safe" superpower. Neither one is offering dependency-free AI.

Franco Calderón · Latambusiness.org

Deploying AI in Latin America?
Know Whose Stack You're Actually Depending On.

Model choice, compute supplier, and data sovereignty rules are being set market by market, right now. Thirty minutes to map where your exposure actually stands.

Book a Call
Chile Intelligence → Critical Minerals Read → Chinese EVs Read →
Sources   CrowdStrike Counter Adversary Operations, "CrowdStrike Researchers Identify Hidden Vulnerabilities in AI-Coded Software" (Nov 20 2025) · RAND, "Open Models, Soft Power, and the Spectrum of U.S.-China Artificial Intelligence Competition," Daniels & Dohmen (Mar 26 2026) · Americas Quarterly, "DeepSeek Reveals Latin America's AI Crossroads," Levy Yeyati & Guilera (Feb 19 2025) · Noema, "China's Open AI Models Are Advancing Its Global Soft Power," Nathan Gardels (Jul 10 2026) · Fortune, on Deutsche Bank's Adrian Cox and the open-vs-closed AI split (Jul 27 2026) · OpenAI, "Introducing gpt-oss" · CENIA, Latam-GPT launch announcement (Feb 10 2026) · Presidencia de Chile, official launch remarks · Euronews, on Latam-GPT (Feb 12 2026) · Eurac, on Kast's 2025 election win and Chile's March 2026 transition.