Latin America spent the week negotiating artificial intelligence capacity in Beijing, one country at a time. The EFE news agency reported that Ecuador, Chile, and El Salvador had high-level delegations in China within the same window of days. Ecuador closed seven agreements for $42.7 million in non-reimbursable funds; Chile ended a five-day tour with cooperation agreed on artificial intelligence, the digital economy, scientific innovation, and industrial chains; and the president of El Salvador’s Legislative Assembly met with his Chinese counterpart.
What did not happen matters as much as what did. No binding technology treaty was signed: what was agreed are intentions. Chilean Foreign Minister Francisco Pérez Mackenna met with Wang Yi, with Vice Premier He Lifeng, and with Science Minister Yin Hejun (with whom he specifically discussed innovation, robotics, and artificial intelligence), but the only thing signed during the entire tour was a phytosanitary memorandum on fresh fruit. At the same time, the Chile-China Express submarine cable remains frozen under review by Chile’s ministries of Defense and Foreign Affairs. Chile offers the “Choose Chile” program to Chinese investors and holds back a Chinese cable in the same week.
The same day, on the other side of the continent, Reuters put a consolidated figure on the Brazilian compute package: 2.3 billion reais, about $444.2 million. Of that total, 1.3 billion reais (more than half) go to the Rio de Janeiro supercomputer built with Huawei and iFlytek, and 1 billion to the Rio Grande do Norte tender, which is expected to be awarded to Nvidia. Brazil is splitting public budget between one American supplier and two Chinese ones, and the official justification is not depending on a single company, technology, or country.
Together, the two stories describe the real state of the region: four countries negotiating separately, with both blocs at once, and no regional body (not CELAC, not ECLAC, not the development bank CAF, not the OAS) at any of the tables. The question is not whether it makes sense to talk to both blocs, because it does. It is whether ten bilateral conversations without a common position get better terms than a single one with the entire region behind it.
Also today
- Nvidia pays $6 billion to license Poolside’s models and hires 109 of its engineers, without buying the company — A non-exclusive license to the system Poolside uses to build its open models, plus $1 billion in investment at a pre-money valuation of $12 billion. The three cofounders stay. It is the third deal of the year with the same structure, which is exactly the one no merger control regime registers as a transaction.
- Nvidia shows that Claude Opus 5 goes from 30% to 100% on a reasoning test without touching a single model weight — The difference came from the harness: the software that surrounds the model and decides how to use it. If seventy points come from systems engineering and not from training, the barrier to competing on autonomous tasks drops by several orders of magnitude. It is worth saying out loud that the research is published by the company that sells the compute.
- Harvey, funded by OpenAI, builds its first legal model on Chinese open weights and reports that it is 90% cheaper per query — The model is trained on top of Kimi K3, from China’s Moonshot AI. The economic argument the region has been making out of budget constraints has just been made, with numbers, by someone who does not have those constraints.
- Gemma passes one billion downloads, and a compressed version is already analyzing images from a satellite in orbit — The Jet Propulsion Laboratory flew a 4-billion-parameter Gemma model on a Loft Orbital satellite: the first in-orbit demonstration of a vision-language model reading its own spacecraft’s sensor. On the ground, India integrated it into a public health app with more than 100 million downloads.
- More than a third of the songs uploaded to Apple Music are generated entirely with artificial intelligence, and they account for less than 0.5% of listens — The figure comes from Apple Music’s own vice president, Oliver Schusser, and arrives alongside the announcement that “Made With AI” labels go from voluntary to mandatory. Volume and consumption differ by two orders of magnitude: it is the best argument available against legislating out of panic.
In the region
Beyond the trips to China and the Brazilian budget, the week recorded no new publications from UNESCO, the OECD, ECLAC, the IDB, CAF, the OAS, or the global GPAI partnership, and the ministries and data authorities of Mexico, Colombia, Argentina, Peru, Uruguay, and Panama were checked with nothing new. It is worth noting, instead, a piece of data by its absence. The annual tech talent report from the consultancy CBRE, a reference for the hemisphere now in its thirteenth edition, measures 50 North American markets with thirteen metrics and does not cover a single Latin American market. Its findings are telling: artificial intelligence jobs grew 45% year over year and already make up 31% of tech job postings in the United States, compared with 11% in 2022; there are 751,000 workers in the sector in the United States and Canada; and New York displaced the Bay Area for the first time with 394,300 workers versus 375,730. Without a comparable measurement of its own, no country in the region can negotiate its place in the talent chain with data, which is precisely what it went to Beijing to do.
Launches
- DeepSeek-V4-Flash-Vision-Exp — An experimental variant that adds reading of images and screenshots to the text model, while keeping the reasoning. Up to 600 images per request and, the detail that matters, it is charged at the same rate as text. It is compatible with the OpenAI and Anthropic interfaces, so migrating means changing one line of configuration. The benchmark results are self-reported: it is advisable to test it with real invoices, forms, and documents in Spanish and Portuguese before believing the table.
- Roblox Sentinel v2, PII Classifier v2.0, and Voice Safety Classifier v3 — Three child safety classifiers in production, released as open source. The personal data detector goes from 17 to 189 languages and improves its precision from 63.41 to 90.52 points; Sentinel detects early signs of risk to minors before the conversation becomes explicit and contributed almost 70% of the cases Roblox identified in twelve months. It is the first time a personal data detector of this level has been available for free in Spanish and Portuguese: it serves regional platforms that could never train it and gives a concrete technical reference to regulators who today require detection without saying with what. It is worth noting that Roblox faces lawsuits over the harassment of minors, so the release also works as reputational defense.
Threads we’re following
The Brazilian compute package has been growing one piece of data a day. Yesterday the tender specifications became known, which were the first complete computational sovereignty strategy written by a state in the region; today came the breakdown of the money, and it shows that 56% of the public budget goes to the machine built with Chinese partners and the rest to the tender that an American supplier is expected to win, both through staggered disbursements from the federal science fund. On the same day that three neighboring countries were negotiating separately in Beijing, Brazil chose the other available strategy: not choosing a bloc, and paying for both.
Outside the region, money moved in the same direction. Nvidia took a minority stake in Cloverleaf Infrastructure, the intermediary between utilities and data centers (the bottleneck stopped being the chip and became the plug), while Broadcom is seeking more than $60 billion in debt to finance Anthropic’s chips, and Anthropic in turn hired Amir Salek, founder of Google’s custom chip program, to set up its internal semiconductor division.
If no one has ever calculated what Latin America would have gotten by negotiating together (price, technology transfer, access to compute already installed), what does that missing calculation say about what we really believe regional bodies are worth when the time comes to sign?
Correction (September 30, 2026). The original version said in the description that four countries in the region brought delegations to Beijing; the correct number is three (Ecuador, Chile and El Salvador), according to the EFE wire story published by Infobae.
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.