For years, the Latin American conversation about computing capacity consisted of negotiating the arrival of other people’s infrastructure. This week Brazil changed strategy: on Thursday the 20th, at the Augusto Severo Technology Park in Macaíba, in Rio Grande do Norte, Lula opened the tender for Brazil’s first state supercomputer dedicated to artificial intelligence, for 1.06 billion reais and 7,200 petaflops of power. That figure would place it among the ten most powerful AI machines in the world. For scale: Santos Dumont, the largest Brazilian supercomputer to date, operates at 27 petaflops.
What is interesting is not the equipment but what the tender specifications require. The document calls for technology transfer, hiring of Brazilian suppliers, and training of some 5,000 people; delivery is estimated between 12 and 18 months after the contract is signed. Rio Grande do Norte was chosen over São Paulo and Campinas on criteria of regional inequality and clean energy availability. And the announcement did not come alone: on the same day, a second supercomputer was presented in Rio de Janeiro, for 1.276 billion reais over five years, built by the National Education and Research Network (Rede Nacional de Ensino e Pesquisa) together with the Chinese companies Huawei and iFlytek to train models in Portuguese; a 10- and 15-year chip strategy based on the open RISC-V architecture; and a National Center for Algorithmic Transparency and Trustworthy AI housed at the Federal University of Minas Gerais. Four pieces that, together, form something we have not found in writing from any other state in the region: the machine, the model in its own language, the chip, and the institutional oversight framework.
The tension was visible from the first day, and Lula himself practically named it when he said that “neither the Chinese nor the American is better.” The only supplier capable of meeting the specifications for the first machine is American; the partners for the second are Chinese. Declared sovereignty consists, for now, of choosing between two dependencies instead of one. Even so, the tender specifications are the only replicable part of the package, and that is why they matter outside Brazil: while Chile proposes a premium token bank in the 2027 Budget and Mexico sets up a national laboratory, Brazil chose to buy hardware with technology transfer and training clauses inside the contract. Which of the three bets ages best will depend on whether those clauses are audited or diluted.
Also today
- Pew measures how much of the web was written with AI — 35% of the pages published since ChatGPT appeared show signs of automated authorship. The useful finding is the breakdown by domain: around 10% for .com, 5% for .org, and just 1% for .edu and .gov. The study analyzed about 490,000 Common Crawl pages and covers only English; we do not know the equivalent in Spanish or Portuguese, where the available human corpus is an order of magnitude smaller.
- Marvell gives Google the option to buy $12.2 billion in shares — The rights are triggered for every $500 million in silicon purchased, through 2033. Broadcom, Google’s previous partner, fell more than 5%. Those who buy outside that circle (universities, states, the Brazilian supercomputer itself) pay list price with no equity in return.
- Researchers hide encrypted instructions in a web page and get Grok to leak chat history — The content filter cannot read encrypted text; the model can, because it has the key. Adversa AI notified xAI on June 3, August 4, and August 10: as of the 19th it was still unpatched and without a mitigation timeline.
- Grok spent days responding with word salad and xAI said there was no incident — Entire paragraphs of meaningless text, which the company attributed to a “rare temporary generation glitch” on the same day that the vulnerability, unpatched since June, was published.
- The joke about cooling data centers with urine turned out to be real public policy — In Loudoun County, Virginia, with more than 250 data centers, recycled water covered about 757 million liters a day in 2025: 43% of demand. The remaining 57% still came from the drinking water network.
In the region
The day brought a single regional development, but a big one, and it is entirely Brazilian. The package announced on Thursday makes Brazil the first state in the region to buy frontier computing capacity for itself, instead of waiting for a global provider to install it. It is worth noting the three questions the announcement leaves open, because they are the ones that will decide whether this was public policy or ceremony: there are still no published rules on who gets access to the machine and by what criteria; the clauses on technology transfer and training 5,000 people are not yet in a signed contract with verifiable metrics; and it is not clear what the new algorithmic transparency center will do that the ANPD (Brazil’s data protection authority, now a regulatory agency, with AI among its four oversight priorities and twenty inspections scheduled for 2026 and 2027) is not already doing, nor why evaluation capacity is housed at a university and not at the regulator. In parallel, and in a much more modest but equally concrete register, Serpro opened a free introductory AI course for the general Brazilian public, in vertical, downloadable video: a format designed for phones and offline use, which is the right decision for the region and one almost no one makes. What was not discussed: the state-owned company developed it with Microsoft, the state’s own cloud and AI provider.
Launches
- Meta AI for Mac, in beta — Meta AI’s first native desktop application: point it at a window and it answers about what it sees on screen, takes voice dictation in any program, and, with a Facebook or Instagram professional account, connects to Google Workspace. Free, 16 MB, in selected regions that Meta does not specify. The declared target audience (small businesses and creators) describes the region’s commercial fabric well; it remains to be seen how it performs in Spanish and Portuguese, and what screen permissions it asks for.
- Binance Agent OS — An access layer that connects AI agents with Binance’s trading, market data, wallets, and payments, compatible with various assistant tools through the Model Context Protocol, an open standard for connecting models with external services. The safeguards are architectural (the agent cannot withdraw funds to external addresses), not behavioral: within its permissions it decides on its own. Given the region’s level of retail crypto adoption, and with none of its financial authorities having published a framework for this, it is advisable to try it in a test environment before using real money.
- ChatGPT plugin for Apple Messages — It reads, sorts, drafts, sends, and deletes messages on the user’s behalf; OpenAI says it runs locally and does not index everything. It matters less for its direct use in the region (here the dominant messaging app is WhatsApp) than for the company’s own warning: it is best not to turn on persistent approval, because it removes the last chance to review a message before it is sent in your name.
Threads we’re following
This adds to something we have been observing: the distance between what the region’s states sign and what they can later verify. In recent days we saw Brazil debut rules for the use of AI in election campaigns and discover, in the first independent measurement, that the country’s most-used assistants did not comply with them; and we saw Amnesty International reconstruct Argentina’s state surveillance inventory by reading public purchase orders that had been in plain sight from the start. The Macaíba tender belongs to the same family of problems: the document says the right words (technology transfer, local suppliers, 5,000 people trained), and everything depends on whether someone checks them after the photo op. That is exactly what almost never happens in the region.
Brazil wrote into a tender what the whole region says it wants, for a machine that only a foreign supplier can deliver. Who today has the mandate and the budget to audit those clauses eighteen months from now, and why has no other country in the region even tried to draft specifications like these before signing a memorandum of understanding?
Correction (September 30, 2026). The original version said the Rio Grande do Norte supercomputer cost 1.6 billion reais; the correct figure is 1.06 billion reais, according to Brazil’s National Laboratory for Scientific Computing (LNCC).
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.