OpenAI launched its commercial operation in Brazil, with its own team in São Paulo. It is the company’s first office in the Americas outside the United States, and it did not come alone: along with the announcement came a memorandum of understanding with Prodam, São Paulo’s municipal technology company, to bring artificial intelligence to the public administration of a city of 11.5 million inhabitants. Until yesterday, a frontier lab sold to the region remotely; as of today it has an address, a point of contact and a public counterpart.
The figures that explain the choice are published by the company itself—they are worth reading with that caveat—and they are, even so, the most granular measurement of AI use in a country in the region published to date: Brazil is one of ChatGPT’s three largest markets in the world by weekly active users, with a base that nearly doubled in a year and about 215 million messages a day. It is also the world’s second-largest country by developers using the API and the largest Latin American market for Codex, the company’s coding tool, whose weekly users have multiplied more than elevenfold since January. In June, 35% of messages on Brazilian individual accounts were work-related, five points above the global average.
The detail that turns this into more than an expansion story is the order of events. The memorandum with Prodam—with no public amount and no information about a competitive process—covers the largest municipal administration in the region and, according to the company, is expected to extend to other Brazilian administrations. It comes while PL 2338/2023, Brazil’s AI legal framework, still awaits a vote in the Chamber of Deputies. First comes the supplier with a municipal contract; then the law that was supposed to define its obligations. It is worth comparing that agreement with what ChileCompra or SECOP II would require for an equivalent public purchase. Added to the announcements are agreements with ITA, IMPA, the Hospital das Clínicas of USP, ENTER and Estímulo. And the same day, not by coincidence, OpenAI began showing ads on ChatGPT’s free and Go plans in India, the third market in the sequence after the United States and Europe, always where the mass of free users is largest. Brazil, with a base dominated by those same plans, is the next obvious candidate.
Also today
- OpenAI and more than a hundred companies sign a letter calling for a global defense against AI-powered cyberattacks — Visa, Mastercard, General Motors and Capital One sign alongside the labs. There is not a single Latin American signatory or government on the list.
- The memory shortage caused by AI data centers reaches Android apps — Google says so bluntly, and the blow falls on budget phones, which are the best sellers in the region.
- The same invented authors sign hundreds of papers — a preprint from the Samsung AI Center in Warsaw and the University of Warsaw documents 1,655 records with ghost authorship on Zenodo, CERN’s repository, several with backdated fabricated publication dates and with real DOIs that Google Scholar and ResearchGate are already harvesting. It has not been peer-reviewed.
- Barret Zoph joins Google as vice president of research — four employers in 22 months within the same three-company circuit; he now brings reinforcement learning and post-training to Gemini.
- Hugging Face sells a $399 open-source robot duck — 25 centimeters, lidar, a beak that lifts 800 grams, and it can skate. All this while the company negotiates its sale to Nvidia.
In the region
The day’s other Brazilian event is worth reading alongside the first. The ANPD, Brazil’s data protection authority, fined ByteDance 153.7 million reais—about $29.7 million—for processing the data of children and adolescents on TikTok without a legal basis. The sanction closes an investigation opened in 2021 and rests on the LGPD, Brazil’s general data protection law. But the amount is the least interesting part: what matters are the corrective measures. TikTok must delete the irregularly collected data and automatically activate the most restrictive privacy settings on the accounts of users under 16, changeable only with the express authorization of parents or guardians; expand parental controls; tighten content filters; and, for those who use it without an account, limit the experience to 12 hours, suspend advertising, disable direct messaging and reduce data collection. It is the first time a data authority in the region has told a global platform how its product must be configured by default, instead of merely sanctioning processing that has already taken place; it stands as a reference standard available to Mexico’s Secretariat of Anti-Corruption and Good Governance, Colombia’s SIC, Argentina’s AAIP and Chile’s future authority under Law 21.719. The contrast with the north is instructive: the same day it emerged that Meta’s settlement with the U.S. attorneys general includes $459 million to settle the Cambridge Analytica claims, divided among 46 states and two territories, with a clause under which they waive all future claims, even for facts not yet known. An order of magnitude more money, but conditions negotiated with the company; the ANPD imposed its own. Outside Brazil, the institutional silence continues: neither UNESCO nor the OECD, ECLAC, the IDB, CAF or the OAS published anything within the window, and neither did the ministries and data authorities of Chile, Colombia, Mexico, Argentina or Peru.
Launches
- Qwen3.8-Flash-Next, from Alibaba — a multimodal mixture-of-experts model (an architecture that activates only part of the network for each query, which makes it cheaper to run): 125 billion total parameters and about 6 billion active. Open weights on Hugging Face and ModelScope, including an FP8 version, with no waitlist. Alibaba claims performance competitive with Claude Opus 4.6 and DeepSeek V4-Flash. What no one has measured yet is precisely what matters to us: Spanish and Portuguese.
- Built-in browser in Claude Cowork — Claude’s desktop app (Mac, Windows and Linux) adds its own browser in the side panel: it browses, reads, clicks and fills out forms without using the personal session or a Chrome extension. Active in Enterprise since August 26 and rolling out to Pro, Max and Team. The interesting case in the region is public portals without a programming interface—public procurement, tax filings, case-law searches—which is where it delivers the most and also where it carries the most risk.
- Model Hardware Standard, from Anthropic — a specification for AI agents to operate physical laboratory and manufacturing instruments, with simple commands, device discovery and natural-language labels that declare safety limits. The company says it cuts integration from weeks to hours. It is a closed, application-only preview, with Genentech, the University of Washington, Carnegie Mellon, HHMI Janelia and QuEra as the only partners, none of them Latin American.
Threads we’re following
Week after week, Brazil has been concentrating almost everything the region produces on AI: three days ago we looked at fifty videos of synthetic voters campaigning on TikTok, most of them correctly labeled as AI and illegal nonetheless. Today the same platform receives the first major sanction from Brazil’s data authority, with concrete orders on how a teenager’s account must be configured. And the same day, the Brazilian state shows up on the other side of the counter: as a client of a foreign lab, with no known competitive process and no AI law in force. It is the same country doing both things at once, and that tension promises more chapters.
The same week, two ways of treating a global platform: one is told how its product must be configured; the other gets a signed memorandum. Can a state be a demanding regulator of one foreign platform and an unconditional client of another at the same time, or does the second relationship end up defining how far the first can go?
Correction (September 30, 2026). The original version named the INAI as Mexico’s current data authority; the correct name is the Secretariat of Anti-Corruption and Good Governance, which took over personal data protection after the INAI was abolished in 2025, according to the Secretariat itself. It also said that the $459 million in Meta’s settlement was three orders of magnitude more than the ANPD fine; the correct figure is one order of magnitude (about 15 times), according to the figures from Tech Policy Press, the ANPD and The Next Web.
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.