Brazil · In development
In the third quarter of 2026 Brazil is the focus of the AI Middle Way Coalition, but current efforts are restricted by its upcoming presidential election. What follows is the context of Brazilian AI policy — foreign investment, and the possible role of US foundations in supporting an equitable AI policy.
Briefing current as of 22 August 2026
This document presents the context of Brazil’s AI policy development in detail. Brazil is not yet a formal member of the AI Middle Way Coalition — it is, with Mexico, one of the two nations where the Middle Way is currently in development.
PL 2338/2023 is still stuck in the Chamber’s special committee; the rapporteur is now trying to strip fixed high-risk categories out of the statute, and a copyright/training-data fight is unresolved. Realistic in-force date: 2027–2029.
Filed 9 December 2025 to be grafted onto PL 2338 — creating a national AI system (SIA) with an AI Council modeled on the central bank’s monetary committee. This is the horizontal structure, now on paper.
A deepfake of the imprisoned Jair Bolsonaro endorsing his son Flávio was played at the PL’s 25 July convention — a live case at the electoral court.
President Lula’s Big Tech decrees took effect on 20 July 2026 and have become a campaign axis, entangled with a US tariff dispute.
Maíra Junqueira became Regional Director for Brazil on 7 July 2026, succeeding Átila Roque.
Brazil is governing AI through three moving instruments at once: a stalled framework bill, a national investment plan disbursing slowly, and a data-protection authority acting as de facto regulator. Nothing binding has yet been enacted.
The Senate approved PL 2338 unanimously on 10 December 2024 and sent it to the Chamber of Deputies. There a special committee (comissão especial) was formed, chaired by Dep. Luísa Canziani (PSD-PR) with Dep. Aguinaldo Ribeiro (PP-PB) as rapporteur. The floor vote has slipped repeatedly — from 2025 to “February 2026,” then May, and as of August 2026 there is still no confirmed date. Two fights are holding it: (1) Ribeiro wants to remove fixed high-risk categories from the statute and delegate risk classification to sector regulators under ANPD coordination — which rights researchers argue makes protection unstable; and (2) copyright compensation for artists whose work trains models remains deadlocked. Because the text is being amended, it must return to the Senate after the Chamber, pushing a realistic in-force date to roughly 2027–2029.
The Executive sent Congress a governance bill — signed by ministers Esther Dweck (Management/Digital Government), Ricardo Lewandowski (Justice) and Secom chief Sidônio Palmeira — to be appended to PL 2338. Its purpose is partly constitutional: only the Executive can create or expand a federal agency’s powers, and the Senate text had assigned new competencies to the ANPD. The bill creates the SIA (National System for the Development, Regulation and Governance of AI), with a Brazilian AI Council (CBIA) modeled on the central bank’s COPOM, plus advisory committees for industry/civil society (CRIA) and independent scientists (CECIA), and the ANPD as “residual regulator.” See Section 5.
The Plano Brasileiro de Inteligência Artificial, “IA para o Bem de Todos,” commits about R$23 billion (≈US$4 billion) through 2028, coordinated by MCTI, with Finep as principal financier (a R$5 billion “IA Brasil” program under the FNDCT was set up in May 2025). Execution lags the headline: the flagship sovereign supercomputer (~5,000 GPUs, RISC-V roadmap, ~R$1.8bn) has had its tender slip into 2026 with only ~R$200 million committed so far, and the marquee BNDES–Finep AI venture fund is a modest R$205 million. Academic economists at Unicamp describe a mood shift “from euphoria to disquiet.”
Pending the framework, the ANPD already governs AI through the LGPD (automated decisions, transparency) and launched an AI regulatory sandbox in 2025 (participants Metatext, Synapse AI and IA Greenworld, with USP support, running under supervision until December 2026). Two other tracks matter: the TSE’s election rules on AI (Section 7), and the Redata data-center tax regime — the provisional measure lapsed in February 2026 and a replacement bill (PL 278/2026) passed the Chamber and awaits the Senate.
Luciana Santos (Minister of Science, Technology & Innovation, MCTI) owns the PBIA and is the public face of federal AI strategy. Esther Dweck (Management & Digital Government) is the lead signatory of the December 2025 governance bill and drives the sovereign-cloud agenda; Ricardo Lewandowski (Justice) and Sidônio Palmeira (Secom) co-signed it. Rui Costa (Casa Civil) is the central coordinator, and Henrique Miguel (MCTI secretary) runs the supercomputer procurement. At the ANPD, director-president Waldemar Ortunho leads the agency; board member Miriam Wimmer has been its most AI-focused voice.
The pivotal figures now are rapporteur Aguinaldo Ribeiro and committee chair Luísa Canziani. Government leader Reginaldo Lopes (PT-MG) wants coordinated Chamber-and-Senate passage. In the Senate, Eduardo Gomes steered PL 2338 to approval in 2024; former Senate president Rodrigo Pacheco created the jurists’ commission but left the presidency in February 2025 (succeeded by Davi Alcolumbre), so his role is now historical.
The intellectual spine of the bill is the Senate jurists’ commission chaired by STJ Minister Ricardo Villas Bôas Cueva, with Laura Schertel Mendes (UnB/IDP) as its principal author and members including Ana Frazão, Bruno Bioni, Claudia Lima Marques and Juliano Maranhão (USP). The technical heart of national compute is the LNCC in Petrópolis (director Fábio Borges de Oliveira), operator of the Santos Dumont supercomputer.
Data Privacy Brasil (founder Bruno Bioni) is the leading AI-governance think tank; ITS Rio (director Ronaldo Lemos) is the most internationally visible voice arguing for a pluralist “third way”; the CTS-FGV center in Rio (Prof. Luca Belli, of the CyberBRICS project) anchors Global-South AI-governance networks. These three — plus the Coalizão Direitos na Rede — are largely defending the Senate’s rights-based model against the rapporteur’s changes.
Lemos’s and Belli’s work maps almost directly onto the Middle Way thesis. Both are natural intellectual allies — see Sections 9 and 12.
Foreign capital has arrived overwhelmingly at the infrastructure layer, and the national pattern is clear: the United States is present across the full stack — infrastructure through consumer — while China is infrastructure-heavy and export-oriented. Brazil forecasts nearly US$200 billion for AI and data centers by 2029.
| Investor | Origin | Commitment | Focus / status |
|---|---|---|---|
| Microsoft | US | US$2.7B (2024, 3 yrs) | Cloud/AI data centers in São Paulo state; pledge to train 5M Brazilians; Azure + Copilot. Largest single Brazil tech investment to date. |
| AWS / Amazon | US | US$1.8B (to 2034) | São Paulo cloud region expansion (on top of ~US$3.4B since 2011); Bedrock frontier-model access. |
| Google Cloud | US | Not disclosed | São Paulo region now runs Trillium TPUs; Gemini for Government/Education; Portuguese localization; clients incl. Receita Federal. |
| Nvidia | US | Enabler | GPUs behind Santos Dumont; Jan 2026 “AI Day São Paulo”; named Claro its first LatAm sovereign-AI cloud partner. |
| Oracle | US | Conditional | Two cloud regions; says it has a “check ready” but is waiting on the Redata tax regime. |
| OpenAI | US | Office (2025) | Brazil is a top-3 ChatGPT market (50M+ users); first Latin American office in São Paulo announced Aug 2025. |
| ByteDance / TikTok | China | US$9.8B phase 1 (up to ~US$39B) | Pecém (Ceará) data-center megacampus on renewable power; compute-for-export; operations targeted 2027. |
| Huawei Cloud | China | Undisclosed | Brazil region + 100% YoY revenue growth; AI Token/Agent services; also a hardware vendor inside the government cloud. |
| Alibaba Cloud | China | Undisclosed | Enterprise AI push as part of international expansion; no confirmed dedicated Brazil data center yet. |
On frontier-model reach, OpenAI has the strongest consumer footprint (50M+ monthly users, a new São Paulo office, and regulatory engagement); Microsoft Copilot and Google Gemini reach Brazil through their cloud regions, with Gemini offering explicit government/education/Portuguese tiers. No dedicated Brazil presence or government contract surfaced for Anthropic (reachable via API/AWS Bedrock) as of this research. A new entrant to watch is South Korea’s NAVER, reportedly pursuing a large GPU-cloud campus.
Welcome the capital and the training — but keep the enterprise and civic layers, and the terms, Brazilian.
Read “indigenous” two ways: home-grown sovereign capacity, and literally Indigenous peoples. Brazil is moving on both.
The Santos Dumont supercomputer at LNCC was upgraded fourfold in 2025 to ~18.85 petaflops as the PBIA’s “first step,” with a much larger exascale machine in (delayed) procurement. Note a correction to the July memo: SoberanIA is not the federal program but a Piauí state sovereign-AI ecosystem (Governor Rafael Fonteles; state AI secretary André Macêdo), nationally launched with MCTI in December 2025 with R$40 million — and it is one of the clearest cases of home-grown AI already reaching low-income users (WhatsApp crime reporting, teacher training, public health). On models, Maritaca AI (Unicamp spin-off, CEO Rodrigo Nogueira) builds the Portuguese-optimized Sabiá family; the open Bode model comes from UNESP’s Recogna lab (João Paulo Papa and colleagues); and Widelabs (CEO Nelson Leoni) built Amazônia IA, billed as the first Brazilian-made LLM.
The flagship academic hub is C4AI — the Center for Artificial Intelligence (USP + FAPESP + IBM), directed by Prof. Fabio Cozman, with IBM’s Claudio Pinhanez as vice-director. Add CEIA at UFG (Anderson da Silva Soares), an EMBRAPII-accredited applied-AI unit that made Goiás a national hub, and funding channels through CNPq, FAPESP (whose new ProciêncIA program launched a first call in December 2025) and Finep. Research strength is genuine — Brazil ranks second in Latin America on the regional AI index (Section 11).
The concrete flagship is the Yẽgatu / Nheengatu language-revitalization project led by IBM Research’s Claudio Pinhanez with USP linguists and, crucially, FOIRN (the Federation of Indigenous Organizations of the Rio Negro) — deliberately designed around Indigenous data sovereignty: communities own the data they generate and set the terms of use, with software open-sourced for other languages. This mirrors, almost exactly, the “custody” dimension of the Mexico vertical path framework. The national institutional anchors are APIB (the Indigenous peoples’ articulation) and FUNAI, amid an active Brazilian discourse on Indigenous data sovereignty and bem viver digital.
Brazil now has a horizontal structure in two layers — one operating, one proposed.
Formed in July 2025 and coordinated by MCTI, the PBIA working group brings together roughly eleven ministries and five institutional bodies: MCTI (coordinator), Communications, Casa Civil, Secom, Finance, Management/Digital Government, Development-Industry-Commerce (MDIC), Education, Justice, Foreign Affairs (Itamaraty) and Health — plus Finep, BNDES, CNPq, Capes and EMBRAPII, with Anatel and Telebras also seated. This is, in effect, the interministerial council called for in the July memo — though it sits under MCTI rather than the Casa Civil.
The December 2025 government bill would formalize coordination in a Brazilian AI Council (CBIA) — up to five ministries plus the ANPD, deliberately modeled on the central bank’s monetary-policy committee (COPOM) to give AI strategy standing and continuity — with advisory committees CRIA (industry/civil society) and CECIA (independent scientists). Until it is enacted, GT-PBIA is the operative body.
The public-sector backbone is substantial: LNCC (Santos Dumont); BNDES, building a R$500M–R$1B AI/data-center fund; Finep (with a mandate to direct 30% of the AI fund to the North, Northeast and Center-West); and the Nuvem Soberana (“sovereign cloud”) run by Serpro and Dataprev — launched June 2025, over R$1 billion, more than 250 agencies.
The “sovereign” cloud runs on foreign vendor equipment (AWS, Huawei, Oracle, Google), drawing criticism that sovereignty has been “bought as a product.” That gap — political sovereignty without technical sovereignty — is precisely the space a Middle Way argument occupies.
Brazil’s advantage is that the rails already reach ordinary people; AI can ride on them.
Two mature systems give Brazil a vertical spine most countries lack: PIX, the central bank’s instant-payment rail (with AI now central to its anti-fraud roadmap through 2026), and Gov.br, the digital-identity layer with well over 130 million accounts and ~45,000 services — an obvious surface for an AI assistant layer and for benefit targeting. These are the Brazilian equivalents of the cooperative formalization described in the Coalition’s core thesis.
Norte Conectado has laid five subfluvial fiber infovias under Amazon rivers, reaching about 6.1 million people with a target of 7.5 million by 2027; the FUST digital-inclusion program put R$2.8 billion (via BNDES) into reaching 3.5 million people across 1,223 municipalities and roughly 680 favelas. These are the physical preconditions for any vertical AI reach.
AI is being layered onto health (a “first intelligent SUS services network,” pursued partly through a 2026 health mission to China), education (MEC teacher-training and a public consultation), and courts. Two states lead: Piauí (SoberanIA — the clearest case of AI already touching low-income users) and Ceará (a state AI policy presented in May 2026, plus the Cientista-Chefe program embedding scientists in state institutions). São Paulo is studying a billion-real state AI supercomputer with ~30% capacity reserved for public use. Regionally, SUDENE (Northeast) is funding municipal innovation. At community level, PerifaCode, Olabi and favela-based programs (e.g., in Heliópolis) put a quebrada at the center of AI inclusion — natural partners for the vertical, population-facing pillar.
Brazil votes on 4 October 2026. AI shows up in two registers — as a weapon and as a policy cleavage.
The defining fact: Jair Bolsonaro was convicted in September 2025 (27 years, 3 months) for the coup attempt and is under humanitarian house arrest, ineligible until 2030. His son, Senator Flávio Bolsonaro (PL), was confirmed as the party’s presidential candidate on 25 July 2026 (with Argentina’s Javier Milei as guest). Tarcísio de Freitas declined the presidency to seek re-election as São Paulo governor. Lula (PT) is the front-runner: a Quaest poll in early August put the first round at Lula 39% / Flávio 30%, and the runoff at roughly Lula 44% / Flávio 39% — but pollsters diverge sharply (AtlasIntel and Gerp show a much tighter race), so treat any single number with caution.
At the 25 July PL convention, organizers played an AI-generated deepfake of the imprisoned Jair Bolsonaro declaring “o futuro é Flávio Bolsonaro presidente.” The PT-led federation filed a representation at the TSE; analysts say it configures both an illegal electoral deepfake and early campaigning, and — because Bolsonaro is under house arrest — a possible breach of his detention terms. The case was live as of early August and could rebound onto Flávio’s candidacy.
Lula’s Big Tech decrees (12.975 and 12.976/2026), in force since 20 July 2026, impose platform accountability and target AI-generated non-consensual sexual deepfakes. The opposition brands them “censorship,” and 20 US Republican congressmen wrote the USTR objecting — so the government now runs “digital sovereignty” (data centers, sovereign cloud, national AI capacity) as a campaign theme, framing US/Big-Tech pushback as electoral interference. AI governance is thus not a fringe issue but a live dividing line — useful context for timing any outreach (Section 12).
The electoral court bans harmful deepfakes, mandates highlighted labeling of AI content, prohibits new synthetic content from 72 hours before to 24 hours after the vote, and sets fines of R$5,000–R$30,000 with possible loss of candidacy. A May 2026 precedent (the Evandro Leitão case) established that an electoral deepfake is punishable even if it had no potential to mislead — a strict standard heading into October.
Ford has been in Rio de Janeiro since 1962 — one of its oldest overseas offices. As of 7 July 2026, Maíra Junqueira is the new Regional Director for Brazil, succeeding Átila Roque; the VP for International Programs is Martín Abregú, and the global president is Heather Gerken (the addressee of the elevation package). Ford Brazil’s named programs are civic engagement/government and natural-resources/climate justice; its AI and digital-rights work flows mainly through Ford’s global Technology & Society program and pooled vehicles rather than a Brazil-specific AI line — which is exactly the gap the “elevate to the president’s level” strategy targets.
Two large vehicles matter, and Ford is in both: Open Society’s “AI in the Public Interest” initiative (2023; ten funders incl. Ford, MacArthur, Mozilla, Omidyar; ~US$200M in the field) and MacArthur-convened “Humanity AI” (US$500M over five years, pooled grants from 2026). Both are US-framed with no published Global-South or Brazil earmark — the strategic opening for the Coalition is to give them a credible Global-South destination.
Luminate (Omidyar) funds Brazilian independent media and accountability; Mozilla runs trustworthy-AI grantmaking. Domestically, Data Privacy Brasil (Bruno Bioni) is the key AI-governance recipient — it joined the first UN Global Dialogue on AI Governance in Geneva in July 2026 via a Global-South alliance; Instituto Alana leads on children’s digital rights; Fundação Lemann works on AI in education; and CGI.br/NIC.br runs the OBIA observatory tracking the PBIA. (Note: the funder recalled as “Betty & Jacob Lagnado” is the Instituto Betty e Jacob Lafer, a São Paulo justice-focused philanthropy — not primarily an AI funder.)
At the regional level the scaffolding is multilateral rather than philanthropic: the IDB’s fAIr LAC+ platform (responsible AI, country hubs), CEPAL/ECLAC with Chile’s CENIA (the ILIA index), UNESCO (Montevideo office; the Ethics Recommendation and readiness assessments) and CAF as regional financier. These are the bodies a São Paulo Declaration would plug into (Section 11).
For a São Paulo Declaration, the strongest architecture pairs an academic host with a diplomatic convener and substantive drafters. In rough order of fit:
| # | Institution | Lead contact | Why it fits |
|---|---|---|---|
| 1 | USP — C4AI + IEA-USP | Prof. Fabio Cozman (C4AI) | The on-brand São Paulo academic host: deepest AI+policy stack (technical center + Institute of Advanced Studies’ governance program), FAPESP-funded, public-good mandate. |
| 2 | CEBRI (Rio) | Julia Dias Leite, CEO | Brazil’s premier IR think tank and best pure convener; runs an AI/competitiveness workstream; diplomatic reach (Celso Amorim). Neutral ground; pair with a São Paulo venue. |
| 3 | CTS-FGV (Rio) | Prof. Luca Belli | Strongest AI-governance/legal track record; CyberBRICS and IGF Data & AI Governance Coalition — Global-South framing built in. |
| 4 | ITS Rio | Ronaldo Lemos | Best intellectual fit: Lemos publicly argues for “divergent approaches” to AI governance — the Middle Way thesis. Ideal door-opener and drafter. |
| 5 | MCTI + FAPESP / SP State | Min. Luciana Santos; FAPESP | Governmental weight and staged-policy follow-through (owns the PBIA); São Paulo state + FAPESP give official “São Paulo” sponsorship. Engage after the October election. |
USP (IEA-USP + C4AI / Cozman) as academic host; CEBRI (Julia Dias Leite) as diplomatic co-convener; FGV-CTS (Belli) and ITS Rio (Lemos) as substantive drafters; MCTI/PBIA and São Paulo State/FAPESP as sponsors once the election passes. Unicamp (home of Maritaca) and the Sociedade Brasileira de Computação are natural technical endorsers.
AI Middle Way development in Mexico is not a parallel track — it is the mechanism that could make Brazil reachable. Four concrete channels:
Mexico gives Brazil a partner to join, a language to share, a method to copy, and a coalition to complete.
Brazil is the lever because it is the largest economy and already holds the regional gavel. The regional governance space is real but only lightly occupied — an open lane for a Brazil-led soft-law declaration.
There is a clear ministerial lineage to slot into: the Santiago Declaration on ethical AI (2023, ~20 countries, UNESCO/CAF) → the Cartagena ministerial (2024, 19 countries) → a São Paulo node could be the next step rather than a rival track. Crucially, Brazil hosts the next eLAC Ministerial Conference (2026) under CEPAL’s digital agenda — a calendared vehicle for exactly this. Brazil already ranks second in Latin America on the ILIA regional AI index (behind only Chile), so it leads from strength. The IDB’s fAIr LAC+, CAF and UNESCO Montevideo supply the implementation pipeline that turns a declaration into staged policy.
Two facts define the opening: the EU–Mercosur agreement (in force 1 May 2026) contains no AI provisions, and Mercosur’s most concrete digital step so far is a cross-border digital-ID agreement (Asunción summit, 2026). There is no binding regional AI framework — so a credible São Paulo blueprint (a public compute commons, a Portuguese-and-Spanish model ecosystem, an ethics-and-transparency standard, a BNDES-style financing template) would set the reference architecture others adopt.
Chile (regional pace-setter, eLAC chair, risk-based AI bill), Uruguay (AGESIC’s AI strategy explicitly seeks regional leadership) and Peru (the region’s clearest AI statute, Law 31814 with 2025 regulation) are the natural core, with Colombia (Cartagena host) close behind. Argentina under Milei — pursuing radical deregulation and “unregulated AI hub” positioning — is a likely spoiler; engage it tactically, don’t build consensus around it. A Brazil–Chile–Uruguay–Peru–Colombia core is realistic and ideologically coherent.
This is the most important question. The research points to a clear, low-risk opening move — and to what not to do yet.
Open, informally, with the one Brazilian who already argues the thesis — Ronaldo Lemos (ITS Rio) — and, in the same window, with Prof. Fabio Cozman at USP’s C4AI as the São Paulo academic host. Request a short, no-agenda conversation (virtual is fine), with no funding ask — exactly the pattern that worked in the UNAM approach. Use the freshly signed Mexico City Declaration as the proof point and the invitation for Brazil to become the Coalition’s fourth founding nation.
Lemos is the lowest-friction, highest-credibility door-opener: his public writing on “divergent approaches to governing AI” is the Middle Way, and his global network (Columbia, the former Meta Oversight Board) lends instant standing. Cozman gives the on-brand São Paulo institutional host. Together they can pull in CEBRI (Julia Dias Leite) as diplomatic convener and FGV-CTS (Luca Belli) as drafter — the host architecture in Section 9 — without approaching five institutions cold.
Keep the first phase academic and civil-society only until after the 4 October election. Government engagement (MCTI, ANPD, São Paulo State) is essential for phase two, but electoral-law restrictions, cabinet uncertainty, and the live “sovereignty vs. censorship” fight make government outreach premature and politically risky now. The academic track has no such constraint and can begin immediately.
Mirror the Mexico playbook precisely: propose a single, low-commitment 90-minute invitation-only working session in São Paulo (hybrid), hosted by USP’s IEA or ITS Rio, with Prof. Soraj Hongladarom joining from Bangkok — the Brazilian equivalent of the C3 CITA session. That session, not a signed declaration, is the real first step: it lets the room test the idea and produces the constituency that a São Paulo Declaration later formalizes.
Nations, foundations, and enterprises that share the vision of a third path for AI governance are welcome.
Selected primary and press sources, grouped by section. All were consulted August 2026; some figures from trade press may shift.