Leverage outside the chain Converting what Brazil holds into power in the global AI order — a multi-actor systems analysis

The capability that matters most in AI is concentrating in two states. The United States and China hold the compute, the frontier labs and the capital behind them; everyone else largely consumes what they produce, on access they do not control. Concentration on that scale is its own risk. The worry is not only that a few firms and two governments set the terms of a general-purpose technology. It is that every other country inherits those terms: the prices, the rules, the choice of what gets built and for whom, with no seat where they are decided.

A country outside that frontier has one realistic move. It will not win the race; the gap is measured in orders of magnitude and in a chip industry it does not have. What it can do is refuse to be a price-taker — by pooling leverage with states in the same position, and by writing multilateral rules before the terms harden.

Leverage in the AI order comes in two kinds, and the distinction carries this paper. The first is positional: a seat in the supply chain the frontier already cannot do without. Taiwan holds it because the advanced fabs are there; the Netherlands holds it because ASML is; South Korea holds it in memory. Nobody needs a strategy paper to tell them Taiwan matters — the fabs are the strategy. Brazil holds essentially none of this: no chips, no frontier lab, a research supercomputer two orders of magnitude below a frontier training run.⁶ The second kind is convertible: assets the frontier will want on terms someone still gets to set. Clean energy and siting for a datacentre build-out that is already hunting for both; rare earths and lithium for the hardware chain; a market of two hundred million people behind regulated access; courts that platform business models answer to; a diplomatic network that can assemble the other holders. Convertible leverage does not bite by existing, as Taiwan’s does. It bites only if it is converted — priced, conditioned, and pooled — and that conversion is the strategy this paper designs. This is why the argument has to look beyond the immediate AI supply chain: for a country like Brazil, the strongest cards are not in the chain at all.

Middle powers are the states that rank below the great powers yet still shape outcomes through diplomatic, economic and multilateral weight; the category is contested, and it already stretches from Canada and South Korea to Brazil and Indonesia.³⁶ Brazil is a middle power in that traditional register, not in an AI one — by positional measures it barely registers, and the paper does not pretend otherwise. The coalition it has in mind is therefore defined by holdings, not by rank: the traditional middle powers; the states with positional stakes in the stack, upstream in energy, water, land, critical minerals and datacentre siting, or in the middle layers in model and application development; and the countries whose leverage sits outside the chain entirely, in the large populations whose markets and data the frontier needs. A stake the frontier wants is leverage, whether or not it sits in the chain. And the coalition is not only a bargaining bloc; it is a vehicle for peer-learning and cooperation among countries outside the frontier: shared compute and data, pooled safety research, regulatory approaches copied rather than reinvented.

The Arq Foundation made a version of this case for Europe in June 2026: a continent that consumes frontier AI it does not control, holds roughly five percent of global compute against the United States’ near-eighty,²¹ and keeps losing bargaining power unless it pools what leverage it has.¹ The argument is not specifically European. This paper is a first application of it to one country, Brazil, chosen because it sits at several points of the stack at once — minerals, clean energy, a continental market, a credible diplomatic network. One correction travels with it: where Arq writes in the register of frontier-catastrophic risk and European strategic autonomy, Brazil’s politics responds to sovereignty, near-term harm and development. The structural logic is the same, and the method is built to travel. The same means–ends and actor analysis could be run for a coalition of these countries, traditional middle powers and supply-chain powers together.

The client is the Brazilian federal government, taken at the level of the Presidency and the Casa Civil, which coordinates but does not command. That framing matters for the method: line ministries are not instruments of a unified actor but distinct actors with their own mandates, resources, and incentives. The Ministry of Science, Technology and Innovation (MCTI) owns the national AI plan and the supercomputing stock; Itamaraty owns multilateral reach; the economic bloc of MDIC, Fazenda, and BNDES owns the investment levers; the data-protection authority (ANPD) is slated to coordinate regulation it does not yet have the statute to enforce; and the Supreme Federal Court (STF) and the Superior Electoral Court (TSE) have become the country’s de facto technology regulators while Congress stalls.^5,16,17 The fragmentation is itself part of the problem to be analysed.

AI sovereignty as context: the wave Brazil is joining

A leverage strategy for Brazil does not enter an empty field; it joins a queue. Since 2023 governments with the fiscal room have published AI industrial policies in close succession — investment, procurement, tax credits, regulatory posture — and the AI Now Institute’s AI Nationalism(s) collection surveyed the field, from the US CHIPS Act and the EU’s champion-building to the strategies of India, the UK, South Africa and the UAE.³⁹ The survey’s finding is uncomfortable for every government in the queue, Brazil included: states use AI industrial policy to buy geopolitical leverage and competitiveness while cloaking those objectives “underneath thinly defined ‘AI for social good’ aims,” and the policies mostly end up expanding the concentrated, largely US-based private power they claim to escape. Two of its mechanisms map onto Brazilian instruments directly. The first is what Kak, citing Gabor, calls the de-risking state: government builds an ecosystem lucrative for private capital while the public bears the risk. REDATA has exactly this structure — roughly R$5.2 billion in tax expenditure to attract hyperscale datacentres^7,29 — and only its contrapartidas decide whether Brazil is conditioning capital or subsidising it. The second is the national champion trap: the intention of autonomy cannot wish away dependency. France championed Mistral as a sovereign alternative and watched Microsoft buy into it within a year; Davies describes the UK’s public compute blitz as “a fantasy of independence” resting on Silicon Valley’s funding and infrastructure.³⁹ That is the international base rate against which the PBIA and REDATA should be read, and it is why this paper scores autarky (M5) as harshly as it does.

The word sovereignty itself has become contested ground, and Brazil has particular reasons to notice. Grohmann traces the term to the dependency debates of 1970s Latin America, where sovereignty named the opposite of technological dependence, and documents its capture: since around 2023 the hyperscalers have sold “sovereignty as a service” — “they install a local cloud in the country and say, ‘You are sovereign now because the cloud and the data centers are in your own territory, but they are owned by us.’”⁴⁰ ARTICLE 19’s “data colonialism” warning about REDATA (Section 4) is the same diagnosis in Brazilian terms.³² This paper adopts the distinction as analytic vocabulary: nominal sovereignty is hardware in the territory under someone else’s control; genuine sovereignty is capacity whose terms Brazil sets. Criteria C4 and C6 are built to tell them apart, and M1’s conditions exist to purchase the second rather than the first. Grohmann also names what the state-centred debate leaves out: Brazil already holds a bottom-up counterpart, the popular digital sovereignty of the Movimento dos Trabalhadores Sem-Teto’s technology sector, which has organised tech workers and built worker-led platforms since 2018.⁴⁰ The actor analysis of Section 5 puts these movements in the room.

Reframing impact: what the strategy must be able to show

The second correction comes from the series AI Now, the Aapti Institute and The Maybe published around the 2026 India AI Impact Summit.⁴¹ The summit rebranded the global AI conversation around impact, and the series documents what the word did in practice: “the ‘right’ words were being used to have the wrong conversations.” Impact lingo — AI for good, human capital, frugal AI — dressed a familiar agenda; the summit’s concrete deliverables were datacentre investments announced on the sidelines and India’s alignment with the US Pax Silica declaration.⁴¹ Brazil’s own vocabulary is exposed to the same capture: the PBIA is subtitled IA para o Bem de Todos, AI for the good of all, and nothing in the plan’s R$23 billion obliges anyone to demonstrate that the good arrives.³

Deshmukh and Samdub distill the series into three disciplines, and each translates into a test this analysis can carry. Question: demand evidence of public benefit rather than assertions of it — the gap between promise and delivery is visible when abstract futures of work are set against the documented experience of data workers, or when datacentre siting meets the resistance it has already met in Chile, the US and Canada.⁴¹ Concretize: map who owns the infrastructure and scrutinise the transnational deals — their question, whether Global South countries holding data, labour and raw materials can strike fair deals with hyperscalers or end up giving the cake away for free, is the question REDATA’s conditions answer or fail. Build: federate the community efforts that already exist — linguistic datasets, small task-specific models, cooperative platforms, what Grohmann calls prototypes of struggle — rather than replicating hyperscale from above.^40,41

Read against Brazil, the reframe widens what this analysis must count as impact. Labour: an ILO-based estimate puts 31.3 million Brazilian jobs in exposure to generative AI.³⁴ Opportunity cost: REDATA’s tax expenditure and the BNDES credit lines compete, real by real, with the project the government itself declared first priority, the end of hunger²⁹ — AI Now’s question “should we invest in the advancement of edtech at the cost of providing more students school lunch?” is, in Brazil, a budget line, not a thought experiment.³⁹ Participation: Kapczynski and Michaels’ standard, cited throughout the collection, is that industrial policy must answer to publics beyond elite interest groups, with organised capacity for structurally disadvantaged groups to shape it.³⁹ And existing capacity: the MTST tech sector and the data-worker collectives are not petitioners waiting for policy; they are builders whose prototypes a coalition strategy can federate.⁴⁰

The reframing shapes the analysis in three places, and the rest of the paper carries it. First, criterion C6 carries two tests beyond its socio-environmental composite: demonstrated public benefit and a participatory process (Section 1.3). Second, the problem statement carries a guard against benefit that is asserted rather than shown (Section 1.2). Third, the conditionality of M1 and the design of the coalition instruments M2, M3 and M6 are written against nominal sovereignty and toward federation of existing capacity (Sections 3, 4 and 6).

Brazil’s national project, and where AI sits inside it

Brazil is not short of plans. In close succession the Lula government has published a four-year master plan (the PPA 2024–2027), a neo-industrial strategy to 2033 (Nova Indústria Brasil), an ecological-transformation plan run from the Finance Ministry, a climate plan with a binding emissions band, a R$1.7-trillion investment programme (Novo PAC), a decade-long science-and-technology strategy (ENCTI), the national AI plan (PBIA), and a datacentre-incentive regime (REDATA).^{23,24,25,26,27,28} Read together they cross-reference each other and converge on a recognisable project. Six macro-objectives recur, each anchored in named instruments:

• Neo-industrialization and productive sovereignty — rebuild domestic capacity on green, digital, health and defense bases. NIB, ENCTI, Novo PAC, PPA Eixo II.
• Ecological transformation and just climate leadership — reconcile growth with decarbonization. Plano Clima (59–67% cut by 2035, net-zero 2050), Plano de Transformação Ecológica, TFFF.²⁵
• Technological, digital and data sovereignty — control over data, compute and critical technologies. PBIA, REDATA/PNDC, Nuvem Soberana, NIB Missions 4 & 6.
• Social inclusion and the end of hunger and inequality — the declared first priority. Brasil Sem Fome, Bolsa Família, Minha Casa Minha Vida, PPA Eixo I.²⁹
• Energy transition as a development asset — leverage the clean matrix for industry and exports. PNE/PDE, Combustível do Futuro, the ~88% renewable grid.³³
• National sovereignty and an autonomous foreign policy — strategic autonomy, multipolarity, middle-power leadership. Active non-alignment, G20/BRICS/COP30, critical-minerals policy.

The strategy advances four of the six macro-objectives and is bounded by two. It advances technological, digital and data sovereignty, its home; neo-industrialisation, through AI-modernised industry; the energy transition, where the clean grid becomes a bargaining and export asset; and the autonomous foreign policy the coalition is built to serve. In the government’s own framing the means map onto real instruments rather than inventing new ones. It collides with the other two, ecological transformation and social inclusion, wherever leverage is pursued by extraction. Mining the critical minerals that give Brazil its bargaining power means opening lithium and rare-earth frontiers in the Jequitinhonha valley and toward the Amazon, on Indigenous and quilombola land and scarce water;^30,31 siting the datacentres that host sovereign compute means drawing energy and water the climate and water-security goals have already claimed.³² The strategy is therefore bounded. It is worth pursuing only where it stays consistent with the emissions band, the territorial rights and the just-transition commitments Brazil has set itself — and where it can show Brazilians what they get for it. Section 4 takes those conflicts up directly; criterion C6 carries them through the scoring. What this paper does not do is deliver the energy transition, end hunger, or run ecological transformation as ends in themselves: those are separate branches of the national project, and they enter here only as assets the strategy draws on or bounds it must respect.

The analysis follows the sequence set out by Enserink and colleagues.² Section 1 demarcates the problem through a means–ends analysis, selects a focal objective, and derives criteria — now carrying the public-benefit and participation tests the introduction set out. Section 2 builds the system diagram. Section 3 scores the policy options qualitatively and tests them against scenarios. Section 4 surfaces the conflicts between the strategy and Brazil’s other commitments, and the bounds they impose. Section 5 turns to actors and institutions. Section 6 integrates the views in a multi-actor system diagram. Section 7 sketches how the same analysis would generalise to a middle-power coalition. Section 8 reflects on the roles played and the limits of the exercise. An appendix sets out the axes of each national plan.

1Problem demarcation, objectives and criteria

1.1Means–ends analysis and focal objective

The fundamental objective from which everything else hangs is a prosperous, sovereign, and secure future for Brazilians, and the national project’s six macro-objectives serve it (Figure 1). The branch this paper develops is technological, digital and data sovereignty, the objective of securing Brazil’s agency over its AI-driven future. It decomposes into three means: building sovereign AI capacity, governing AI and containing its harms, and building international AI leverage and coalitions. The other five macro-objectives stay in the tree as context: three the strategy advances, two it must not set back.

Selecting the focal objective is a choice, and it carries the argument. Building sovereign capacity at the frontier is the option Brazil is most tempted by and least able to afford. The flagship asset, the upgraded Santos Dumont supercomputer, reached 18.85 petaflops in 2025, the fastest research machine in Latin America.⁶ A frontier training run is two orders of magnitude beyond it. The domestic chip programme collapsed with CEITEC, so there is no closing that gap soon. And the national AI plan’s R$23 billion is real money spread thin across five axes.³ Governing harms is necessary, but it is already being done, and not by the client: the STF rewrote platform-liability law in 2025 and the TSE wrote some of the world’s most detailed rules on electoral AI without waiting for a statute.^16,17 What remains underexploited, and what Brazil is comparatively well placed to do, is to convert the assets it actually holds into leverage and to organise partners around it. That is the focal objective — with the introduction’s caveat attached: leverage is an instrument for the fundamental objective, not a substitute for it, and the criteria below are built to keep the difference visible.

1.2Focal means, side-effects and the problem statement

Five means serve the focal objective, each a distinct theory of how Brazil might generate leverage (Table I). Four are coalition- or asset-based; the fifth, autarky, is the deliberate foil — leverage through independence rather than through partners. Each carries a characteristic side-effect, and reading those side-effects together produces the problem statement.

None of the five is invented for this paper. The option space is assembled from instruments Brazil already operates and from proposals already on the table, then stretched to its corners as the method requires.² M1 generalises the contrapartida logic REDATA and PL 2780 wrote into law: the conditions exist, they are merely weak. It is also what Mazzucato and Rodrik propose as industrial policy with conditionalities — public support that carries enforceable terms — with the CHIPS Act guardrails as the great-power precedent.^7,30,43 M2 extends what Itamaraty has already begun: the BRICS Leaders’ Statement on AI governance negotiated under Brazil’s 2025 presidency and the G20 Maceió digital declaration are its first moves.^12,14 M3 is the UN’s own proposal in Brazilian clothes: the High-Level Advisory Body recommended a Global Fund for AI and a capacity-development network in 2024.¹⁵ Current AI — US$400 million launched at the Paris summit, with Chile, India, Kenya and Nigeria among its nine partner governments — shows the pooled-fund form already exists; the TFFF supplies the financing template Brazil knows how to run.^13,44 M4 is not proposed but observed: the STF ruling, the TSE resolutions and the ANPD’s 2024 suspension of Meta’s AI training on Brazilian personal data are the lever in use. Von Thun makes the general case for wielding existing regulation, the EU’s DMA, as AI industrial policy.^16,17,39 M5 is the corner the option space must contain and the pitch Brasília is most often sold — the full-stack “sovereign AI” programme; keeping it in the table is what lets Section 3 show it dominated rather than merely assert it. The sixth means, the commons the partners introduce in Section 6, has a working regional prototype in LatamGPT.⁴⁵

The side-effects share a shape. Every route to leverage either provokes the powers Brazil depends on, deepens the dependency it is trying to escape, or strains a domestic system that is already incoherent. The introduction added a fourth failure mode, quieter than the other three: a strategy that succeeds as leverage and still cannot show the public what it bought. Together they give the problem statement.

How can the Brazilian federal government maximise its leverage and coalition power in the global AI order without provoking great-power retaliation, deepening technological dependency, fragmenting domestic governance, compromising its own climate, territorial-rights and social commitments — or spending public money on AI whose public benefit it cannot demonstrate?

1.3Objectives tree, criteria and indicators

The problem statement names one objective to maximise and four conditions to respect. Adding coalition strength, which the statement folds into leverage but which the coalition logic of the case makes worth isolating, yields six criteria (Figure 2). The sixth, C6, carries the fourth guard: beyond Brazil’s climate, territorial-rights and just-transition commitments, it scores the two tests the introduction imported from the impact critique — whether an option’s public benefit can be demonstrated, and whether affected publics take part in shaping it.^39,41 C6 is marked in clay throughout as the bound the national project imposes. Scores express each option’s contribution toward the desired direction of a criterion, so a plus is always favourable — including for the criteria to be minimised, retaliation exposure (C3) and dependency (C4).

• C1 — Strategic leverage (maximise). Indicator: share of inbound AI-infrastructure and critical-mineral capital placed under enforceable sovereignty conditions [% of CapEx], read alongside a qualitative leverage index.
• C2 — Coalition strength (maximise). Indicator: number of states aligned in a Brazil-anchored AI bloc, weighted by the depth of their commitments.
• C3 — Retaliation exposure (minimise). Indicator: trade and technology value exposed to coercion [US$ bn at risk].
• C4 — Technological dependency (minimise). Indicator: reliance on foreign-owned compute and frontier models [% of national stack] — counted by control, not by geography, so that a foreign-owned datacentre on Brazilian soil still counts as dependency (the nominal-sovereignty distinction of the introduction).⁴⁰
• C5 — Governance coherence (maximise). Indicator: number of overlapping AI authorities and conflicting mandates [count, inverse].
• C6 — Socio-environmental integrity & public benefit (maximise). Indicator: consistency with the Plano Clima emissions band, ILO 169 consultation and territorial rights, water-security limits, and the just transition; plus the two tests of the introduction: demonstrated public benefit (each public real spent names the public it serves and the evidence it will be judged by) and participation (channels for labour, Indigenous and community organisations in design, not only in consultation after siting) [composite, against named commitments].^39,41

2System diagram

The system diagram (Figure 3) connects the five means, on the left, to the six criteria, on the right, through the internal factors the client can influence and the external factors it cannot. External factors enter from the environment as dashed slate nodes: the intensity of US–China technology rivalry (X1), the pace of frontier AI progress and the compute demand it generates (X2), and global demand for critical minerals and clean energy (X3). The last is the quiet engine of the whole case — it is what gives Brazil’s minerals and grid their bargaining value.^7,9,11 Two factors are drawn in clay: mineral and datacentre expansion enlarge a socio-environmental footprint that erodes Brazil’s socio-environmental licence, the chain that feeds C6.

A correction on the mineral portfolio is owed here, because niobium leads nearly every sovereignty speech — Brazil holds the world’s largest reserves and roughly nine tenths of production — and it buys less than it appears to. About ninety percent of niobium leaves the economy as ferroniobium, an additive for high-strength low-alloy steel, in a global market on the order of US$2–3 billion a year;^9,42 its AI-stack uses (superconducting magnets, capacitors) are niche; and the deposit at Araxá is operated by CBMM, a private company whose contracts the federal government does not hold. A monopoly over a small, steel-directed market is a fine business and a poor bargaining chip — in the introduction’s terms, not even positional leverage, because the AI chain does not run through it. The minerals that carry AI-relevant leverage are different: rare earths, where Serra Verde and the second-largest reserves on earth sit inside the permanent-magnet supply chain the US and EU are anxious about;^9,10 lithium, where the Jequitinhonha frontier is opening;³¹ and the in-country beneficiation play of PL 2780, which converts ore into processed value rather than exporting the leverage raw.³⁰ Wherever this paper says mineral bargaining power, those are the weights: rare earths and lithium carry it; niobium is a headline, not a lever.

Three structures in the diagram do most of the analytical work. The first is a balancing feedback loop, marked in heavy clay, which this paper names the accommodation trap. A more cohesive coalition raises Brazil’s leverage, which makes the bloc more threatening, which raises great-power coercion, which fractures cohesion as members are picked off bilaterally. It is the dynamic Mark Carney named at Davos in 2026, when he warned that middle powers who bargain alone with a hegemon end up competing to offer the most accommodating terms — that, in his phrase, “if we’re not at the table, we’re on the menu.”¹⁸ The loop is balancing, not reinforcing, which means coalition gains decay unless something offsets the defection pressure. Institutionalisation offsets it: a facility with rules and disbursements, not a communiqué. That is the structural argument for M3.

The second structure is the sovereignty–dependency paradox. Inbound capital builds sovereign compute, which reduces dependency; the same inbound capital, if it arrives as foreign-owned hyperscale infrastructure, deepens reliance on foreign models and operators, which increases it. Whether the conditions bite decides which effect dominates. This is the paradox AI Now documents across seven national strategies and Grohmann compresses into “sovereignty as a service”: hardware in the territory is not capacity under national control.^39,40 M1 is the edge that bends the paradox toward sovereignty; without it, the REDATA incentive regime risks subsidising the very dependency it is meant to cure.^7,8

The third structure is the extraction–licence chain. Mineral bargaining power and inbound datacentre capital both enlarge a socio-environmental footprint: lithium and rare-earth frontiers on Indigenous and quilombola land, datacentres drawing scarce water and clean power. That footprint erodes the socio-environmental licence C6 tracks.^31,32 M1 and M3 are the levers that protect that licence: renewable-energy and water contrapartidas, free, prior and informed consultation under ILO 169, the TFFF’s Indigenous safeguards. An unconditioned extraction push has none of them. The chain is why the same minerals that raise C1 can lower C6, and why the strategy has to be conditioned rather than merely pursued.

3Qualitative impact assessment with scenarios

3.1Means–criteria scorecard and recommended policy

Table II scores each means against the six criteria. The pattern is clear before any weighting. M1, M2, and M3 are complementary: M1 generates leverage from assets and bends the dependency paradox, M2 builds the coalition, and M3 institutionalises it against the accommodation trap. M4 is powerful on leverage but is the only option that scores doubly adverse on retaliation. The United States’ July 2025 tariff action named the 2025 platform-liability ruling and the PIX dispute explicitly.¹⁹ M4 also worsens governance coherence by deepening reliance on the judiciary. M5 is dominated: it sacrifices the coalition for an independence Brazil cannot manufacture without a chip industry, and it carries the heaviest extraction footprint.

C6 does double work in the table, and two readings depend on it. The first concerns M5: under the public-benefit test autarky scores strongly adverse, because a national-champion stack concentrates public money on capacity no public was asked about — the pattern the AI Now survey finds wherever champions are funded on competitiveness grounds alone³⁹ — and because it carries the heaviest extraction footprint in the table. The second concerns what the recommended means’ positive scores assert. They hold only for the conditioned versions: capital traded for safeguards and audited local benefit (M1), a coalition that travels with shared safeguard norms (M2), a facility built on the TFFF’s Indigenous-protection template with participation in its governance (M3). The conflict the criterion names does not disappear; the conditions hold it off. Strip the conditions, pursue the same minerals and datacentres without the renewable-energy, water, consultation and benefit contrapartidas, and M1 collapses to strongly adverse on C6. The recommendation, then, is not “leverage” in the abstract but conditioned leverage, and Section 4 sets out which conditions.

† M5 reduces dependency only in the long run and only if a domestic chip and model capacity materialises, which the CEITEC experience suggests is implausible on any policy-relevant horizon.³ ‡ M1 scores positive on C6 only where its conditions actually carry the safeguards and the benefit tests: renewable-energy and water contrapartidas, ILO 169 consultation, and auditable local benefit — jobs, R&D, technology transfer stated in advance and checked. Strip those and the same lever falls to −−. That contingency is the subject of Section 4.

The recommended policy is the bundle M1 + M2 + M3, with M4 used selectively as a bargaining instrument rather than a posture, and M5 rejected. The three reinforce each other across exactly the criteria the problem statement protects, and M3 is what keeps the coalition from decaying under the accommodation trap. M4 is held in reserve because its leverage is real but its retaliation cost is the highest in the table and falls on the whole economy, not only the AI file.

3.2External-factor scorecard

Table III reads the other way: it asks how the criteria move if each external factor intensifies, holding the recommended policy fixed. The reading is mostly encouraging for the strategy and explains its timing. Scarcer compute and hungrier mineral and energy markets raise the value of precisely the assets M1 conditions and M2–M3 pool. The exposure runs through rivalry: a harder US–China split raises both the value of Brazil’s neutrality and the cost of exercising it. The same table carries the warning. The mineral and energy demand that scores ++ on leverage (C1) scores adverse on integrity and benefit (C6): the hotter the market, the harder the pressure to open frontiers faster than the safeguards can be built — and, the impact critique adds, the easier it becomes to announce benefits instead of demonstrating them, because a hot market supplies its own headlines.⁴¹ C1 and C6 pull against each other through X3, and that is the conflict the strategy has to manage rather than ride.

3.3Scenario analysis

Combining the external factors gives four scenarios. S0 is the baseline. S1, bipolar lock-in, is a hard US–China split into technology blocs (X1 high). S2, frontier acceleration, is rapid capability growth with compute and resource demand spiking together (X2, X3 high). S3, multilateral spring, is a world where a UN-centred regime and the Global Digital Compact gain real traction and middle-power coordination becomes cheap.^14,15 Table IV scores the recommended bundle in each.

The bundle is robust in the weak sense that matters for strategy: it never fails outright, and the instrument that carries it shifts predictably with the world. In a splitting world the facility (M3) does the work of holding the coalition together; in an accelerating world the conditionality (M1) does the work of keeping leverage from curdling into dependency. Autarky (M5) fails in every column of every scenario, which is the point of having kept it in view.

4Conflicts and bounds

The strategy meets Brazil’s other commitments at seven points, and at five it cuts against the project unless bounded (Table V). None of the conflicts is hypothetical. Each is a live dispute with a paper trail: Ministério Público investigations, ILO 169 complaints, state-assembly inquiries. Each also has a place in the plans where Brazil set its own line.

Read down the last column and the bounds are already written. Brazil has set itself an emissions band, signed ILO 169, embedded a just transition in the Plano Clima, attached contrapartidas to REDATA, and subtitled its own AI plan “for the good of all.” The strategy does not need new safeguards invented for it; it needs the existing ones treated as binding rather than aspirational. Seven commitments bound the means:

The bounds. (1) The Plano Clima emissions band — compute and mining expansion must fit sectoral carbon budgets, and new clean generation for datacentres must be additive, not diverted. (2) Indigenous and quilombola protection — ILO 169 consultation, with demarcation status never used to bypass it. (3) Water security — enforce REDATA water standards and require EIA/RIMA for large datacentres and mines in stressed basins. (4) Just transition — reskilling, social protection, and a say for workers over technology. (5) Genuine sovereignty contrapartidas — domestic capacity, R&D and interoperability, not nominal data residency. (6) The green-industrial logic — channel mineral and compute value into domestic processing and the bioeconomy, not raw export. (7) Demonstrated public benefit and participation — every public real spent on compute, attraction or minerals names the public it serves and the evidence it will be judged by, and affected publics get a channel into design, not only a consultation after siting.^39,41

The practical consequence is sequencing, not abandonment. The leverage is real and the window (Section 3) is open, so the strategy is worth pursuing — but staged behind the safeguards rather than ahead of them. Five triggers should change the posture outright: a datacentre cluster whose projected demand exceeds local clean-generation additionality should be relocated or deferred; a breach of a Plano Clima sectoral budget should cap the compute and mining electricity draw; an adverse ILO 169 or Funai consultation finding should halt licensing; unmet REDATA contrapartidas should claw back the incentive; and a beneficiary that cannot demonstrate its promised local benefit at audit — the jobs, the R&D, the transfer it was credited for — should lose the incentive on the same claw-back logic. None of these is exotic. Each is enforcement of a commitment Brazil has already made.

5Actors and institutions

The client is a coalition, not a principal. The analysis so far has treated “Brazil” as if it chooses; Section 5 drops that fiction and looks at the actors whose alignment the strategy requires — including the environmental authorities, the affected communities, and the organised movements that the conflicts of Section 4 put squarely in the room.

5.1Interests, objectives and possible actions

Table VI lists the principal actors, their core interests, the resources they bring, and the actions open to each; the lower block holds the external actors.

5.2Formal institutional context

The formal rules constrain who may do what, and several of them are unsettled in ways that bear directly on the strategy (Table VII). Two are worth stating in the Crawford–Ostrom grammar,³⁸ because they are the rules most likely to break.

ADICO — rule on the regulator’s mandate. The Executive (attribute) must (deontic) initiate legislation to vest AI-governance powers in the ANPD (aim) whenever those powers exceed its existing statutory mandate (condition), or else the resulting act is exposed to challenge for unconstitutional initiative (or-else).⁵

ADICO — rule on platform liability. An application provider (attribute) must (deontic) act expeditiously to remove manifestly unlawful content (aim) once it is aware of it, even absent a prior court order (condition), or else it may be held civilly liable (or-else).¹⁶

5.3Resource dependency

Figure 4 places the client at the centre of the flows it depends on. The client’s hardest dependencies (compute, chips, frontier models) run to actors outside its authority, and the same actors hold the coercive instruments (tariffs, export controls) drawn in heavy clay. What the client can offer in return is real but upstream: minerals, clean energy, market access. One flow is easy to miss: civil society and the movements supply legitimacy and consent — and working prototypes of community-run technology — in exchange for participation channels and funding the client mostly has not offered.⁴⁰ The strategy tries to improve the terms of trade along these arrows, and to add the bottom channel, Itamaraty to middle-power partners, as a counterweight to the great-power channel on the left.

5.4Power–interest matrix

Figure 5 sorts the actors by power and interest. The key-players quadrant is crowded, and that crowding is the finding. AI governance in Brazil has too many high-power, high-interest actors and no settled hierarchy among them. That is the coordination problem the strategy must solve internally before it can project anything outward. Three placements carry most of the implication. Congress sits as a powerful but disengaged context-setter — it holds the veto over the AI law yet has let it stall, which is what leaves the judiciary filling the vacuum. The ANPD, the natural regulator, sits among the subjects, high in interest and low in present power, waiting on a statute. And the middle-power partners sit there too: individually weak, collectively the entire premise of M2 and M3. Moving those partners from subjects toward players is, in one image, the external task of the strategy; empowering the ANPD is the internal one. The conflicts of Section 4 add three more placements. The environmental bloc (MMA, Ibama) holds real licensing power and sits among the context-setters the strategy must keep satisfied. Indigenous peoples and affected communities sit low in formal power but high in interest; through ILO 169 and the courts they can halt a project outright, which makes moving them from contestation toward consent a precondition of the strategy, not a courtesy. And the civil-society and tech-labour movements sit beside them, low in formal power and high in interest. They hold what bound (7) requires and no other actor in the matrix can supply: participation, and the working prototypes to ground it.^40,41

6Multi-actor system diagram

The final diagram (Figure 6) returns to the causal map of Section 2 and colours each recommended means by the actor that must own it. The picture is a division of labour. The economic bloc owns conditionality (M1); Itamaraty owns the coalition and the facility (M2, M3); the judiciary’s leverage over Big Tech (M4) is available but kept to the side. The diagram also adds what the partners bring rather than what Brazil does to them: a partner-owned means, pooling a shared compute and data commons among outside-frontier peers (M6), and a partner criterion, the capacity gains those partners would need to see to stay in the coalition (C7). C7 is the answer to the accommodation trap. Partners defect when the bloc costs them more than it returns; a facility that visibly raises their capacity, and a commons they learn and build from together, is what makes staying worth more than the bilateral deal a great power will offer them.

The commons is also where the reframing’s third discipline, build, lands in the causal map. The efforts worth federating largely exist: linguistic datasets for languages the frontier ignores, small task-specific models, cooperative platforms, pooled compute among aligned organisations — and, in Brazil’s own region, LatamGPT, the open model CENIA coordinates with some sixty institutions across fifteen countries, already framed by its builders as a regional public good.^40,41,45 A commons assembled by federating that capacity — from member states and from the movements inside them — raises C7 at a fraction of hyperscale cost, and it is the one instrument in the bundle that cannot be bought back by a hyperscaler, because there is no single owner to buy. The causal structure of Section 2 survives the reframing otherwise intact; what changed is what C6 demands of every path that reaches it.

Read together, the actor analysis and the causal map give the same instruction. The strategy is feasible only as a coalition — internally, among ministries that currently optimise separately, and externally, among partners who will stay only if the facility pays. The client’s distinctive job is not to build any single means but to hold that coalition together long enough for the assets to be converted before the accommodation trap pulls it apart. And it has to carry the safeguards rather than externalise them: leverage assembled by breaking Brazil’s own climate, territorial and social commitments (C6) buys a new dependency, not sovereignty — and leverage that cannot show its public what it bought will not survive the first change of government.

7From Brazil to a middle-power coalition

Brazil is one client. The same analysis runs for a different unit: a coalition of middle and supply-chain powers as the client itself, negotiating the global AI order together rather than each on its own. Most of the structure carries over. The fundamental objective is shared: agency over an AI-driven future without winning the frontier. The focal means survive almost unchanged: condition inbound capital on common terms (M1), lead and widen the coalition (M2), institutionalise it in a shared facility (M3), and hold collective market and regulatory power over the frontier vendors (M4). Autarky (M5) fails for the bloc as it fails for Brazil. The coalition is also where peer-learning and cooperation happen: a shared compute and data commons (M6), pooled safety research, and regulatory lessons traded across members rather than each learned the hard way — and, following Section 6, a federation of the community-built capacity its members already hold, rather than a bloc-scale replica of the hyperscale model.⁴¹

What changes is the actor map. The client is no longer one government but a set of them, each carrying the defection incentive the accommodation trap describes, plus a coordinating secretariat that has to be built rather than assumed. The great powers and the hyperscalers keep their roles. The members divide along the stack the introduction set out: compute and capital from some, energy, minerals and land from others, large consumer markets from a third — a few holding positional leverage, most holding convertible.³⁶ The coalition’s leverage is the sum of those stakes; its fragility is that each member can be offered a separate bilateral deal — and the sovereignty-as-a-service pitch is precisely that bilateral deal, made attractive one capital at a time.⁴⁰

The criteria carry over, with one of them turned outward. C6, the socio-environmental and public-benefit bound, becomes a shared standard the coalition has to hold — otherwise members compete each other’s safeguards down, and a race to the bottom on environmental, labour and benefit terms becomes the coalition’s version of the accommodation trap. This is a sketch, not the analysis. A proper treatment would name the members, score the bundle for the bloc rather than for Brazil, model the defection dynamics directly, and add the criteria that appear only at coalition level — burden-sharing, the seat of the secretariat, the rule for admitting members. That is the next paper; this one establishes the method on a single, concrete case first.

8Reflections

Placed on Mayer’s hexagon of policy-analytic activities,³⁷ this paper sits mainly in two corners: research and analyse, in the system and actor models, and design and recommend, in the scored bundle. It also works a third corner, clarify values and arguments: the introduction states openly which conception of sovereignty and which conception of impact the scoring rewards, and cites the critique that forces the choice.^39,41 It still does not attempt to mediate or democratise; those would require the partners — and the movements — in the room, not modelled on a page.

The style is correspondingly the rational and the client-advice style. The analysis is laid out so its judgements can be checked and argued with, and it is addressed to a client with a decision to make. It is not a participatory exercise, and it should not be mistaken for one. The actor analysis is a model of interests, not a substitute for asking the actors — a limit the reframing itself sharpens, since its standard is that affected publics shape the policy rather than appear in its tables.³⁹

The recommendation is a single package, conditioned and sequenced. The bracketed codes tie each item back to the analysis; the plain sentences stand on their own:

• Lead with leverage, not a frontier race. Brazil cannot out-build the United States or China at the cutting edge, and should stop spending scarce public money trying. Its leverage is what it already holds: minerals read soberly — rare earths and lithium, not the niobium of the speeches — clean power, a large market, and a diplomatic network.
• Attach conditions to foreign investment (M1). Trade access to that market, those minerals and that power for things that build domestic capacity: computing capacity kept under Brazilian control, data stored in-country on Brazilian terms, real technology transfer, and local benefit stated in advance and audited — rather than investment that deepens dependence behind a sovereign flag.
• Build a coalition, and bargain inside it rather than alone (M2). Brazil should act as one member of a wider group of middle and supply-chain powers, so that a great power cannot pick countries off one at a time.
• Give the coalition an institution (M3). Fund a standing facility modelled on Brazil’s Tropical Forests Forever Facility (TFFF), so the coalition is held together by rules and money rather than communiqués.
• Keep regulatory and court power over Big Tech in reserve (M4). It is real leverage, but it carries the highest retaliation cost of any option, so use it to bargain rather than as a standing posture.
• Do not try to go it alone (M5). Full self-sufficiency is too slow and too costly without a domestic chip industry — and a national champion built on competitiveness grounds alone reproduces at home the concentration the strategy resists abroad.
• Show the public what it gets. Treat IA para o Bem de Todos as a test, not a subtitle: for each incentive and each facility project, publish who benefits, by how much, on what evidence, and how affected publics took part. A leverage strategy that cannot answer the question the impact critique keeps asking — why AI, and for whom? — is spending public money to build someone else’s leverage.^39,41
• Stay inside Brazil’s own commitments. Keep the strategy within the national climate targets, Indigenous consultation (ILO 169) and just-transition promises, and stage it behind those safeguards rather than ahead of them.
• Pass the AI law and fix the regulator’s mandate. Until a statute exists the courts set technology policy by default — the same behaviour the great powers cite when they retaliate.^5,19

The scores are qualitative and ordinal; they record direction and rough magnitude, not quantities, and they have not been weighted into a single ranking, which is deliberate — the weighting is the client’s, not the analyst’s. The causal diagram fixes signs that in reality are conditional, and the accommodation trap in particular depends on parameters this paper does not estimate: how fast coercion bites, how much the facility pays. The actor analysis is a snapshot of a system in motion: the AI law, the regulator’s mandate, and the mineral deals were all unsettled as this was written.^5,10 A fuller treatment would quantify the leverage indicator, model the defection dynamics, and test the recommendation against the actors’ own responses rather than against scenarios alone — starting with the movements of Section 5, whose response cannot be modelled from a desk. The argument offered here is the prior step: a defensible account of why, for Brazil, leverage is the objective worth maximising and a coalition the only way to hold it — and why neither stands without a public benefit Brazil can show.

 Appendix — the national plans, by axis

The consolidation in the introduction draws on the plans below. Official axes, missions and targets are given as published (Portuguese, with English gloss); lead body, horizon and headline budget are noted where fixed. Items still in consultation or pending enactment are flagged.

A1 · PPA 2024–2027 — Plano Plurianual

Planejamento e Orçamento · 2024–2027 · ~R$13.3 tn · the constitutional master plan (Lei 14.802/2024).

• Three eixos: Desenvolvimento social e garantia de direitos (social development and rights); Desenvolvimento econômico, sustentabilidade socioambiental e climática (economic development, environmental and climate sustainability); Defesa da democracia e reconstrução do Estado e da soberania (defence of democracy, reconstruction of the State and sovereignty).
• Six priorities: combate à fome e redução das desigualdades; educação básica; saúde; Novo PAC; neoindustrialização; combate ao desmatamento e emergência climática.
• Visão 2027: a democratic, just, developed and environmentally sustainable country. 35 strategic objectives, 69 indicators, five cross-cutting agendas (Children; Women; Racial Equality; Indigenous Peoples; Environment).

A2 · Nova Indústria Brasil (NIB)

MDIC / CNDI / ABDI · to 2033 · Plano Mais Produção R$300 bn through 2026 (BNDES, Finep, Embrapii).

• Mission 1: Cadeias agroindustriais sustentáveis e digitais — sustainable, digital agro-industrial chains for food, nutritional and energy security.
• Mission 2: Complexo econômico-industrial da saúde — resilient health industrial complex to reduce SUS vulnerabilities (70% of needs made domestically).
• Mission 3: Infraestrutura, saneamento, moradia e mobilidade sustentáveis — sustainable infrastructure, sanitation, housing and mobility.
• Mission 4: Transformação digital da indústria — digital transformation of industry (digitalize 90% of firms). Houses REDATA / data-centre policy.
• Mission 5: Bioeconomia, descarbonização e transição e segurança energéticas — bioeconomy, decarbonization, energy transition and security.
• Mission 6: Tecnologias para a soberania e defesa nacionais — sovereignty and defence technologies (autonomy in 50% of critical defence tech).

A3 · Plano Clima — Plano Nacional sobre Mudança do Clima

MMA / Comitê Interministerial sobre Mudança do Clima · approved Dec 2024; sectoral strategy Mar 2025.

• Two pillars: mitigação (an Estratégia Nacional de Mitigação plus eight sectoral mitigation plans) and adaptação (adaptation plans).
• NDC targets: net emissions −59% to −67% below 2005 by 2035 (0.85–1.05 GtCO₂e); interim 2030 of 1.2 GtCO₂e; net-zero 2050.
• Cross-cutting: transição justa (just transition) as an explicit axis; economic instruments include Fundo Clima, sovereign sustainable bonds, Eco Invest Brasil, the Taxonomia Sustentável and the TFFF.

A4 · Plano de Transformação Ecológica (PTE)

Ministério da Fazenda · launched COP28, Dec 2023 · the greening/financing counterpart to NIB.

• Six axes: finanças sustentáveis (sustainable finance); adensamento tecnológico (technological densification); bioeconomia (bioeconomy); transição energética (energy transition); economia circular (circular economy); nova infraestrutura verde e adaptação (green infrastructure and adaptation).
• Instruments: regulated carbon market; sovereign sustainable bonds (first issuance ~US$2 bn); national sustainable taxonomy; reformed Fundo Clima; Eco Invest Brasil.

A5 · Novo PAC

Casa Civil · Aug 2023 · ~R$1.7 trillion.

• Nine axes: Cidades Sustentáveis e Resilientes; Transição e Segurança Energética; Transporte Eficiente e Sustentável; Inclusão Digital e Conectividade (R$27.9 bn, a new axis); Água para Todos; Infraestrutura Social Inclusiva; Inovação para Indústria e Defesa; Educação, Ciência e Tecnologia; Saúde.
• Digital component: 4G/5G expansion across 5,570 municipalities; connectivity for 138,000 schools and 24,000 health units. Tension: R$335 bn allocated to oil & gas.

A6 · ENCTI 2024–2034 — science, technology & innovation

MCTI · ten-year horizon · public consultation Dec 2025 (in finalization).

• Four axes: expansão e integração do Sistema Nacional de CT&I; inovação empresarial e reindustrialização (tied to NIB missions); projetos estratégicos para a soberania tecnológica; CT&I para o desenvolvimento social.
• Priority technologies needing intensive effort: IA, semicondutores, tecnologias quânticas, transição energética. Detailed targets set in five-year PACTI action plans.

A7 · PBIA 2024–2028 — Plano Brasileiro de Inteligência Artificial

MCTI · July 2024 · R$23.03 bn, 54 initiatives · “IA para o Bem de Todos”; successor to the 2021 EBIA.⁴

• Axis 1 — Infraestrutura e Desenvolvimento de IA (R$5.79 bn): Santos Dumont upgrade, renewable-powered datacentres, Portuguese-language models.
• Axis 2 — Difusão, Formação e Capacitação (R$1.15 bn): skills and talent, including marginalized groups.
• Axis 3 — IA para os Serviços Públicos (R$1.76 bn): Nuvem Soberana, Infraestrutura Nacional de Dados.
• Axis 4 — IA para a Inovação Empresarial (R$13.79 bn): national datacentres in the North/Northeast on renewable energy, startups fund. Aligned with NIB.
• Axis 5 — Apoio à Regulação e Governança (R$103 m): OBIA, algorithmic-transparency centre.

A8 · Energy strategy

MME / EPE · long-term planning and the fuel transition.

• Planning: PNE 2050 (revised toward 2055, ~88% renewable capacity, ~R$2 tn investment); PDE 2035 (public consultation Feb 2026); PLANTE consolidating transition actions.
• Combustível do Futuro (Lei 14.993/2024): green-diesel (PNDV), sustainable aviation fuel (ProBioQAV), biomethane; ethanol blend to E30 (2025); biodiesel toward B20 by 2030.
• Other: Marco Legal do Hidrogênio de Baixa Emissão; regulated offshore wind; PROINFA; Luz para Todos in the Amazon.

A9 · Critical & strategic minerals

MME / ANM / CNPM · governance under construction.

• Framework: strategic-minerals list (Resolução MME 2/2021); Política Pró-Minerais Estratégicos (Decreto 10.657/2021), prioritizing licensing of critical-mineral projects.
• New governance: Conselho Nacional de Política Mineral installed 16 Oct 2025; Plano Nacional de Mineração 2025–2050 mandated; PL 2780/2024 (Política Nacional de Minerais Críticos e Estratégicos) approved by the Câmara 6 May 2026, pending in the Senate; creates a Fundo Garantidor (R$2 bn) and ~R$5 bn in fiscal credits for in-country beneficiation.
• Assets, in order of AI-relevant leverage: rare earths (Serra Verde, Goiás; second-largest reserves on earth); lithium (Vale do Jequitinhonha); niobium (CBMM/Araxá) — globally dominant but a small, steel-directed market operated by a private company; see the correction in §2. Socio-environmental and Indigenous-land conflicts: see §4.

A10 · Digital & data sovereignty

MDIC / MGI · 2025.

• REDATA (MP 1.318/2025): zero-rates federal taxes on datacentre equipment in exchange for contrapartidas — 100% clean energy, water efficiency, ≥10% capacity for the domestic market, 2% of investment in local R&D; part of the Política Nacional de Data Centers (PNDC).
• Nuvem de Governo / Nuvem Soberana (Serpro + Dataprev): public data kept under state control; 250+ federal bodies onboarded since 2025.
• Diagnosis: only ~40% of Brazilian data processed in-country; US firms ~70% of the global cloud market.

A11 · Social inclusion

MDS and others · the declared first priority.

• Brasil Sem Fome: 80+ actions integrating 32 programmes; goal to exit the Hunger Map. Severe food insecurity fell to 3.2% of households (IBGE, 2024).
• Bolsa Família (R$600 minimum, 21.4 m families, R$685 bn in the PPA); Minha Casa Minha Vida (2 m+ homes by 2026); plus Farmácia Popular, Desenrola, Luz para Todos.
• Future of work: transição justa and the PBIA’s human-centred mandate against ~31.3 m AI-exposed jobs.

A12 · Foreign policy

Itamaraty / Presidency · active non-alignment.

• Posture: não-alinhamento ativo and strategic autonomy;²² South-South cooperation; reform of global governance; multipolarity.
• Platforms: G20 presidency (2024, Global Alliance Against Hunger and Poverty); BRICS presidency (2025); COP30 in Belém (2025), the “implementation COP” (Belém Package / Global Mutirão).
• Signature deliverable: the TFFF — US$4/ha of standing tropical forest, ≥20% earmarked for Indigenous Peoples and Local Communities; >US$6.7 bn announced and 66-country Launch Declaration by 22 Nov 2025.