AI sovereignty for a Latin American company is a list of dependencies you can price
On 10 February 2026 Chile's National Center for Artificial Intelligence, CENIA, presented Latam-GPT at an event in Santiago, a language model built for Latin America. Reports differ on the details, and the most common account describes a model of about 70 billion parameters, built on Meta's Llama 3.1 architecture, trained mainly on Spanish and Portuguese text, with plans to add Indigenous languages later. Funding came to about $550,000 from CENIA and the Development Bank of Latin America, and initial training ran on Amazon Web Services. A project launched in the name of regional AI sovereignty therefore began on an American open-weights base and an American cloud. I do not offer that as criticism. It shows that sovereignty comes in degrees, and that a useful discussion starts by asking which dependency, exactly, a company wants to reduce.
Brazil's government has set larger goals. The Brazilian AI Plan, presented in July 2024 for the period 2024 to 2028, foresees about R$23 billion of investment across infrastructure, training, business innovation, public services and regulation. It includes a language model in Portuguese and an expansion of the Santos Dumont supercomputer at LNCC. Press reports in August 2026 describe further announcements, including the purchase of an AI supercomputer for around R$1 billion and a partnership on a Portuguese and Spanish model, and I would treat those as reported plans until they are delivered. Public programs deliver on public timetables. A company that needs to ship this year has to reason about its own exposure.
Marketing talk of sovereignty tends to bundle together things that are separate. My argument is that a company gets more from a list: the distinct ways in which its AI capability depends on parties and conditions it does not control, each with a measurable cost and a possible hedge.
Six dependencies that can be separated
The first is legal jurisdiction. Which country's authorities can compel access to the data, and under which law? A provider incorporated in the United States may be subject to US legal process for data it controls, including data stored abroad, under the CLOUD Act of 2018, while the company processing Brazilian personal data is subject to the LGPD. The LGPD regulates international transfers in articles 33 to 36, and in August 2024 the ANPD issued Resolution CD/ANPD 19/2024, which established standard contractual clauses and gave organizations that relied on contract-based transfers twelve months, until 23 August 2025, to adopt them. A company sending customer data to a foreign model API is making an international transfer in the legal sense and needs a documented basis.
The second is location. Where is the data stored, and where does inference run? Some deployments of foreign models can be pinned to a region such as São Paulo, and some cannot. Latency, contractual commitments and regulatory expectations in sectors like banking can make location a requirement and not a preference. The facts differ by provider and by model, change often, and need checking in writing.
The third is control of the model. A closed API can change under you, as the published retirements and silent behavior changes of hosted models show, and it can be withdrawn or repriced. Open weights can be hosted by you, at the cost of keeping the hardware busy, hiring the staff to run it and adopting improvements more slowly.
The fourth is compute supply. Advanced accelerators are made in a few places, the rules on exporting them have changed more than once in recent years, and countries without domestic fabrication or large national clusters depend on cloud allocation. For most companies this matters indirectly, through prices and availability of the services they rent.
The fifth is language and cultural fit, and the sixth is skills. The sixth is the most under-appreciated. A company whose staff cannot evaluate a model, tune one or swap it has outsourced its judgment along with its compute, and will accept whatever the vendor says about quality. A simple measure is the number of people who can run the evaluation suite on a new model and read the results, and in a small team the answer can be one. The remedy is training and an evaluation habit, and it costs far less than any of the infrastructure projects in the news.
Language and the token premium are the most measurable of these, so the next section takes them up.
The Portuguese premium
Language models read and bill in tokens, and tokenizers were built mostly on English text. Aleksandar Petrov and colleagues, in a NeurIPS 2023 paper, compared 17 tokenizers and found that the same text can differ in length by as much as 15 times across languages, with disparities persisting even in tokenizers designed for multilingual use. They noted that this produces unequal costs, latency and context limits.
Portuguese sits at the mild end of that range, and the premium is still measurable. I tested it on material at hand: ten of the essays on this site, each written in English and in Portuguese with the same content, with the tables removed. With OpenAI's current tokenizer (o200k), the Portuguese version used on average 17.7% more tokens than the English one, with a range from 16.6% to 19.8% across the pairs. With the older tokenizer used by GPT-4 (cl100k), the average premium was 41%, with a range from 36% to 45%. Part of the difference comes from the Portuguese texts having more words than the English ones, since translation expands text, and the rest from the tokenizer breaking Portuguese words into more pieces. In both cases, the figure that matters for a bill is the one measured: the same meaning, in tokens.
The result is a surcharge on every call. A pipeline that would use a billion tokens a month in English uses about 1.18 billion in Portuguese on the newer tokenizer, so the bill is 18% higher at the same price per token, and the context window holds about 15% less content. Other providers' tokenizers will differ, and I could only measure OpenAI's, so the sensible response is to measure your own: take a hundred representative documents, run them through your provider's token counter in both languages where a parallel version exists, and use the ratio in the cost model.
Currency adds another layer. Most frontier APIs are priced in dollars. A bill of $50,000 a month at an exchange rate of R$5.50 is R$275,000. At R$6.50, which I use only as an illustration, the same usage costs R$325,000, an increase of about 18% with no change in volume or price. A cost-model that ignores exchange rates understates the risk for any Brazilian company with a large AI bill. Some local providers invoice in reais, which removes this exposure and substitutes another.
What the regional options offer
Maritaca AI, a Brazilian company, sells the Sabiá family of models through an API priced in reais. Its published price list shows Sabiazinho-4 at R$1 per million input tokens and R$4 per million output tokens, and Sabiá-4 at R$5 and R$20, each with a 128,000-token context window. The company reports that its reasoning model, Sabiá-4 Thinking, scores an average of 90.8% across a suite covering function calling and agents, legal tasks and general tasks, compared with about 92.4% to 92.8% for the leading global models it tested, and higher than those models on a Brazilian legal-writing test. These are the company's own benchmarks, and I found no independent evaluation. They are worth taking seriously enough to run on your own cases.
The purpose of such models is a fit with Portuguese legal and administrative language, pricing in local currency and the possibility of keeping data in the country. Whether they beat a global model for a given task is an empirical matter, and the answer can differ by task: a model may be better at reading Brazilian court filings and worse at writing code. Open models that were trained or tuned for the region offer a different bundle: you can host them yourself, at your own cost, with full control. Latam-GPT is one. Another is the Tucano family, whose second generation (Nicholas Kluge Corrêa and colleagues, March 2026) is a fully open suite of Portuguese models from 0.5 to 3.7 billion parameters, with base, instruction-following and reasoning variants, an openly released pretraining corpus and training recipes. Models of that size are small, and their appeal is that they run cheaply on modest hardware and can be adapted to a narrow task, like classifying documents in Portuguese, where a large general model would be an expensive overkill.
Fit to language is a dependency in its own right, and generic benchmarks hide it. Brazilian Portuguese carries its own legal and administrative vocabulary, from labor law to tax documents to the names of the forms a bank must file, and global models trained mostly on English text may handle it well on one document and drift into European Portuguese or invent a plausible regulation on the next. The only reliable check is to test the candidates on your own documents, graded by the people who know what a correct answer looks like.
What a map looks like
To make the exercise concrete, take an insurer, an invented example for illustration, with five uses of AI. A claims-triage model runs on a hosted global API, processes personal and health data, runs in the provider's São Paulo region, bills in dollars and could switch to a fallback in about six weeks. A customer-service assistant uses a regional model billed in reais, with no sensitive data and a one-week switch. Marketing copy generation uses a global API in a US region, with no personal data, and could switch in a day. Underwriting analysis runs on self-hosted open weights, with all data inside the company's cloud account and a two-month switch. Document summarization for the legal department uses a global API with a contractual no-training clause and a three-week switch.
Laid side by side, the map shows that the claims model carries the most exposure: personal health data, a dollar bill and the longest switching time among the API-based uses. That is where legal review of the transfer basis, a regional fallback and a currency hedge belong. The marketing tool, which looks like the most "foreign" dependency, needs nothing.
A portfolio, not a flag
I would organize the choices by the class of data and task, with a default for each.
| Task and data | Default approach | Why | Check |
|---|---|---|---|
| Non-sensitive, hard reasoning or coding | Global frontier API, in-region if available | Best capability per unit of effort | Written terms on retention and training use, exit plan |
| Personal or regulated data, moderate difficulty | In-region deployment of a capable model, or a regional provider | Transfer rules and sector expectations | Legal basis for any cross-border transfer, evaluation on your cases |
| High-volume Portuguese tasks with stable requirements | Smaller model, regional or open, possibly self-hosted | Token premium and currency exposure compound at volume | Break-even against API pricing, utilization |
| Strategic or classified data | Hosted in your own environment with open weights | Control is the requirement | Staffing and security of the stack |
Around the table sit two habits. Keep a tested fallback for every critical use, so a change in terms, price or availability costs days and not months. And keep enough in-house skill to run your own evaluations, because the ability to judge a model on your tasks is the one sovereign capability that no provider can sell.
Be skeptical of state projects as dependencies
Public projects such as the national plan are valuable, and a company should follow them and use them where they offer something concrete, such as shared research infrastructure, datasets or a model it can evaluate. It should not depend on them for plans with a delivery date. Timelines in public technology programs move, a Portuguese model described in some coverage as due by 2026 sits in a plan that runs to 2028, and the sensible way to treat such projects is as options. If one is delivered and passes your evaluation, you gain a choice. If it is late, you have lost nothing you depended on.
When sovereignty is not worth the cost
A fair objection is that all this is a luxury, and that the best strategy for most firms is to buy the strongest model available and move on. For non-sensitive work the objection is largely right. The premium for a regional or self-hosted option is a real cost, and paying it for a task that handles no protected data and has no continuity risk is wasteful. Sovereignty belongs at the margin: where law requires control, where a loss of access would stop the business, and where the dependency is large enough that price changes would hurt.
Start with a map
The first step costs a day. List every AI use case and record, for each, the provider and model, where inference runs, the class of data involved, the legal basis for any transfer abroad, the currency of the bill, and how long it would take to switch to a fallback. The map will show which dependencies are concentrated and which uses carry the exposure. And it will provide the denominator for the decision, since a company that knows 80% of its spend sits in two non-sensitive use cases can relax about them and spend its effort where the map shows red.
Sources
- CENIA, Chile. Latam-GPT, presentation on 10 February 2026; coverage by AP, Intellinews and AI Business (February 2026). Reports differ on model size and participation.
- Ministério da Ciência, Tecnologia e Inovação (2024). Plano Brasileiro de Inteligência Artificial 2024-2028. mcti.gov.br; coverage of August 2026 announcements by the specialized press.
- Maritaca AI. Sabiá API price list and model announcements, maritaca.ai (consulted October 2026); benchmark figures are reported by the company.
- Petrov, A., La Malfa, E., Torr, P. H. S., Bibi, A. (2023). Language Model Tokenizers Introduce Unfairness Between Languages. NeurIPS 2023. arXiv:2305.15425.
- Brasil. Lei nº 13.709/2018 (LGPD), arts. 33 to 36; ANPD, Resolução CD/ANPD nº 19/2024 (Regulamento de Transferência Internacional de Dados), 23 de agosto de 2024.
- United States. Clarifying Lawful Overseas Use of Data (CLOUD) Act, 2018.
- Token measurement: ten English and Portuguese essay pairs from this site, tokenized with OpenAI's o200k_base and cl100k_base encodings (October 2026); exchange rates in the example are illustrative.
- Corrêa, N. K., Sen, A., Fatimah, S., Falk, S., Landgraf, L., Kastner, J., Flek, L. (2026). Tucano 2 Cool: Better Open Source LLMs for Portuguese. arXiv:2603.03543.