Latam-GPT, unveiled in February as a sovereignty initiative, is built on Meta’s Llama 3.1 and was trained using Amazon’s infrastructure. Eight months later, the World Bank recommends not entering the race for state-of-the-art models. Dependence on underlying technology is already a given. What determines which companies capture value is how well each company uses AI, and that depends on the people involved and how things are managed.
A report with incomplete answers
The Economic Outlook for Latin America and the Caribbean, which the World Bank presented on October 6, is explicit: the region «does not need to join the race to build cutting-edge AI models that cost billions of dollars» to reap most of the benefits. Its fundamental constraint, according to the report, is absorption capacity, which requires investment in human capital, digital infrastructure, and management practices.
The World Development Report 2026, published in August, acknowledges the cost of that recommendation. It warns that a handful of companies in just a few economies control the most advanced models and the chips on which they depend, and that That concentration creates risks of dependence. Their answer is that national systems should be able to work with multiple providers at the same time. It’s a good answer for a government with technical teams, but an incomplete one for a company that still can’t scale even a single tool.
Building something doesn't eliminate dependence either
The option to build is no exception to this pattern. According to CENIA, Latam-GPT was developed on the Llama 3.1 architecture, with 70 billion parameters, and more than 300 billion regional text tokens were added through continuous pre-training. The AWS infrastructure reduced training time from 25 to 9 days. The project’s own website acknowledges that its first version underperforms compared to other models in some tests.
None of that detracts from the project's merit, but it shows that Sovereignty in AI is negotiated in layers: The base model comes from Meta and Amazon Web Services, while the corpus and the expertise of those who fine-tuned it are unique to the country. The same applies to IFX, which announced an investment of nearly $169 million in Colombia for 2026 and 2027 in data centers and connectivity. More physical capacity in the country does not change who manufactures the chips or who trains the models.
Adapting it does not eliminate the dependency either. The concept note from the World Development Report believes that having open-source base models, developed in advanced economies, is a prerequisite for adapting AI to the local context. This is the logic I described in «AI that works doesn't cross the border», with one difference: there, the dependence stemmed from waiting for someone else's turn, whereas here it is inherent in either of the two paths.
Where the value is actually determined
The report itself points out that a medium-sized manufacturing company in São Paulo can access the same models via the cloud as a Silicon Valley company. Even so, among the Brazilian companies in its sample that have already adopted AI—60% of which at least experimented with it—only 37% reached the pilot phase or beyond, and just 28% are scaling up their use. With the same tools available, more than half fail the test.
The difference compared to advanced economies lies in organization. Among these companies, 38% provide AI training to their employees, compared to about 75% in the G7, and only 17% have created a dedicated team or a senior management position for AI, compared to about 50% in the G7. According to the report, the bottleneck lies in moving from using generic tools to adapting models to each company’s data and workflows.
The argument for building first
Those who advocate building first have solid arguments. Without energy, connectivity, and computing power, scaling is impossible, and an open regional model gives Latin America something no single provider can take away. I concede both points: there is a physical limit, and Latam-GPT has created capabilities that did not exist before. But those capabilities are spread across more than 60 institutions, and Brazilian companies are already operating below the current ceiling. Infrastructure and a proprietary model are necessary conditions, but neither of them trains people nor changes the management approach required for integration.
The obstacle is one of knowledge rather than money. In Costa Rica, 63% of companies that do not use AI cite a lack of in-house expertise as their main barrier, compared to 29% that cite costs. The report finds the same pattern in Argentina and Mexico. Chile is the exception, where financing tops the list (42%), followed closely by staff skills (36%).
It is the gap in design and oversight that I analyzed in «We're adopting more AI than the rest of the world. Why isn't it noticeable?», now with multilateral backing. The World Bank warns that delaying action would entrench the region's structural dualism and delay productivity gains.
The region will continue to purchase the base layer, whether through licenses or models adapted to third-party architectures. The decision that does rest with the region is how much to invest in ensuring its companies know how to use it. If, in two years, we add more data centers and Brazil’s adoption curve remains the same, we will have financed capacity that no one has learned to utilize, and the region will face a twofold dependency: on those who manufacture the technology and on those who know how to use it.
Sources: World Bank, Economic Outlook for Latin America and the Caribbean: Capitalizing on the AI Transition (October 2026), World Development Report 2026 and its concept note; CENIA and LatamGPT.org; Portafolio (IFX).
