Brazil can profit from AI without creating its own ChatGPT

Brazil’s Artificial Intelligence Plan allocates around R$23 billion between 2024 and 2028, distributed across infrastructure, training, public services, business innovation, and governance. The budget isn’t limited to building a chatbot, suggesting a broader strategy for technology adoption.

Infrastructure and Portuguese-language data limit a national model

Telma Woerle, executive director of the Center of Excellence in Artificial Intelligence (CEIA) at the Federal University of Goiás (UFG), says competing with the largest language models is hindered by infrastructure and a lack of public data in Portuguese. According to her, current datasets aren’t sufficient to train a system exclusively in this language at the scale of international competitors.

“There are other types of Brazilian AIs that we could develop,” she said, citing health, environmental issues, and biodiversity.

The researcher suggests that combining other languages would be necessary for training entirely in Portuguese, but this doesn’t prevent the development of national solutions focused on local problems.

Training costs rise, while execution becomes cheaper

A 2024 study by researchers from Epoch AI and Stanford estimated that the computing cost for frontier training has grown by about 2.4 times per year since 2016. If this trend continues, the largest training rounds could exceed US$1 billion in 2027.

In contrast, Stanford’s AI Index 2025 showed that the cost of inference—running a trained model—of systems with performance equivalent to GPT-3.5 fell more than 280 times between November 2022 and October 2024.

Opportunity for companies to adopt AI without building infrastructure

Paulo Ernani, operations director at consulting firm AVEX AI Lab, sees this difference as an opportunity for Brazil. He argues that companies can automate customer service, business processes, or operations without needing to manufacture chips, train their own models, or build data centers.

“Brazil doesn’t need the infrastructure thesis to work in order to capture value,” he said.

Competition among suppliers can benefit purchasing companies, although a sector correction also exposes the country to financial and exchange rate risks. For Ernani, the biggest risk is the lack of technology adoption.

Argument for local development in specific areas

Telma emphasizes that solutions created abroad reflect the problems and conditions of the markets where they were developed. Creating AI in Brazil allows working with local needs and data, as well as leveraging the country’s diversity to develop products that can find customers abroad.

“You create your own intelligence, you solve a regional problem, a problem of the country,” she said.

Her proposal is to focus efforts on areas such as health and the environment, where Brazil has specific challenges and knowledge to address them. This requires research capacity and infrastructure, but sets a different goal than overcoming a general model in all tasks.

Frequently Asked Questions

• Does Brazil need to create its own ChatGPT?

  Experts distinguish between adopting existing tools and national development. Paulo Ernani sees benefits for companies using AI, while Telma Woerle advocates for Brazilian solutions to specific needs, such as health and the environment.

• What are the obstacles to a Brazilian model?

  Telma Woerle points to limitations in computational infrastructure and public data in Portuguese for training a model exclusively in that language at the scale of major international competitors.

• Is artificial intelligence becoming more expensive or cheaper?

  It depends on the goal. A study by Epoch AI identified an increase in the cost of training frontier models. Stanford, on the other hand, recorded a decrease of more than 280 times in the cost of running systems with performance equivalent to GPT-3.5 between November 2022 and October 2024.