The launch of ChatGPT at the end of 2022 kicked off the hype surrounding generative artificial intelligence . And it also led to the checks filled with digits collected by OpenAI , the tool's owner, and other Large Language Models ( LLMs ), the engines behind this technology.
Up to this point, however, the billion-dollar LLM race has been concentrated in a few names. And in two geographies: the United States, with OpenAI itself, which has already raised more than US$200 billion, Anthropic , Google , and xAI . And China, with players like DeepSeek .
But there are those willing to challenge the predictions of this race. A study by the Federal University of Goiás (UFG) identified more than 50 LLMs developed in Brazil, which raises the following question: is there, in fact, room for "Brazilian ChatGPTs" on this map, and will they survive the AI race?
“The US and China have resources on an unparalleled scale,” Patricia Magalhães, co-founder and CEO of NeuralMind, tells NeoFeed . “It’s not about replacing these global models. But about creating a position that gives us some advantage and sovereignty. We cannot be 100% dependent on external AIs.”
Founded in 2017, originating from Unicamp's laboratories and long before the buzz generated by ChatGPT, NeuralMind is an example of this David and Goliath struggle. The startup has raised R$ 20 million, including a R$ 1 million investment from Fundepar and funding from organizations such as Finep.
But this substantial difference in resources doesn't prevent the company from being one of the local deep tech companies willing to prove that there is a viable path for LLMs made in Brazil , whether created from scratch or from pre-trained models. A thesis that finds an echo in other voices in the market.
According to Anderson Soares, vice president of AI Brasil and coordinator of the Center of Excellence in Artificial Intelligence at UFG, it's difficult to compete with generalist LLMs like ChatGPT and Claude, from Anthropic. But there's a good shortcut in models specialized in niches and local particularities.
“Brazil has unique problems and is a very fertile ground for validation for almost everything, in different sectors,” says Soares. “Healthcare, for example, is a data machine that few countries have, as are the financial and legal markets, which are very advanced in digitization.”
According to Soares, although Brazil is a significant country, going into that level of detail here wouldn't be a priority for major global models. At least not at this moment, because the country lags behind markets like North America, Europe, and Asia.
Gustavo Araújo, co-founder and Chief AI Officer of Distrito, reinforces this sentiment. “You can’t just take a US healthcare model and copy and paste it ,” he says. “And the fact that we have local LLMs allows companies here to avoid being held hostage by the terms and conditions of one or two large providers.”
Meanwhile, a movement spearheaded by a player who dictates the rules of the game in another essential link of the generative AI chain promises to shake up the LLM landscape. This could also benefit Brazilian models.
“There’s a significant shift in Nvidia’s strategy, which has been encouraging the development of other models and driving decentralization in this arena,” says Araújo. “After all, it doesn’t want 60% of its revenue dependent on four or five players.”
Patricia Magalhães, co-founder and CEO of NeuralMind
Seeing this gap and with a portfolio of models, a proprietary search engine, and applications, NeuralMind is focused on sectors such as healthcare, energy, and legal. It already serves clients such as the Brazilian Federal Court of Accounts, the National Agency for Supplementary Health, and the Hospital das Clínicas in Porto Alegre.
In addition to a spin-off from its healthcare vertical, the company plans to position itself as a legal tech company, which involves developing Jurema, a family of templates for legal tasks, in partnership with Escavador. Magalhães has the final say on these strategies.
“We are talking about highly regulated sectors and critical infrastructure with little margin for error,” says the CEO of NeuralMind. “So, relying on a foreign model is a real risk. You can have price fluctuations, in dollars, variations in usage policies, and impacts from geopolitical instability, which are not improbable.”
Local accent
This local accent is also reflected in the names of some local LLMs and deep tech companies. Among them are Sabiá, from Maritaca AI, one of the pioneers in the field, as well as Cabrita, Bode, Samba, and Amazônia IA. And, with this accent, Clarice.ai is one of the startups seeking to prove that it can be a different kind of beast.
Founded by Felipe Iszlaji, who holds a degree in computational linguistics and a post-doctorate in creative computing and AI, the company was born in 2020, the centenary year of the writer Clarice Lispector, who inspired its name. It focuses on models for writing, editing, revising, and humanizing texts in Portuguese.
After creating two Small Language Models (SLMs), Clarice.ai is developing a third, more robust version, aiming to improve its accuracy by at least 30%. This, on the one hand, seeks to reinforce its differentiation. And, on the other hand, it shows how far the startup has already come with its proposal.
“We are training and refining this model with our own data,” says Iszlaji. “Every month, we have 1.5 million corrections that are validated by our users.”
Clarice.ai has accumulated over 500,000 users since its inception – 80,000 of whom are currently active in the B2C segment, where it offers two subscription plans. In the B2B segment, a market the company is beginning to develop, it has approximately one hundred subscribers, including publishers and media groups such as Itatiaia, based in Minas Gerais.
NeoSpace is another company betting on proprietary data. But from a different perspective, using a concept called the Data Language Model (DLM), which, unlike LLMs and SLMs, is not fed solely by language mechanisms, but by any data and document format.
Based on this principle, the company provides platforms for companies such as banks, insurance companies, and telecom operators to train their models using their own data. The idea is to open up opportunities for applications ranging from improving customer service to anticipating customer behavior.
“We hand over the brain, the DNA, the newborn child. But it will study and learn from that company's data,” says Bruno Pierobon, co-founder of NeoSpace. “That's something that OpenAI and Anthropic don't have. We're not competing for the same data, which gives us good protection.”
NeoSpace also differentiates itself in another aspect, here, in relation to most of its local peers, which still have limited access to investment. Created at the end of 2023 by the founders of Zup, acquired by Itaú in 2019, the startup attracted an investment in 2025 led by the bank itself.
In the $18 million funding round, Itaú was joined by angel investors such as Martin Escobari of General Atlantic, Micky Malka of Ribbit Capital, and Nigel Morris of QED. The company was valued at $100 million.
The figure helped to overcome one of the major challenges for deep tech companies behind these LLMs: the high cost of training and improving the models. Until then, capitalized by the exit from Zup, NeoSpace's partners had invested R$ 10 million of their own money.
“It’s difficult to compete with the computing power of these big players ,” says Pierobon. “While they make mistakes in a hundred training sessions a month, we only make one.”
The funding round also helped NeoSpace advance in its talks with clients. Besides Itaú itself, the startup has 12 advanced negotiations, including some abroad – two are already in the United States, where the company is setting up an office, and one in Europe.
Pierobon emphasizes that having a history with Zup and the investors helped open doors for the funding round, even though the investment thesis was more complex and in a sector that, at first glance, seemed quite risky.
“I think this is the first time we’re seeing a startup movement that truly unites academia with business ,” he says. “So, sometimes it can be difficult to structure both ends and package that thesis for investors.”
From academia to the marketplace
Everton Gago is one of the names that blends these two worlds. With over 20 years of experience in AI, divided between academia at Unicamp and companies like Ci&T, he founded infinity6 in 2024, alongside Leonardo Chaves.
Since then, the deep tech company has raised two rounds of funding from an undisclosed fund, totaling R$10 million. What attracted the mysterious investor is a thesis also centered on models trained with customer data – from stored histories to real-time information.
This portfolio involves three engines dedicated to three tasks: recommendation, forecasting, and pricing. It is already used by industries, particularly pharmaceuticals and food; medium-sized banks; and retailers, mostly in e-commerce operations.
“Banks hire us, for example, to help understand which product should be recommended to a particular client and at what time,” says Gago. “Our proposition is pure ROI. In the pharmaceutical sector, we have achieved significant results, with an increase of around 23% in the average transaction value.”
Infinity6 also invests in models created from scratch, which support other data and document formats, just like NeoSpace. And, at the same time, in smaller models that require less computing power to be trained and operated.
“Training a large engine like the GPT 5.0 requires several billion dollars,” says Gago. “A small, well-targeted model costs much less, from US$1,000 to US$2,000.”
The pursuit of greater efficiency was also one of the pillars that guided one of the newest local names in this market: Lua Vision, created by Paulo Câmara, David Kang, and Plínio Ceccon at the end of 2025, after a three-year structuring process.
With an initial investment of R$100,000 from the trio and stemming from Câmara's doctoral thesis in AI at Tel Aviv University, the deep tech company has built five versions of its LLM from scratch – a sixth is on the way.
“Generalist LLMs are like large libraries, aiming to have as much knowledge as possible,” says Kang. “Our thesis is to be the best brain. The reasoning power of our model is focus, not the amount of information it contains.”
The models are also trained with customer data and run on -premise , meaning they are installed on company servers. This, among other things, reduces risks, as well as expenses related to cloud providers.
This thesis is already being tested with proof-of-concept projects in Brazil and abroad. There are about 20 pilot projects in sectors such as health and education, ranging from an initiative with more than 20,000 students in Africa, funded by Microsoft, to a project with a consulting firm in Latin America.
“With this pipeline, we imagine we’ll already have revenue traction in the second half of the year,” says Kang, “That’s what we’re pursuing now.”
