Petrobras $PETR3 (+2,45%)
$PETR4 (+0,54%) has developed a new generative artificial intelligence (AI) model to speed up maintenance processes for its oil platforms and other industrial facilities such as refineries and gas processing plants.
By 2029, the company expects to save R$20 million through efficiency gains with the AI assistant called Petronemo, said Otavio Ciribelli, the company's general manager of information technology. The executive presented the project, which was developed in collaboration with Deloitte, at GTC 2025 in San Jose, California, organized by Nvidia.
"We have considered conservative metrics," Ciribelli told Valor, shortly after his presentation at GTC 2025. The AI assistant is currently in the certification phase. In the coming months, it will be integrated into the evaluation process of parts in need of repair on the platforms.
"We carry out 30,000 inspections per year, and around 20% of these result in discrepancies that require repair recommendations," explained the Petrobras manager. The current process for analyzing repair and maintenance recommendations on the platforms takes two to three weeks.
Petronemo will make recommendations that will be evaluated by the company's 400-strong team of reliability engineers. "It's not a very large group, but it's absolutely important," Ciribelli emphasized.
The work of these engineers has a direct impact on the lifespan of Petrobras' production platforms, he said. "When we put a platform into operation, we want it to stay healthy until the last minute. The saddest story is when a platform that is designed to last 25 years has to be shut down prematurely."
In the project developed with Deloitte, Meta's large open-source language models, Llama 1 and 2, were used as the training basis for the AI assistant. The more robust data load and training were developed in-house and hosted on Petrobras' computers.
The Petronemo model is processed on four supercomputers with graphics processing units (GPUs) from Nvidia. These include the Pegasus, Tatu and Gaia machines from Petrobras and the Santos Dumont supercomputer from the National Laboratory for Scientific Computing (LNCC).
The company has also set itself the goal of developing an AI assistant that seamlessly integrates into processes and speaks the jargon of Petrobras engineers, jokingly referred to as "petrolês" in Portuguese, as Mr. Ciribelli mentioned in his presentation.
Tim Wiesel, Partner and Head of AI and Data at Deloitte Brazil, who presented Petronemo together with Ciribelli at GTC 2025, said the project is an example of the trend for companies to internalize their computing infrastructures, a model known as "on-premises", with the aim of reducing development costs for AI applications.
"We did a study at Deloitte that demonstrated a 30 percent increase in [compute] performance and 50 percent reduction in cost per token," Wiesel said, referring to the data fragments used to train AI models and billed by cloud AI providers.