When Jensen Huang, CEO of Nvidia, speaks, the tech sector pays attention. His recent defense of open AI models in his first post on X is not just a statement of principles, but a clear strategy that redefines the future of enterprise artificial intelligence. At a time when generative AI is consolidating as critical infrastructure, Nvidia's decision to open its Nemotron models raises fundamental questions about ownership, sovereignty, and access to technology.

Kari Briski, vice president of generative AI software at Nvidia, delves into this vision in an exclusive interview. For her, open models are not a simple philanthropic gesture, but a strategic tool that allows companies and countries to take control of their technological destiny. 'Open models allow companies and countries to own, inspect, and adapt models with their own data,' explains Briski. This approach, which combines the publication of weights and data, is transforming the way organizations approach AI, moving from being passive consumers to active creators.

Beyond the Model: The Value of Open Data

Nvidia's strategy goes beyond releasing model weights. By also publishing the datasets used for training, the company is opening a new paradigm of participation. 'Companies reached out to us saying, “Thank you, but I want to understand why you published this dataset,”' comments Briski. This transparency has led many organizations to realize that they can select, create, and control their own data, a crucial aspect in regulated sectors such as healthcare or banking, where differential privacy and anonymization are essential.

For CIOs, this openness represents an unprecedented opportunity. They no longer depend on closed models that cannot be audited or adapted. Instead, they can build AI systems that align with their specific needs, whether to automate business processes or to create more efficient virtual assistants. As we noted in our article on implementing generative AI in workflows, the ability to fine-tune models to proprietary data is a key differentiator.

Sovereign AI: The Impetus for Underserved Regions

The connection between open models and digital sovereignty is another pillar of Nvidia's strategy. Briski is clear: 'It's about boosting a region that would otherwise not have the computing capacity needed to reach a foundational model.' Historically, access to high-performance computing clusters has been limited, often dependent on grants that forced work interruptions. Open models, combined with open data, offer a starting point that eliminates the need to gather all the knowledge of the Internet from scratch.

This approach is especially relevant in regions like South Asia, where social and linguistic dynamics differ notably. 'Voice and word of mouth are very important in South Asia, while chatbots are not so much,' notes Briski. Adapting models to local dialects and contexts is a challenge that can only be addressed with open models that allow continuous fine-tuning. This directly connects with the need for automating processes with adaptive AI, a topic we explore in depth.

From Master Model to SLM: A Flexible Architecture

When asked about the future of models, Briski responds with refreshing pragmatism: 'It depends.' AI is infrastructure, and as such, it must adapt to different environments. 'It could be a large master model that, at the edge, is an SLM,' she explains. This flexibility is key for companies that need to deploy AI on devices with limited resources, a trend we are already seeing in CRM for workshops and SAT, where efficiency is critical.

Nvidia offers different model sizes (Nano, Super, Ultra) so developers can iterate on smaller GPUs and scale later. Additionally, the company does not just release checkpoints into the void: 'We don't just throw the checkpoint over the wall; we help them traverse that last mile,' assures Briski. This support includes early access for partners and optimization for different platforms, which reduces adoption barriers.

Collaboration and Community: The New Open Source

The tech industry has shown that collaboration around open standards like Linux can transform the sector. Can this model be applied to AI? Briski believes so, and points to the Nemotron Coalition and the Open Secure AI Alliance as examples of this vision. 'Bringing together the brightest minds committed to open source and collaboration' is the goal, with members contributing in areas like pretraining, post-training, or data.

Regarding the debate between performance and quality, Briski is emphatic: 'It's about both.' A slow model is not useful, but a fast one that gives incorrect answers is not either. That's why Nvidia focuses on three pillars: efficiency, cutting-edge, and openness. Furthermore, models no longer work in isolation; 'a planning agent routes a query to the best model to complete the task,' a trend that is redefining the architecture of enterprise applications.

Is the Latest Model Always the Best?

A frequently asked question is whether companies should constantly update to the latest models. Briski compares this situation to traditional software: 'You do an update and it just doesn't work like before.' Therefore, continuous evaluation and adaptation are essential. 'Companies and countries need the skills to quickly evaluate, update prompts, and adopt new models,' she advises.

Nvidia allows its models to be forked and modified freely, and the company closely follows these adaptations. 'I love seeing different forks of our models,' confesses Briski. This community feedback is invaluable for improving models and understanding what real needs users have. Performance benchmarks, while important, are just the starting point: 'Benchmarks are the minimum essential, not the maximum limit we should reach.'

In a landscape where generative AI is transforming entire industries, Nvidia's bet on open models could be the catalyst that democratizes access to technology. As we have analyzed in other articles, such as the paradox of defensive AI, openness is not without challenges, but the benefits in terms of innovation and sovereignty are hard to ignore.

Original source: ComputerWorld. Analysis and adaptation by ForgeNEX.