When Jensen Huang, CEO of Nvidia, speaks, the tech industry listens. His recent defense of open AI models in his first post on X was no coincidence: he noted that these 'strengthen security and cybersecurity, accelerate innovation and diffusion, and favor sovereignty.' But what lies behind this statement? We spoke with Kari Briski, vice president of generative AI software at Nvidia, to unravel the implications for businesses and governments.

What do open models really bring?

Open models allow companies and countries to own, inspect, and adapt models with their own data. According to Briski, 'the pace of adoption varies by region and company, but all models have made enormous strides in closing the gap.' What's interesting about Nemotron, Nvidia's open-weight model, is that they also publish their data, which has opened a new world of participation. 'Companies reached out to us saying: thanks, but I want to understand why you published this dataset. That made them realize they could select, create, collect, and control their own data.'

Approach for CIOs and sovereign uses

Briski notes that both chief information officers (CIOs) and sovereign uses are on her mind. 'At the highest level, it's practically the same: use cases, datasets, outcomes you want to achieve.' The evolution has gone from simple question-and-answer to autonomous workloads that require specific tools. 'A company might have 2,000 tools; a region, 2,000 regional tools. They are fundamentally the same, but with different go-to-market strategies: reinforcement learning environments, synthetic data generation, or anonymization and differential privacy for healthcare and banking data.'

Why does Nvidia make its models public?

Nvidia's strategy is not altruistic. 'We are developing Nemotron for ourselves especially, to understand how our systems work at scale—not just for training, but also for inference—which allows us to iterate on aspects like model architecture to optimize token efficiency.' Additionally, they believe in a thriving ecosystem: 'When you publish a model, more startups emerge, more developers, and barriers to entry are reduced.'

Open models and sovereign AI

The connection is direct. 'It's about boosting a region that otherwise wouldn't have the computing capacity needed to reach a foundational model,' explains Briski. Historically, access to computing clusters depended on grants, with a back-and-forth that prevented constant iteration. 'These are the laws of AI scalability: the more computing power, the more intelligence; the more access, the faster you can boost. Open models and open data are that initial boost. You don't need to re-collect all the knowledge on the Internet to create a pretraining model.'

Regional adaptation and SLMs

Regional differences are key. 'Voice and word-of-mouth are very important in South Asia, while chatbots are less so,' she notes. What's the future approach? 'It depends,' she responds. 'Language is constantly evolving. Let's go back to basics: AI is infrastructure. Not just the model, but the framework, skills, runtime environment. All of this constitutes a new computing platform.' And she adds: 'It could be a large master model that, at the edge, is an SLM.' This flexibility is essential for integrating AI into everyday applications, understanding dialects and specific areas, driving the virtuous data cycle.

Limited hardware and smaller models

Companies in developing countries, like Nepal, express frustration over lack of access to GPUs. Briski responds: 'If you know Nvidia, you know it's all about the ecosystem, and we want to make it easy for developers. That's why we have many different sizes: Nano, Super, and Ultra in the Nemotron family, not only based on where it's deployed, but so developers can iterate on smaller GPUs and then scale to a more robust model like the Ultra.' Additionally, they work as a model-as-a-service across all cloud providers, with partners getting early access.

Open collaboration and performance vs. quality

Like Linux, the industry collaborates on open technologies. Briski states: 'In terms of performance versus quality, it's both. You can't have a high-quality model that's slow or heavy, nor a fast model that's terrible.' They focus on three aspects: efficiency, cutting-edge, and openness. Today, models work together: 'A planning agent routes a query to the best model to complete the task.' The Nemotron Coalition aims to bring together the brightest minds to collaborate on model architectures, token efficiency, and new techniques like latent mixture of experts.

Is the latest model always the best?

Briski compares to software: 'You do an update and it just doesn't work like before.' That's why the coalition incorporates partner evaluation criteria to ensure they don't regress. 'You have to change your mindset regarding AI and how you test it. Is the latest model necessarily the best? I'd say yes. Companies and countries need the skills to evaluate quickly, update prompts, and adopt new models.'

Open licenses and forks

Nvidia allows forking its models and publishing them on Hugging Face. 'We have a very open license,' says Briski. They publish NVFP4 reduced-precision checkpoints, popular for their smaller footprint. 'I love seeing different forks of our models. We also track them, which gives us insight into what people care about.' Benchmarks are the minimum, not the maximum. 'In the last 90 days, the type of workload has changed drastically, from question-answer pairs to agent-type workloads.' They want to show that you can be equally intelligent with computational and token efficiency.

For businesses, adopting open models is not just about cost, but sovereignty and control. As we explained in our guide on implementing generative AI in workflows, security is a critical factor. Open models allow inspection and adaptation, but they also require continuous evaluation, as we saw in the analysis of Claude's gaps. The flexibility of these models is comparable to that offered by solutions like NEXGestión for finance, where adaptation to customer needs is key. Ultimately, Nvidia's bet on open models not only democratizes AI but redefines digital sovereignty in an increasingly interconnected world.

Original source: ComputerWorld. Analysis and adaptation by ForgeNEX.