Kari Briski is making sure Nvidia can handle the agentic AI boom

AI agents are only as powerful as the models behind them, which must process vast amounts of data. No company is more vital in helping process that data than Nvidia. “All these agents are spinning off subagents that are doing multiagent swarming, and all kinds of tokens are being generated,” Kari Briski says, referring to the basic units of data that models process and produce. As vice president of generative AI software product, Briski leads her team to make sure Nvidia’s technology stack manages the tokens efficiently. As the generative artificial intelligence boom has progressed, Nvidia has managed to remain the dominant seller of the graphics processing units (GPUs) that run the AI models. That’s partly because developers have bought into building, training, and running models on the Nvidia hardware and software stack. For example, Nvidia’s CUDA software (Compute Unified Device Architecture) gives developers fine-grain control over how the chips manage the work of the models. But it’s not just CUDA. “It’s everything from optimizations that we’re putting at the chip level,” Briski says, “all the way up to reference architecture for agents,” the blueprints for how those systems are built. Much of it happens out of sight, but it’s what makes the entire agent ecosystem possible. Nvidia does build, train, and deploy its own AI models. But the reason is a bit different from companies like OpenAI or Anthropic. Nvidia’s main goal isn’t to sell subscriptions for AI models and apps to consumers and enterprises; it’s to drive demand for its chips and software platform. But it’s more than that, Briski explains. By staying close to core AI research, Nvidia can more easily steer development of its chips to make models run better. And, she says, it lets Nvidia researchers search for whole new model architectures that might better exploit the speed and power of the GPUs. This profile is part of Fast Company’s AI 20 for 2026, our roundup spotlighting 20 of AI’s