NVIDIA engineers Kyle Kranen and Nader Khalil detailed the company's latest advances in AI inference infrastructure during a live recording at NVIDIA GTC. Kranen, one of the lead architects behind NVIDIA Dynamo, explained how the datacenter-scale inference framework optimizes model serving through techniques like prefill/decode disaggregation, intelligent scheduling, and Kubernetes-based orchestration. The approach prioritizes cost, latency, and quality tradeoffs to efficiently handle the computational demands of modern large language models at enterprise scale. Khalil, head of NVIDIA Brev, discussed the company's efforts to democratize GPU access for developers by reducing barriers to entry for high-end hardware. The conversation centered on NVIDIA's "Speed of Light" (SOL) philosophy—CEO Jensen Huang's first-principles approach to optimization—and explored critical emerging challenges including long-context model limitations and agent security. The latter addresses how to safely enable AI agents with file access, internet connectivity, and code execution capabilities without introducing critical vulnerabilities. The discussion reflected NVIDIA's evolution from a chip manufacturer into a comprehensive AI infrastructure provider, with the company introducing internal model APIs through its Build platform and planning dedicated sessions on Dynamo and agent technologies at GTC. These developments position NVIDIA not merely as a hardware vendor but as an orchestrator of the entire AI inference stack.