Earlier this month at NVIDIA GTC, telecom operators showed what becomes possible when AI runs inside their networks, close to the people who use it. One of the projects NVIDIA featured as a benchmark case study was a distributed inference architecture for voice AI. In a Comcast benchmark, it reached 42,362 tokens per second, an 80.9% gain, and cut the cost per token by 76.1%, while voice replies stayed under 500 milliseconds. In practical terms, the same GPUs produce nearly twice as much output, each answer costs about a quarter of what it did, and a caller hears the reply without a noticeable pause.
The architecture was designed by Miroslav Volek, Head of Engineering at Personal AI. For Volek, the result is the latest step on a long road that began with a question he first asked as a boy: what does a machine actually do when a person uses it? He reached AI through machine tools and government software, teaching himself each new field along the way.