The frontier labs are about to spend $100 billion on a single training run. In December, they got beaten on the hardest benchmark in AI by a company that spent nothing on one: Zoom. No frontier model. No gigawatt data center. It topped Humanity's Last Exam above Gemini 3 Pro, then GPT-5.2 Pro.
I sat down with Zoom's CTO Xuedong Huang, the man who started Microsoft's speech group in 1993 and ran Azure AI, to find out how a "meeting company" beat the labs at their own benchmark. The answer isn't a bigger model. It's the layer nobody benchmarks: the router that decides which model gets used at all.
So here's my question for you: does the router become the moat? Or do frontier models get so cheap that routing stops mattering?