Javed Khan knows the codec business from the inside. Before joining Neat as CEO in March, he ran Cisco Collaboration as SVP and General Manager, overseeing a multi-billion-dollar portfolio that included its room device line. Neat's co-founder and CTO, Ivar Johnsrud, was one of the engineers behind the Tandberg C90, later the Cisco C90, still widely cited as the high point of codec design.
Neither is arguing those systems were poorly built. Khan's argument, set out in a blog post this week, is that the model behind them has hit a ceiling. The industry's response has been to scale the problem rather than solve it: more compute, more cabling, more sensors around the same central architecture. Costs and complexity grow. What the room can actually do does not.
"The best version of a legacy model is still the old model," he writes. "The world has changed, and the potential of AI has shifted the goalposts."
Why Distributed Architecture Works Better in Large Meeting Rooms
Khan's proposed alternative draws on his time at Aptiv, where he ran Intelligent Systems and built distributed edge compute platforms for automotive, aerospace, and robotics. The comparison he reaches for is a self-driving car: cameras, sensors, and compute distributed across the vehicle, processing in real time, co-ordinating without routing everything through a single central unit.
Large conference rooms, he argues, have the same structural problem. People move. Conversations shift across a space. A centralised system was not designed to track that dynamism. Adding more hardware around it does not change that. "A centralized system can't keep up," he writes. "It wasn't designed to."
Neat's answer is to distribute compute, sensing, and AI inference across every device in the room. Cameras, microphones, and companion units each process locally and co-ordinate over IP. Rooms scale by adding or removing devices rather than replacing core hardware. Khan frames this as evolution rather than rejection: "This isn't abandoning the codec. It's redefining what it becomes."
He also distinguishes Neat's approach from partial attempts by legacy vendors. Some distribute devices but still route processing back to a central unit. That, he argues, just recreates the bottleneck in a different form.
