Follow The Frontier S1 Room 1: Models, Chips and Datacentres

​Follow The Frontier is a small Friday series from BoldCap where we sit with a few people and talk through one hard area of AI at a time. This is the first of four rooms.

​This room brings models, chips and data centres into one conversation: what would it take to make useful intelligence cheaper, and what is holding us back? Sometimes the answer is a better model. Sometimes it is getting more out of the machine we already have. And sometimes the limit is moving data or finding enough power to run it.

​We want to talk about alternatives to attention, including recurrent networks and state-space models; test-time scaling, where models spend more compute working through a problem; and cheaper training and serving approaches, from zeroth-order optimization to speculative decoding and quantization. The question is how these choices change what we can build and what it costs to use.

​Those choices also change what we need from hardware. An instant answer and an agent working overnight place different demands on the same chip. We'll explore the memory wall, in-memory compute, chiplets, optics and the links between chips, asking where each approach helps and where the bottleneck moves next. Underneath it all is the data center: power, cooling and the practical limits of getting these systems built and running.

​If you work on any of these, or you have hit one of them as a wall, this room is for you.

​There are only 10 invites. We keep it small so everyone in the room can actually talk.

​The room is at Lavelle Road, Bangalore. Please RSVP and tell us a little about what you work on.