Approach hardwires model weights into transistors, and uses older 6nm process. Targetting 70b model sizes (presumably 16 bit) by year end. It should cost much less than a 140gb card. but I don't know details.

other discussions
 

Taalas HC1: 17,000 tokens/sec on Llama 3.1 8B vs Nvidia H200's 233 tokens/sec. 73x faster at one-tenth the power. Each chip runs ONE model, hardwired into the transistors.