Eight of these "Sohu" devices can replace a 160xH100 – quite a bold claim. 

However, the catch is that Sohu supports only transformers and only their inference. This means sacrificing versatility (like what NVIDIA GPUs offer) in favor of speed for a very specific set of operations needed for transformer inference.

Once mass production of Sohu begins, the situation will be as follows: to create a fundamentally new architecture that could potentially replace transformers in production, it will need to be demonstrated that the new architecture runs on universal GPUs faster than transformers on specialized hardware like Sohu. Alternatively, additional resources will need to be invested in developing new specialized chips for the new architecture.