[nomel]

> near zero chance for you main customers workload

What percent of this hardware is running inference for ReLU models? ;)