Using cheap flash memory instead of HBM could slash AI inference costs

Original: Meta really fumbled this guy.

Source: x.com ↗