Ok, so the authors apparently used atlas cloud hosting, which charges $1.80 per h100/hr, which would change the overall cost to around $0.08/ million input and $0.18/million output, which seems much more in line with massive inference margins for major providers.