Because they actually know how to implement fault tolerant scale on AWS. Most companies don't have simian army randomly wrecking infrastructure. Many companies/engineers think they can spin up a few (or few hundred) ec2 instances and think they have solved resiliency.

AWS is a tool, not a bullet made of silver.

They are shifting capex to opex. The complexity of solution that can perform OK on aws is significantly higher then complexity of solution that runs on metal. There are use cases for which aws will never be a viable option. AWS in general is a bad option for anyone but Netflix (Netflix consumes over 30% of aws resources if they make a mistake that causes even 30-40% spike all AWS customers are f#$ed)