Am I missing something? If the performance scales linearly, they are at 1000 computers internally (1/10 * 10,000), and it was said to take 8 hours. That would only be 80 hours if they hadn't have used this service.

This makes me believe someone is lying about something in this article.

Perhaps their internal capacity is already tied up in other tasks, so while they have 1000 cores internally, they can't all be monopolized for 80 hours for a single task like the AWS machines can.