The short answer is that none of our regions are so big that we have to own our own top-of-rack switching yet (though some are getting close), so we can take a random spot in (say) an Equinix rack, rather than having to engineer a network (we run our own BGP for Anycast, but our uplinks are just DAC connections to our upstream providers switches.
For what it's worth: we've been on our own hardware ever since we launched, long before the 14MM Intel/Dell round. It's really the only way we can see to make the margins make sense. A big part of the premise of AWS and GCP is that they're allocating hyperscaler-grade resources to the task of making sure they claim most of the margins in hosting things on their own platform, not middlemen --- though that may be more true for some services (like EC2) than others (like S3).
For what it's worth: we've been on our own hardware ever since we launched, long before the 14MM Intel/Dell round. It's really the only way we can see to make the margins make sense. A big part of the premise of AWS and GCP is that they're allocating hyperscaler-grade resources to the task of making sure they claim most of the margins in hosting things on their own platform, not middlemen --- though that may be more true for some services (like EC2) than others (like S3).