I have seen a variety of datacenter designs. Some are layer 2 based, some lean more towards layer 3 based. To see what the 'big guys' are doing, I've been watching the release of some recent Internet Drafts.
One of them is from Microsoft/Arista/Facebook employees, and titled: Use of BGP for routing in large-scale data centers. draft-lapukhov-bgp-routing-large-dc-07.
It discusses using a Clos Network physical architecture with EBGP as the routing protocol, without any IGP protocol. It is said able to scale up to thousands and thousands of servers in a redundant, load balancing fashion. This type of network is good for captive, single tenant, style networks.
Another document references Framework for DC Network Virtualization. This document describes some vendor neutral common terminology and concepts. It also provides the ground work for talking about the layer-2/layer-3 design of the underlying network, the mechanism of load balancing multiple tenants over a common hypervisor group, and how to deal with any particular tenant when spread over multiple datacenters.
Another document, VXLAN: A Framework for Overlaying Virtualized Layer 2 Networks over Layer 3 Networks goes into some specifics of an overlay network. 2018/08/26: this has been released as rfc 7348.