Fog computing plays a crucial role in satisfying the requirements of delay-sensitive applications such as connected vehicles, smart grids, and actuator networks by moving data processing close to end users. Passive optical networks (PONs) are widely used in access networks to reduce the power consumption while providing high bandwidth to end users under flexible designs. Typically, distributed fog computing units in access networks have limited processing and storage capacities that can be under or over utilized depending on instantaneous demands. To extend the available capacity in access network, this paper proposes a fog computing architecture based on SDN-enabled PONs to achieve full connectivity among distributed fog computing servers. The power consumption results show that this architecture can achieve up to about 80% power savings in comparison to legacy fog computing based on spine and leaf data centers with the same number of servers.