Nowadays, the exponential growth of LLMs in size, increasingly stringent data sovereignty regulations, and growing power and energy constraints in modern datacenters are driving the adoption of multi-datacenter infrastructures to distribute the training process. However, standard training frameworks prove ill-suited for these heterogeneous environments, lacking awareness of inter-datacenter constraints or stretching communication patterns across sites without differentiating between network tiers. This approach causes severe bottlenecks on high-latency inter-datacenter links, hindering scalability. To address this challenge, we propose a hierarchical communication strategy designed to decouple intra-site aggregation from global synchronization through the election of "leader model replicas". Experimental results show that the proposed hierarchical strategy reduces training iteration time by up to 32% and exposed communication time by up to 65% compared to the baseline Megatron-LM strategy, providing an effective solution for scaling training workloads across geographically distributed datacenters.