A city AI fabric adds a second GPU rack that skews leaf uplink utilization under ECMP. What load-distribution action best restores balance?

A city AI fabric adds a second GPU rack that skews leaf uplink utilization under ECMP-only load distribution. The solution is to rebalance leaf uplinks and update ECMP membership so the new GPU rack shares spines properly—new uplinks from the new leaf to spines are added, and ECMP is updated so flows from both racks distribute across all available paths. That rebalancing removes the skew and lets both racks operate at fair utilization. Enabling adaptive routing on spine switches so the new GPU rack’s traffic is redirected across uncongested leaf paths only reacts to congestion at the spine layer and leaves the skewed leaf uplink ECMP membership unchanged, so utilization stays imbalanced and collective flows risk reordering. Configuring weighted ECMP on spine uplinks so the new GPU rack’s traffic is redistributed toward underused links only biases existing spine paths without adding balanced leaf uplinks and updating ECMP membership for the new leaf, so fair utilization is not restored. Advertising revised spine costs from the new rack so the GPU traffic is redistributed across available ECMP paths steers preference by cost without rebalancing leaf uplinks and ECMP membership, which undermines equal-cost forwarding and leaves the skew in place. Exam caveat: adding new uplinks is not enough; ECMP must be reconfigured to include the new paths, or the new capacity sits unused while old paths remain overloaded. Operational check: add uplinks from the new leaf to all spines, update ECMP membership on the new leaf, run load tests with both GPU racks active, and show spine uplink utilization now balances between the two racks.