Even the largest Airscale baseband configuration (6BB+2CU) is not a power hog, at a maximum of 2kWh, typically < 1kWh. The Nokia outdoor unit you referred to achieves a 70% relative advantage with the same traffic profile; power consumption naturally increases as traffic grows back to the previous level.
If we now consider that one or more H100/H200 units should be installed in an Airscale (indoor unit) subrack, cooling becomes a challenge: the hardware is dimensioned for 50% lower consumption. Installing GPUs in outdoor units is difficult due to passive cooling.
So, AI brings energy savings primarily by running algorithms in the NOC and scaling RAN resources up/down centrally with better precision. Bringing AI into the baseband processing itself does not save energy; instead, it increases energy consumption.