The stochastic fluctuations of AI computational loads inject harmonic currents into the DC bus, amplifying bus voltage ripples and weakening the power quality. Existing strategies typically rely on high-gain control strategies to minimize harmonic output impedance, aiming at full absorption of harmonic currents. However, such designs rarely consider engineering constraints such as capacity and current boundaries, which impose inherent limits on harmonic absorption. To address these issues, this paper proposes an adaptive harmonic impedance control and flexible compensation method for AI data centers. By integrating DC bus voltage feedforward with output current feedback, a virtual harmonic impedance control channel is constructed to enable real-time impedance shaping. Then, an adaptive gain regulation mechanism is developed to adjust harmonic impedance according to the available capacity and current margin. Compared with traditional strategies relying on fixed high gains or resonant links, the proposed method allows for the continuous regulation of harmonic impedance over a wide range. This enables the dynamic matching of harmonic absorption capability with the available capacity, effectively suppressing the risks of overcurrent, saturation, and stability degradation. Simulation and 8 kW experimental results verify the correctness and effectiveness of the proposed analysis and control strategy.