The vast heterogeneity of brain tumors—spanning patient populations, imaging acquisitions, and the fundamental biological differences between primary and metastatic disease—poses a significant obstacle to developing universal AI-based segmentation models. Addressing the MICCAI BraTS 2025 “Generalizability of Segmentation Methods Across Tumors” (GoAT) challenge, we developed a highly tailored framework within the nnU-Net v2 architecture. This features a deep, six-level residual encoder U-Net with a Focal Loss objective for complex features. Our core contribution is a novel dynamic batching strategy (3 \(\rightarrow \) 2 \(\rightarrow \) 1). This approach maximizes GPU memory utilization, enabling high-resolution training of our large-capacity model on a single consumer-grade GPU—obviating the need for multi-GPU compute clusters—while improving training efficiency and promoting generalization. After 5-fold cross-validation on the official training data (n=1,351), our model was evaluated on the blind validation set (n=451). Our solution achieved exceptional mean and median Dice scores of 0.8704 and 0.9367 (WT), 0.8542 and 0.9382 (TC), and 0.7759 and 0.8989 (ET), with a corresponding median 95th percentile Hausdorff Distance of 2.00 mm for the tumor core. These results validate our method’s robust generalization across a wide spectrum of tumor morphologies and prove the power of strategic, resource-efficient training innovations in creating a single, clinically-relevant universal model.