Heterogeneous AI Compute: Why Mixed NVIDIA, AMD, and TPU Clusters Are Harder Than They Look

A practitioner's look at the runtime, scheduling, and deployment challenges of running AI training and inference across different architectures.