NVIDIA’s Open Software Trap: The Real Cost of the New Inference Stack