The great NPU failure: Two years later, local AI is still all about GPUs