Capital allocation in compute is misinvested, with generalist GPUs (such as Nvidia A100, H100, H200, GB200, GB300, and AMD MI300x) categorized as over-built for training. Conversely, specialized ASICs focused on inference (including SambaNova, Cerebras, Etched, Taalas, d-Matrix, Positron, Furiosa, Tenstorrent, MatX, TensorDyne, Fractile, Groq, and Google TPU) are categorized as under-built.