TPUs have a lower marginal cost advantage compared to NVIDIA GPUs, but lack external ecosystem support and all optimization costs must be borne by Google alone.