OAT introduces the first step towards foundation models for topology optimization by combining a neural field auto-encoder and latent diffusion models with large scale training on a new general dataset of 2M optimized topologies called OpenTO.