You don't need data center for inference. Even largest models run on a single server. You need data center for training; or for serving millions of users. Evaluating the model is neither of those.