I imagine Microsoft like any other bean counter company heavily quantizing the GPT models or whatever they use to serve it at scale with minimal cost.