I am increasingly asked during candidate interviews how much dedicated inference compute they will have to build with Codex. Pairing this with usage per user growing significantly faster than the number of users, it's pretty clear that compute will be something that is scarce.