Only explanation I would have is that a higher number of users would cause load balancing to serve lower parameter models. That way you’d have less reliable answers at peak usage times, but could enable more simultaneous users.