Large language models often weigh tens of gigabytes or even hundreds of gigabytes, placing pressure on traditional centralized hosting in terms of bandwidth costs and download speeds.