The large models are actually MoE these days so they're usable on ordinary hardware with weights streaming from SSD, just very slow. You're nonethess right that it makes the cloud-based offering more popular, since you can use that for convenience after testing a few inferences locally.