[mchen076]

"Isn’t it annoying / time intensive / costly to rely solely on cloud GPUs when trying to learn or have hobby projects in ML?"

No, it's all of those things when you use them professionally as well. There's a couple of things that people (usually business/operations people) overlook with respect to the cloud. Let's take a look:

1. The cost of compute clouds grows exponentially with every addition. Need an SSD? Extra money. Need an extra large HDD? Extra money. Need it on a secure private network? Extra money. You can very easily rack up the cost of a single local GPU system per-user per-year if you work on the cloud as a large organization.

2. Security: The cloud causes a lot of privacy issues and security issues. Are you a government agency with confidential information? Do you need to keep exact track of where it is at all times? The cloud will never let you do this.

3. Connectivity: I've yet to have any cloud provider provide uptime on the level that a local GPU box can provide. Our record cloud down-time was over a week.

For these reasons, in larger organizations, cloud tends to get a lot of push back. Unfortunately, this is largely ignored by executives who have drank the cool-aid from Microsoft and think "cloud is the future"