[pleoxy]

You can adopt a centralized system. And it's fashionable to do so. Lots of people here and other places advocate doing everything you can on the server.

But it's not required. Additionally, there are plenty of good incentives not to, including cost. Client CPUs are fast, free to use, and don't require a high latency network hop. Local ram, disk, and network transfer are also available in abundance. You can totally reverse the architecture where the client is doing the heavy lifting and provide the user a good experience.

Intel CPUs are inefficient when pegged, but more efficient at idle. On AMD your using the same cores the cloud is. Without the massive IO attached on the server side the client CPUs are more efficient for a unit of work. Apple is more efficient than the PC side presently, and Qualcomm is entering the fray and appears to be quite efficient.

So I'm really not certain where you're efficiency claim about client hardware comes from. All that hardware is just sitting there and sales vary, but they are not cratering for high performance local compute.

You have a point about GPUs, but only a very few very specialized applications need those.