It's not a good bet that data centers will need to build out at a constant rapid pace into the future to keep pace with demand. A lot of the work in servers in

It's not a good bet that data centers will need to build out at a constant rapid pace into the future to keep pace with demand. A lot of the work in servers in recent years has been about improving utilization of available resources, and I expect that there is still a lot of headroom left for improvement. It's probably safe to say that across all datacenters, the vast majority of heat is generated while servers are mostly idle and have too little to do. I don't think that it's safe to anticipate reaching some kind of fundamental limit on power efficiency, either. Nobody has come up with a viable replacement for the silicon transistor, despite decades of trying, and current CPU technology has practically stopped generating any improvements for the past several years. I think the important quantity to measure is the energy that goes into powering up the cpu for one clock cycle, because clock cycles are what drive computation forward and their individual energy usage basically just add together. For decades after the invention of the transistor, this energy dropped rapidly due to shrinking the wires. However, the wires reached a point where their capacitance began to fall off with decreasing size, and there have been no significant energy improvements since. Solving that problem requires either a revolutionary discovery about the physics of conductivity, a completely novel mode of doing computation that isn't based on electronics (photons?), or a fundamentally novel principle of organizing a cpu that somehow avoids being clock-bound (I heard about an effort over ten years ago, but I have not heard of it since). Any of those improvements could have surfaced at any time in the past few decades and been eagerly pursued, but none did, so the odds don't favor the miraculous appearance of one of them in the foreseeable future.