Data center costs for dedicated air cooling for a rack over ~8kw is significant ie physical ducts are needed to be installed. Water cooling allows for more even heating controls as you can get hot spots within a rack even when drawing air through it.As an example, a modern rack nowadays will usually consume around 10-30kw of power, whereas a high-capacity AI rack can consume well above that and is expected to be around 50-100kw of power.
Just to put this into perspective the average American home uses 11kw, so it’s entirely possible that in the near future a single AI rack in a data center will reach the equivalent of 10 homes worth of power. A typical data center can have hundreds of racks, reaching essentially a small suburb’s worth of power. The cost of powering such a data center could easily reach into the 10’s of millions of dollars per year in electrical and maintenance costs.
On top of the power consumption, other resources such as water may also be required if the servers or GPUs are water-cooled.
Nvidia will have a rack of AI/GPU processors that will cost ~USD3m each and are expected to sell USD30B of them next year and are water cooled. Individual GPU thermal point of ~1kW each
https://nvidianews.nvidia.com/news/nvidia-blackwell-dgx-generative-ai-supercomputing
Many data centers are built around power availability ie current centers have maxed out available power input.
Let's not forget the other data center costs of security, redundancy (dual electricity providers, dual back up diesel generators, static switches and battery strings and the physical building. Many have 5G connectivity for remote server management/power cycling, timing synchronisation and also 5G towers on top.
I recall the time many years ago working in telco manufacturing when switches of POTS lines went from electro-mechanical to electronic and the biggest selling point was not needing active cooling ie passive cooling in the switch sites was all that was required plus five 9s uptime. The low power requirements were designed in and the biggest issue was the short product life cycle of Intel processors when carriers insisted on 30 year support for the technology! Clearly the world moved on with Cisco saying redundancy meant buying a second device with four 9s uptime and needing air conditioning. We are lucky if a product has a 5 year life cycle now.
Last edited:
Upvote
0