With the new demands for Data Centers that can run Generative AI-based applications appearing in more and more areas of the market, are existing Data Centers that were built five to ten years ago or more, technologically obsolete?
Data center designs were stable for about twenty years. If you built one twenty years ago, the one someone else built ten years later would have had pretty much the same design concepts. That is not the case today. Today, you need to have full redundancy for both power and broadband connectivity coming into the data center with totally diverse routing. The data center has to address the Nine Rs of Smart Grids/ Communication Networks.
The new Generative AI (Gen AI) Data Centers need to have much more power compared to traditional data centers built in the past twenty years. The need for higher density power per square foot as well as more rack space are new design criteria not easily retrofitted into an existing data center.
Flooring also requires a heavier distribution of weight per square foot and these types of physical design elements are not easily added onto existing data centers which were built with different “rules-of-thumb”.
There are many issues that need to be addressed when you add in so much power to a Data Center. With a Gen AI Data Center, you are almost required to include a liquid-cooled approach for cooling rather than the traditional airflow designs. You cannot have such dense GPUs (Graphics Processing Units) with having power consumption issues that generate so much heat, without having better methods to dissipate that heat as it fills the facility.
Another question that arises when you try to add very dense power applications to a data center is, does the data center have enough power coming into it to provide the actual amount of power
necessary to operate the applications?
If you are a company looking to buy a data center for your corporate requirements, you might be much better off looking at building a new facility from the ground up, rather than buying off on an existing one that is technologically obsolete. If you are looking at it as an asset with a useful life of fifteen to thirty years, you need to insure it is not going to fail in five or six years due to having inadequate power coming into it.
Some companies are trying to sell off existing data centers that may have obsolete designs and have no way to retrofit it to provide enough power to run new power-hungry applications. Do not fall for their bargain price. Obsolete junk is still obsolete junk even if it’s at a discount.
New demands for tenfold increase in power and the use of GPU (Graphics Processing Units) which take ten-to-fifteen times the amount of power per chip has made designing data centers a whole new science.
Some companies are talking about retrofitting older data centers to adapt to the new demands of Gen AI applications, but the reality is, that could be a supreme waste of time and money.
It’s like going back to old desktop PCs and trying to add a new processor onto an older PC chassis. Adding a turbo card to get faster processing was not the best thing to do because you still had older storage memory that was slow as well as a chassis that only provided so much power. Instead of trying to switch cards out, the smart idea was to just go buy something new, where the processing chip was mated with better and faster storage memory as well as all of it being put into a stronger, higher-powered chassis that could support more powerful cards and memory.
At this point, data centers are no different. You may not be able to add all the new amenities into an existing data center because you will be constricted by many limitations including ceiling heights, floor weight distribution maximums per square foot, and other physical attributes like power distribution and cooling capacities.
One thing is obvious, with all the extra heat from all the processors running, temperatures run a lot higher than the older data centers. Airflow cooling is not enough. You will need liquid-cooled facilities and that is not something that can be easily added-on to existing data centers.
The way I see the market going and the demand for new data centers hitting higher levels every quarter, I would focus on building new. Trying to squeeze so much more performance out of a maxed out data center would not be the best approach to take, or the most cost-effective.
