Source: IT Daily
As demand for AI rises and Europe imposes strict sustainability regulations, are in-house data centers still a viable solution? Schneider Electric thinks so, and certainly in Belgium. A thoughtful approach is needed, though.
On-premises data centers of companies face major challenges. After all, the EU launched the Energy Efficiency Directive: a regulation that stipulates, among other things, that operators of data centers with a capacity above 500 kW must report on their consumption and efficiency.
Data centers also face stricter standards. For example, the Climate Neutral Data Centre Pact targets a power usage effectiveness (PUE) of 1.4, and Germany wants to get data centers to 1.2 by 2026. That means that for every watt of IT equipment a data center should consume a maximum of 1.2 watts, with only 0.2 watts of overhead for cooling, among other things. Since data centers are large consumers, that push for greater efficiency is a welcome one.
Bigger is not always better
Hyperscalers and colocation specialists have the scale to build new data centers with low PUE, but what about proprietary on-premises at enterprises? Such a data center must also run efficiently so as not to cause the owner problems with ESG reporting. That becomes all the more difficult given the increasing density of IT, driven by AI. Where server racks used to measure two to four kilowatts, that now rises to sixty or even 120 kW in extreme cases.
Colocation specialists and hyperscalers want to attract IT infrastructure from such data centers to their sites, but that is not always necessary or even desirable, according to Martijn Aerts, VP Secure Power BENE cluster at Schneider Electric. “In Europe’s data center hotspot, there is currently a problem with permits and access to power. In Dublin, citizens have even filed a complaint against high-consuming data centers.”
Large sites powered by AI and HPC consume immense amounts of power. This involves not only power availability, but also grid capacity. Data center specialists compete for real estate that can be connected to an adequate power supply. Aerts: “In Belgium, that problem does not prevail in that way.”
Small and dispersed
“We actually have a lot of small data centers in Belgium,” Aerts notes. “Just think about universities and hospitals.” We’re not going to turn all of those data centers into huge AI training centers right away. “The racks of 60 kW or more are for Microsoft, Google and Meta. With us, we’ll see smaller increments, and that’s where modularity is possible.”
Aerts believes in brownfield solutions, converting parts of existing data centers to accommodate AI servers. As long as these do not become too large, the impact on the power grid is limited and the Belgian grid can cope.
Such data centers are quite accessible, especially if you rework existing sites or opt for modularity. “We are already talking very concretely with customers to make part of the server room AI-ready” Aerts said. “And there are modular prefab modules.”
Learning from the telecom sector
The latter is very important. “Building a new data center has a lead time of two to three years, and costs a lot of money. With prefab solutions, you can add 10 MW, 5 MW or even 500 kW. That costs less, is readily available, and has a smaller impact on the power grid.“
