AI consumes computing power and thus energy. The AI models usually run on GPU hardware, graphics cards that have a very high power consumption. When we compute on servers, in terms of energy, it is really only electrical energy that is converted into heat. As a consequence, this means that additional energy is required to cool the servers. In German financial center Frankfurt alone, for example, the heat generated in data centers could theoretically heat the entire city. Data centers as the basis for cloud and AI solutions constitute a rapidly growing industry that we need to be thinking about now to ensure we are well positioned for the future.

From our perspective, it’s not enough. Renewable energy, of course, is better than a conventional electricity mix. But our fundamental goal should be that of using as little energy as possible. This starts with software development at the applications level and extends through the various software levels right up to the data centers. The question, then, is this: How can a data center be operated as efficiently as possible? Cooling is a major factor. Ten years ago, the same amount of energy necessary to run the servers was needed to then cool them. Today, we only need around 20 to 30 percent. Still, we must continually improve. The second major issue is waste heat utilization. As I mentioned, from an energy point of view, we are basically inputting electrical energy into data centers and getting heat out – even if, of course, there is some computing that takes place as well. This waste heat needs to be put to good use, and there are many different approaches for doing so, such as feeding it into the heat supply of buildings or connecting it to district heating networks.

Our focus is on what is called direct hot water cooling. We siphon off the heat produced by the components of the server – the processor or the graphics cards – directly to utilize it for other purposes, such as heating a building, for example, as we are doing in our pilot project in the former data center of the European Central Bank in a high-rise building in Frankfurt. This is a technical challenge, because we must tap the heat at a high temperature – otherwise, little can be done with it. Direct hot water cooling is more energy efficient than air cooling, which requires air to be cooled using traditional cooling systems. Doing so consumes significantly more energy than simply running a pump that circulates water, allowing it to flow over the servers and the hot components. That already results in lower energy use. Second, we then use the waste heat by feeding it into the building – into the heating system, for example. On the basis of this technology, we are deploying an open source-based cloud solution that AI customers can use to perform AI model training and inference in the most sustainable way possible.

Together with Vattenfall, we are building a cloud for AI companies or for other firms that require high computing power and for which the issues of sustainability and energy efficiency are vital. We set up our data center containers with direct hot water cooling on the site of a biomass power plant. We are, in other words, right at the renewable energy source. Our facilities there are so efficient that we only need about seven percent additional energy to run the data center relative to the immediate server power. We are also connected to the district heating network, into which we feed the heat generated during the cooling process. It is then routed onward to surrounding households near Stockholm.

Classic data centers are still operated with air cooling, and this makes waste heat utilization difficult, because the temperature level of the air is not particularly high and heat transport over long distances is difficult. To change the status quo with air cooling, a great many stakeholders would need to coordinate and move forward together. The large data centers are mostly colocation data centers where customers rent access to the servers. So, customers would have to bring in water-cooled hardware themselves, and the data centers would have to create the necessary infrastructure. This is the reason that Cloud&Heat exists in its current form. We have our own data center, and we can offer customers our own cloud. The biggest obstacle is that a great many stakeholders in the market tend to think conservatively in this area. To retool, data centers must first invest and recoup the investment costs through savings in operating costs or compensation provided for waste heat.

First, it is also a question of habits and availability. There is still far more air-cooled hardware out there than water-cooled hardware. Although the number of providers and models is increasing, the breadth of offerings hasn’t yet reached the same level. Some say it is the manufacturers who need to make a move. The manufacturers, though, say that demand for water-cooled systems isn’t great enough for them to switch their product lineups. So progress is generally only being made in small increments. That’s why we are providing support on this front and have, for example, developed a water-cooled server together with Thomas-Krenn. AG, which was one of the first systems to receive the Blue Angel label for environmentally friendly products.

There should also be greater political support for an issue that is so important to society. There are models from other countries in which, for example, the waste heat discharged from such processes is remunerated. Essentially, it is CO2-free heat, since the CO2 has already been “consumed” during the computation step anyway. So, we can either destroy this heat or put it to good use. It would make perfect sense for society to create an incentive to reward the extraction of CO2-free heat with a premium. Then the necessary technologies would be developed more quickly.

The third point is transparency toward users. On the one hand, this concerns the use of the cloud by companies. The fact that certain computing steps cause a certain amount of CO2 emissions needs to be transparently communicated. But end users also lack any awareness of the fact that using certain apps or services produces CO2 emissions. Awareness of such CO2 footprints has already developed in other areas.

It is, in fact, not easy. In the cloud context, for example, different customers share a server, so it isn’t immediately possible to say how much customer X consumed on the server or how much energy was consumed in this or that computing process. But there are solutions for that, too. We are addressing precisely these kinds of questions in the context of a major European research project (IPCEI-CIS). The important thing here is to arrive at the solution as quickly as possible, even if a high degree of accuracy cannot be immediately achieved. Even the CO2 footprints that are found on food products aren’t accurate down to the decimal place. That’s not really the point, though. What is important is that we have an informed sense that some foods lead to large CO2 emissions and others do not. By the same token, we also need to take a step forward with cloud solutions to create more transparency and, looking ahead, move computing tasks to the point where CO2 is lowest.

I believe we will see increasing convergence here. We are seeing right now that as electricity prices rise, it is becoming increasingly attractive to improve energy efficiency. In some countries, the cost of electricity is still so low that no one has to worry much about it. In such places, the environment and the economy are diverging. Of course, the solution isn’t to raise the price of electricity as an innovation incentive for energy-efficient technologies. That would just result in data centers relocating to other countries. At the same time, the environment and the economy are no longer mutually exclusive.

It is good that the problem has been recognized at the political level, but it is also extremely important to act quickly, because megawatt-scale data centers are now being planned and built. If we don’t take countermeasures now, the data centers will operate as designed for decades. The local government of Amsterdam no longer allows new data centers to be built unless they include a plan for waste heat. So, there is already movement. But governments continue to struggle with setting criteria aimed at sustainability.

This is a good approach for establishing a certain framework that everyone can adhere to. But I don’t think there are many data centers out there yet that can meet the requirements. However, we do orient ourselves on these kinds of ambitious goals. In addition, we also need to make sure that there is enough breathing room to allow new solutions that are not yet reflected in any sustainability label to develop their full impact.