In cloud computing research, many models are used to predict server power. However, a lot of these models are not sufficiently tested on industry hardware due to lack of access to this type of hardware in academia. In this work, we address this need for model evaluation and lack of data from real data centers. We obtain a large dataset from a data center of the company AEB SE, located in their headquarters in Stuttgart. The dataset contains hardware utilisation data on the averages of CPU-frequency, CPU utilisation, server power consumption, and ambient temperature as well as peak power consumption. These metrics are measured at five-minute intervals over the span of a year, for all 73 servers. We use the information on average CPU utilisation and average server power to train four server power models that use CPU utilisation to predict the power consumption of a given server. Two of these power models are from literature, and the other two are our own work. We form server groups, based on a combination of hardware characteristics the servers have, such as CPU models, server types and storage sizes. We then train the models on them and compare the accuracy the models have. This answers the question which hardware characteristics should be considered when grouping servers as a basis for training the power models and which distinctions are unnecessary. We also compare the models to each other, based on their accuracy, generalisability and speed of training. We find that one of our models was in all but a few cases the most accurate one. It also generalises better than the other three models and is one of the two fastest models in training. However, it does have the issue of predicting inaccurate and sometimes even semantically incorrect results in higher CPU utilisation areas. In plotting the server power samples at specific CPU utilisations, we observe that the general shape of these plots resembles a horizontal asymptote. Therefore we propose a model that tries to imitate this general shape. Unfortunately, the dataset we obtain is heavily biased towards lower CPU utilisation areas, which may introduce an error in our evaluation of the two least accurate models, one of ours and one from literature, both of which are dependent on using power measurements obtained at full utilisation. The dataset we obtain is freely available for research and can be used to evaluate other power models, or in other research that requires hardware utilisation data from a data center.