Buildings are a substantial, and sometimes overlooked, consumer of energy, accounting for roughly a third of global primary energy consumption, with roughly half of that being used for heating and cooling. While the replacement rate of buildings is low, and energy retrofits are challenging, if possible at all, the installation of an advanced control system is comparatively quick and simple. This may be done in an effort to reduce a building's energy consumption by avoiding excessive heating or cooling, but also to harness a building for ancillary grid services. Our efforts in this regard are focused on model predictive control, because of its ability to deal with slow-acting systems, incorporate input and output constraints, as well as general versatility in terms of control objectives. However, the commissioning of model predictive control may require enough work-hours to make it be prohibitively expensive. This applies in particular to obtaining a sufficiently accurate model of the plant, since buildings are usually unique, and thus require an individual model. Hence, we restrict ourselves to data-driven methods, that are able to learn a stable and accurate model within a few days of training. We find linear input-output models to be well-suited to this purpose. As our findings indicate that there is little room for improvement in terms of model accuracy and energy efficiency, compared to well-established methods, and considering the increasing prevalence of electric chillers and heat pumps in buildings, we emphasize the use of buildings as virtual power plants for ancillary grid services as a promising and less exhausted avenue of research. Since it rarely feasible to control the power of a chiller or heat pump in a building directly, and we demonstrate the limitations of controlling them indirectly, we explore different methods of aggregating many buildings into a cohesive unit to provide ancillary grid services on a large scale. We show practically usable levels of tracking accuracy with simple methods, relying exclusively on centralized computation and indirect control of the individual buildings. Therefore, we recommend further exploration in this direction, using real buildings in the field, to enable the continued integration of wind and solar power into the electric grid.