Massive market interest in AI has driven unprecedented investments in Special HW and runtimes for Neural Networks. Tensor computations are emerging as the de-facto API for all these special hw and runtimes. In this talk, we show how we can automatically transform and optimize relational queries and Classical ML pipelines into tensor computations, and run on special hardware. Interestingly the performance we obtain significantly outperform classical systems and even custom-build GPU DBMSs. At the same time, this approach retains very low engineering costs, thanks to a minute code footprint (<10k LoC) and free portability---as we piggyback on tensor runtimes getting ported to all the new HW coming out. We conclude touching on further research directions that emerge once both queries and ML models are uniformly represented as tensors computations.