The Internet of Things (IoT) is expected to interconnect a myriad of devices. Emerging networking and backend support technology not only has to anticipate this dramatic increase in connected nodes, but also a change in traffic patterns. Instead of bulk data such as file sharing or multimedia streaming, IoT devices will primarily exchange real-time sensory and control data in small but numerous messages. Often cloud services will handle these data from a huge number of devices, and hence need to be extremely scalable to support conceivable large-scale IoT applications. To this end, we present a system architecture for IoT cloud services based on the Constrained Application Protocol (CoAP), which is primarily designed for systems of tiny, low-cost, resource-constrained IoT devices. Along with our system architecture, we systematically evaluate the performance of the new Web protocol in cloud environments. Our Californium (Cf) CoAP framework shows 33 to 64 times higher throughput than high-performance HTTP Web servers, which are the state of the art for classic cloud services. The results substantiate that the low overhead of CoAP does not only enable Web technology for low-cost IoT devices, but also significantly improves backend service scalability for vast numbers of connected devices.