Several Internet of Things (IoT) applications are booming which rely on advanced artificial intelligence (AI) and, in particular, machine learning (ML) algorithms to assist the users and make decisions on their behalf in a large variety of contexts, such as smart homes, smart cities, smart factories. Although the traditional approach is to deploy such compute-intensive algorithms into the centralized cloud, the recent proliferation of low-cost, AI-powered microcontrollers and consumer devices paves the way for having the intelligence pervasively spread along the cloud-to-things continuum. The take off of such a promising vision may be hurdled by the resource constraints of IoT devices and by the heterogeneity of (mostly proprietary) AI-embedded software and hardware platforms. In this paper, we propose a solution for the AI distributed deployment at the deep edge, which lays its foundation in the IoT virtualization concept. We design a virtualization layer hosted at the network edge that is in charge of the semantic description of AI-embedded IoT devices, and, hence, it can expose as well as augment their cognitive capabilities in order to feed intelligent IoT applications. The proposal has been mainly devised with the twofold aim of (i) relieving the pressure on constrained devices that are solicited by multiple parties interested in accessing their generated data and inference, and (ii) and targeting interoperability among AI-powered platforms. A Proof-of-Concept (PoC) is provided to showcase the viability and advantages of the proposed solution.