Artificial Intelligence (AI) computation workloads are very challenging for resource-constrained Internet-of-Things (IoT) devices. Offloading of the AI workloads to edge nodes introduces network transport latencies that may make the offloading infeasible for low-latency applications. For applications that allow for anticipatory (speculative) computation with partial application datasets, e.g., scalable encoded images, we introduce and evaluate the concept of No further delay. We define No further delay as the optimal tradeoff time instant between waiting for more application data (which increases the latency) and starting the speculative computation (whose success probability increases with more data). We evaluate the expected negative latency achieved by anticipatory computing with partial application-layer datasets compared to computation on the device (with full datasets) with a novel adaptation of the speedup factor (ratio) from Amdahl’s Law. Previously, speculative execution has been limited to the instruction-level in microprocessor program execution; in contrast, we are the first to propose and examine speculative execution for application-layer datasets. For a computer vision inference workload with progressively coded JPEG images, we find that our No further delay approach with edge node offloading achieves three-fold speedups (despite the network latency to the edge node). This article serves as problem definition for a new class of AI mechanisms that should be developed in future research to estimate the optimal No further delay time instant based, e.g., on historical success rates of application-layer tasks computed with partial datasets.