The Tactile Internet (TI) sets a sub-millisecond tactile feedback round-trip goal, which may become unattainable once sensing, radio, and computing delays are considered. Negative-latency control mitigates this gap by predicting user intent ahead of time, shifting part of the delay budget into the past. We demonstrate that neuromorphic many-core hardware can deliver anticipatory inference within strict energy and latency constraints at the network's edge. A calibrated ten-member Convolutional Neural Network (CNN) ensemble for reach-to-grasp prediction is ported, via an automated microTVM flow, to the many-core SpiNNaker-2 (S2) chip and compared with a Google Coral Edge-TPU baseline. With one glove stream, S2 achieves 11.4ms end-to-end inference at 192mW (2.18mJ). Exploiting fine-grained dynamic-voltage-frequency scaling and temporal multiplexing, a single S2 can serve 15 concurrent streams while retaining 20.0ms latency and reducing energy to 0.48mJ per inference - a 6.8× reduction over the single-stream case on SpiNNaker and 4.6× lower than Coral's 32.82mJ at 10.59ms for 15 streams served sequentially. These results indicate that energy-centric scaling, rather than raw TOPS, is the primary bottleneck for TI wearables, and that arrays of low-power processing elements outperform monolithic accelerators on the metric of energy-per-predicted-action.