New study finds 6G can enable cloud-based autonomous driving by 2028

A new study from researchers Pouya Parsa, Kawon Han, and Seongjin Choi, published on arXiv, systematically analyzes whether cloud infrastructure can feasibly offload real-time autonomous driving tasks over 5G and 6G networks. The paper focuses on frontier vision-language-action (VLA) models, which require approximately 60 trillion floating-point operations per forward pass—too computationally heavy for economical onboard deployment. The researchers develop an analytical framework coupling communication limits, a roofline GPU service model, stochastic latency, and utilization-aware cost, applied to New York City traffic patterns. They introduce two decision tiers: a reactive 100 ms budget and a deliberative 300 ms budget, assuming an onboard reactive fallback.

The results reveal three nested binding regimes: communication binds first in dense cells (5G fails early, 5G-Advanced becomes the practical threshold, and 6G adds headroom), compute binds next under the reactive budget (near-term VLA latency-infeasible due to memory-bandwidth-bound autoregressive FP16 decode at ~114 ms on 2025 hardware, clearing 100 ms around 2027), and cost binds last (utilization-pooled cloud GPUs undercut onboard hardware for VLA, which costs up to $8,500 per vehicle-year and sits idle most of the day). Feature-level offloading (S2) is where VLA cost crossover concentrates. The study concludes that latency decides which model is admissible in which year; cost decides whether it is economical.