OneRail uses Nvidia AI for real-time last-mile delivery optimisation
OneRail launched OmniSTAR, an AI-powered delivery optimization platform built on Nvidia's cuOpt and cuDF, enabling real-time evaluation of multiple fulfillment modes (owned fleets, couriers, parcel carriers) to select the lowest-cost option meeting service requirements The GPU-accelerated architecture reduces computation times by up to 10x, bringing calculations that previously took 20 minutes down to under two minutes and weekly computations down to approximately two days This speed enables liv
Analysis
TL;DR
- OneRail launched OmniSTAR, an AI-powered delivery optimization platform built on Nvidia's cuOpt and cuDF, enabling real-time evaluation of multiple fulfillment modes (owned fleets, couriers, parcel carriers) to select the lowest-cost option meeting service requirements
- The GPU-accelerated architecture reduces computation times by up to 10x, bringing calculations that previously took 20 minutes down to under two minutes and weekly computations down to approximately two days
- This speed enables live, order-level decision-making rather than batch planning, allowing retailers to evaluate multiple fulfillment combinations before assigning an order
- OmniSTAR is already deployed with enterprise customers, including US Foods (which adjusted pricing and delivery patterns after identifying margin-eroding configurations) and an unnamed tire distributor that achieved $40 million in run-rate savings over three years
- OneRail and Nvidia collaborated for three years on the project, with direct engineering engagement, and OneRail expects OmniSTAR to exceed $6 billion in gross merchandise volume by Q4 2026
Why It Matters
This represents a significant practical application of GPU-accelerated mathematical optimization to last-mile logistics, demonstrating that the gap between research-grade optimization and production deployment is narrowing. For AI practitioners and logistics professionals, it validates the commercial viability of combining ML-based prediction with real-time optimization at scale, and shows how GPU acceleration can transform operations that were previously constrained to offline, batch-mode planning.
Technical Details
- Architecture: OmniSTAR combines Nvidia cuOpt (an open-source, GPU-accelerated optimization library for vehicle routing and mathematical optimization problems) with cuDF (a GPU-accelerated tabular data processing library for filtering, joining, and aggregating datasets), integrated with OneRail's proprietary delivery pricing and performance data
- Optimization approach: cuOpt uses GPU-accelerated heuristics to generate candidate solutions and iteratively improve them rather than exhaustively testing every possible route, producing high-quality results within set computation time windows; it accounts for vehicle costs, capacities, travel times, operating windows, starting locations, and weighted cost models (distance, time, monetary cost)
- Prediction layer: OneRail's broader ML models estimate service time, lateness risk, first-attempt delivery success probability, and expected price ranges, which feed into the optimization systems that determine order execution
- Data scale: The platform draws on millions of deliveries across a network of over 12 million drivers and 1,000+ logistics partners, covering pricing and performance across different transportation modes
- Dynamic re-optimization: Because cuOpt is stateless, the system can rerun delivery scenarios as variables change (fuel costs, weather, shipping conditions, vehicle breakdowns, driver absences, traffic, new high-priority orders), enabling continuous recalibration rather than static planning
Industry Insight
- GPU-accelerated optimization is moving from niche research to enterprise production: The OneRail-Nvidia collaboration demonstrates that complex logistics optimization, once the domain of specialized operations research teams running overnight batch jobs, can now run in near real-time, opening the door for order-level optimization at scale across retail and distribution
- The "prediction + optimization" pipeline is becoming a standard architecture for operational AI: OneRail's separation of ML-based forecasting (service time, lateness risk, delivery success probability) from optimization-based decision-making mirrors best practices in dynamic vehicle routing research and suggests this two-stage approach will become increasingly common in supply chain AI
- Margin recovery through delivery optimization is a measurable, high-value use case: The reported $40 million in run-rate savings and the US Foods case (identifying low-margin products being shipped long distances on high-cost equipment) show that delivery optimization directly impacts profitability, making it a compelling ROI story for enterprise adoption and likely accelerating investment in similar platforms across the logistics sector
Disclaimer: The above content is generated by AI and is for reference only.