OneRail just made a move that should worry anyone still running last-mile delivery on static rules and spreadsheets. The company has launched OmniSTAR, an AI-powered platform built on Nvidia’s cuOpt decision optimization engine and cuDF data processing library, designed to tell retailers, wholesalers, and distributors exactly how each individual order should be delivered — and at what cost.
The pitch is simple: evaluate owned fleets, couriers, parcel carriers, and every other delivery mode on the table, then pick the cheapest option that still hits the required service level. OneRail CEO Bill Catania didn’t mince words with CNBC: “If you don’t have the ability to make lightning-fast decisions, you’re giving up margin. Last-mile fulfilment is expensive.” The performance numbers back that urgency. A routing calculation that used to take 20 minutes now finishes in under two. A week-long computation drops to roughly two days. That’s a 10x speedup, which matters because it pushes optimization from a batch process into live operations where multiple fulfilment options can be evaluated before an order ever gets assigned.
The architecture is worth understanding. Nvidia’s cuOpt is a GPU-accelerated optimization library that doesn’t brute-force every possible route — it generates candidate solutions and iteratively improves them using heuristics. That’s a meaningful distinction for anyone who assumes AI means exhaustively testing everything. It doesn’t. It gets good enough, fast enough, within a set time budget. OneRail layers its own delivery data on top: pricing and performance metrics from millions of deliveries across a network of more than 12 million drivers and over 1,000 logistics partners. The company says OmniSTAR can identify delivery rules that quietly inflate costs and assess how delivery choices affect item-level profitability.
What separates OmniSTAR from generic routing software is the tight coupling between prediction and decision. OneRail’s machine learning models estimate service time, lateness risk, first-attempt delivery probability, and expected price ranges. Those predictions then feed directly into the optimization layer. A 2024 review in the European Journal of Operational Research makes the same distinction: travel-time prediction and real-time re-optimization are separate research areas, and combining them well is genuinely hard. OneRail is betting Nvidia’s hardware makes that combination fast enough to actually run in production.
OmniSTAR is already live with selected enterprise customers, including US Foods, where OneRail says the system identified delivery configurations that were quietly eroding margin. That’s the pattern worth watching: companies don’t know what their delivery decisions actually cost until an optimizer shows them the alternatives. The fact that cuOpt is stateless — meaning every change in conditions requires re-modeling and re-submitting the problem — could be a limitation or a feature. OneRail’s answer is to rerun scenarios as fuel costs, weather, and shipping conditions shift. Whether that holds up at scale across thousands of live orders is the open question.
