Getting a parcel from a warehouse to a doorstep costs retailers a lot of money, and the choice of how to do it is usually made by fixed rules or a planner with a spreadsheet. OneRail, a delivery-technology company, says it has built a system that makes that choice order by order, using Nvidia software to do the maths fast enough for live operations. The product is called OmniSTAR, and the core claim is that it picks the lowest-cost delivery option that still meets the required service level.
From “which route” to “which kind of delivery”
Most routing software answers the question of how a vehicle should visit a list of stops. OmniSTAR starts one step earlier. According to OneRail, for each order it compares options such as a retailer’s own fleet, local couriers, parcel carriers and other delivery modes, then selects the cheapest one that meets the service promise. The target customers are retailers, wholesalers and distributors.
That difference matters because the delivery mode often decides the cost more than the exact route. A heavy, low-margin item sent a long way on expensive equipment can quietly wipe out the profit on an order, and a rules-based system may never notice.
The Nvidia pieces under the hood
OmniSTAR combines two Nvidia tools with OneRail’s own delivery data. The first, cuOpt, is an open-source library that uses graphics processors to solve routing and other optimization problems. Nvidia’s documentation says it can account for vehicle costs, capacities, travel times, operating windows and starting locations, and that costs can be measured in distance, time, money or a weighted mix. The second, cuDF, is a GPU library for processing large tables, covering tasks like filtering, joining and totaling data.
Nvidia says cuOpt does not test every possible route. It generates candidate solutions and keeps improving them with GPU-accelerated heuristics, which are shortcuts that find strong answers within a set time. In plain terms, the result is a very good answer fast, not a mathematically proven best one.
The speed claim: twenty minutes down to two
OneRail says the system can cut computation time by as much as 10 times. A calculation that once took 20 minutes would finish in under two, and one that took a week would take about two days. The company argues that this speed is what lets the optimization run inside live operations, where several fulfillment options can be weighed before an order is assigned.
One technical detail explains how the system stays current. Nvidia notes that cuOpt does not hold state between runs, so when conditions change, such as a breakdown, a driver absence, a road closure or a new priority order, the problem has to be set up and submitted again. OneRail says OmniSTAR reruns delivery scenarios as fuel costs, weather and shipping conditions change.
The data behind the decisions
The software is only as good as the data it sees. OneRail says its dataset covers millions of deliveries across a network of more than 12 million drivers and over 1,000 logistics partners, with pricing and performance for different transport modes. Its machine-learning models estimate things like service time, lateness risk, the odds of a successful first delivery attempt and expected price ranges, and those estimates feed the optimization step.
US Foods and an unnamed tire distributor
OneRail points to two customers. At US Foods, it says the system found delivery setups that were hurting margins, including low-margin products traveling long distances on higher-cost equipment, and that US Foods used the findings to adjust pricing and restructure some delivery patterns. OneRail also says an unnamed large tire distributor reached $40 million in run-rate savings over three years, and that it expects OmniSTAR to pass $6 billion in gross merchandise volume in the fourth quarter of 2026.
In March, FedEx launched FedEx SameDay Local in collaboration with OneRail, linking customers to a national network of more than 1,000 delivery providers.
Numbers to hold at arm’s length
Every figure here, from the 10-times speedup to the $40 million, comes from OneRail. The tire distributor is unnamed, so nobody can check its savings, and “run-rate” is a projection of the annual savings rate, not money already counted. The speedup also depends on the size and shape of each problem, and the company gives only two examples.
What is credible is the direction. Faster solvers let a company evaluate more scenarios and respond to changes during the day, instead of planning once and living with the result. Retailers weighing a platform like this should ask for results on their own order data before relying on a vendor’s averages.
