OneRail launches OmniSTAR using Nvidia cuOpt and cuDF for live last-mile decisions
The platform claims up to 10x faster optimisation and cites early enterprise deployments plus a Q4 2026 GMV target above $6B.
OneRail has launched OmniSTAR, a GPU-accelerated last-mile delivery optimisation platform built on Nvidia’s cuOpt and cuDF. The company says the speedup is enough to shift routing and delivery-mode selection from batch planning into live, order-by-order execution.
OmniSTAR Debuts as a GPU-Accelerated “Decision Layer” for Last-Mile Delivery
OneRail’s OmniSTAR is positioned as a decision system that sits inside last-mile fulfilment, the final leg from a local hub to the customer where costs and service failures tend to concentrate. The product’s core job is not just mapping a route. It is choosing how each order should be executed across multiple fulfilment modes.
Mechanically, OmniSTAR evaluates options including owned fleets, couriers, parcel carriers, and other delivery modes, then selects the lowest-cost option that still meets the required service level, OneRail said. That “service level” constraint is what turns this into an optimisation problem rather than a pure cost-minimiser, because the cheapest mode is often the one that misses the promised window.
Under the hood, OmniSTAR combines Nvidia’s cuOpt decision optimisation engine and cuDF data processing software with OneRail’s delivery pricing and performance data. The routing and delivery-mode calculations run on Nvidia accelerated computing infrastructure, per OneRail. The company framed the build as a deeper integration than a logo-level partnership, saying it engaged directly with Nvidia’s cuOpt engineering team and participated in the Nvidia Inception programme. The project took three years of work with Nvidia before launch, CNBC reported.
OneRail also tied the product to a broader stack that mixes prediction and optimisation. It said its machine-learning models estimate service time, lateness risk, probability of first-attempt delivery success, and expected price ranges, and that those predictions feed into optimisation systems that decide how an order should be executed.
From Batch Planning to Live Optimisation: The 10x Speed Claim and Margin Framing
The claim doing the work here is latency. OneRail said OmniSTAR can reduce computation times by as much as 10x, with examples including a calculation that previously took 20 minutes completed in under two minutes, and a calculation that took a week reduced to about two days. The company’s argument is that this is the difference between optimisation as a planning exercise and optimisation as an operational control loop.
In practice, batch planning tends to lock decisions early because recomputing is too slow to do repeatedly as conditions change. OmniSTAR is pitched as fast enough to run “within live delivery operations,” where multiple fulfilment options can be compared before an order is assigned, OneRail said. That is a different usage pattern: instead of producing one plan for the day, the system keeps re-solving the vehicle routing problem (VRP), a class of optimisation problems that searches for efficient routes under constraints like time windows, capacity, and costs.
Nvidia’s own description of cuOpt matters because it defines the tradeoff. Nvidia describes cuOpt as an open-source, GPU-accelerated optimisation library for vehicle routing and other mathematical optimisation problems, and says it uses heuristics rather than exhaustively testing every possible route. The solver generates candidate solutions and iteratively improves them to produce high-quality results within a set computation time, Nvidia said.
That time-bounded approach fits last-mile operations where the question is often “good enough, now,” not “perfect, later.” OneRail CEO Bill Catania framed the business consequence directly in a CNBC interview: “If you don’t have the ability to make lightning-fast decisions, you’re giving up margin,” and added, “Last-mile fulfilment is expensive.”
OneRail said OmniSTAR is already deployed with selected enterprise customers and cited US Foods as an example. In that deployment, OneRail said the system flagged delivery configurations that were reducing margins, including low-margin products transported long distances using higher-cost equipment, and that US Foods adjusted pricing and restructured some delivery patterns based on those findings.
The bigger ROI numbers are harder to underwrite from the packet alone. OneRail told CNBC that an unnamed large tire distributor achieved $40 million in run-rate savings over three years using the platform. The customer was not identified, and no methodology or third-party verification was provided alongside the figure.
OneRail also told CNBC it expects OmniSTAR to exceed $6 billion in gross merchandise volume (GMV) during Q4 2026. That is a forward-looking target, and the assumptions behind it were not disclosed in the provided material.
Signals to Watch for OneRail launches Nvidia-powered last-mile
The first confirmation point is customer naming and case-study quality. US Foods is a concrete deployment reference, but the largest savings claim in the packet is tied to an unnamed tire distributor. If OneRail starts publishing independently verifiable case studies, the “up to 10x” compute-time claim and the $40 million run-rate savings figure become more than a sales narrative.
The second is benchmarking that separates speed from decision quality. cuOpt’s design is to deliver high-quality solutions within a fixed compute budget, not guaranteed optimality. What matters operationally is whether OmniSTAR’s lower latency holds at scale while still meeting service constraints and producing better cost outcomes under re-optimisation events like traffic, road blockages, driver absences, or new high-priority orders, which Nvidia lists as triggers for dynamic re-optimisation.
The third is whether OneRail can substantiate its volume ambition. The company’s stated expectation to exceed $6 billion in GMV in Q4 2026 is a clean milestone, but it needs interim disclosures or third-party reported volume to be tradable as an adoption signal rather than a projection.
The fourth is distribution expansion that increases the reachable delivery network. In March 2026, FedEx launched FedEx SameDay Local in collaboration with OneRail, connecting customers to a national network of more than 1,000 delivery providers. More partnerships of that type would widen the set of fulfilment modes OmniSTAR can realistically compare for a given order.
My Take: This Is an Enterprise AI Adoption Signal, Not a Crypto Catalyst—Yet
The mechanism that decides whether OmniSTAR matters is not “AI in logistics.” It is whether GPU-accelerated optimisation actually moves from periodic planning runs into live order assignment, because that is where mode selection becomes a margin lever instead of a reporting tool. If the 20-minutes-to-under-two-minutes example holds in production, OneRail’s pitch that you can compare fleets, couriers, and parcel options per order starts to look operationally plausible.
The catch is that the packet’s biggest numbers are company-provided and lightly specified. The threshold that matters is independently verifiable evidence that the speedup and savings persist at enterprise scale, plus credible GMV progress toward the Q4 2026 target, because that is what would turn this from a vendor story into a durable adoption signal.
