OneRail announced OmniSTAR on 1 September 2026, a delivery decisioning platform built on Nvidia’s accelerated computing stack, and gave CNBC the exclusive. The pitch is clean: evaluate every available delivery option for every order — owned fleet, courier, parcel — and pick the cheapest one that still meets the promised service window, in the moment the order lands. The company says Nvidia’s hardware and software cut the computation time by as much as ten times, turning a twenty-minute routing problem into a two-minute one.
That is a real product built on real technology, and the underlying machine learning problem is a genuinely hard one. But the launch arrived in three different documents — a Business Wire release, a CNBC exclusive and a wave of trade pickup — and those documents do not all say the same thing. The headline speed multiple is the best of three published examples. The Nvidia engine at the centre of the stack is free and open source, which moves the interesting question somewhere else entirely. And the word “first” in the press release headline runs into Nvidia’s own published customer list.
This article works through what OneRail actually announced, what can be verified from primary sources, and where the numbers need a second look. Nothing here suggests the product does not work. The argument is narrower and more useful: the claims that travelled furthest are not the claims best supported by the underlying record, and anyone evaluating this category needs to know which is which.
Table of contents
- What OneRail Actually Announced on 1 September 2026
- The OneRail Speed Claim: One Number, Three Different Shapes
- Nvidia cuOpt Is Free, and That Changes Where OneRail’s Advantage Sits
- Testing the Word “First” Against Nvidia’s Own Customer List
- The Two Deployments OneRail Has Disclosed
- The Network Behind OneRail’s Model: 2020 Versus 2026
- What OneRail Is Really Selling to Retailers
- A Claim-by-Claim Ledger for the OneRail Launch
- What This Means If You Run Delivery Operations
- References and Further Reading
What OneRail Actually Announced on 1 September 2026
OmniSTAR is a decisioning layer. It sits above the fulfilment options a retailer or distributor already has and chooses between them per order, rather than planning a fixed route set in advance. The Business Wire release describes the target as “enterprise retailers, wholesalers and distributors” and frames the outcome as same-day delivery at a cost that works.
The technical stack is named explicitly in the CEO’s own quote, which is unusually specific for a launch release. Bill Catania, co-founder and CEO, credits “NVIDIA cuOpt GPU-accelerated decision optimization engine, NVIDIA cuDF for data processing, and NVIDIA accelerated infrastructure” combined with “the proprietary data and operational models we’ve built over millions of deliveries.”
The platform, the stack and the people quoted
Three people carry the announcement. Catania handles the business framing. David Daeschler, OneRail’s head of AI, handles the technical framing and told CNBC the Nvidia relationship began three years ago. Azita Martin, Nvidia’s vice president and general manager for retail and consumer packaged goods, supplies the partner quote describing “a real-time decision layer that can route an order to the right carrier and delivery mode at the right cost, rather than relying on static rules or manual planning.”
| Element | What the launch says | Primary source |
|---|---|---|
| Product | OmniSTAR, a delivery decisioning platform | Business Wire, 1 Sep 2026 |
| Optimisation engine | Nvidia cuOpt | Catania quote, press release |
| Data processing | Nvidia cuDF | Catania quote, press release |
| Relationship | Nvidia Inception; direct cuOpt engineering engagement | Press release |
| Duration | “Multi-year”; three years per Daeschler | Press release; CNBC |
| Named customer | US Foods | Press release |
| Unnamed customer | “A large tire distributor” | CNBC exclusive |
| Commercial forecast | Over $6bn gross merchandise volume in Q4 | CNBC exclusive |
Where the announcement appeared, and where it did not
Here is the first oddity, and it is checkable in about a minute. OneRail’s own newsroom does not carry this launch. The company’s sitemap indexes 158 newsroom URLs and was last modified on 3 September 2026 — two days after the announcement. Searching that index for “omnistar” or “nvidia” returns zero matches. The public press-releases page tells the same story: its newest item is dated 5 March 2026, about a Gartner Market Guide.
So the flagship product launch of 2026 went out over Business Wire and to CNBC, and never landed on the company’s own press page. That is not evidence of anything improper. It is a reminder that the vendor’s own site is not always the authoritative record, and that a launch can be simultaneously well-covered and thinly documented at source.
The OneRail Speed Claim: One Number, Three Different Shapes
The speed figure is the claim that travelled. Every trade pickup leads with it, and it is the single number a busy operations director will remember. It is also the claim that fragments most under inspection, because the launch published it in three forms and they do not agree.
20 minutes to two minutes, or 20 minutes to two and a half
The press release states that Nvidia infrastructure reduces computation time “by as much as 10 times — compressing a 20-minute problem down to under two minutes, and a week-long calculation to roughly two days.”
CNBC, working from the same briefing on the same day, wrote something different: “Where choosing the best routing for a package may have previously taken 20 minutes, OneRail said its platform can do it in 2½ minutes.”
Twenty into two is a factor of ten. Twenty into two and a half is a factor of eight. Both numbers came from OneRail on 1 September 2026, describing the same operation. The gap is not large in practical terms — nobody’s fulfilment strategy changes between eight and ten — but it shows the headline multiple is a rounded ceiling rather than a measured result.
The week-long calculation that only improves 3.5 times
The second example in the press release is the one worth pausing on. A week-long calculation compressed to “roughly two days” is a speedup of about 3.5 times, assuming a seven-day week. That is a genuinely useful improvement. It is also barely a third of the headline number, published in the same sentence as the headline number.
None of this makes the product slower than advertised. It means the honest summary of the OneRail claim is “between roughly 3.5 and 10 times faster depending on the problem,” and only the top of that range made it into the headlines.
Nvidia cuOpt Is Free, and That Changes Where OneRail's Advantage Sits
This is the part most of the coverage skipped, and it is the most important thing to understand about the announcement.
What cuOpt does and what it costs
Nvidia cuOpt is an open-source, GPU-accelerated engine for decision optimisation. It is released under the Apache 2.0 licence and distributed through GitHub, pip, conda, Docker and the NGC catalog. Nvidia describes it as “free for developers to unlock real-time optimization at an unprecedented scale,” and open-sourced it in collaboration with the COIN-OR Foundation.
It solves vehicle routing problems including TSP, VRP and pickup-and-delivery, alongside linear and quadratic programming, with mixed-integer programming still in beta. It needs CUDA 12 or 13 and a Volta-generation GPU or newer. Nvidia’s own documentation notes it can run a 10,000-location problem in about 30 seconds on a single A100.
| Layer | Who supplies it | Is it exclusive to OneRail? |
|---|---|---|
| cuOpt optimisation engine | Nvidia, Apache 2.0 | No — free to anyone |
| cuDF data processing | Nvidia, open source | No — free to anyone |
| GPU infrastructure | Nvidia hardware | No — purchasable or rentable |
| cuOpt engineering access | Nvidia Inception | Partly — relationship-dependent |
| Delivery pricing and performance dataset | OneRail | Yes |
| Carrier and courier network | OneRail | Yes |
Why the data, not the GPU, is the moat
Read the table and the strategy becomes obvious. Four of the six layers in the OneRail stack are available to any competitor with a credit card and a CUDA-capable GPU. The two that are not are the dataset and the network.
To OneRail’s credit, the company says this itself. Catania’s quote pairs the Nvidia components with “the proprietary data and operational models we’ve built over millions of deliveries,” and Daeschler’s analogy to CNBC — that OneRail is doing “for delivery what ChatGPT and Anthropic have done for words,” training “on data that we have, just like words on the internet” — is a data-moat argument, not a hardware one.
The comparison is loose, since natural language processing and constrained vehicle routing are not the same discipline and cuOpt is a heuristic solver rather than a language model. The underlying point survives the looseness: the corpus is the asset, and OneRail is the only party that owns this one.
The trade coverage inverted this. Headlines led with Nvidia, because Nvidia is the name that draws clicks. The defensible asset is the delivery history, and any buyer evaluating this space should be interrogating dataset coverage and carrier depth, not GPU model numbers.
Testing the Word "First" Against Nvidia's Own Customer List
The press release headline reads: “OneRail Launches OmniSTAR, the First AI-Powered Delivery Decisioning Platform Built with NVIDIA AI.” That is a strong word, and Nvidia publishes enough of its own material to test it.
Domino’s, clicOH and Blue Yonder got there earlier
Nvidia’s cuOpt product page names Blue Yonder as a partner whose platform “optimizes last-mile delivery, enabling thousands of deliveries daily across hundreds of vehicles.” It names Domino’s Pizza for vehicle routing with sub-second runtimes — a collaboration Nvidia presented publicly at GTC in autumn 2021, five years before this launch. And Nvidia’s developer blog published a spotlight in August 2024 on clicOH, which reported a 20x speedup in cluster route planning and a 15% reduction in overall operating costs using cuOpt for last-mile delivery.
| Organisation | cuOpt application | Published figure | When |
|---|---|---|---|
| Domino’s Pizza | Real-time vehicle routing | Sub-second runtimes | GTC, autumn 2021 |
| clicOH | Last-mile route planning | 20x; 15% lower operating cost | Aug 2024 |
| Blue Yonder | Last-mile delivery at scale | Thousands of deliveries daily | Listed by Nvidia |
| Lowe’s | Supply chain re-optimisation | Not stated | Listed by Nvidia |
| OneRail | Delivery mode decisioning | Up to 10x | Sep 2026 |
What the claim survives on
The “first” claim survives only on the exact phrase “delivery decisioning platform” — a category OneRail has effectively defined for itself. There is a real distinction underneath the marketing: Domino’s and clicOH are optimising routes for a fleet they control, while OmniSTAR chooses between fulfilment modes, including third-party couriers and parcel carriers, before routing is even relevant. That is a different problem, and framing it as a separate category is defensible.
What is not defensible is reading the headline as “nobody has used Nvidia AI for last-mile delivery before.” Nvidia has been publishing last-mile cuOpt case studies since at least 2024.
The Two Deployments OneRail Has Disclosed
A launch is only as strong as its deployments, and this one discloses exactly two. Both are worth examining, because neither is quite the customer the pitch describes.
US Foods is a $39.4 billion foodservice distributor, not a retailer
The press release names US Foods and describes a specific finding: OmniSTAR “identified delivery configurations that were eroding margins, including shipping low-margin items long distances on high-cost equipment,” after which the company adjusted pricing and restructured delivery patterns.
That is a credible and unglamorous result — the kind of finding that suggests real deployment rather than a demo. US Foods reported $39.4 billion in net sales for fiscal 2025, up 4.1%, and runs more than 6,500 trucks from 76 distribution facilities to roughly 250,000 customer locations. An operator at that scale finding margin leakage in its own delivery configuration is a meaningful reference.
But note what it is not. US Foods is a foodservice distributor supplying restaurants, hospitals and hotels. It is not a retailer competing with Amazon for consumer same-day delivery, which is how CNBC framed the platform’s purpose.
The unnamed tire distributor and the $40 million
CNBC reports that the platform “has already been deployed with some customers, including a large tire distributor that saw OmniStar save the company a run rate of $40 million over three years.”
Two things are worth flagging. First, the phrasing mixes metrics: a “run rate” is an annualised figure, while “over three years” is cumulative. Those are different numbers, and the sentence does not resolve which one is meant.
Second, the customer is not hard to narrow down. OneRail’s own newsroom carries two entries about American Tire Distributors — a partnership announcement for final-mile automation and a FreightWaves feature on ATD adopting the OneRail platform for on-demand deliveries. The ATD relationship dates to November 2020, when OneRail expanded ATD Express nationally after a phase-one run in 19 of ATD’s 115 markets, delivering tyres within 90 minutes of order.
If that is the same customer — and it is the obvious candidate on the public record — then the three-year savings window opened years before OmniSTAR launched on 1 September 2026. The savings would then belong to the OneRail relationship broadly, not to the platform announced this month. That is an inference from the vendor’s own newsroom rather than a stated fact, and OneRail has not confirmed the identity.
The Network Behind OneRail's Model: 2020 Versus 2026
The dataset is the moat, so its growth is the number that matters most. Usefully, OneRail has published network figures at two points six years apart, and the comparison is stark.
Drivers, partners and what “proprietary data” means here
| Measure | Nov 2020 (ATD release) | Sep 2026 (OmniSTAR release) | Change |
|---|---|---|---|
| Drivers in network | 4.5 million | Over 12 million | Roughly 2.7x |
| Logistics partners | 100+ delivery companies | Over 1,000 | Roughly 10x |
| Founded | 2018, Orlando, by Bill and Lisa Catania | — | |
| Total funding raised | $109m, including a $42m Series C in Dec 2024 | — | |
A tenfold increase in logistics partners is the figure that supports the product thesis. Mode selection is only valuable if there are genuinely different modes to select between; a decisioning engine choosing among 100 carriers is a much weaker proposition than one choosing among 1,000. On this measure the six-year build-out is more impressive than the speed claim.
The 12 million driver figure deserves a lighter touch. It counts drivers reachable through partner networks, not drivers employed or contracted by OneRail, and courier rosters overlap heavily. It is a reach number, not a capacity number.
What OneRail Is Really Selling to Retailers
Strip out the launch language and the proposition is more specific, and more interesting, than the general case for integrating AI in logistics usually allows.
Decisioning versus routing
Most delivery software answers “what is the best route for this set of stops?” OmniSTAR answers an earlier question: “which fulfilment mode should carry this order at all?” Owned fleet, contracted courier, parcel carrier — each with different cost curves, service guarantees and failure modes, evaluated per order against live market conditions. Modes such as out-of-home delivery shift those economics again.
That earlier question is where the margin actually sits. A perfectly routed van is still the wrong answer if a parcel carrier would have met the promise for a third of the cost. The press release makes this point well, noting that OmniSTAR surfaces “internal rules and constraints that drive unnecessary cost, such as delivery commitments or routing logic that conflict with margin goals.” The US Foods finding is exactly that pattern.
Where 2.5 minutes is fast enough, and where it is not
Two and a half minutes is genuinely fast for a mode-selection decision made against live pricing across a thousand carriers. It is not fast in the sense Domino’s means when it reports sub-second routing, and the two should not be compared — different problems, different scopes, different stakes.
The practical test is whether the decision window fits the operational one. For an order that will be picked, packed and dispatched within the hour, a two-minute decision is invisible. For a checkout page promising a delivery slot while the customer waits, it is not. OneRail’s framing — a decision layer behind fulfilment, not in front of the customer — suggests the former, and the distinction matters for anyone scoping intelligent automation around it.
A Claim-by-Claim Ledger for the OneRail Launch
| Claim | Verdict | Basis |
|---|---|---|
| Built on Nvidia cuOpt and cuDF | Supported | Named in the CEO’s own quote; Nvidia supplies a partner quote |
| Three-year Nvidia collaboration | Supported | Daeschler to CNBC; “multi-year” in the release |
| Up to 10x faster computation | Partly supported | One of three published examples reaches 10x; another is 3.5x |
| 20-minute problem in 2 minutes | Contested | CNBC published 2.5 minutes the same day |
| “First” delivery decisioning platform on Nvidia AI | Narrow | True only for that self-defined category; Nvidia lists earlier last-mile users |
| 12m drivers, 1,000+ partners | Vendor-stated | Consistent across release and CNBC; reach not capacity |
| US Foods deployment | Supported | Named in the release with a specific, unglamorous finding |
| $40m run rate over three years | Unverifiable | Customer unnamed; mixes annualised and cumulative metrics |
| Over $6bn GMV in Q4 | Forecast | Company estimate; no methodology published |
| Lets smaller retailers match Amazon | Unproven | Both disclosed deployments are large B2B distributors |
What This Means If You Run Delivery Operations
The useful conclusion is not “OneRail overstated its launch.” Most launches do, and this one is more restrained than most — a 10x claim in a field where Nvidia’s partners have published 240x is not an aggressive number.
Questions to ask before a pilot
The conclusion is that the interesting due diligence is not about Nvidia at all. cuOpt is free, and a capable in-house team with GPU access can run it. What cannot be replicated is six years of delivery pricing and performance data across a thousand carriers. So the questions worth asking are about coverage, not silicon.
Ask which carriers are live in your specific postcodes and product categories, because a thousand-partner network averages out very differently region by region. Ask what the decision latency is against your own order volume rather than a reference benchmark. Ask whether the reference customers resemble your operation — as of this launch, the public ones are a foodservice giant and a tyre distributor, both B2B, neither an e-commerce retailer. And ask how the platform behaves when its recommendation conflicts with a delivery promise already made at checkout, since the release itself flags that surfacing those conflicts is part of the product.
Those are the questions that separate a genuine predictive analytics capability from a fast solver with a good story attached. The technology here is real and the direction of travel is right. The claims just deserve reading in their original form rather than in the version that reached the headlines.
References and Further Reading
CNBC: OneRail launches AI platform with Nvidia for retailers to make faster delivery decisions
AI News: OneRail uses Nvidia AI for real-time last-mile delivery optimisation
Engineering.com: OneRail launches AI-powered delivery decisioning platform
Nvidia: cuOpt GPU Optimization Engine
Nvidia: NVIDIA Open-Sources cuOpt, Ushering in New Era of Decision Optimization
Nvidia Developer Blog: clicOH Accelerates Last-Mile Delivery 20x with NVIDIA cuOpt
OneRail: OneRail Partners With ATD for Final Mile Automation
Business Wire: OneRail Secures $42 Million Series C Investment
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.