NVIDIA and Palantir Launch Sovereign AI Stack for Critical Supply Chains

NVIDIA and Palantir have introduced a jointly developed sovereign AI stack for complex, critical supply chains. NVIDIA is deploying it first in its own operations—effectively serving as the product’s initial real-world customer—before the companies offer the architecture to other organizations. 11

The central idea is not simply to add a chatbot to supply-chain data. It is to combine operational data, optimization software, AI models and expert planner judgment in a governed environment that an organization controls.

What NVIDIA and Palantir announced

The stack brings NVIDIA’s open Nemotron models into Palantir Foundry and the Palantir Artificial Intelligence Platform (AIP), with the Palantir Ontology providing the operational model underneath. NVIDIA and Palantir said the system is designed to help organizations identify constraints, codify operational expertise and guide supply-chain decisions. 11

For NVIDIA, the first focus is accelerating the path from semiconductor wafer to first token—the point at which an AI system is deployed and producing output. The announcement describes this as a way to codify operational intelligence inside NVIDIA’s supply chain. 11

Why NVIDIA’s supply chain is the first test case

The supply chain behind NVIDIA’s Vera Rubin systems illustrates the scale of the challenge. A single rack involves roughly 1.3 million components, spanning compute, memory, networking, power, cooling and mechanical systems. 37

That level of interdependence makes material allocation more than a spreadsheet problem. A plan can be constrained by manufacturing capacity, component availability, inbound timing or commitments already made to customers. NVIDIA is using Palantir’s Ontology and its supply-chain Command Center alongside NVIDIA cuOpt to evaluate scenarios and allocate scarce materials across sites. 16

NVIDIA describes its decision objective as reducing Time of Ownership: the time materials remain in NVIDIA’s control before they can contribute to delivered systems. In practice, the system is intended to identify which constraint is limiting throughput and which allocation choice best addresses it. 1

How the AI stack works

The components have distinct roles:

• Nemotron open models provide the reasoning layer that can be specialized for supply-chain decisions.
• Palantir Foundry and AIP provide the data and application environment in which teams can work with models and operational workflows.
• Palantir Ontology represents the important real-world entities and relationships: materials, sites, capacity, allocations, commitments and production outcomes. 116
• NVIDIA cuOpt supports optimization and scenario analysis for decisions such as material allocation. 1

Together, the companies position the stack as a governed operating environment rather than a general-purpose model connected to isolated datasets. NVIDIA’s developer account emphasizes that proprietary supply-chain data, model weights and inference remain within a single governed environment. 1

Why planner judgment was part of the model training

NVIDIA found that mathematical optimization alone did not capture all the context that experienced planners use. Decisions may depend on unstructured and fast-changing information, such as emails, weather forecasts, geopolitical developments, supplier debriefs and institutional knowledge. 1

NVIDIA used planners’ allocation choices, their rationales and later outcomes as post-training data for a specialized Nemotron 3.5 Lightning model. The model has 30 billion parameters in total and uses a mixture-of-experts design with roughly 3 billion active parameters per forward pass. 1

On NVIDIA’s development allocation-decision benchmark, the company reported the following results:

Model	Reported allocation-decision accuracy
Post-trained Nemotron 3.5 Lightning	86.7%
Nemotron 3 Ultra	55.5%
Base Nemotron 3.5 Lightning	17.5%

These are NVIDIA-reported results on its own development benchmark, not an independent measure of performance across every supply chain or deployment. 1

Why people still review the recommendations

The companies are not presenting the system as an autonomous replacement for supply-chain planners. Allocation decisions can involve consequential tradeoffs, and new information may emerge that was not represented in prior training data.

Human planners therefore remain responsible for reviewing, adjusting and improving recommendations. Those interventions can become new governed decision records, creating a feedback loop for continued post-training and operational refinement. 1

How the stack will be deployed

NVIDIA and Palantir said other organizations can use the architecture for their own supply chains through the Palantir Sovereign AI Operating System Reference Architecture, deployed on cloud or on-premises infrastructure. The stated focus includes complex industries and government environments where organizations want to retain control of proprietary data. 11

The related reference architecture is built on NVIDIA Enterprise Reference Architectures and is designed to run Palantir software including AIP, Foundry, Apollo, Rubix and AIP Hub. 29

Dell is part of the surrounding infrastructure ecosystem: its related architecture combines Dell AI Factory infrastructure, including PowerEdge, ObjectScale and PowerFlex, with Palantir and NVIDIA technologies. 17 Reporting on the September announcement also named Cisco, Rackspace and Nebius as partners for hardware or cloud deployment options. 25

Where it was showcased

The supply-chain deployment and wider sovereign-AI applications were showcased at Palantir’s AIPCon 11 event. 6

The significance of the announcement is that NVIDIA is using the combined stack on its own highly constrained supply chain first. That makes the deployment a practical test of whether domain-specific models, optimization and governed operational data can improve decisions in one of the most complex industrial planning environments. 111