NVIDIA and Palantir Technologies have deployed an AI stack to support supply chain operations. The system combines NVIDIA’s Nemotron open models with Palantir Foundry and its Ontology data layer.
It will help NVIDIA identify supply constraints, assess allocation choices, and capture operational expertise. The deployment starts within NVIDIA’s own supply chain, which spans millions of parts, thousands of suppliers, and manufacturing partners worldwide.
The companies aim to shorten the time from wafer production to a completed AI system reaching production. Each NVIDIA Vera Rubin rack contains about 1.3 million parts, according to the companies’ statement.
NVIDIA said the system creates a shared view of inputs involved in material allocation decisions. It connects materials, manufacturing sites, production capacity, allocations, and other operational information through Palantir’s Ontology.
“Supply chains are the operating system of the physical economy,” NVIDIA CEO Jensen Huang said in a statement. He added that the companies are combining software to “reason, plan and orchestrate” the journey from wafer to token.
The system also uses NVIDIA cuOpt for allocation and scenario planning. cuOpt models the supply chain as a mixed-integer linear programme and aims to minimise Time of Ownership. It can also show which constraints limit a particular allocation decision.
NVIDIA and Palantir found that quantitative optimisation alone could not capture all factors affecting allocation decisions. Planners also use emails, weather risks, geopolitical events, and supplier discussions.
The companies therefore built a workflow that records each allocation decision, its rationale, expected result, and actual outcome. This information becomes training data for a customised language model.
NVIDIA post-trained its Nemotron 3.5 Lightning model to recommend allocation ranges and explain the factors behind them. Supply chain planners still review the recommendations and make the final decisions.
On the development benchmark, the post-trained model achieved 86.7% allocation-decision accuracy. Nemotron 3 Ultra scored 55.5%, while the base Lightning model scored 17.5%.
The post-trained model also recorded 58.6% balanced accuracy and 57.5% Macro-F1. NVIDIA said the results show the value of specialising a smaller model for a defined decision-making task.
The system keeps proprietary data, model weights, and inference within a governed environment. Organisations can deploy the stack on-premises or in the cloud, depending on their requirements.
Palantir CEO Alex Karp said the collaboration would bring sovereign AI capabilities to NVIDIA’s supply chain. He described the partnership as a cornerstone of the companies’ sovereign AI strategy.
Palantir customers can also post-train Nemotron models using their own operational data. The companies said this approach can support supply chain operations across manufacturing, energy, healthcare, automotive, and aerospace.
The companies plan to use feedback for future reinforcement learning, while keeping retraining under governance.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.