From Wafer-Out to First Token: Codifying Supply Chain Expertise with Nemotron and Palantir Foundry

NVIDIA and Palantir Codify Supply Chain Expertise with Specialized AI

NVIDIA has developed a specialized AI model on Palantir Foundry to manage its intricate global supply chain for systems like the Grace Blackwell NVL72. The system captures the complex decision-making of human planners and codifies it into a specialized Nemotron model, creating a continuous learning loop. This approach addresses the critical material allocation problem, a weekly combinatorial challenge that quantitative solvers alone could not fully optimize, by incorporating qualitative signals that were previously siloed in emails and operational debriefs.

From Optimization to Augmentation

The initiative began with the goal of minimizing "Time of Ownership" (TOO)—the period components are held at a manufacturing site. While NVIDIA cuOpt provided a powerful quantitative baseline by solving the weekly mixed-integer linear program, analysis revealed that human planners consistently outperformed the solver. They integrated unstructured data like geopolitical forecasts and supplier communications, which the model couldn't see. To capture this expertise, NVIDIA and Palantir built a Digital Supply Chain Intelligence command center where all decisions, rationales, and outcomes are recorded in Foundry's Ontology, forming a rich training dataset.

• Base Model: Nemotron 3.5 Lightning (30B parameters).
• Post-Training: Fine-tuned on expert decisions using NeMo tools within a governed Palantir Autopilot lifecycle.
• Performance Uplift: The specialized model reached 86.7% allocation-decision accuracy.
• Comparative Accuracy: It outperformed its own base model by 69.2 percentage points and the larger Nemotron 3 Ultra by 31.2 points on the development benchmark.

An Enterprise Flywheel for Institutional Knowledge

The resulting workflow presents planners with an AI-generated recommendation, complete with its reasoning and identified risks. Planners can accept, edit, or override this suggestion, and every interaction is fed back into the Ontology. This closed-loop process enables future governed retraining runs, compounding institutional knowledge and shortening the learning curve for new team members. The project demonstrates how smaller, open-weight models like Nemotron can be efficiently adapted for high-value, agentic enterprise workflows without exposing sensitive operational data to external services.

Strategic Takeaway: This collaboration proves that the most significant operational gains come not from replacing human experts with AI, but from building systems that explicitly learn from the nuanced, qualitative judgments that make those experts valuable in the first place.