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

NVIDIA and Palantir collaborated on a Digital Supply Chain Intelligence command center built on Palantir Foundry to unify material, capacity and qualitative signals used in critical material allocation decisions. The Ontology in Foundry connects materials, manufacturing sites, commits, capacity, allocations, production outputs and unstructured qualitative signals into a governed data layer.

On the quantitative side, NVIDIA cuOpt solves a weekly mixed-integer linear program that minimizes Time of Ownership across manufacturing sites and reports which constraints are binding each period. NVIDIA states that human planners consistently outperformed the quantitative model by incorporating emails, weather forecasts, geopolitical events and supplier debriefs that the solver could not see.

NVIDIA then post-trained Nemotron 3.5 Lightning on captured allocation decisions, rationales and outcomes, using NeMo Anonymizer, NeMo Data Designer and NeMo AutoModel within a Palantir Autopilot lifecycle. The model is a 30-billion-parameter mixture-of-experts model with roughly 3 billion active parameters per forward pass. NVIDIA says training used LoRA adapters on 2x NVIDIA B200 GPUs.

On the development benchmark, the post-trained Nemotron 3.5 Lightning reached 86.7% allocation-decision accuracy, which NVIDIA says leads Nemotron 3 Ultra by 31.2 percentage points and its own base model by 69.2 points. Ultra reached 55.5% and base Lightning 17.5%. NVIDIA reports balanced accuracy of 58.6% versus Ultra's 42.0%, and macro-F1 of 57.5% versus 39.5%.

NVIDIA states that future production risk forecasting remained difficult despite fine-tuning, and that the smaller model's gains are concentrated in the domain it was post-trained on. Accepted, edited and overridden recommendations feed back into the Ontology to enable future governed retraining runs. NVIDIA says future feedback will be used for reinforcement learning, and that the model never retrains itself in production.