NVIDIA Trains Nemotron 3.5 to Manage Its Supply Chain

NVIDIA partnered with Palantir to build a Digital Supply Chain Intelligence command center on Palantir Foundry, aiming to streamline its complex global supply chain for platforms like the Grace Blackwell NVL72. While NVIDIA cuOpt solved quantitative allocation problems by minimizing Time of Ownership, human planners still made better decisions by factoring in qualitative data like weather forecasts and emails. To capture this human expertise, NVIDIA post-trained its 30-billion-parameter Nemotron 3.5 Lightning model on past allocation decisions and their rationales.

The training pipeline utilized NeMo Anonymizer, Data Designer, and AutoModel within Palantir Autopilot. Running a Low-Rank Adaptation fine-tuning process on two NVIDIA B200 GPUs took only minutes. On the development benchmark, the specialized Nemotron 3.5 Lightning model achieved an impressive 86.7% allocation-decision accuracy. This outperformed the much larger Nemotron 3 Ultra by 31.2 percentage points, which scored 55.5%, and beat its own base model by 69.2 percentage points, which scored 17.5%.

The fine-tuned model also excelled in balanced accuracy, scoring 58.6% compared to Ultra's 42.0%, and achieved a macro-F1 score of 57.5% against Ultra's 39.5%. These metrics are crucial because supply cuts are more common than increases, meaning a naive model could easily skew results. For enterprise practitioners, this demonstrates that a lightweight model with roughly 3 billion active parameters per forward pass can easily beat general-purpose models when tailored to a specific domain.

The system operates as a continuous feedback loop. When planners accept, edit, or override the model's recommendations, the data writes back into the Palantir Foundry Ontology to power future retraining runs. This setup allows organizations to run sovereign AI entirely within their own secure compute boundaries, keeping proprietary supply-chain data safe while steadily compounding institutional knowledge.