NVIDIA and Palantir Technologies have created a unified supply chain command center that automates material allocation decisions by combining quantitative optimization with codified human expertise (Nvidia blog). NVIDIA's supply chain spans millions of parts and thousands of suppliers globally, with performance measured from wafer-out to first token. The critical material allocation problem—deciding what components to allocate to each manufacturing site—was previously reworked manually every week. The system now integrates Palantir Foundry's Ontology (a governed data layer connecting materials, manufacturing sites, capacity, and production outputs), NVIDIA cuOpt (a GPU-accelerated optimization solver), and a custom Nemotron 3.5 Lightning model trained on historical allocation decisions (Nvidia blog).
The key innovation is how the system captures human expertise. Back-testing revealed that human planners made better decisions than pure mathematical optimization because they integrated qualitative signals—emails from partners, geopolitical events, weather forecasts, and years of accumulated knowledge—that the solver could not see (Nvidia blog). NVIDIA and Palantir captured each allocation decision, its rationale, expected result, and actual outcome, then used this institutional knowledge to post-train Nemotron 3.5 Lightning, a 30-billion-parameter open-weight model with a mixture-of-experts architecture that enables efficient inference (Nvidia blog). For commerce practitioners, this approach demonstrates how to build AI systems that augment rather than replace expert judgment: the model recommends an allocation range with reasoning, and a human planner makes the final decision. The training pipeline includes anonymization, synthetic data expansion, supervised fine-tuning of LoRA adapters, and point-in-time backtesting before deployment, all managed by Palantir Autopilot (Nvidia blog).
The open-weight nature of Nemotron allows organizations to post-train the model within their own compute boundary without exposing operational data externally, making this architecture applicable to other complex supply chain and procurement workflows (Nvidia blog).