Palantir Foundry and cuOpt drive NVIDIA supply chain allocation

Details

NVIDIA's operations team is managing a hardware supply chain of unprecedented complexity: a single Grace Blackwell NVL72 rack contains 18 compute trays, each requiring 2 Grace CPUs, 4 Blackwell GPUs, and 32 HBM3e memory packages sourced across thousands of suppliers, OEMs, and contract design partners. The coming Vera Rubin architecture will require a supply chain network twice that size. To keep pace, NVIDIA built a command centre called 'Digital Supply Chain Intelligence' using Palantir Foundry as the operational data backbone. Foundry's Ontology layer models facilities, supplier commits, component stocks, and production targets as interconnected objects. NVIDIA's open-source cuOpt library — GPU-accelerated and designed for logistics optimization — reads this layer directly and formulates the allocation problem as a mixed-integer linear program. The objective is minimizing 'Time of Ownership' (TOO): the duration from when a factory receives raw materials to when finished sub-assemblies ship. Factory allocations are recalculated weekly across rolling two-quarter planning horizons, identifying active bottlenecks like regional assembly capacity versus raw memory availability. Mathematical optimization alone couldn't handle soft operational variables — supplier call transcripts, regional weather forecasts, partner emails, geopolitical developments. NVIDIA addressed this by post-training Nemotron 3.5 Lightning on these qualitative records, creating a hybrid system that combines formal MILP optimization with LLM-based soft constraint reasoning. The architecture is a practical blueprint for enterprises facing supply chain problems too complex for pure optimization but too structured for pure LLM reasoning.