Palantir Foundry and cuOpt drive NVIDIA supply chain allocation
ainews ·
NVIDIA is using Palantir Foundry and the open-source NVIDIA cuOpt library to automate hardware supply chain allocation decisions across its global manufacturing sites. The company tracks operational delivery from wafer-out to first token, a window it splits into time-to-rack (fab output to assembled data centre system) and time-to-token (power, cooling, networking, and day-one software readiness). The complexity is significant: an NVIDIA Grace Blackwell NVL72 rack holds 18 compute trays, each needing two Grace CPUs, four Blackwell GPUs, and 32 HBM3e memory packages drawn from thousands of suppliers, OEMs, and contract design partners, while the supply chain planned for the Vera Rubin architecture is twice as large. Assembly depends on parts arriving via direct inventory, consignment stock, and external suppliers, and allocations are reworked weekly over rolling two-quarter horizons. To coordinate this, NVIDIA's operations team built a 'Digital Supply Chain Intelligence' command centre on Palantir Foundry, whose Ontology models facilities, supplier commits, component stocks, and production targets, with cuOpt formulating distribution as a mixed-integer linear program that minimises Time of Ownership and surfaces constraints such as regional assembly capacity caps. Because mathematical optimisation alone could not capture unstructured variables like supplier call transcripts, weather forecasts, partner emails, and geopolitical events, NVIDIA post-trained Nemotron 3.5 Lightning, an open-weight mixture-of-experts model with 30 billion total parameters and roughly three billion active per forward pass. The pipeline uses NeMo Anonymizer, NeMo Data Designer, and NeMo AutoModel with LoRA, while Palantir Autopilot handles data lineage and model tracking. Evaluated on historical allocation records, the post-trained model reached 86.7 percent decision accuracy, 58.6 percent balanced accuracy, and a 57.5 percent macro-F1 score, outperforming Nemotron 3 Ultra and the untuned base model, with fine-tuning completed on two B200 GPUs in minutes. Operational choices and overrides are continuously written back to the Ontology to form preference pairs for reinforcement learning, with production models kept isolated from live retraining.
ainews