Nvidia is using Palantir Foundry and its own cuOpt optimization library to automate weekly hardware-allocation decisions across a global manufacturing network. The system combines mathematical planning with a customized language model intended to capture operational information that does not fit neatly into inventory tables.
Nvidia measures delivery from “wafer-out to first token.” That interval includes the journey from semiconductor fabrication to an assembled rack, followed by the power, cooling, networking and software work required before the system can produce useful computation.
Component Scale Turns Allocation Into a Systems Problem
A Grace Blackwell NVL72 rack contains 18 compute trays. Each tray requires two Grace CPUs, four Blackwell GPUs and 32 HBM3e memory packages, with inputs flowing through thousands of suppliers, equipment manufacturers and contract-design partners. Nvidia says the supply chain planned for its Vera Rubin architecture will be twice as large as the network supporting Grace Blackwell.
Assembly depends on parts arriving through three channels: Nvidia’s direct inventory, consignment stock and external suppliers. A missing component can leave earlier deliveries waiting and extend what the company calls Time of Ownership, the period between a facility receiving materials and finished subassemblies leaving it.
Operations staff recalculate factory allocations every week across a rolling two-quarter horizon. They must balance part availability, production throughput and customer schedules. A locally efficient decision may still delay an entire rack when another tier of the bill of materials cannot support it.
Foundry and cuOpt Build the Weekly Schedule
Nvidia created a Digital Supply Chain Intelligence command centre in Palantir Foundry. Foundry’s Ontology represents facilities, supplier commitments, inventories and output targets as connected operational objects. CuOpt reads that layer and formulates distribution as a mixed-integer linear program designed to minimize Time of Ownership.
The solver produces delivery schedules and identifies active constraints, such as a regional assembly limit becoming more important than available memory stock. That makes the result more than a ranked recommendation: it identifies which constraint currently controls achievable output.
Structured optimization could not capture every factor human planners use. Supplier calls, partner emails, weather forecasts and geopolitical events arrive as unstructured information. Nvidia therefore post-trained Nemotron 3.5 Lightning, an open-weight mixture-of-experts model with 30 billion total parameters and about 3 billion active parameters per forward pass.
The training pipeline anonymizes sensitive operational fields, balances examples with synthetic disruption scenarios and applies low-rank adaptation while leaving the base model frozen. Palantir Autopilot manages data lineage, model tracking and delivery of recommendations.
Historical Tests Show Gains and Limits
On historical allocation records, Nvidia says the tuned Lightning model reached 86.7% decision accuracy. The larger Nemotron 3 Ultra achieved 55.5%, while the untuned Lightning base model reached 17.5%. Balanced accuracy was 58.6% and macro-F1 was 57.5%, compared with 42% and 39.5% for Ultra.
Those results are company-reported and depend on Nvidia’s historical dataset, labels and decision definition. They do not establish performance at another manufacturer or under an unseen disruption. The source also notes that forecasting production risks further into the future remained difficult.
Fine-tuning reportedly completed within minutes on two Nvidia B200 GPUs. Planner edits, overrides and factory outcomes are now written back into the Palantir Ontology. Nvidia plans to turn that record into preference pairs for reinforcement learning, scoring allocation accuracy, policy compliance and evidentiary grounding.
Production models will remain isolated from live, unmonitored retraining. That boundary matters because operational feedback can contain errors as well as expertise. Nvidia’s system demonstrates a concrete path from optimization to AI-assisted allocation, while its own metrics show that accuracy alone does not remove uncertainty, human overrides or the challenge of looking further ahead.
