NVIDIA is using artificial intelligence and optimization technology to automate complex supply chain allocation decisions across its global hardware manufacturing network.
The company has combined Palantir Foundry with NVIDIA cuOpt to coordinate the movement of components needed to build increasingly complex AI systems. The goal is to reduce the time required to move hardware from semiconductor fabrication through final assembly and ultimately make systems ready to run workloads.
Tracking Hardware From Factory to Data Center
NVIDIA measures its operational performance using a metric that tracks the journey from wafer production to the point when a system can process its first token.
This process includes two major stages. The first is time-to-rack, which measures the time required to move components from fabrication facilities through assembly and into a completed data center system. The second is time-to-token, which also accounts for requirements such as power, cooling, networking, and software readiness.
As NVIDIA’s hardware becomes more sophisticated, managing these dependencies has become increasingly difficult.
Coordinating Thousands of Components
NVIDIA’s Grace Blackwell NVL72 systems contain 18 compute trays. Each tray requires two Grace CPUs, four Blackwell GPUs, and 32 HBM3e memory packages.
Those components come from a massive network involving thousands of suppliers, original equipment manufacturers, and contract design partners.
The supply chain supporting NVIDIA’s upcoming Vera Rubin architecture is expected to be even larger, roughly twice the size of the network supporting Grace Blackwell.
Manufacturing cannot begin until the necessary parts arrive through one of three channels: direct inventory, consignment stock, or external suppliers. A shortage of even one component can delay an otherwise ready shipment.
NVIDIA refers to the resulting period as “Time of Ownership,” which measures the time between a facility receiving materials and completed sub-assemblies leaving the facility.
Factory allocations are therefore repeatedly adjusted across rolling two-quarter planning periods to account for component availability, manufacturing capacity, and customer demand.
Foundry Connects the Supply Chain
To manage these relationships, NVIDIA built what it calls the Digital Supply Chain Intelligence command center using Palantir Foundry.
Foundry’s Ontology provides a way to represent facilities, supplier commitments, component inventories, and production targets as interconnected objects.
This gives NVIDIA’s operations teams a unified view of the relationships between different parts of the supply chain rather than treating each constraint independently.
NVIDIA cuOpt can then use this operational information to determine how hardware should be allocated.
Using Optimization to Make Allocation Decisions
cuOpt is an open-source library designed to accelerate decision optimization using NVIDIA GPUs.
For NVIDIA’s supply chain, the system formulates allocation as a mixed-integer linear programming problem. The objective is to minimize Time of Ownership while accounting for constraints throughout the bill of materials.
The system does more than generate weekly delivery schedules. It can also identify the constraints preventing additional production, such as regional assembly capacity limits or shortages of memory components.
This allows NVIDIA to determine whether a delay is being caused by a manufacturing bottleneck, a shortage of materials, or another limitation elsewhere in the supply chain.
AI Handles Information Traditional Optimization Misses
Mathematical optimization alone cannot account for every variable that human planners encounter.
Important information can exist in unstructured sources such as supplier conversations, email exchanges, weather forecasts, and geopolitical developments. These factors can affect whether a particular component will arrive as expected even when the underlying numerical supply-chain data appears healthy.
NVIDIA addressed this problem by post-training its Nemotron 3.5 Lightning model specifically for these operational decisions.
The model is an open-weight mixture-of-experts system with 30 billion total parameters, although approximately 3 billion parameters are active during each forward pass.
NVIDIA’s training pipeline processes historical operational information while protecting sensitive data. It uses NeMo Anonymizer to redact sensitive fields, NeMo Data Designer to create balanced training examples and simulated supply disruptions, and NeMo AutoModel to apply low-rank adaptation while leaving the underlying model weights frozen.
Palantir Autopilot is then used to manage data lineage, model tracking, and the delivery of recommendations.
Smaller Model Outperforms Larger Model
The results suggest that domain-specific training can be more important than simply using a larger AI model.
When evaluated against historical allocation decisions, the post-trained Nemotron 3.5 Lightning model reached an 86.7 percent decision accuracy.
That compared with 55.5 percent for the larger Nemotron 3 Ultra model and just 17.5 percent for the untuned Lightning model.
The post-trained model also recorded a 58.6 percent balanced accuracy and a 57.5 percent macro-F1 score. Nemotron 3 Ultra recorded 42 percent balanced accuracy and a 39.5 percent macro-F1 score.
The model was fine-tuned using two NVIDIA B200 GPUs in just minutes, demonstrating how relatively small amounts of specialized training can significantly improve performance on a narrowly defined operational problem.
Building Toward Reinforcement Learning
NVIDIA is also creating a feedback loop around its supply chain decisions.
Operational decisions, planner changes, overrides, and actual factory outcomes are continuously recorded in the Palantir Ontology.
That information can eventually be used to create preference pairs for reinforcement learning. Recommendations can be evaluated based on factors such as allocation accuracy, compliance with operational policies, and the quality of the evidence supporting each recommendation.
NVIDIA says production models will remain isolated from live, unmonitored retraining.
The broader approach demonstrates how AI can move beyond generating text or analyzing documents and instead become part of the decision-making infrastructure behind large-scale manufacturing operations. For a company producing increasingly complex AI hardware, automating the allocation of thousands of interconnected components could become just as important as improving the chips themselves.
