NVIDIA uses AI to improve complex chip supply allocations

NVIDIA is using Palantir Foundry, GPU-accelerated optimization and an open-weight language model to help allocate scarce components across its manufacturing network. The work is intended to reduce the time needed to turn finished silicon into working data-centre infrastructure, according to a post on the NVIDIA Technical Blog at developer.nvidia.com.

NVIDIA measures that interval in two stages. The first, time-to-rack, covers the journey from a chip leaving a fabrication plant to a completed system reaching a data-centre floor. The second, time-to-token, begins once the hardware arrives and includes the power, cooling, networking and software needed before it can perform useful work. The project described in the post focuses mainly on shortening time-to-rack.

A supply chain with shifting bottlenecks

The scale of the challenge is considerable. NVIDIA says its Grace Blackwell NVL72 platforms depend on millions of components and thousands of suppliers worldwide. Dozens of original equipment manufacturers and original design manufacturers take part in assembling the final systems.

Even a single compute tray is highly complex. Each rack contains 18 trays, while one tray includes two Grace CPUs, four Blackwell GPUs and 32 HBM3e memory stacks. NVIDIA says the supply chain being developed for Vera Rubin is twice the size of the Grace Blackwell network.

Availability does not remain constant. CPUs, GPUs and memory can each become the limiting component at different times, with the part delaying production one week potentially easy to obtain the next. Every major component also has its own bill of materials, suppliers and delivery schedules. A manufacturing site cannot begin an assembly until all required items have arrived, whether those materials come directly from NVIDIA, from NVIDIA-held consignment stock or from external suppliers. Early arrivals may therefore remain idle while a missing item holds up the entire build.

NVIDIA tracks the time that materials remain under its control between arriving at a manufacturing location and leaving as part of a sub-assembly or finished product. The company calls this measure Time of Ownership, or TOO.

Material allocation is reviewed weekly and extends across the rest of the current quarter and into the following one. Near-term decisions are generally already committed, meaning new information has its greatest effect on later weeks. The task is to determine which manufacturing sites should receive limited materials, and in what quantities, while accounting for each site’s capabilities and capacity.

Turning operational data into decisions

NVIDIA’s supply-chain operations team worked with Palantir to assemble a shared view of the information behind those choices. The company refers to the result as its Digital Supply Chain Intelligence command centre. It brings together warnings, production obstacles and other signals that might previously have been scattered among separate systems.

Palantir Foundry supplies the underlying operating environment. Its Ontology links materials, sites, supplier commitments, production capacity, allocations and outputs. It also accommodates qualitative information, creating a governed representation of the relationships involved in the supply chain rather than treating each item as an isolated table entry.

That structure allows planners to test alternative futures. They can examine the consequences of receiving less memory, or assess how the network might change when another manufacturing site becomes available. The aim is to move beyond receiving a single calculated answer and instead understand the trade-offs around it.

NVIDIA cuOpt, described in the post as an open-source library for GPU-accelerated decision optimization, performs the allocation calculation. It takes data from the Ontology and returns an allocation recommendation. The problem is expressed as a mixed-integer linear program whose objective is to reduce TOO.

The software also identifies the constraints that are limiting the result. This can show planners whether a decision is being held back by a particular regional manufacturing capacity, memory availability or another factor. Because the calculation is fast, users can repeatedly adjust assumptions and investigate the effect of potential changes.

Capturing what experts know

Historical testing showed that numerical optimization alone could not reproduce the full basis of human decisions. Planners were also considering partner correspondence, weather risks, geopolitical developments, supplier-call transcripts and experience accumulated over years. Those details could influence how much material a site should receive even when they were not represented in the solver’s structured inputs.

NVIDIA and Palantir therefore designed the workflow to record not only an allocation, but also the reasoning behind it, the expected outcome and what eventually occurred. The company says this turns previously informal expertise into information that can be reviewed and used to train a model.

After assessing open Nemotron models, the teams selected Nemotron 3.5 Lightning for the execution part of the system. The model uses a mixture-of-experts design, has 30 billion parameters and activates about 3 billion during each forward pass. NVIDIA presents that relatively compact footprint as useful for a focused model that can be trained and deployed with less computing capacity than a larger general-purpose system.

The model is intended to estimate production risk, suggest an allocation range, identify the evidence behind its recommendation and explain the result to supply-chain staff. Since Nemotron is open, NVIDIA says organizations can post-train it within their own computing environment using internal operational data.

Past decisions also provide an evaluation method. The system can replay a historical choice using only information available at the time, conceal the eventual result and compare the model’s recommendation with the planner’s decision and the outcome. Palantir Autopilot manages the associated workflow, including jobs launched from Ontology data, monitoring of the customized model and lineage linking the source data to model versions and results.