Nvidia and Palantir Pair AI With Human Review for Supply-Chain Decisions

Nvidia and Palantir have partnered to deploy AI for supply-chain management, beginning with Nvidia’s network of millions of parts, thousands of suppliers and global partners. Their system pairs mathematical optimization with a model trained on past expert decisions, but Nvidia says a planner still reviews each recommendation and makes the final allocation call.

The first deployment is a difficult proving ground. Nvidia says a single Vera Rubin rack contains 1.3 million components, while its wider operation has to match changing material availability with manufacturing capacity, supplier commitments and customer obligations across multiple sites. A component that blocks production one week may not be the constraint the next.

One allocation problem, two kinds of input

On the quantitative side, Nvidia cuOpt solves a weekly mixed-integer linear program, a method for choosing among constrained options. Its objective is to minimize the time materials remain under Nvidia’s ownership across manufacturing sites. The tool also identifies the binding constraint behind an allocation, so planners can test alternatives rather than receive only one answer.

Nvidia says historical back-testing showed why optimization alone was not enough. Human planners were factoring in partner emails, weather forecasts, geopolitical events and supplier debriefs that the solver could not see. The partnership’s core bet is that those rationales can become a governed record for future recommendations instead of remaining informal knowledge held by individual planners.

Turning planning history into a model

Palantir Foundry provides the shared operating layer for the system. Its Ontology, Palantir’s name for a connected data layer, links materials, manufacturing sites, commitments, capacity, allocations, production outputs and qualitative signals. Nvidia says it built a Digital Supply Chain Intelligence command center in Foundry to bring those inputs together for allocation decisions.

Nvidia then post-trained its open Nemotron 3.5 Lightning model on recorded allocation decisions, the rationales behind them and their outcomes, using a governed Palantir Autopilot workflow. The company says that process anonymizes personal information, expands training examples with synthetic data and fine-tunes a limited set of model parameters while leaving the base model weights frozen.

How Nvidia says the feedback loop works

• The model receives current operational context and returns an allocation recommendation, rationale and risks.
• A planner reviews the recommendation and makes the final allocation call.
• Accepted, edited and overridden recommendations, along with production outcomes, are recorded for later governed retraining.
• Nvidia says the model never retrains itself in production.

A benchmark measures past calls, not live results

Nvidia reports that its post-trained 30-billion-parameter Lightning model reached 86.7% allocation-decision accuracy on its development benchmark. Nvidia says Nemotron 3 Ultra scored 55.5% on the same task and the base Lightning model scored 17.5%. Those company-reported results compare models against Nvidia’s historical decision task; they are not evidence yet of improved live supply-chain outcomes.

The companies are also positioning the arrangement as an enterprise offering beyond Nvidia’s own operation. Palantir is integrating open Nemotron models into Foundry so customers can fine-tune them with their operational data. The stack is designed to run on customer hardware or in the cloud through the Sovereign AI OS reference architecture.

The stated goal is not a general promise to automate every supply-chain decision. Nvidia and Palantir say the system is designed to spot bottlenecks earlier, speed material allocation and evaluate alternatives. Its unresolved test is whether institutionalizing past planner judgment can improve future calls without reducing the scrutiny that complex, fast-changing decisions require.