NVIDIA and Palantir Launch Sovereign AI for Supply Chains, Reaching 86.7 Percent Accuracy on Parts Allocation

NVIDIA and Palantir unveiled a sovereign AI effort aimed at critical supply chains on September 10. The first deployment is NVIDIA's own manufacturing operation, where an optimisation solver and a custom-trained language model now support the weekly decision of which parts go to which site. The post-trained model reached 86.7 percent accuracy on allocation decisions.

1.3 Million Parts in a Single Rack

The starting point is the sheer part count of AI infrastructure. An NVIDIA Vera Rubin rack is built from 1.3 million parts. On the current GB200 NVL72, a single compute tray needs 2 NVIDIA Grace CPUs, 4 NVIDIA Blackwell GPUs and 32 HBM3e stacks, and the Vera Rubin generation doubles that footprint. With thousands of suppliers and manufacturing partners worldwide, sending one extra component to one site always pushes back a completion date somewhere else.

Alex Karp, cofounder and CEO of Palantir, described NVIDIA as having the most valuable and intricate supply chain in the world. Jensen Huang, founder and CEO of NVIDIA, framed supply chains as the operating system of the physical economy and called AI factories among the most complex systems ever built.

The Target Metric Is Time of Ownership

What makes the design interesting is the metric chosen for optimisation. Time of Ownership measures the span from the moment a manufacturing site receives material to the moment it leaves as part of a sub-assembly or finished product. In other words, it attacks the time inventory spends sitting on the floor.

The metric splits into 2 intervals: time-to-rack, covering silicon through to an assembled system, and time-to-token, covering installed infrastructure through to productive output. Squeezing the first interval alone does not pay off if the second one stretches. Putting both on a single axis is what separates this from ordinary production efficiency work.

NVIDIA cuOpt handles the computation. Each week it solves the minimisation of Time of Ownership as a mixed-integer linear program, working through thousands of variables and constraints and surfacing which constraint is currently holding everything back. Planners can then explore what changes when that constraint moves.

Capturing What Human Planners Actually Know

Mathematical optimisation alone did not match the judgement made on the floor. By NVIDIA's account, experienced planners consistently outperformed the quantitative models, and the gap came from information that never reaches a table: supplier call transcripts, email exchanges between staff, regional weather forecasts and geopolitical events. They were reading context that sits outside the numbers.

The answer was to structure that operational context in Palantir Ontology and let a language model draft recommendations on top of it. Training ran through the NeMo tool chain. NeMo Anonymizer stripped personally identifiable information, NeMo Data Designer generated synthetic data covering scenario variations, and NeMo AutoModel performed supervised fine-tuning with LoRA. The run used 2 B200 GPUs and finished within minutes.

That produced the headline figure. The post-trained Nemotron 3.5 Lightning, which carries 30 billion total parameters with 3 billion active per pass, recorded 86.7 percent accuracy on allocation decisions. Nemotron 3 Ultra, used as a comparison, reached 55.5 percent, and the un-tuned base model 17.5 percent. Balanced accuracy came in at 58.6 percent and Macro-F1 at 57.5 percent. Recommendations were validated through a point-in-time backtest against the decisions actually taken.

People Decide, the System Learns

The operating loop draws a clear line. Ontology supplies the context, cuOpt produces the quantitative optimum, and Nemotron assembles a recommendation together with its rationale. A human planner makes the final allocation call, and both the decision and its outcome flow back into Ontology as material for the next training cycle.

Rather than handing decision-making to an agent, the design moves the work of assembling evidence onto the machine. In a domain where a single mistake echoes for weeks, that boundary is a practical one. It also has a side effect: because human judgement is continuously captured as training data, knowledge that used to live with individuals accumulates as an organisational asset.

Built to Run on Your Own Infrastructure

Deployment shape matters here as much as the model. The stack follows Palantir's Sovereign AI Operating System Reference Architecture, known as SAIOS, and can be deployed immediately on-premises, in the cloud or in colocation. Dell Technologies, Cisco, Rackspace and Nebius are named on the infrastructure side.

Supply chain data is a dense collection of commercially sensitive material, from contract terms to production capacity. Precisely because it is hard to hand wholesale to an external managed service, AI work in this area has tended to stop at spreadsheets and rules of thumb. Designing the stack to sit inside a company's own environment should matter to anyone bringing the same idea elsewhere.

That said, the reported accuracy is a measured figure from a controlled setting, and how it holds up during an actual supply disruption is a separate question. There is also the open point of how far training padded with synthetic data can stretch to cover situations with no precedent. Starting with its own operation means NVIDIA takes on that validation first.

Summary

NVIDIA and Palantir released a sovereign AI stack for supply chain parts allocation on September 10. cuOpt solves the weekly minimisation of Time of Ownership, a post-trained Nemotron 3.5 Lightning builds recommendations at 86.7 percent accuracy, and a human planner makes the final call. Facing a Vera Rubin generation built from 1.3 million parts, it is a notable attempt to translate expert judgement into something a machine can read.