The most interesting figure in Nvidia and Palantir's September 10 supply-chain announcement is not a dollar amount or a customer count. It is an accuracy gap: on the job of deciding which scarce components go where, a 30-billion-parameter model that had been taught Nvidia's own past decisions scored 86.7%, while a 550-billion-parameter general-purpose sibling managed 55.5%.
Key takeaways
- Post-trained Nemotron 3.5 Lightning (30B) reached 86.7% on Nvidia's materials-allocation task; Nemotron 3 Ultra (550B) reached 55.5% on the same work.
- Nvidia's engineers explicitly capped the claim, reporting that production-risk forecasting stayed unreliable even after the same fine-tuning run.
- Palantir customers can reproduce the pattern on private data and keep weights, data, and inference inside their own infrastructure.
Why the parameter count lost
Roughly eighteen-to-one is the size ratio between the two models, which makes the result read like an argument against scale. It is better read as an argument about scope. Nvidia solutions architects Nell Barber, Rana Haber, and Aastha Jhunjhunwala attached an unusually candid caveat to their own benchmark: the advantage is confined to the one domain the model was post-trained on, and forecasting which production lines were about to run into trouble remained hard regardless. In other words, fine-tuning moved a decision task and left an adjacent prediction task roughly where it was.
That asymmetry is the practical finding for anyone sizing a deployment. Allocation is a bounded problem with a long internal record of correct answers, which is exactly the shape that rewards specialization. Forecasting is not.
How the stack is wired
Nemotron is the smallest component in the system rather than its engine. Palantir's Foundry and Artificial Intelligence Platform supply the substrate, and the company's Ontology keeps a live structural map of parts, plants, spare capacity, and commitments already promised to customers. Nvidia's cuOpt solver does the combinatorial work of splitting a short supply across competing demands.
The language model's contribution is judgment at the end of that chain, reading the surrounding situation and proposing an action for a human planner to accept or override. Both companies are selling the assembly as a single command surface for supply teams, with materials allocation as the opening use case rather than the whole product.
Why Nvidia tested on itself
A $5.4 trillion market capitalization buys an unusually punishing test environment. Thousands of suppliers and a worldwide manufacturing footprint feed racks that Nvidia counts in the millions of parts, roughly 1.3 million for a single Vera Rubin unit, and the failure mode is brutally simple: one absent component freezes assembly while every part that did show up depreciates in a warehouse.
Jensen Huang, Nvidia's founder and chief executive, pitched that coordination burden as the real bottleneck in AI buildouts, describing supply chains as the operating system of the physical economy and AI factories as among the most complex systems ever assembled. Palantir chief executive Alex Karp, predictably, preferred to emphasize the difficulty of the customer he had landed, characterizing Nvidia's supply chain as arguably the world's most intricate.
What a buyer can actually do with this
Neither the partnership nor the sovereignty framing started here. Nvidia and Palantir tied their models and software together last October, broadened that into controlled-environment deployments in June, and have now produced the reference build with a published score on it. The sequence matters commercially: the June announcement was a capability, and this one is a case study with a number.
Reproducing it means training Nemotron on proprietary operational history through Foundry and AIP, then hosting the result on-premises, in a colocation facility, or with a cloud provider, depending on how tightly the data has to be held. Agriculture, manufacturing, pharmaceuticals, retail, technology, and government are the named targets, and the stack is slated for a turn at Palantir's AIPCon 11.
None of that portability would work without Nvidia publishing weights and, for much of the Nemotron family, training recipes as well. The company has spent a year lobbying for frontier open-weight models on exactly these grounds, then spent $12.9 billion acquiring Hugging Face in early September. A supply-chain deployment where the small open model wins is, conveniently, the best possible advertisement for that position. The primary source is the companies' joint newsroom release; The New Stack's report adds the benchmark detail.
FAQ
Is Nemotron 3.5 Lightning open source?
Open weights is the accurate term. Nvidia distributes the weights, and for much of the family the training data and recipes too, which is what makes private fine-tuning possible. The governing terms are Nvidia's own model license rather than an OSI-approved one.
Does this mean companies should stop using large models?
No. The 86.7% figure covers one narrow allocation task with abundant labeled history, and Nvidia's own write-up says the gains do not generalize. A forecasting task inside the same project resisted the same treatment.
What does sovereign mean in this context?
It refers to control rather than geography: the customer's proprietary data, the model weights, and the inference all stay inside infrastructure the customer governs, whether that is an on-premises cluster or a dedicated colocation footprint.
