NVIDIA has spent the AI boom selling the hardware that makes artificial intelligence possible. Now it is turning AI back onto one of the hardest problems created by that boom: getting all of that hardware built and delivered on time.
On September 10, NVIDIA and Palantir announced a new supply-chain AI stack that is being deployed first inside NVIDIA’s own operations. It combines NVIDIA’s Nemotron open models and cuOpt optimisation software with Palantir Foundry, AIP and its Ontology data layer.
The announcement sounds like enterprise software at first glance. The more interesting story is what the system is being asked to do. NVIDIA is trying to capture the judgement of experienced supply-chain planners — including information that conventional optimisation tools can miss — and turn it into a system that can help identify constraints, model trade-offs and recommend how scarce components should be allocated.
AI is being pointed at the machinery behind AI
Modern AI systems do not arrive from a single factory. NVIDIA says its supply chain spans millions of parts, thousands of suppliers and a global network of manufacturing partners. A rack-scale system requires processors, high-bandwidth memory, networking, power, cooling and mechanical components to arrive in the right places at the right time.
The complexity is increasing. NVIDIA says each Vera Rubin rack involves 1.3 million parts, while its technical account of the project describes the Vera Rubin supply chain as twice the size of Grace Blackwell’s.
That creates an unusually difficult allocation problem. A component sitting at a factory early is not necessarily useful if another critical part is late. A shortage can also move from one component to another as supply conditions change.
NVIDIA already uses mathematical optimisation to help make these decisions. But in a technical post accompanying the announcement, the company said a back-test found that human planners were incorporating signals the quantitative model could not see: supplier emails, severe-weather forecasts, geopolitical events, debrief calls and years of accumulated experience.
That gap is what NVIDIA and Palantir are trying to close.
Nemotron is learning the reasoning behind the decision
The system creates what NVIDIA calls a Digital Supply Chain Intelligence command centre in Palantir Foundry. Palantir’s Ontology connects materials, manufacturing sites, capacity, allocations and production data with less structured information that can influence a decision.
Instead of simply recording what planners decided, the workflow can capture why they made the decision and what happened afterwards. NVIDIA then uses that history to customise Nemotron models for the organisation’s own operating environment.
The distinction matters. This is not a general-purpose chatbot being dropped into a factory. It is an attempt to build a specialised model around proprietary operational data and the judgement of the people who already understand the supply chain.
NVIDIA says a specialised 30-billion-parameter Nemotron 3.5 Lightning model reached 86.7% allocation-decision accuracy on its development benchmark. That is a company-reported benchmark rather than an independent evaluation, but it gives some indication of what the project is trying to achieve: reproduce useful expert decisions rather than merely generate plausible text.
Human planners are not being removed from the loop. According to NVIDIA, the models recommend actions, explain trade-offs and flag emerging risks while supply-chain experts retain control of final decisions. Accepted, edited and rejected recommendations can then become feedback for later model training.
Why this partnership matters beyond NVIDIA
There is an almost circular quality to the project. The company at the centre of the AI infrastructure boom is now using AI to improve the supply chain that produces AI infrastructure.
But the larger significance is where this kind of technology could go next.
NVIDIA and Palantir say the same architecture can be adapted for industries including manufacturing, energy, healthcare, automotive and aerospace. Companies can customise Nemotron with their own data while keeping control of their models, data and deployment environment, with deployments available on premises, in colocation facilities or in the cloud.
That emphasis on control explains the companies’ use of the term “sovereign AI.” In this case, it is less about a consumer-facing AI product and more about organisations being able to use specialised models without handing their most sensitive operational knowledge to a generic external system.
For ordinary technology users, none of this will produce a new app to download tomorrow. Its importance is more fundamental. The next phase of AI adoption is increasingly happening behind the interfaces we see — inside factories, logistics networks, data centres and the systems deciding where scarce hardware goes.
Chatbots made AI visible. Projects like NVIDIA and Palantir’s supply-chain system show where it may become harder to see, but potentially more consequential.
Sources: NVIDIA Newsroom and NVIDIA Technical Blog.
