Nvidia runs its own materials allocation on the Palantir product it now wants to sell

Before this, a planner at Nvidia pulled numbers by hand out of several systems of record and decided who got which parts, a process Nvidia's vice president of enterprise AI, Justin Boitano, describes as manual across multiple systems ^[8][3]. Palantir's Ontology, a data layer that maps physical operations into permitted actions, unifies those sources ^[9]. That removes the copying, which is real work saved. The allocation call itself is a separate skill, made by the planner. Boitano says human planners still make the key decisions ^[11], and that design choice is also why the software's contribution will be hard to isolate from the planner's.

The pitch is wider than the install. Boitano describes sequencing the entire build: when power comes online, how the building gets constructed, when cooling and physical compute show up, then automating allocation of compute in tokens through that infrastructure ^[23]. What is actually running is one decision class inside one company. The partnership behind it began last October ^[2]; the Palantir team arrived at Nvidia four weeks ago, under the company's practice of sending its own employees into customer offices to stand the software up ^[12]. Neither side disclosed financial terms ^[13]. Nvidia has said it plans to produce up to $500 billion of AI infrastructure in the US through partnerships and federal initiatives ^[17], which is the scale both firms point at when describing who else might buy this.

The metric is already named, and it is a good one. Boitano calls time of ownership the North Star, meaning less inventory sitting on Nvidia's books as it moves through the chain ^[7]. Fast Company's account carries no baseline for it and no post-deployment figure ^[22]. Any company that carries inventory can compute that number from its own books, which is what makes its absence the first thing to ask about.

The rack arithmetic explains why allocation went first. At roughly 1.3 million components, a hypothetical per-part rate of 99.999 percent arriving on time and on spec still leaves about 13 parts per rack that miss ^[21], and Boitano's stated concern is that any single piece can break the whole system ^[6]. Five nines is an illustrative figure; the actual rate is undisclosed. The shape holds at whatever rate you assume, and it is why the cheapest lever is which part ships where rather than how fast anything runs.

Separating a reference deployment from a purchase comes down to three checks. The first is whether the bottleneck is data scattered across systems of record, or a judgment your best planner makes and cannot explain; Ontology addresses the first case. The second is whether you can state today's number on the metric you would judge the rollout by, before an install team walks in; if not, the vendor supplies that number afterwards. The third is reading the post-training terms. Palantir does not build its own large language models and instead sells the harness, controls, audit logs and fine-tuning tooling around someone else's ^[15]. Kawasaki, quoted by Fast Company, argues that what distinguishes this system is not optimization but that planners' decisions, expectations and outcomes feed back into the model's weights ^[19]. When your planners' judgment becomes training data, ownership of those weights amounts to ownership of a part of your operation that was never written down anywhere else.