Nvidia has begun running its own supply chain on Palantir software, making the chipmaker at the center of the AI boom the first customer of a product the two companies built together and now plan to sell across industry and government.

The collaboration, which builds off a partnership that began last October, turns Palantir’s platform into a shared command center for Nvidia’s global supply chain, deploying Nvidia’s own open-weight models to fix its bottlenecks. “We’re really at the center of the world’s largest infrastructure buildout in human history,” says Justin Boitano, Nvidia’s vice president of enterprise AI.

Few supply chains are as complex, or consequential. The speed of Nvidia’s governs how quickly AI infrastructure gets built anywhere in the world. The company has said it plans to produce up to $500 billion of AI infrastructure in the U.S. through partnerships and federal initiatives, alongside building supercomputers and secure AI factories for the government.

The first focus is materials allocation, the decisions that determine which parts move where, and how fast. Each Vera Rubin rack requires Nvidia to coordinate thousands of suppliers and roughly 1.3 million components across compute, memory, networking, power, cooling and mechanical systems. “Any single [piece] could essentially break the whole system,” Boitano tells Fast Company, “so how you allocate those materials really matters.”

Nvidia’s aim is to cut the time that inventory sits idle, he says. “Our North Star is time of ownership, making sure that we’re not carrying all this inventory on our books through our supply chain, and minimizing that time.”

Amid a chip shortage, a power shortage, years-long grid interconnection queues, and a national pushback against data center construction, every minute gained counts. The CEOs of both companies and the Trump administration say the stakes of that buildout are existential, and with roughly 1.5 percent of US GDP now tied to AI data centers and hardware, the economy depends on it too.

Boitano says Nvidia’s open-weight models and Palantir’s software can help streamline every level of AI deployment, including, through a preexisting partnership, the construction of the data center itself. “When does power come online, how does the building get constructed, when do I have different cooling show up, when does the physical compute infrastructure show up, and then automating everything to allocation of compute in tokens through that infrastructure.”

But the approach can help any company or government body who is “thinking through a complex supply chain, doing inventory management to build, which you can imagine really is a large portion of the GDP of our economy.” The companies did not disclose customers, though last December they announced an initial collaboration to help Lowe’s optimize its global supply chain, covering over 1,700 stores and 7,500 vendors.

Previously at Nvidia, Boitano says, a planner assembled data by hand to decide allocations, “a manual process across multiple systems of record.” Palantir’s Ontology, a data layer that maps a company’s physical operations into permitted actions, unifies those sources. Nvidia’s cuOpt software models supply constraints and weighs tradeoffs; its Nemotron models, post-trained on a user’s internal data, recommend actions, explain them and flag risks. Human planners still make key decisions, says Boitano.

Palantir doesn’t develop its own large language models. Instead, it sells software that sits on top of other companies’ AI models, offering harnesses, controls, audit logs, and tools for fine-tuning and post-training. Its main platforms—Gotham, Foundry and AIP—wrangle data, analyze trends and automate processes for companies and government agencies, including Airbus, Walmart, NATO and ICE. The company sends employees into customers’ offices to set up the software; a Palantir team arrived at Nvidia four weeks ago. (Neither company disclosed financial terms of the deal.)

What Kawasaki argues distinguishes the system is not optimization, something many software companies try to do, but how the system learns from each decision. Decades of judgment held by experienced planners—about which supplier slips, which substitution is safe—along with each recommendation and outcome, feed back into the weights of the model.

“If I can take those decisions and understand why they were made, what was the expectation of that decision, and then what happened—and think about that more like a training trace, as opposed to just essentially friction in the machine—this is now an asset,” Kawasaki says.

In the process, more decisions can be automated. Boitano compares it to how programmers came to trust coding agents, reviewing every output until they stopped. “Initially you want people in the loop,” he says, “but as you trust the system, you can let it run in a longer horizon.”

Amid fears of hacks and agents going rogue, or both, safeguards trained into models should be assumed removable. “The question is how does the system still fail safe,” he says, and Palantir’s Ontology “acts as the rule set of what the AI can do.”

‘I did not think this was possible a year ago’

The companies call their offering “sovereign AI,” echoing a growing demand among governments and companies for more control over their AI stacks—and a recent fixation of Palantir CEO Alex Karp, who has called on enterprises to end their dependence on hyperscalers like Google, Microsoft and OpenAI.

During a fiery appearance on CNBC in July, to announce another Nvidia tie-up, Karp blasted the enterprise business model of the large AI labs, and warned that the AI labs business model was forcing enterprises and governments to give up “their means of production.” “Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane.”

Kawasaki says sovereignty isn’t just about data custody, or freedom from the frontier model companies, but how well your own custom AI stack works. “Your primary source of sovereignty is your own outperformance,” he says. “Forget about your data if your business isn’t succeeding.”

That case rests on open-weight models having improved fast enough to beat frontier models on the narrow benchmarks a given company cares about, at a fraction of the cost. For most customers, frontier models are still essential, he says, “but in a lot of cases, you can use these open models and post-train them to achieve beyond frontier capabilities.” Kawasaki says the post-training runs at Nvidia took minutes rather than requiring massive clusters. “I did not think this was possible a year ago… that a custom post-trained model was even worth the time.”

The partnership originated in work for military and intelligence agencies that operate their own classified clouds, where Nvidia sells chips and Palantir sells its battlefield-focused Maven Smart System. “This is where we learned it was possible, and we learned it was effective, and we learned it would work,” Kawasaki says. “That was where Palantir learned that this was going to be a big market.”

Government interest now, he says, is “urgent and extreme.” “You can’t use the closed source models on a submarine, down range. It has to be, you’re either disconnected or it’s not allowed.”

The push for “sovereign AI” picked up steam last year as governments around the globe questioned their reliance on the US and its cloud providers. More than 180 government-backed sovereign AI projects are now underway worldwide, according to the Center for a New American Security. On the corporate side, says Kawasaki, “I don’t have a customer that isn’t interested in this.”

While Nvidia and Palantir have announced no joint sovereign deployments outside the US, in January a UK data center developer named Sovereign AI selected Palantir and Accenture to build AI infrastructure across Europe and the Middle East, using Dell servers with Nvidia chips. The two companies also unveiled a system in March, Sovereign AI OS, for deploying AI in organizations’ own data centers, aimed at customers with latency-sensitive workflows, data sovereignty needs and wide geographic distribution.

Nvidia’s own sovereign business abroad is large and growing. In July, Nvidia, Japan’s trade ministry and Noetra Corp. announced an “AI factory” billed as the world’s first national AI infrastructure for physical AI, with 27,500 Rubin GPUs across 140 megawatts of data center capacity. Nvidia has struck similar arrangements with HUMAIN in Saudi Arabia, G42 in the UAE, NAVER in South Korea, Mistral in France and BharatGen in India.

But the pushback on US giants overseas has also extended to Palantir’s own contracts. In Germany, the Bundeswehr is testing a number of European alternatives to Palantir’s Maven—including Almato, Orcrist and ChapsVision—and a broader European backlash is gathering, including inside NATO. Palantir’s European business has grown regardless: it inked a £1.5 billion UK partnership last September, then a £240.6 million Ministry of Defence contract in December, its largest with the agency to date.

For Nvidia, the Palantir deal further extends the chipmaker’s reach across the AI ecosystem. Jensen Huang has backed open models and invested aggressively in the labs and neoclouds that buy his chips, a strategy that has drawn scrutiny for its circularity but that SemiAnalysis founder Dylan Patel reads as insurance against being squeezed. “A world where OpenAI, Anthropic, and Google models are the only models is one in which he’s screwed,” he told podcaster Dwarkesh Patel in March. “A world where hyperscalers are the only ones building compute is one he’s screwed in.”

Both companies collectively hold billions of dollars in US government contracts and have strong ties to the Trump administration. Nvidia and competitor AMD last year arranged a deal with the White House to share 15 percent of their China-related revenue with the government in exchange for resuming exports of specific chips like the H20.

President Donald Trump’s financial disclosures also showed dividend earnings from significant stakes in many companies, including Palantir and Nvidia. Both companies’ market values have soared since late 2024: to roughly $300 billion for Palantir and $5.6 trillion for Nvidia, making it the most valuable company in history. 

As the Palantir system proves itself at Nvidia, Kawasaki argues that more automation—across the AI buildout and the rest of the economy—is not optional.

“We haven’t done a lot of building like this in the United States in quite some time,” he says. “When you think about how many more people can be employed to do complex construction jobs, the number is somewhat limited. And so therefore, if you want to produce more with the same amount of inputs, the computer really does have to do more work.”