Booz Allen Moves to Nvidia's Nemotron; OpenRouter Share Under 1%

Booz Allen Hamilton, a government contractor with a $10 billion market capitalization, moved its cybersecurity and defense software off Anthropic's models and onto Nvidia's Nemotron several months ago. The firm's CTO Bill Vass said the change has produced significant cost savings, though the company declined to quantify them. Booz Allen spends tens of millions of dollars annually on AI, according to a person familiar with the company.

“What we've been doing is the opposite of tokenmaxxing,” Vass said, using the term for consuming as much AI as possible.

What did Booz Allen actually switch?

The move covers the models that run Booz Allen's software, not the ones that write it. Developers at the contractor still mostly use Anthropic's Claude Code unless a customer asks for a different coding tool, Vass said. Claude Code still draws the strongest praise of any tool for software engineering, and Vass said he wants his developers running Nemotron for coding as those models improve.

Nemotron's progress with contractors and government agencies owes much to Palantir. The two companies have drawn closer together, and both now market AI that lets enterprises avoid dependence on the frontier labs Anthropic and OpenAI. Nvidia recently signed a deal to use Palantir's software inside its own operations to manage its supply chain. Palantir launched a product in late June aimed at helping U.S. government agencies and enterprises run Nvidia's open-source models through its software. Some government customers have already moved off closed-source frontier models to Nemotron, Palantir CEO Alex Karp said this summer.

How much of the coding market uses Nemotron?

Adoption runs beyond large buyers. Kilo Code, an open-source coding platform, said Nvidia's Nemotron models handled nearly 17% of the roughly 50 trillion tokens that passed through its software in October, up from 8% in the previous month. Individual developers and small and large enterprises use the platform mainly to reach open-source models for coding work.

Scale elsewhere tells a different story. OpenRouter, the Stripe-owned service that lets businesses switch between models from one place, processed about 700 trillion tokens last month across a broader set of tasks than coding. Nemotron accounted for just about 4% of those tokens over the past month, down slightly from a peak of 7% in one week in August.

Spending tells the same story. So far in November, Anthropic accounted for a higher share of OpenRouter customer spending, at 31%, than any other AI provider, while Nvidia accounted for less than 1%.

Why does Anthropic still dominate spending?

Most companies keep directing the bulk of their AI budgets to Anthropic's models. That has pushed firms to moderate spending, and they have diverged on whether to set firm token caps for their software engineers. Apple, Microsoft and Disney have restricted AI spending further. The caps have become a sticking point for engineers themselves.

Maria Colacurcio, CEO of human resources software startup Syndio, said talent acquisition leaders at the company's large enterprise clients have begun receiving requests from job candidates seeking higher token budgets as part of their employment packages. The requests typically follow an offer that includes higher or unlimited token budgets and access to the latest OpenAI and Anthropic models.

“So there's all of these sort of side addendums being created where candidates are getting basically agreements to these carve outs,” Colacurcio said.

None of it is formal policy yet. “Nobody's writing it down. It's not really part of policy,” she said. If AI prices stay high and enterprises keep restricting spending, as Apple, Microsoft and Disney have done, tokens could turn into a formalized, coveted employee benefit.

Where does Nvidia's own reputation sit?\n

Nvidia is meanwhile managing what has been described as the worst hit to its reputation in its three-decade history: the revelations that chief executive Jensen Huang had an affair with a woman reportedly working for China's spy agencies. The controversy has not stopped the open-source model push. Palantir and Nvidia keep expanding joint offerings, and Kilo Code's October share shows Nemotron models gaining ground in coding workflows.

For Nvidia, the near-term case for Nemotron rests on price rather than reach. With Anthropic still taking 31% of OpenRouter spending and Nvidia under 1%, the cost savings Booz Allen cites should keep pulling government and contractor workloads onto open-source models as enterprises tighten AI budgets.