Nvidia pushes AI token budgets as employee perk

Nvidia CEO Jensen Huang has floated a new perk for software engineers: not just salary, equity and bonuses, but a budget of AI “tokens” to spend on models like Claude, ChatGPT and Gemini. TechCrunch reports that Huang suggested token allocations could reach roughly half of base salary, implying hundreds of thousands of dollars a year in compute for top staff.

The proposal lands at a moment when token spending is no longer a rounding error. “Agentic” systems—tools that run continuously, spawn sub-agents and work through task lists—can burn through millions of tokens a day without a human typing much at all, TechCrunch notes. That changes what “tools” mean in knowledge work: the marginal cost of producing an email, a pull request, or a research memo becomes a metered expense tied to a platform.

Once compute is priced per word, per action, per day, compensation starts to look like procurement. A token allowance is framed as empowerment—more compute, more output—but it also makes work measurable in a way that salary is not. The New York Times, cited by TechCrunch, describes engineers “tokenmaxxing” on internal leaderboards at companies including Meta and OpenAI, competing over consumption as if it were performance. When usage becomes a metric, managers get a new dial to turn: quotas, dashboards and budget cuts can all be presented as “efficiency” rather than supervision.

It also shifts bargaining power toward whoever controls the meter. A company can raise or lower token budgets overnight; a platform can change pricing, rate limits, or model access on its own schedule. What used to be a locally installed tool becomes a dependency with terms of service. In that world, an employee’s productivity is partly determined by a line item controlled by finance and an external vendor.

There is a second-order effect on headcount decisions. TechCrunch quotes the concern that when token spend per employee approaches or exceeds the employee’s salary, the arithmetic changes for CFOs. If the organisation is effectively paying for “another worker” in compute, the question becomes how many humans are needed to coordinate and review what the compute produces. Tokens can be pitched as a benefit while quietly normalising the idea that one engineer should supervise a small fleet of machine labour.

The token allowance also blurs the meaning of “compensation”. If tokens are required to meet expectations, they function less like a bonus and more like a mandatory input—comparable to being “paid” with the right to use the company’s machinery. In boom times, generous budgets help recruit. In lean times, they become an easy place to cut without touching headline salaries.

Huang’s suggestion was framed as a recruiting tool. It also describes a workplace where output is metered, tracked and priced—one prompt at a time.

The Times’ reported internal leaderboards turn token consumption into a status game. The bill still lands with whoever controls the budget.