Thought Leadership
Attracting top software engineers has never been easy. Nvidia CEO Jensen Huang thinks he has found a new way to sweeten the deal: tokens.
During his keynote at this year’s GPU Technology Conference, Huang laid out a vision in which companies would routinely provide their engineers with generous token budgets on top of their regular salary.
The productivity pitch
On stage, Huang sketched out the math behind his idea. An engineer earns a few hundred thousand dollars a year in base pay. He would add roughly half of that again in the form of tokens, arguing that access to AI compute allows developers to multiply their output many times over. In his view, that is a bargain for everyone involved.
According to Huang, token budgets are already becoming a talking point in Silicon Valley job interviews. Candidates want to know how much AI compute they will have at their disposal in a new role.
For those unfamiliar with the term: tokens are the basic unit large language models use to process text, roughly equivalent to a word fragment. AI providers typically charge for their services based on the volume of tokens consumed.
A trend beyond nvidia
Huang appears to have put words to something that has been brewing for a while. Business Insider has reported that Silicon Valley companies are already experimenting with AI inference compute as an additional component of compensation packages. Investors reportedly see it as a fourth pillar alongside salary, bonuses and stock options, and are encouraging companies to list token budgets directly in their job postings.
The trend has reached OpenAI as well. Thibault Sottiaux, who leads engineering for the company’s Codex service, recently wrote on X that AI compute is becoming increasingly scarce and valuable. Job candidates now routinely ask how much dedicated inference capacity they would get to work with.
Nvidia wins either way
Huang’s proposal also happens to align neatly with his company’s business model. More token budgets for developers mean greater demand for compute power, and that compute overwhelmingly runs on Nvidia hardware. During the same keynote, Huang projected that orders for the Blackwell and Vera Rubin chip generations alone would reach one trillion dollars by 2027.
