Nvidia CEO Jensen Huang told a packed arena at GTC 2026 on March 16 that his company plans to give engineers annual AI token budgets worth roughly half their base pay. For someone earning $300,000, that means an extra $150,000 in compute credits to run AI agents, automate tasks, and, in Huang's telling, become ten times more productive.
Days later, on the All-In Podcast, he went further. A $500,000 engineer who only burns through $5,000 in tokens? "I will go ape something else," Huang said. When a host asked whether Nvidia is spending $2 billion annually on tokens for its engineering team, his answer was blunt: "We're trying to."
The math, and why it matters for Nvidia specifically
The framing is clever. Huang is repackaging an infrastructure cost as a human capital investment. Instead of arguing that companies should spend more on compute (a pitch that sounds self-serving coming from the CEO of the world's largest GPU maker), he's arguing that companies should spend more on their people, who just happen to need Nvidia hardware to consume those tokens.
Follow the logic: normalize token budgets as an employee benefit, push engineers to consume as many tokens as possible, watch inference demand compound, then sell the chips that handle that inference. Nvidia's Vera Rubin and Blackwell platforms exist for exactly this workload. Huang projected $1 trillion in purchase orders through 2027 during the same keynote. The token compensation pitch and the trillion-dollar demand forecast aren't separate announcements. They're the same argument.
"Every engineer is going to have a token budget," Huang said at GTC. "What used to be a thing for engineers is when you come to work, they give you a laptop. Now when you come to work, they give you a laptop and tokens." The comparison to CAD tools for chip designers is his favorite analogy: an engineer who refuses to use AI tokens is like a chip designer insisting on pencil and paper.
Who else is buying this?
Huang isn't operating in a vacuum. Business Insider reported earlier in March that Silicon Valley firms are already experimenting with token budgets alongside traditional salary, bonuses, and equity. Tomasz Tunguz of Theory Ventures called tokens a potential "fourth component" of pay packages, which sounds about right for 2026 venture capital discourse.
Thibault Sottiaux, an engineering lead on OpenAI's Codex team, noted on X that candidates are asking about compute access in interviews. That tracks. If you're an AI researcher choosing between two offers, the one with generous inference budgets lets you iterate faster, test wilder ideas, and ship more. Compute scarcity is real, and access to it is becoming a genuine differentiator.
But there's a gap between "engineers want more compute" and "companies should formalize token budgets as compensation." One is an observation about the job market. The other is an accounting decision with tax, governance, and retention implications that nobody has really figured out yet. Tokens aren't equity. They don't vest. They're not portable. If you leave Nvidia, your token budget stays behind.
The productivity question nobody can answer
Huang's claim that token access makes engineers 10x more productive is doing a lot of heavy lifting. Nvidia has 42,000 employees. If the company actually hits that $2 billion annual token spend, that works out to roughly $47,000 per employee across the whole company, or considerably more if you limit it to the ~17,700 engineering staff. The implied productivity gains would need to be enormous to justify that at scale.
And the track record for enterprise AI projects isn't encouraging. Andreas Welsch, founder of Intelligence Briefing, has pointed out that roughly 80% to 85% of AI projects have failed since 2018. Giving every engineer a fat token budget doesn't automatically solve the problem of figuring out which tasks AI actually does well. It just means more tokens get burned.
Nvidia posted $215.9 billion in revenue for fiscal 2026, up 65% year-over-year, so the company can afford the experiment. Whether other firms can is a different question. A mid-stage startup with 200 engineers isn't writing $2 billion token checks. The model, as described, works best for companies that are already printing money.
Nvidia expects Vera Rubin to begin shipping later this year with 10x inference performance per watt over its predecessor. If token compensation becomes a recruiting norm, every new GPU generation that lowers cost-per-token makes the benefit cheaper to provide while keeping the nominal budget impressive. That's a nice flywheel if you're the one selling the GPUs.
