Jensen Huang has introduced a metric for engineering productivity that ties token consumption directly to salary. Speaking at GTC 2026, the Nvidia CEO proposed that a $500,000 engineer should consume less than $250,000 in annual AI tokens to justify their cost. Engineers exceeding this threshold face potential removal from the team.

The test reflects how AI infrastructure costs now factor into workforce decisions at major tech companies. Token consumption measures how much computational power engineers use when running models, training systems, or developing AI applications. It's a concrete way to quantify AI resource usage against headcount spending.

Huang's framework suggests companies can maintain team size while reducing overall token budgets if engineers improve efficiency. An engineer using fewer tokens accomplishes more with less compute, which signals either better coding practices, smarter model selection, or more efficient development workflows.

This metric creates practical pressure for engineering teams to optimize. Developers might shift toward smaller models, batch processes more intelligently, or reduce redundant model runs during development. Better prompt engineering and fewer experimental iterations also lower token costs.

The approach carries risk. Token consumption varies wildly depending on role. A researcher exploring novel architectures naturally burns more tokens than a backend engineer. An engineer working on large-scale simulations faces different constraints than one building APIs. Applying a uniform threshold across different domains could eliminate specialists doing legitimate high-compute work.

Nvidia's framing also reflects the company's business interests. As an AI infrastructure provider, Nvidia profits from high token consumption. Yet Huang advocates for efficiency metrics that might reduce token spending. This suggests he believes efficiency gains and architectural improvements will drive faster AI progress, ultimately expanding the total market for compute.

The token budget test likely becomes standard at other enterprises managing AI costs. Companies spending millions on language models need ways to measure whether individual engineers deliver value proportional to infrastructure spending. Huang's metric translates vague concerns about AI efficiency