Bryan Catanzaro's admission highlights a growing cost imbalance that most companies chasing AI aren't prepared for
Here’s a number that should make every CFO : at Nvidia’s applied deep learning division, the cost of running computers now exceeds the cost of paying the humans who operate them.
Bryan Catanzaro, Nvidia’s vice president of applied deep learning, made the revelation in an Axios interview on April 26, 2026. His team’s compute expenses, spanning GPUs, electricity, and cloud infrastructure, have officially overtaken what the company spends on salaries for the same group.
The 23% problem #
A 2024 study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) found that AI automation is economically viable in only 23% of jobs focused on vision tasks. That means in 77% of those roles, human labor is still the cheaper option.
When your compute bill rivals your payroll #
Training large models requires thousands of GPUs running for weeks or months. Inference, the process of actually using those models in production, adds another layer of ongoing expense. Then there’s the electricity to power it all, the cooling systems to keep data centers from melting, and the cloud infrastructure fees that stack up faster than a streaming service adding surcharges.
For many tech firms, IT budgets driven by AI infrastructure are now rivaling their total payroll costs. Software companies used to be attractive precisely because their margins were fat: build it once, sell it forever, pay mostly for people. AI flips that model on its head by making compute the dominant line item.
What this means for investors #
The MIT study’s finding that only 23% of vision-heavy jobs are cost-effective to automate suggests the timeline for that payoff might be longer than expected.
There’s also a second-order effect worth watching. Nvidia itself benefits enormously from this spending spree, since it sells the GPUs that make these compute bills so expensive. But even Nvidia’s own VP is acknowledging the cost imbalance within his own team.
The firms most likely to generate positive returns from AI investments are those targeting use cases with clear, demonstrable cost advantages over human labor, not the 77% of vision tasks where humans are still cheaper, but the specific domains where automation genuinely delivers.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our