Nvidia’s Bryan Catanzaro: AI compute costs beat employee costs right now

If you manage budgets, the uncomfortable line in the AI story is this: Nvidia’s own Bryan Catanzaro says compute is more expensive than paying human workers, today. Catanzaro, Nvidia vice president of applied deep learning, told Axios in April, “For my team, the cost of compute is far beyond the costs of the employees.” In other words, the supposed labor replacement is not yet winning on the spreadsheet.

This matters because the rest of the market is acting like the savings are real already. Big Tech has announced $740 billion in capex this year so far, according to Morgan Stanley, a 69% increase from 2025. And the spending is running ahead of clear productivity proof, at least in the way executives traditionally want it, with AI yet to show evidence of widespread increased productivity.

Layoffs might look like the “AI is replacing workers” storyline playing out in real time. Meta announced in a memo earlier this year that it plans to lay off 10% of its workforce, about 8,000 employees, and also scrap plans to hire for 6,000 open positions. The memo framed it as “run the company more efficiently and to allow us to offset the other investments we’re making,” which is a telling budget priority: reduce payroll costs while funding the AI push.

Microsoft has also offered thousands of employees a voluntary buyout, the largest the company has ever offered. But the counter-signal from Nvidia is blunt: AI is not yet cheaper than labor in the situations Catanzaro is dealing with. And an MIT study from 2024 provides an economic baseline that supports that lived experience. Researchers analyzed the technical requirements of AI models needed to perform jobs at a human level and found AI automation would be economically viable in only 23% of roles where vision is a primary part of the work. In the remaining 77% of the time, it was cheaper for humans to continue their work.

There is also the reliability tax, which is harder to quantify but brutal in practice. The source notes that in one instance an engineer said an AI agent destroyed his database and network due to what he called “overuse.” That is not just a one-off horror story. It is the kind of operational risk that increases the human oversight burden and raises the total cost of deploying AI systems in production.

So why is the capital still pouring in? The article points to a gap between spending and outcome evidence. Despite no clear evidence of AI improving productivity and, according to the Yale Budget Lab, no widespread data to support the idea of AI displacing jobs, Big Tech firms have continued to announce massive AI capex. The scale is so large that some companies are already scrambling their own internal plans. Uber chief technology officer Praveen Neppalli Naga told The Information earlier this month, “I’m back to the drawing board because the budget I thought I would need is blown away already.” He was referring to Uber’s pivot to AI coding tools, including Anthropic’s Claude Code, plus the speed at which adoption changed the numbers.

And budgets do not just move, they get restructured. The article says Uber burned through its entire 2026 AI coding tools budget by April after it incentivized adoption through leaderboards. Meanwhile, Microsoft is cancelling most of its direct Claude Code licenses, The Verge reported last month, and pivoting to GitHub Copilot CLI. The reason given is that the technology became too popular too fast as the tech firm pushed employees to integrate AI into workflow. That is a second-order effect boards often miss: adoption incentives can turn a pilot into a cash register.

Meanwhile, workforce cuts keep rising alongside the AI spend. According to data from Layoffs.fyi, there have been more than 118,000 layoffs in tech in 2026 so far across nearly 100 companies. The rate of these workforce reductions already outpaces last year’s total of about 120,000 layoffs. This combination, spending more on AI while cutting labor, exposes a discrepancy in the economics of AI, according to Keith Lee, an AI and finance professor at the Swiss Institute of Artificial Intelligence’s Gordon School of Business. Lee told Fortune, “What we’re seeing is a short-term mismatch.” The short version for executives: the cost structure has not caught up with the ambition.

When could it? Lee says AI has remained less efficient than human labor because hardware and energy raise operating costs for providers. At the current pace, AI expenditures may reach $5.2 trillion by 2030, with $1.6 trillion from data center spending and $3.3 trillion from IT equipment, according to McKinsey data. Spending could surge to $7.9 trillion by 2030 at an accelerated pace. Fees for AI software have increased by 20% to 37% over the past year, Tropic noted in December 2025. And AI companies may also be losing money because their fixed subscription model may not cover operating costs for heavy AI users.

Lee’s core point is not just that AI needs to become cheaper than humans. It needs to become both cheaper and more predictable at scale. He suggests warning signs to watch for: performing inference for a large language model with 1 trillion parameters could plummet by more than 90% over the next four years, according to a report last month from Gartner. AI infrastructure improvements and hardware supply could also shift the economics, and model designs plus a pricing shift from flat subscription to usage-based pricing could reduce the “pay whether you use it or not” mismatch.

Until those tipping points show up, the strategic stakes for boards and CFOs are immediate. You are paying for compute now, not later. If AI is still a cost center relative to labor, then the “AI transformation” story has to be managed like a capital program with milestones, not like a straight-line replacement narrative. Federal Reserve data shows about 18% of companies had adopted AI tools as of the end of 2025, a 68% growth since September 2025. Adoption is moving fast. The reckoning is whether reliability improves with fewer hallucinations, less need for human oversight, and whether AI can integrate into company infrastructure in a way that survives real operations.

The big takeaway: executives can keep spending, but the justification changes. In the near term, AI may be a complementary tool rather than a labor-savings substitute, at least “until the cost structure stabilizes.” The winners will be the teams that treat compute and deployment costs as first-class metrics, and that demand proof of both economic viability and operational reliability, not just flashy capability.