The Cost of AI Is Outpacing Your Payroll

The quote worth paying attention to

Bryan Catanzaro, Nvidia's VP of Applied Deep Learning, told Axios something that most enterprise leaders probably didn't want to hear: 'the cost of compute is far beyond the costs of the employees.' That's not a complaint from someone struggling to make AI work. It's a data point from the person paid to push AI harder than almost anyone else on the planet - and he's finding it expensive.

The comment landed in late April and got picked up by Fortune, TechSpot, Tom's Hardware, and Entrepreneur. The real cost of AI has been a topic most boardroom conversations have avoided, where the narrative still centers on labor replacement and efficiency gains. Catanzaro's statement didn't fit that story - which is exactly why it matters. For enterprise teams scaling AI workloads in 2026, the math he's describing is more common than the industry likes to admit.

Nvidia isn't the only one hitting this wall

Catanzaro's experience isn't isolated. At Uber, AI spending has ramped up fast enough to exhaust the company's entire planned 2026 AI budget before the year is over. Uber's CTO Praveen Naga told Axios he had gone back to the drawing board because the budget he had expected to need was already blown. Token-based pricing - the dominant model for large language models today - turns inference usage into a metered operating cost that scales with every API call. It doesn't behave like a software license. It behaves like a utility bill.

That dynamic hit Swan AI just as sharply. Founder Amos Bar-Joseph posted on LinkedIn about a $113K Anthropic bill for a four-person team. The rough math works out to around $28K per person per month in AI compute spend - a figure that likely exceeds each team member's monthly salary.

These aren't edge cases from companies using AI carelessly. They're the honest result of building real products on frontier models in 2026. And they show that AI infrastructure cost is rapidly becoming a meaningful operational line item for teams doing serious AI work - not a one-off project expense that rolls off once the build is done.

What the MIT study actually says about AI replacing workers

A 2024 MIT study analyzed the technical requirements for AI models to perform jobs at a human level across different roles. The conclusion: AI automation was economically viable in only 23% of roles where vision tasks are central. In the remaining 77%, human labor was still the cheaper option.

That's a more nuanced finding than the headline suggests. It's not arguing that AI doesn't work - it's saying that the cost of running it at production scale, with the reliability and human oversight most production environments require, still favors people in most cases. The 23% where AI wins financially tend to be the most repetitive, most structured roles with the least judgment involved. For everything else, the AI replacing workers narrative is running well ahead of the actual economics.

This is a useful corrective for any organization that has built its AI business case around replacing a headcount budget with an inference budget. The substitution is real in some narrow categories. It isn't yet broadly true across professional roles.

$740 billion in, and AI ROI is still an open question

None of this data has slowed down capital deployment. Big Tech committed $740 billion in AI-related capital expenditures in 2026 so far, according to Morgan Stanley - a 69% jump from the prior year. That's a structural bet on the cost curves inverting. But the timeline on that inversion is unclear, and the companies making those bets are spending now on the assumption it happens fast enough to matter.

Meanwhile, the Yale Budget Lab has found no widespread data supporting AI job displacement at scale. Federal Reserve data from late 2025 shows only about 18% of US companies had adopted AI tools - a meaningful jump in adoption rate since September 2025, but still a minority of the US market.

The disconnect here is real and worth naming. Companies are cutting headcount and framing it as AI-driven efficiency. But if AI spending on compute exceeds the cost of the employees being replaced, the efficiency story isn't about technology doing more with less - it's about headcount reduction while operating costs quietly migrate from payroll to compute bills. Total cost of operations doesn't automatically shrink. It shifts categories, and often becomes harder to track.

The enterprise AI spending problem nobody's modeling

Most enterprise teams are still working with an outdated cost framework. AI is being treated like a project investment or a software license - a fixed or predictable expense with a clear scope. What Catanzaro, Naga, and Bar-Joseph are all describing is something structurally different: compute as a variable operating cost that scales with usage, fluctuates with pricing decisions made by AI providers upstream, and accumulates in ways that don't show up cleanly in traditional IT budgets.

Keith Lee, an AI and finance professor at the Swiss Institute of Artificial Intelligence, described the real benchmark: AI needs to become not just cheaper than human labor, but predictable and cost-controllable at scale. Right now, neither condition reliably holds for most production workloads. Frontier model pricing changes periodically. Usage creep is real. Rate limits and token costs create spending behavior that looks more like a cloud infrastructure bill than a software subscription - with all the unpredictability that implies.

If AI spending increasingly resembles payroll rather than a capital purchase, US enterprise teams need to manage it accordingly: with real budgets per team or product, governance structures, and cost attribution that CFOs can actually track. The organizations treating token spend as a line item they watch weekly will fare better than those absorbing it into a generic IT budget and hoping quarterly earnings tell a clean story. AI productivity gains are real in some contexts, but they don't automatically cover a compute bill that's scaling faster than the outputs.

What comes next on the compute cost curve

Compute costs are still declining at the unit level, and hardware improvements on Nvidia's own roadmap should continue to reduce per-token costs over the next few years. The question isn't whether the unit price drops - it's whether that decline will outpace the growth in usage. The answer, historically, is no.

Cloud computing is the template here. Cheaper compute consistently enabled more usage, and total AI spending tended to rise even as unit costs fell. The same dynamic is likely to play out in AI inference. Enterprise teams assuming that falling per-token costs will solve their AI budget problem are probably underestimating how much usage expands when the cost drops. Cheaper access and more ambitious AI applications typically move together.

For now, the AI ROI equation for most US enterprise teams depends less on whether AI is technically capable than on whether the organization can manage the operational cost of running it at scale. That's a finance and governance problem as much as a technology problem - and most organizations aren't treating it that way yet. The companies that get serious about compute cost management now, before the bills get larger, will have a meaningful operational advantage over those that wait.

Sources