AI compute costs now surpass human salaries, senior tech executives warn
Senior executives at Nvidia and Uber have disclosed that AI compute costs are outpacing human payroll, casting doubt on the financial logic of rapid AI adoption as big tech companies pursue sweeping layoffs and token bills mount.
- Nvidia VP Bryan Catanzaro confirms compute costs exceed employee costs on his team.
- Uber exhausted its full-year AI budget within the first four months of 2026.
- A 2024 MIT study found human workers remain cheaper in 77 per cent of roles examined.
The cost of deploying artificial intelligence tools has begun to outstrip what companies pay their human workers, according to senior executives at two of the technology industry's most prominent firms.
Bryan Catanzaro, Nvidia's Vice President of Applied Deep Learning, told Axios that for his team, compute costs now far exceed employee costs.
The admission carries particular weight coming from Nvidia, the company whose graphics processing units form the hardware foundation of the global AI build-out.
Catanzaro's assessment is shared by Uber's Chief Technology Officer, Praveen Neppalli Naga, who disclosed to The Information that the ride-hailing platform had exhausted its full-year AI budget just four months into 2026.
The disclosures arrive as major technology companies are simultaneously cutting headcount while increasing AI investment, a combination attracting growing scrutiny over the industry's stated rationale for replacing workers with automated tools.
Meta Platforms announced last month that it would lay off approximately 8,000 employees. Microsoft has separately offered voluntary retirement packages to nearly 9,000 workers. Both companies have characterised the decisions as efficiency measures linked to AI adoption.
Evidence that productivity gains have materialised is, however, mixed. Amazon employees told The Guardian in March 2026 that AI tools had in some cases reduced their productivity rather than improved it.
Academic research lends weight to those concerns. A 2024 study by the Massachusetts Institute of Technology (MIT) found that in 77 per cent of roles examined, human workers remained both preferable and more cost-effective than AI alternatives.
The cost of continuous AI deployment
The pricing model driving many of the largest bills is token-based charging, under which costs accumulate with each AI interaction rather than being capped at a flat fee.
Autonomous AI agents configured to run continuously — executing tasks such as code review, data processing, or customer queries on a recurring schedule — can generate substantial expenditure with limited human oversight.
Swan AI chief executive Amos Bar-Joseph illustrated the scale of such costs in a LinkedIn post, disclosing that his four-person team had incurred a monthly bill of US$113,000 from Anthropic, the maker of Claude.
Simplified arithmetic puts that figure at approximately US$28,000 per person per month, a sum likely exceeding most team members' monthly salaries.
Current pricing from major AI providers is widely understood to be subsidised by venture capital and investor funding rather than reflecting actual operational costs.
Industry analysts have warned that enterprises building deep dependencies on these services may face significantly higher bills once subsidised pricing is withdrawn.
Observers have drawn comparisons to previous technology adoption cycles, notably cloud computing, in which accessible early pricing gave way to significant cost increases once enterprise lock-in had been achieved.
Diverging views on value
Not all senior figures interpret the rising bills as a warning sign. Nvidia chief executive Jensen Huang has argued that high token expenditure is itself a proxy for productivity.
Huang drew attention by suggesting he would be concerned if an engineer earning US$500,000 a year failed to consume at least US$250,000 in AI tokens annually.
The comment prompted a trend known as tokenmaxxing, in which workers compete to maximise their AI token expenditure as a visible signal of activity, with some employees reportedly accumulating bills of US$150,000 per month.
In a social media post dated 18 May 2026, Huang argued that AI would assume routine tasks, freeing workers to concentrate on higher-value responsibilities.
He predicted that AI-driven productivity would multiply global gross domestic product manifold, and expressed confidence that the resulting economic expansion would generate new employment rather than eliminate it.
Uber's Neppalli Naga shares a similar long-term outlook, envisioning AI agents supervised by other AI agents eventually assuming the roles of software engineers. He noted that approximately 11 per cent of Uber's live code updates are already generated by AI.
Questions of sustainability
Whether current AI spending represents a transitional cost — incurred while automation matures ahead of reducing headcount — or a permanent operating expense is a question without a settled answer.
Recent analysis suggests that companies which deployed AI without clearly defined business objectives have generally incurred net losses on those initiatives rather than savings.
For investors and corporate decision-makers, a central question is how AI providers will manage pricing as the industry matures: sustaining subsidised rates, moving toward cost-reflective pricing, or pursuing efficiency gains through smaller, more specialised models.
Proprietary deployments using locally hosted models represent one cost-reduction pathway available to enterprises with the technical capacity to implement them. Some companies have reported reducing monthly AI expenditure substantially by combining local models with frontier services only for tasks local infrastructure cannot handle.
The answer is likely to depend on competition between providers, and on whether AI capabilities improve sufficiently to justify costs that, in the majority of workplace contexts, currently outpace those of human alternatives.
