AI compute costs are reshaping the economics of tech employment in ways that most boardrooms have not fully reckoned with yet. Bryan Catanzaro, Nvidia’s vice president of applied deep learning, made that blunt when he stated: “For my team, the cost of compute is far beyond the costs of the employees”. That admission, from an executive at the world’s dominant AI chip company, reframes the entire debate about whether AI saves businesses money — or simply redirects where the money goes.
Key Takeaways
- Nvidia’s VP of applied deep learning confirmed compute costs exceed employee salaries on his team.
- Uber’s entire 2026 AI coding tools budget was exhausted before the year was out, per the company’s CTO.
- AI software fees rose 20% to 37% in a single year, according to Tropic December data.
- A 2024 MIT study found AI automation is only economically viable in 23% of vision-primary roles — humans are cheaper in the remaining 77%.
- Jensen Huang envisions every engineer needing an annual token budget as a fourth compensation pillar alongside salary, bonuses, and equity.
How serious are AI compute costs in enterprise deployments?
AI compute costs in enterprise settings are not a rounding error — they are becoming a budget line that rivals payroll. At Nvidia, where the median employee compensation sits at around $300,000 in fiscal 2025, Catanzaro’s team is spending even more on raw compute. At Uber, CTO Praveen Neppalli Naga confirmed the company’s 2026 budget for AI coding tools had been “blown away already”. These are not startups burning venture capital — these are mature, cash-generating businesses watching AI expenditure outpace their hiring plans.
The broader numbers back this up. AI software fees climbed 20% to 37% over the past year, per Tropic’s December data. McKinsey projects total AI expenditures could reach $5.2 trillion by 2030, broken down as $1.6 trillion for data centers and $3.3 trillion for IT equipment — with an accelerated scenario pushing that figure to $7.9 trillion. At those scales, compute is not a cost center. It is the cost center.
Why companies are not treating AI compute costs as a problem
Here is the counterintuitive part: many companies experiencing these ballooning AI compute costs are not alarmed. They are doubling down. Nvidia hired 6,000 additional employees in fiscal 2026 specifically to grow its AI business, even as the compute bill dwarfs its already-generous payroll. The logic is that AI-augmented engineers produce output that justifies the spend — even if that spend is harder to quantify than a salary line.
Keith Lee, an AI and finance professor at the Swiss Institute of Artificial Intelligence’s Gordon School of Business, describes the current situation as “a short-term mismatch” driven by hardware and energy costs. His view is that some firms are beginning to treat AI not as a labor substitute but as a complementary tool — at least until cost structures stabilize. That framing matters. If AI is a tool, its cost is overhead. If it is a replacement for workers, its cost is a failed business case.
AI compute costs vs. actual workforce reductions: the uncomfortable math
The tension between AI spending and workforce decisions is sharpest when you look at what companies are doing simultaneously. Meta plans to cut roughly 10% of its workforce — around 8,000 employees — and scrap hiring for 6,000 additional positions, partly to offset AI investment costs. Microsoft offered its largest-ever voluntary buyout program to thousands of employees. These are not small adjustments. They are structural bets that AI output will eventually justify the displacement.
But the 2024 MIT study complicates that bet. Researchers found AI automation is only economically viable for human-level performance in 23% of vision-primary roles — meaning humans remain the cheaper option in 77% of those cases. That is not a ringing endorsement for wholesale replacement. It suggests the companies laying off thousands while paying more for compute than for people are making a speculative investment, not a proven efficiency play.
Is the token budget the future of AI compute costs management?
Jensen Huang has a specific answer to the compute cost problem: token budgets. The Nvidia CEO envisions a future where every engineer receives an annual token budget as a fourth compensation pillar, sitting alongside salary, bonuses, and equity. The concept treats AI access the way companies once treated travel allowances or equipment budgets — a defined resource allocation per person.
The math behind this is not trivial. OpenAI charges approximately $15 per million tokens for advanced models. Roughly 750 words consume around 1,000 tokens, but coding tasks and AI agents burn through tokens far faster — some engineers incur thousands of dollars per day in usage. Companies like Zapier and Kumo AI are already tracking token consumption per employee to optimize efficiency. Token budgets are not a hypothetical future state. They are an operational reality for organizations paying close attention to where the money actually goes.
Is AI actually cheaper than hiring human workers?
Not always, and often not at scale. The MIT 2024 study found that AI automation is economically viable compared to human workers in only 23% of vision-primary roles. In high-usage enterprise environments, AI compute costs can exceed individual employee salaries — as Nvidia’s own VP confirmed. The cost advantage depends heavily on the specific task, usage volume, and how efficiently a company manages its AI consumption.
What are token budgets and why do they matter?
Token budgets are proposed per-employee allocations of AI compute access, similar to expense allowances. Jensen Huang has suggested they could become a standard compensation element for engineers. With OpenAI’s advanced models priced at around $15 per million tokens, and heavy users potentially spending thousands of dollars per day, token budgets give companies a way to control AI compute costs without cutting access entirely.
How are companies responding to rising AI software costs?
Responses vary. Some firms, like Meta and Microsoft, are reducing headcount to offset AI investment. Others, like Nvidia, are hiring aggressively while accepting that compute costs will exceed payroll. A growing number are tracking per-employee token consumption — Zapier and Kumo AI among them — to find efficiency without sacrificing capability. Tropic’s data showing AI software fees up 20% to 37% in a year suggests the pressure to manage these costs will only intensify.
The AI cost reckoning is real, and it is not going away. Companies that treat compute spend as a temporary growing pain are likely to be surprised by how permanent the new cost structure turns out to be. The smarter bet is to start treating AI access the way Huang suggests — as a budgeted resource with accountability attached — rather than an open tab that finance will eventually demand someone close.
This article was written with AI assistance and editorially reviewed.
Source: Tom's Hardware
