Artificial intelligence has been sold for two years with a very simple promise: to do more with less. Fewer personnel, fewer repetitive tasks, less downtime, and increased productivity. But the business reality is beginning to look much messier. Some tech executives are acknowledging an awkward truth: in certain advanced applications, AI doesn’t turn out to be cheaper than human workers. Sometimes, it’s much more expensive.
The phrase that has ignited the debate comes from Bryan Catanzaro, Vice President of Applied Deep Learning at NVIDIA. According to several American media outlets, Catanzaro explained that in his team, the computing cost of AI is “well above” the expense of employees. He wasn’t talking about a basic chatbot subscription, but about advanced systems—agents, models with large context windows, intensive inference, and tools that consume millions of tokens to review code, execute tasks, or analyze entire repositories.
This data seems to contrast with the narrative of layoffs. If AI is so costly, why do so many companies continue to cut staff and justify part of those cuts with automation? The answer isn’t that companies have found a cheap machine that perfectly replaces humans. It’s more uncomfortable: many firms are firing based on expectations, financial pressure, internal reorganizations, and strategic bets, not because AI has proven to be more efficient across all roles.
The Hidden Cost of Tokens
For individual users, AI appears inexpensive. A monthly subscription can cost less than a business lunch. That price creates a misleading perception when scaled to the business world. In a company, AI use isn’t limited to asking questions in a chat. It involves thousands of calls, long documents, agents executing reasoning cycles, code reviews, repository indexing, automated testing, logs, connected tools, and constant consumption of input, output, and context tokens.
This is where the bill changes. A programming agent doesn’t just respond once. It reads files, makes changes, corrects errors, retests, analyzes dependencies, and repeats the cycle. Each step consumes tokens. If it uses high-level models, the cost skyrockets. The difference between “using AI” and “operating a fleet of agents” is similar to the difference between opening a spreadsheet and maintaining a cloud infrastructure.
Swan AI’s case has become one of the most cited examples. Its CEO, Amos Bar-Joseph, reported that a four-person company had accumulated a monthly bill of $113,421.87 in Anthropic. He presented it as a point of pride, not a failure: his thesis is to build a company with high automation and minimal staff. But this figure clearly shows that the cost of enterprise AI can easily surpass what many imagined.
Uber has also acknowledged similar tensions. Praveen Neppalli Naga, the company’s CTO, explained that using AI tools for development had consumed the planned budget much faster than expected. At the same time, around 11% of backend updates in production are now being written by AI agents. That means the technology is starting to produce real results, but not necessarily with the almost zero marginal cost some assumed.
| Visible Cost | Underestimated Cost for Many Companies |
|---|---|
| Monthly subscription to a tool | Actual token consumption per user, agent, or task |
| Per-seat license | Inference costs for advanced models in long workflows |
| “Automating a task” | Integration, supervision, security, and error correction |
| Human hours saved | Compute costs, auditing, and vendor dependence |
| Fewer staff | More cloud spending, more tools, and increased operational control |
So, why are layoffs happening?
The first reason is financial. Labor costs are visible, recurring, and easy to cut in financial statements. Salaries, benefits, contributions, offices, management, intermediate layers, and entire structures are listed as fixed expenses. AI, on the other hand, is often presented as a strategic investment, innovation budget, or a variable cost that could decrease over time. Although it’s expensive today, many boards believe its unit cost will fall, and future productivity gains will offset this expense.
The second reason is market pressure. Since 2023, investors have rewarded companies promising efficiency, automation, and AI focus. In this context, announcing cuts and reallocating capital to AI sends a signal: the company is adapting. This doesn’t always mean a direct replacement of a dismissed person with an AI agent the next day. Sometimes, it means the company is reducing mature areas to fund data centers, licenses, models, chips, or AI teams.
The third reason is organizational. AI doesn’t replace all roles equally. It might reduce demand for junior roles, support tasks, basic documentation, repetitive QA, initial analysis, first-level customer support, or low-value content creation. At the same time, it increases the need for senior profiles, architects, data specialists, security, infrastructure, AI governance, and human oversight. The outcome isn’t always “less work,” but rather a different composition of work.
The fourth reason is that many companies are acting proactively. Harvard Business Review describes this phenomenon as layoffs motivated by AI potential, not necessarily by proven performance. It’s a gamble: cut now, expecting tools to mature quickly. This gamble may succeed in very narrow tasks but fail in roles where contextual knowledge, customer relationships, or human judgment still weigh heavily.
And a less elegant fifth reason: “AI-washing” layoffs. Some companies might use AI as a convenient excuse for decisions also driven by over-hiring, economic slowdown, margin pressures, strategic changes, or simple cost-cutting. Blaming AI sounds more modern than admitting poor workforce planning.
The Mistake of Comparing a Person to a Tool
The question “Is AI cheaper than an employee?” is poorly formulated if asked in a general way. It depends on the task, volume, model, supervision costs, and potential errors. For simple internal emails, it can be cheap. But replacing a team maintaining critical systems, reviewing sensitive code, or working with complex clients changes the equation.
Additionally, the comparison should consider total cost. It’s not enough to look at tokens. You must also include integration, security, training, permission management, auditing, errors, hallucinations, vendor reliance, regulatory compliance, and human review. In many companies, the first year of AI adoption may be more expensive because costs are doubled: maintaining the existing staff and adding AI infrastructure. Savings, if they appear, come later and not always where expected.
This explains why some executives can simultaneously say AI is more expensive than workers and still invest heavily in it. They’re not after immediate savings alone. They seek strategic options—learning faster than competitors, redesigning processes, capturing data, reducing future dependence on repetitive tasks, and preparing for costs to fall.
The paradox is that this race can lead to layoffs before clear returns are proven. Companies don’t wait for perfect AI; they reorganize staff to adapt. In some cases, this results in more efficient companies. In others, smaller, more pressured teams with AI bills that don’t offset internal knowledge loss.
The serious debate shouldn’t be whether AI “takes jobs” or “creates jobs” in abstract. It’s already changing employment. The real questions are which tasks are eliminated, which skills are amplified, what hidden costs remain, and who bears the risk when a company replaces human expertise with immature automation.
It’s becoming clear that AI isn’t a magic lever for savings. It’s powerful, costly in intensive uses, and hard to measure in real processes. Companies that treat it as just a way to cut personnel may be surprised. Those that see it as an investment involving redesign, cost control, and human oversight will have a better chance of turning it into real productivity.
Frequently Asked Questions
Can AI be more expensive than an employee?
Yes, in intensive applications with agents, advanced models, large volumes of tokens, code review, or tasks requiring many inference cycles. While not true in all cases, some companies are already seeing this happen.
Why do companies lay off workers if AI is still expensive?
Because they aim to reduce fixed costs, reallocate budgets to AI, respond to investor pressures, and anticipate future productivity—even if that future isn’t fully proven yet.
Is the cost of AI only the monthly subscription?
No. Companies also account for tokens, integration, cloud infrastructure, security, supervision, auditing, errors, training, and vendor dependence.
Which jobs are most vulnerable?
Repetitive tasks, documentation, basic support, initial analysis, routine QA, or low-value content creation are easier to automate. Roles requiring judgment, responsibility, context, and human interaction are harder to fully replace.