AI Agents More Expensive Than Employees: Tokenmaxxing Drains Corporate Budgets
Indulging in the excessive use of AI agents without strict cost control is proving to be a very costly mistake for numerous tech companies. According to Axios, some organizations find themselves spending on AI tokens more than they would pay to the human employees that those same agents are supposed to replace or support.
Every request to a language model has a cost measured in tokens, the text units that the model processes. When it comes to agents operating autonomously on complex tasks, often in parallel and without continuous supervision, the consumption is effectively continuous and can escalate to significant levels. For example, Bryan Catanzaro, VP of applied deep learning at Nvidia, stated quite succinctly: "For my team, the cost of compute is far greater than the cost of employees."
Tokenmaxxing - The Performance Anxiety
Adding to the difficulty of managing this situation is a practice that has emerged among developers known as "tokenmaxxing." The term describes the extreme and deliberately exaggerated consumption of tokens, turned into a kind of informal competition. According to the New York Times, an engineer at OpenAI processed 210 billion tokens in a single week, a volume that corresponds to about 33 times the number of tokens contained in a corpus like Wikipedia. A single user of Claude Code has accumulated costs exceeding $150,000 in one month.
These are not isolated cases. Max Linder, a software engineer from Stockholm, told the Times: "I probably spend on Claude more than I earn as a salary," clarifying that the cost falls on the company. Uber engineers using Claude Code have already exhausted the entire 2026 AI budget of the company, according to reports from The Information. Data from Ramp confirm a growth of 13 times in the average monthly spending on tokens among monitored companies, from January 2025 to the present.
The Pragmatic Engineer has systematically analyzed the phenomenon, arriving at uncomfortable conclusions: tokenmaxxing leads to an exponential increase in spending with minimal or no returns. At the highest usage levels, one achieves about double the throughput at a token cost ten times higher, an equation that rarely justifies itself. Worse still, it encourages developers to produce unnecessary work: prompts on already documented topics, starting projects that will never go into production, AI-generated answers to questions whose outcomes are known. At Meta, employees have consumed a total of 60.2 trillion tokens in 30 days; calculated at Anthropic's API prices, the bill would reach $900 million, even discounting enterprise rates.
Tokens as Salary Components
Jensen Huang's response to the problem, as we know, is not to contain spending, but to formalize it as a salary component. At the 2026 GTC in San José, Nvidia’s CEO suggested granting employees an annual AI token budget equal to about 50% of their base salary. Huang explicitly framed this mechanism as a hiring tool, arguing that an engineer with access to an adequate volume of tokens can multiply their productivity by ten times. He added that if a $500,000 engineer claimed to spend only $5,000 on tokens, it would be a warning sign.
Nvidia's proposal highlights a crucial issue for the entire software industry. An AI agent operating via a cloud API at full capacity can cost about $300 per day, or roughly $100,000 per year per agent, according to data discussed in the All-In podcast and confirmed by analyses from CIO.com. When multiplied by entire development teams, the expenses quickly surpass personnel costs. Chamath Palihapitiya, CEO of Social Capital, has already introduced token budgets per developer in his team as a containment measure—a choice that several CTOs are considering.
The Ecosystem Driving Usage and Profit
The pressure towards AI adoption does not just come from below. Boris Cherny, head of Claude Code at Anthropic, stated that "practically 100%" of the code produced by the company is now generated by AI. Google and Microsoft each hover around 25%. At Meta, employee performance reviews also include metrics on the use of AI tools, effectively creating pressure to consume tokens regardless of actual usefulness.
This situation can only benefit service providers and AI model suppliers, and Anthropic itself is facing increased demand with a possible restructuring of its offering, pricing tiers, and what is included in each type of profile. All this pushes priority on the concept of token efficiency, through both active strategies and selecting models known for more restrained token consumption.
However, the fundamental question remains unanswered, at least until now: what is the real value of this spending? More studies indicate that forcing the use of AI tools can make work harder rather than easier, and the episodes of defective code produced en masse by unsupervised agents at companies like Meta and Amazon fuel legitimate doubts about output quality. The speed at which lines of code are generated says nothing about the solidity of the resulting software, and the costs derived from the massive generation of tokens might, in many cases, not yield a measurable return.