Until recently, a token was, to most, an invisible technical unit. A fragment of text processed by models such as ChatGPT, Claude or Gemini. An infrastructure cost, a usage measure, and a figure for engineers and suppliers. But in Silicon Valley that unit begins to take on another meaning and no longer only measures interaction with AI: it begins to function as a marker of productivity, internal prestige and even labor compensation. A twist that has been baptized with its own name: tokenmaxxing.
The expression describes a practice increasingly visible in some of the most influential technology companies: maximizing the consumption of artificial intelligence tokens as a sign of technical sophistication, cultural alignment and performance. Thus, the more tokens are “burned”, the more AI is deduced to be being used; and the more AI is used, the higher the perception of productivity appears to be. A reasoning that is already generating as much traction as skepticism.
The discussion accelerated when Nvidia CEO Jensen Huang proposed during the GTC 2026 conference an idea that, just a few months ago, would have seemed outlandish: offering engineers an annual token allocation equal to about half their base salary. Thus, if a developer earns $500,000 a year, the logic would be to add about $250,000 in computing so that, in theory, they multiply their productivity tenfold. Huang even said that he would “on high alert” if an engineer with that salary barely consumed $5,000 a year in tokens.
The proposal of Huang, who directs the company that manufactures a good part of the chips on which the market’s AI is executed, entails an interest in promoting a growing consumption of it: without computing there are no tokens, and without tokens there is no income growth. Your data centers are, in their own words, “token factories”, So, from that perspective, turning the intensive use of AI into a new cultural norm within engineering benefits the entire chain, starting with Nvidia.
As agentic tools gain capacity, the use of AI is no longer ad hoc
The idea has found fertile ground because it is part of a deeper transformation of technical work. As agentic tools gain capacity, the use of AI stops being punctual and begins to organize entire processes. It’s no longer about asking for help writing a function or reviewing a block of code. Agents can now work for hours without supervision, split tasks, spawn subagents, review entire repositories, and continue producing while their user sleeps. In this new operating regime, token consumption grows exponentially.
There appears the second engine of the tokenmaxxing: anxiety. In companies like Meta, Shopify or OpenAI, the use of AI is beginning to enter performance evaluations, and some information points to internal rankings that show how many tokens each employee consumes. In this context, volume ceases to be a neutral figure and becomes a sign of adaptation to the new environment. Not using AI intensively can be read as lag. Using it a lot, on the other hand, projects the image of someone aligned with the future. It is the logic of a culture that does not yet know how to properly measure the value of artificial intelligence, but does know how to count how much is spent on it.
The figures circulating on the Internet help to understand the magnitude of the phenomenon. An OpenAI engineer would have processed 210 billion tokens in a week, the approximate equivalent of 33 entire Wikipedias. At Anthropic, a single Claude Code user racked up a bill of over $150,000 in one month. A developer in Stockholm summarizes this disproportion with a phrase that has become almost emblematic of the moment: “I probably spend more on Claude than on my own salary.” It is a sign of the type of economy that is beginning to take shape around AI: one in which the cost of computational assistance can approach or even exceed the cost of the human labor that coordinates it.
The point is that this volume does not solve the central problem by itself. Consuming more does not automatically mean producing better. An engineer can use huge amounts of tokens to judiciously automate complex tasks and generate real value. But another may launch parallel processes, open multiple agents, and skyrocket spend without an equivalent improvement in quality, accuracy, or impact. If the organization rewards the volume of use above all, the incentive stops being “work better with AI” and becomes “demonstrate that a lot of AI is used.” It is a small difference in appearance, but enormous in its consequences.
That is why criticism has begun to proliferate. Gergely Orosz, an analyst in the world of software engineering, has pointed out that the problem is not in encouraging teams to use tools that improve their productivity. The problem is taking the price of these tools as a central criterion. Some of the most useful ones are cheap. Some of the more expensive ones may be spectacular from a business perspective, but not necessarily more valuable in the actual work. Productivity, remember that reading, should not be measured by tokens consumed, but by results obtained.
The analogy most often repeated in these critiques is revealing: evaluating employees by the number of tokens they “burn” is similar to measuring soldiers by the number of bullets fired. There is activity, but not necessarily effectiveness. There is movement, but not necessarily progress. In the end, the tokenmaxxing exposes an old weakness of corporate culture: when companies do not know how to measure what is important, they tend to measure what is easy to count.
Unlike salary or share capital, tokens do not generate equity, do not accumulate and are not transferable
In this sense, tokens are entering the salary conversation in an ambiguous way. Investors like Tomasz Tunguz already describe AI inference as a fourth component of engineer compensation, alongside salary, bonuses and stock. On paper, the idea makes sense: if massive access to computing allows a worker to multiply their production, that access is a valuable asset and can be part of the hiring package. But that logic becomes less compelling once you look at what kind of value the employee actually receives.
Unlike salary or share capital, tokens do not generate wealth, do not accumulate and are not transferable as economic security. They are operational capacity, not wealth. A worker may have hundreds of thousands of dollars in computing power, but that does not necessarily improve his or her financial position or future bargaining power. At best, it equips you with tools. At worst, it disguises a remuneration package that seems more generous without increasing the items that do consolidate stability.
There is also a tougher business reading. If a company assigns each engineer the equivalent of a second digital workforce in the form of agents and computing, the implicit expectation is that that worker will produce much more because they now have at their disposal an expensive infrastructure whose use must be justified. At that point, the token stops resembling a benefit and begins to operate as a pressure mechanism.
For now, all of this remains in the cultural experimentation phase. Silicon Valley seeks new metrics for a new era as AI vendors celebrate the growth of consumption. The most ambitious engineers try to prove that they are not being left behind. And managers, meanwhile, observe whether that intensity translates into better products, more speed and greater competitive advantage. It is still not clear whether the tokenmaxxing It is a preview of how technical work will be organized in the coming years or a symptom of an industry that, trapped between fear and euphoria, is confusing activity with value.
What does seem clear is that the conversation has changed. Tokens are beginning to be a cultural unit, a way to measure adoption, distribute status and rewrite, perhaps prematurely, what it means to perform in the artificial intelligence economy.
