What is tokenmaxxing?

The term “tokenmaxxing” reflects the role of tokens as the basic unit of input and output for large language models (LLMs) and the agentic AI systems they power. In this context, one’s token spend—the literal amount of input/output tokens an employee (or amount of money spent on those tokens)—is taken to serve as a reflection of how much that person utilizes AI in their work. 

Tokenmaxxing is typically enforced both formally and informally through measures including tracking teams’ or individual employees’ token consumption through usage leaderboards, penalizing a failure to fully utilize allotted token budgets or ostensibly making employment decisions based on observed AI usage. 

After experiencing its peak in the spring of 2026, tokenmaxxing’s popularity as an approach to AI adoption has largely begun to wane. 

The rise of tokenmaxxing had a number of interrelated causes, most of which have their roots in the increased pressure to realize meaningful return on investment (ROI) on AI spend, directly or indirectly.

The use of generative AI, and particularly the implementation of AI agents, has scaled far faster than the frameworks used to measure their productivity. “In the absence of true metrics, organizations created usage leaderboards, which people quickly learned to game,” said IBM’s Neil Dhar, SVP Consulting, in a Linked post. “[AI] usage soon became a proxy for value.” 

In other words, tokenmaxxing was often borne out of the naive assumption that any increased AI usage increased overall productivity, and therefore that any increase in token spend would eventually correlate to greater productivity. 

Many of the voices most loudly advocating for tokenmaxxing had much to gain from enterprises increasing their AI budgets.

This interplay of economic incentives and industry narratives was often straightforward and intuitive. For example, in March 2026, Nvidia CEO Jensen Huang made headlines after stating on a podcast appearance that he’d be “deeply alarmed” if a software developer with a salary of $500,000 did not spend at least $250,000 on tokens each year. Given that Nvidia’s status as the industry’s preeminent producer of the AI hardware running all this AI inference has made it the world’s largest company by market capitalization, Huang’s comments carry a lot of weight—and he and his company directly profit from using his influence in such a way.

But given the remarkably interconnected web of major AI investment, this kind of economic self-interest is sometimes less obvious. 

For instance, one April 2026 Wall Street Journal story quoted a partner at Sequoia Capital, one of the world’s largest venture capital firms, asserting that “we all should be tokenmaxxing.” She went even further, stating that “what matters most for your company is: ‘Has my employee become insanely AI-pilled?’” It’s imporant to note that Sequoia Capital’s portfolio includes Nvidia, xAI, Anthropic, OpenAI, Google, Hugging Face, LangChain and Github, among dozens of other well-known companies specializing in AI coding agents, data platforms and other AI infrastructure.

While the existence of these economic incentives doesn’t necessarily mean that these prominent individuals don’t genuinely believe what they’re saying, it does explain the eagerness with which many of leading industry figures promoted tokenmaxxing as sound business practice.

The immense economic potential of AI transformation creates intense competition between businesses seeking to gain an early-mover advantage that could translate to future market share. This competition, in turn, created pressure to accelerate a company’s transition to AI-native workflows as rapidly as possible. 

One underlying assumption of tokenmaxxing in this case is that incentivizing (or tacitly forcing) employees to maximize their AI usage and experimentation would both speed up their skill and comfort with AI tools and minimize the time needed to realize transformative change. In practice, tokenmaxxing acts as an effort to internally crowdsource AI-native “killer apps” for the company. 

Many major corporations enforced the company-wide embrace of AI through internal leaderboards, particularly (but not exclusively) for software engineering teams. Meta, for example, had a (now-discontinued) usage leaderboard dubbed “Claudeonomics” (reflecting the popularity of Anthropic’s Claude Code in particular) in which employees competed for elite usage ranks like “Token Legend” and “Session Immortal.” According to reporting from The Information in April 2026, “the Claudeonomics dashboard’s competitive dynamic [had] helped it gain traction internally.”

With the potential for AI-driven productivity increases comes the pressure to realize that potential, whether due to personal goals or fear of threats, direct or indirect, to one’s employment. Tokenmaxxing emerges as a natural behavioral response to such pressure. 

Andrej Karpathy, the influential AI developer and industry veteran credited with coining the term “vibe coding,” articulated his experience with productivity pressure in a conversation with host Sarah Guo in an April 2026 episode of her No Priors podcast. “I feel nervous when I have [Claude or Codex] subscription left over,” he explained. “That just means I haven’t maximized my token throughput.” 

This was in response to Guo’s description of her own experience with this attitude. “I do feel like my instinct is…if I have access to more tokens then I should I just parallelize tasks,” she said. “That’s very stressful, because if you don’t feel very bounded by your ability to spend on tokens, then, you know, you are the bottleneck in this system.”

Despite the intuitive upsides of encouraging employees to experiment with emerging AI tools, maximizing token consumption unto itself as formal practice typically makes increased token burn a vanity metric that prioritizes conspicuous consumption over meaningful AI productivity. 

These fundamental flaws, described in the following sections, have resulted in the general decline of tokenmaxxing as a corporate policy, formally or informally.

Tokenmaxxing, particularly when formally tracked in token leaderboards, often runs into Goodhart’s law: “when a measure becomes a target, it ceases to be a good target.” 

For example, Financial Times reported that Amazon staff admitted to inflating their token spend numbers by using AI for irrelevant tasks in an effort to hit their internal usage targets. Though Amazon leadership had ostensibly stated that token usage tracking wouldn’t be used in performance evaluations, multiple Amazon employees told FT that they suspected otherwise and felt enormous pressure to maximize AI token usage. One employee described how token spend on internal leaderboards created “perverse incentives.”

An emphasis on indiscriminately maximizing token spend can have drastic and predictable financial consequences. Among other reasons, the non-deterministic workflows of agentic AI can result in token spend increasing exponentially and unpredictably.

For example:

• In May of 2026, Uber acknowledged publicly that it had spent its entire 2026 AI budget by the end of April. Uber’s COO noted in public interviews that it was getting “harder to justify” spend on tokenmaxxing, given that the company couldn’t prove a connection between increased AI token consumption and improved customer features or increased developer productivity.

• Halfway through the year, Microsoft cut off Claude Code licenses for its software engineers. The Verge reported that despite public statements to the contrary, this was largely a financial motivated decision in response to spiraling costs. 

• An AI consultant told Axios that one client spent “half a billion dollars in a single month” after failing to impose limits on Claude Code licenses for employees. Follow-up reports implied that the company in question was Amazon, given that they had just recently deprecated their internal leaderboard, and Business Insider reported that a senior vice president at the company had told staff, “Please don’t use AI just for the sake of using AI.”

Incentivizing massive amounts of vibe coding by employees with no experience in software development often introduces serious security risks, as well as broader code quality issues and degraded developer experience for the experienced software engineering staff tasked with reviewing and cleaning up that AI-generated code.