Amazon employees are reportedly participating in a practice known as "tokenmaxxing," a phenomenon where individuals intentionally inflate their usage of internal AI tools to boost their standing on internal metrics, according to a recent report by the Financial Times. This behavior, driven by employer incentives that reward extensive AI utilization quantified by "tokens," has raised concerns among industry analysts about the potential for gamification and the true measure of productivity in the rapidly evolving AI landscape.
The trend is not unique to Amazon. Similar practices have been observed at other major technology companies, or "hyperscalers," including Microsoft and Meta. These tech giants are heavily invested in the very AI technologies they are encouraging their employees to adopt, suggesting a strategic imperative to drive internal adoption and feedback loops. Amazon, for instance, recently reported that the increased valuation of its stake in AI firm Anthropic contributed nearly half of the company’s profits, underscoring the significant financial implications of AI advancements for these corporations.
"Tokenmaxxing" involves employees running AI tools on trivial or repetitive tasks specifically to increase their token counts, thereby climbing leaderboards that measure usage. While the direct impact of these leaderboards on performance reviews at Amazon is reportedly not explicitly tied to compensation or advancement, multiple employees have expressed concerns to the Financial Times that managers may still be observing these metrics. This creates an environment where perceived pressure to demonstrate high AI usage can overshadow genuine productivity gains.
The Rise of "Tokenmaxxing" and its Roots
The concept of "tokenmaxxing" emerged as hyperscalers began to quantify employee engagement with AI tools. Tokens serve as a unit of measurement for the computational resources consumed by AI models during tasks such as text generation, data analysis, or code completion. Companies, eager to demonstrate progress and ROI on their substantial AI investments, began to incentivize usage. This incentive structure, however, has inadvertently created an environment where the act of using AI can be prioritized over the effectiveness of its application.
Gil Luria, head of technology research at D.A. Davidson, voiced his apprehension regarding this dynamic. "That doesn’t sound very healthy," Luria commented to Fortune. "You get the behavior that you create the incentive for. So if you tell people they’ll succeed if they use a resource more, of course they’ll use it more." This sentiment highlights a fundamental challenge in driving technological adoption: ensuring that incentives align with genuine productivity and innovation rather than superficial engagement.
Luria further elaborated that while the power and potential of AI tools to enhance productivity are undeniable, the primary hurdle lies in their widespread and effective diffusion within organizations. "Humans are rigid in how they do things," Luria explained. "So if you don’t create an incentive for humans to change their behavior, try something new, most of us won’t." The "tokenmaxxing" phenomenon, in this context, can be seen as an unintended consequence of attempting to overcome this human inertia.
The issue is further contextualized by Goodhart’s Law, which posits that "when a measure becomes a target, it ceases to be a good measure." In the case of token counts, when the metric of usage becomes the primary target, its ability to accurately reflect true AI-driven productivity diminishes. The pressure to "tokenmax" can lead to a situation where employees are expending computational resources without achieving commensurate business value, creating an illusion of progress.
Broader Trends Across Tech Giants
The observed "tokenmaxxing" behavior at Amazon mirrors similar trends reported at other major technology companies, indicating a widespread challenge in integrating AI effectively. At Meta, for example, an internal leaderboard dubbed "Claudeonomics" was reportedly created by an employee, ranking the company’s approximately 85,000 workers by their token consumption. In a mere 30-day period, the aggregate usage on this dashboard surpassed 60 trillion tokens. Notably, neither CEO Mark Zuckerberg nor CTO Andrew Bosworth appeared among the top 250 users on this leaderboard.
The "Claudeonomics" dashboard was eventually taken down following reporting by The Information. However, Meta’s CTO, Andrew Bosworth, has publicly acknowledged and even endorsed the underlying principle of incentivizing AI usage. He shared an anecdote about a top engineer whose AI token expenditure was equivalent to their salary but resulted in a productivity increase of "5x to 10x." Bosworth’s response, as quoted in Forbes, was "It’s like, this is easy money. Keep doing it. No limit." This perspective suggests a belief that the cost of AI computation is outweighed by the potential for significant productivity gains, even if the measurement of those gains is somewhat indirect.
These developments raise critical questions about how companies are measuring success in their AI adoption strategies. While the intention behind incentivizing AI usage is to foster innovation and efficiency, the current mechanisms may be inadvertently encouraging a form of "gaming the system." The challenge for these hyperscalers lies in designing incentive structures that promote genuine AI-driven value creation rather than simply maximizing token consumption.
Financial Implications and Market Dynamics
The stakes for these hyperscalers are immense, with significant financial investments being poured into AI infrastructure and development. Combined capital expenditure from Amazon, Microsoft, Alphabet (Google’s parent company), and Meta is projected to reach staggering figures. In 2026, this expenditure is already pushing towards $700 billion, with some Wall Street projections exceeding $1 trillion for 2027. This represents a substantial increase from just under $400 billion in 2025, illustrating the accelerating pace of investment in AI.
These companies are reporting to investors that their inference chips, essential for running AI models, are being consumed as rapidly as they are deployed. This demand is fueled by the growing need for computational power to train and operate increasingly sophisticated AI systems. However, a complex dynamic is at play, described by Luria as "circular activity." This refers to the fact that these same technology giants are also investing in their suppliers and customers within the AI ecosystem. For instance, Amazon’s investment in Anthropic positions it as both a user and a stakeholder in an AI company.
This intertwined relationship creates what Luria terms an "overhang around all of the large technology companies, especially Amazon, Google, Microsoft, Meta, Nvidia." The reliance on AI, coupled with cross-investments, can create a self-reinforcing market where the success of one company is intrinsically linked to the success of others, potentially obscuring underlying market fundamentals.
The demand for AI services is currently unprecedented. Leading AI companies like OpenAI and Anthropic are operating at a combined annual revenue run rate exceeding $70 billion, a dramatic increase from virtually zero just two years ago. Luria emphasizes that "Those companies actually represent real economic activity. That is consumers and businesses paying for access to their model." This indicates a genuine market demand for AI capabilities beyond the internal usage within hyperscalers.
While hyperscalers are significant consumers of AI services and developers of AI technology, Luria posits that they are not disproportionately driving this revenue growth solely through their internal programming teams. Instead, the broader market of consumers and businesses is increasingly relying on AI models for a wide range of applications. This external demand is a key driver of the AI boom, separate from the internal adoption metrics being tracked by companies like Amazon.
The Path Forward: Balancing Incentives and Genuine Value
The "tokenmaxxing" trend at Amazon and other hyperscalers highlights a critical challenge in the age of AI: how to effectively incentivize the adoption and utilization of powerful new tools without fostering unproductive behaviors. The pursuit of higher token counts, while intended to encourage AI engagement, risks becoming a proxy for genuine innovation and efficiency.
For companies like Amazon, the goal should be to foster a culture where employees understand and leverage AI to solve complex problems, drive business growth, and enhance customer experiences. This requires a nuanced approach to performance measurement and incentive design. Simply rewarding the quantity of AI usage may not be a sustainable or effective strategy in the long run.
Instead, a focus on the quality and impact of AI applications could yield more meaningful results. This might involve tracking metrics such as time saved on specific tasks, improvements in decision-making accuracy, or the development of novel solutions enabled by AI. Furthermore, fostering an environment where employees feel empowered to experiment with AI responsibly, without undue pressure to artificially inflate their usage, is crucial.
The financial investments in AI are substantial, and the potential benefits are transformative. However, the successful integration of AI into the workforce depends on thoughtful strategy and careful execution. As the technology continues to mature, the industry will need to refine its approaches to measurement and incentivization to ensure that AI adoption leads to sustainable, value-driven outcomes.
Amazon has not immediately responded to requests for comment on this matter. The evolving landscape of AI adoption within large technology firms will undoubtedly continue to be a subject of scrutiny as companies navigate the complexities of harnessing this powerful technology.
This article was originally published by Fortune as "‘That doesn’t sound very healthy’: Amazon’s reported tokenmaxxing might gamify AI usage, analyst warns" and is republished with permission.
