The corporate trend of maximizing artificial intelligence usage, dubbed "tokenmaxxing," is facing a significant slowdown as businesses increasingly scrutinize the costs associated with these technologies. Initially fueled by hype from the tech industry, this approach involved encouraging employees to generate as much AI-driven output as possible, often through tools like OpenAI's ChatGPT and Anthropic's Claude. However, a growing backlash has emerged as companies realize the substantial expenses involved often do not correlate with proportional increases in productivity.
"Tokenmaxxing" refers to the practice of utilizing the maximum number of "tokens"—the fundamental units of text processed by generative AI models. Premium versions of AI products typically offer higher token limits, and a few months ago, high token consumption was often promoted as a sign of high employee performance. Executives from prominent tech companies, including OpenAI's Sam Altman and Nvidia's Jensen Huang, previously expressed enthusiasm for this trend, with Meta even holding internal competitions to reward token usage.
This surge in usage benefited AI developers like Anthropic and OpenAI. However, the strategy's effectiveness for many businesses has become questionable. Microsoft CEO Satya Nadella acknowledged that tokenmaxxing can be habit-forming and noted that companies might be incurring dual costs: paying for token usage and inadvertently sharing proprietary data with AI providers. Palantir CEO Alex Karp has been more critical, stating that U.S. businesses are privately frustrated by the high costs of tokens that do not deliver tangible value, fearing the loss of intellectual property.
Jue Wang, a management consultant at Bain & Company, highlighted that large businesses are now closely examining their AI investment returns. She observed that token costs can double nearly every two months, and for companies with tens of thousands of developers, this represents a significant, unplanned expenditure. This has led to a reassessment of which AI models are appropriate for specific tasks, with companies moving away from using highly capable but expensive models for simple tasks like generating emails.
The industry is now seeing a greater interest in "model routing" solutions, which automatically direct simpler queries to more cost-effective AI systems and reserve powerful models for complex tasks. Hassan El Mghari, who leads developer experience at Together AI, noted that sticker shock over AI product subscriptions is prompting companies to shift focus towards educating employees on effective AI use rather than simply rewarding high usage.
Concurrently, new open-source AI models from Chinese companies are emerging that offer capabilities comparable to leading U.S. models at a lower cost, potentially influencing the tokenmaxxing trend further. However, many industry observers believe the fad is waning. Raffi Krikorian, CTO of Mozilla, likened the focus on token usage to earlier outdated metrics in software development, such as counting lines of code. He suggested that "tokenmaxxing" is likely to be viewed as a short-lived, perhaps amusing, phase in the broader adoption of AI.