Token consumption = amount of work done?···In the AI era, companies wrestle with ‘to-seong-bi’ - 경향신문

The ‘token economy’ that determines competitiveness “Because there is a perception that using more tokens means doing more work, we run agents almost all day.” “We chose a lower-token pricing plan on Claude to control token volume.” These days, among people anywhere in the world who use artificial intelligence (AI) even a little, ‘tokens’ are unquestionably the hot topic. For developers, programmers, and other information technology (IT) workers, token usage is even being used as an objective metric indicating work output or ‘diligence’. Token usage and the resulting economic costs are poised to reshape the frameworks of the economy and industry in the AI era. In a report published at the end of January, ‘AI Tokens: Exploring the New Cost Dynamics of AI’, the global consulting firm Deloitte stated that “AI should be treated as an economic system that operates according to unpredictable, token-based costs.” Jensen Huang, CEO of Nvidia, has predicted the rise of ‘AI factories’, saying that computei.e., token capacitywill be foundational not only for AI data centers but also in manufacturing where ‘physical AI’ is applied. In short, the ‘token economy’ has become what determines corporate competitiveness. A token is the basic computational unit of an AI model. It is the minimum piece of information used by large language models (LLMs) such as OpenAI’s ‘ChatGPT’, Google’s ‘Gemini’, and Anthropic’s ‘Claude’ to understand natural language and to process and generate data. You can think of it as a kind of ‘digital fuel’ needed to run an AI model. In particular, token consumption has been soaring exponentially since the beginning of this year. This trend follows the shift of the AI industry’s center of gravity from training to inference and the full-fledged arrival of the era of ‘agentic AI’ that performs tasks on behalf of users. AI agents are known to use up to one million times more tokens than ‘generative AI’. That is because the volumes of various datatext, images, and morethat agents must input have grown far larger, and the output required to produce real results has grown accordingly. In Silicon Valley, ‘tokenmaxxing’ (tokenmaxxing), meaning pushing token use to the maximum, has become a culture. ‘Tales of derring-do’ such as an engineer using more than 200 billion tokens in a single week circulate openly. Big Tech companies like Meta and OpenAI, as well as software firms such as Salesforce, have built ‘dashboards’ that tally employees’ token usage in real time. They rank how much employees use AI and apply it as a productivity evaluation metric. As a result, tokens are sometimes consumed competitively even when not directly related to work. The Pragmatic Engineer, a newsletter covering Big Tech trends, reported that at Meta and Microsoft (MS), which operated dashboards, tokenmaxxing appeared regardless of rank. One MS developer confessed that, to avoid being criticized for ‘using too little AI’, they asked AI for code that already existed in documents or told it to build features unrelated to actual work, then simply discarded the results. AI, introduced for ‘efficiency’, has ironically ended up amplifying inefficiency within organizations. Meta has ultimately stopped aggregating and publishing token usage. Both Meta and MS have also signaled intense restructuring this year. Some speculate that after increasing AI use across their organizations, they intend to expand the jobs that AI will replace. As ‘AI transformation’ (AX) accelerates not only in companies but also across the public and social sectors, the increase in token usage acts as a direct cost burden. Unless companies and institutions internalize token costs by using their own AI models, they must pay fees to outside firms such as OpenAI, Anthropic, and Google. According to the IT trade outlet The Information, the monthly token usage by Meta employees reached 60.2 trillion. Calculated based on Anthropic pricing, that amounts to a staggering $900 million (about 1.3242 trillion KRW). For companies, strategies to maximize value added per token have emerged as a core task. Emphasizing an industry realignment around tokens, CEO Huang argues that energy efficiency must be maximized to produce more tokens, stressing, “Tokens per watt and tokens per dollar should become the basis for corporate decision-making.” Nvidia’s development of the inference chip ‘Grok 3 LPU’ and more is part of a move to optimize token costs. At last month’s annual developer conference ‘GTC 2026’, he unveiled a new Vera Rubin platform built on the Grok chip, saying, “Our cost per token is the lowest in the world. There is nothing better than this (You can’t beat it).” Startups and small and mid-sized companies are focused on securing ‘to-seong-bi’ (token + value-for-money). A representative trend is mixing high-performance models with comparatively cheaper ones depending on the level of the task. An employee in the domestic IT industry said, “Considering the spending burden from rising token consumption, many small and mid-sized firms adopt pricing plans with token caps or encourage the use of more affordable AI models.”