The shift from focusing on token volume to emphasising token efficiency represents a natural progression for companies adopting AI, according to Yu Zhou, an associate professor at Columbia University and co-founder of Arklex.AI. He shared this view at a private roundtable titled "Beyond the Field: U.S.-China Business, AI and Global Collaboration" held in New York on July 20.
The concept of "Token Maxing" refers to the initial phase of AI adoption where companies monitor employee engagement by measuring token consumption. To drive initial adoption from zero to one, some firms prioritise providing the most powerful models. At this stage, tokens serve more as a metric for adoption rate.
However, as AI scales towards broader application, the focus naturally transitions from "how much is used" to "what value is created." High token consumption does not automatically equate to high efficiency. Enterprises begin evaluating whether the cost of the models justifies the time saved, productivity gains, and revenue improvements, marking the entry of AI into a phase of "return on investment validation." The recently concluded 2026 FIFA World Cup, as a massive AI application scenario, has further amplified this shift.
According to estimates from the white paper "The First AI World Cup: The Victory of Scenarios" by Weijin Research, AI users generated approximately 15.84 trillion tokens daily during the tournament related to World Cup activities, accounting for roughly 3% of global AI computing demand during that period. These tokens primarily originated from match analysis and prediction, as well as the generation of multimodal content like related videos and images. As FIFA's official technology partner, Lenovo Group Ltd. faced the core challenge of transforming this massive AI demand into stable, real-time, and comprehensible services for the event.
One key offering was FIFA AI Pro, made available to all 48 participating teams. This tool generates tactical insights by combining match, tracking, video, and 3D data. FIFA disclosed that the tool was used by all teams and supports 15 languages.
On the operations and broadcast front, Lenovo Group Ltd.'s ThinkSystem servers processed real-time video data from all stadiums, reducing FIFA IPTV broadcast latency to under five seconds. AI-powered systems, such as 3D player digital avatars and referee-view video enhancement, served visualisations for offside calls and live broadcasts. With 104 matches spanning three countries and 16 cities, the challenge extended beyond merely supplying computing power to successfully accomplishing data aggregation, model invocation, and result distribution.
This operational challenge is precisely where Lenovo Group Ltd.'s hybrid AI approach connects with the principle of Token Efficiency. Tasks can be executed on the device, at the edge, in the data centre, or in the cloud: complex tasks invoke more powerful models, while routine tasks are assigned to more cost-effective models. Intelligent routing coordinates this system to balance performance, cost, privacy, and energy efficiency.
Yu Zhou suggests that Token Efficiency does not equate to simply choosing the cheapest model. In the future, enterprises may not need the most powerful model to answer every query. Instead, they might deploy a collaborative system comprising multiple models and agents: foundational models handle routine tasks, while leading-edge models tackle high-difficulty reasoning and critical decision-making, with different agents dividing labour and cross-verifying results.
Therefore, the transition towards Token Efficiency does not signal a cooling of AI demand or a peak in computing power requirements. As the cost per task decreases and returns become clearer, more business processes become viable for AI integration. The 15.84 trillion tokens represent latent demand. What Lenovo Group Ltd., a leader in the global AI ecosystem chain, validated at the World Cup is how to translate computing power into tangible value through tournament-grade infrastructure and AI applications. The next phase will not be about which company "burns" more tokens, but about which can create greater value from the same number of tokens.