AI Token Pricing: The New ‘Water and Electricity’
China’s daily AI token call volume has surged from 100 billion in early 2024 to 140 trillion by March 2026—a more than 1,000-fold increase in just two years. As these digital units become the fundamental currency of the intelligent era, a Xinhua News analysis published Tuesday argues that standardized measurement and transparent pricing are essential for the healthy development of the AI economy.
The Token as Essential Utility
Tokens (词元) are the smallest measurement unit of large language models. Every question asked and every response generated is counted and priced by tokens, much like phone calls are billed by the minute and data by megabytes. The article, written by Wang Yuntao, Deputy Chief Engineer of the AI Research Institute at the China Academy of Information and Communications Technology (CAICT), draws a powerful historical parallel: just as electricity required standardized meters and rate approval to reach every household, tokens need standardized measurement to become the “water and electricity” of the intelligent era.
“High-quality token service is not just about cheap prices,” Wang writes. “Like evaluating tap water, you can’t just look at the price per ton, but also whether the water pressure is stable, how long it takes for water to come out of the tap, and whether it stops flowing frequently.”
The Three Layers of Opacity
The article identifies three distinct layers of opacity in current token billing. Technically, different models tokenize the same text differently, with token counts varying by several times. Output tokens typically cost several times more than input tokens, and multi-turn conversations cause bills to snowball as models re-read previous dialogue. Media investigations found that using the same model to solve a simple math problem can consume up to 10 times more tokens on different platforms.
From a pricing perspective, market prices are disconnected from actual costs. With 50-plus mainstream large model service providers offering varying prices and capabilities, users cannot easily assess value. Marketing complexity compounds the problem—pay-per-use plans, monthly and yearly subscriptions, and membership tiers increase information asymmetry.
“This is like taking a taxi,” the article notes. “The start and end points are the same, but some drivers take a straight line while others take a detour. The fares naturally differ. Worse, the meter itself hasn’t been calibrated—passengers can’t verify the distance traveled or the accuracy of the meter.”
The Cost Structure Argument
Token costs are primarily driven by computing chip depreciation and electricity costs. The article outlines three cost-reduction paths: model architecture innovation (such as sparse activation techniques that only call relevant parameters), inference engineering optimization (caching and batch processing), and scale effects from billion-fold increases in daily call volume.
“Cheap can genuinely be a real skill,” Wang argues. “The domestic model DeepSeek-V3 achieved first-tier international performance with a training cost of only about $5.58 million. Its low price stems precisely from architectural and engineering innovation, not reduced capability.”
However, the article cautions that expensive doesn’t necessarily mean better either—high prices may reflect advanced models or inefficient engineering that makes users pay for waste.
Industry Standardization Efforts
CAICT has been building a comprehensive regulatory framework to address these challenges. On June 16, the institute jointly launched the Token Service Capability Climbing Plan with 10 companies including Huawei Cloud, Ant Digital, Mobile Jiutian, Unicom Digital Intelligence, JD Cloud, and Lenovo Baiying. The plan published the first enterprise-level token service performance baseline for general scenarios: output speed of at least 55 tokens per second, time to first token of no more than 0.9 seconds, and call success rate of at least 99.9%.
The initiative also established the AIIA Token Service Working Group with 13 companies, launched a monitoring platform tracking 30-plus domestic and international platforms across 100-plus models, and released a comprehensive standards system covering service quality, operational capability, production capability, and security capability.
“This is equivalent to installing a calibrated meter for the token market,” the article states. “Users can reference the baseline to identify who offers genuine value and who is inflating their mileage.”
Regulatory and Market Momentum
The push for token standardization extends beyond CAICT. In March 2026, National Data Administration Director Liu Liehong officially designated “词元” as the standard Chinese name for Token at the China Development Forum, positioning it as “the value anchor of the intelligent era.” On July 28, Deputy Director Yu Ying became the first official to explicitly encourage exploring token trading models, calling for market consensus on “paying for quality data.”
All three major Chinese telecom operators have entered the token market. China Telecom launched trial token packages starting at 9.9 RMB per month, while China Mobile introduced nationwide packages with unified computing power measurement starting at just 5 RMB per month—a shift from “traffic management” to “computing power management” as operators seek new growth after traditional revenue stagnation.
Legal Framework Development
Legal scholars are also weighing in on the need for transparent token measurement rules. A legal analysis published in Contemporary Guangxi argues that current laws focus on training data governance, content security, and algorithm filing—but have not directly addressed token measurement. The analysis proposes recognizing tokens as “measurement-type institutional objects” and establishing platform obligations including measurement rule disclosure, dynamic pricing notification, abnormal consumption alerts, and billing log retention.
The Road Ahead
The token market is expected to exceed 50 billion RMB in 2026, with Morgan Stanley projecting China’s AI inference token consumption to grow from 10 trillion in 2025 to 3,900 trillion by 2030—a 227% compound annual growth rate. Unit token prices are expected to drop 40% from 2026 to 2027 as domestic high-performance inference chips and distributed computing architecture optimization take effect.
The upcoming 2026 China International Big Data Industry Expo (August 28-30, Guiyang) has adopted “Token—New Path for Data Element Value Release” as its official theme, with 60,000 square meters of exhibition space covering computing infrastructure, data supply, model-driven, security, and intelligent experience sectors.
“Tokens will eventually follow the same path as water and electricity,” Wang concludes. “Technological iteration brings prices down, and standards and regulations make bills transparent. Both are indispensable.”
As the intelligent era unfolds, the question of how to price its fundamental currency will shape not just China’s AI economy, but the global trajectory of artificial intelligence adoption. The establishment of transparent, standardized token pricing could determine whether AI becomes a utility accessible to all—or remains a privilege of the few who can navigate its opaque billing structures.