Two recent developments have pushed "token" back into the industrial spotlight: Beijing has issued the Action Plan for Accelerating the Development of the Token Economy (2026-2028), using tokens as the lever to systematically shape a new form of the intelligent economy; and the Beijing Economic-Technological Development Area has launched its first batch of "Yiqitoken" loans, with six banks extending nearly 2 billion yuan in credit to companies across the AI supply chain, while banks including China Construction Bank and Shanghai Pudong Development Bank have recently followed suit in quick succession.
Earlier, the Guiyang Big Data Expo took "Tokens — A New Path for Unlocking the Value of Data Elements" as its theme, and tokens have now risen from a technical concept to a new economic form drawing joint attention from policymakers, industry, and finance.
It must be recognized that tokens are a unit of measurement inherent to the technical architecture of large models, and their generation depends on the tokenization system of a specific model — apart from large models, they have no basis at all, and they do not inherently possess the technical foundation to become a cross-entity, cross-industry universal unit of measurement.
Historically, every universal unit of measurement — from the Qin dynasty's unification of weights and measures to the international prototype metre and the kilowatt-hour — has undergone prolonged refinement.
For tokens to truly grow into the unit of measurement for the intelligent economy, three hurdles must be cleared.
The first is the standards hurdle. The number of tokens generated from the same text varies significantly across different models, and after some vendors adjust their tokenizers, the token count for the same text can rise by about 30%. Without a recognized conversion relationship between one "ruler" and another, cross-vendor cost accounting and usage comparisons lose their baseline, and token-based industry statistics and horizontal rankings lose their premise for comparability.
The second is the credit hurdle. The same "10 billion tokens" could represent solid consumption from deep reasoning, or it could be the cumulative bookkeeping of cache calls. On billing rules, how much of a discount caching gets and whether the thinking process is charged vary from vendor to vendor, and prices for similar services can differ severalfold. When metering cannot be verified and pricing cannot be compared, it is difficult to carry serious uses such as performance assessment and credit extension.
The third is the value hurdle. On one hand, a considerable share of current token consumption is concentrated in general scenarios such as customer service Q&A and content generation, while penetration into the core links of the real economy remains limited; new paradigms such as AI agents and multi-turn reasoning are also multiplying the token consumption of a single task. On the other hand, as model compression and inference optimization technologies iterate, the higher the efficiency, the fewer tokens are consumed to complete the same task. To treat such an input metric directly as a measure of output and value would distort not just the numbers, but the direction.
With the three hurdles uncleared and the hype running ahead, risks may follow. Some localities have renamed and rebranded traditional computing centers as "token factories," with no change in the substance of the business, yet seek policy resources under a new concept; some organizations simply link token usage to rankings and performance, inducing a tendency to prioritize scale over effectiveness; and some financial institutions use token volume as the basis for credit extension, even though usage data still lacks means of verification — once distorted, risks may transmit from the industrial sphere to the credit sphere.
Pointing out that tokens have "not yet cleared the hurdles" is not a denial of their value. Tokens and computing power consumption run through the entire industrial process and are core process indicators for observing the vitality of the intelligent economy. The problem is that one must not pull up seedlings to help them grow: they can be used as a reference in statistical observation, but should be treated with extreme caution in performance assessment and credit extension; business activities that genuinely provide model inference services and charge by token should be supported, while behavior that packages old businesses under new concepts should be explicitly constrained.
Next, to let tokens truly shoulder the role of the intelligent economy's unit of measurement, one can draw on the idea of purchasing power parity, using standard corpora as the anchor to measure each model's token compression ratio and form regularly published, recalculable conversion coefficients, so that different vendors' "rulers" can be mutually recognized. Tokens should be promoted as a universal pricing unit for computing power service transactions, billing disclosure should be standardized, third-party metering verification should be introduced, and exploring the release of an authoritative token price index would let bills withstand scrutiny. A mechanism for evaluating the contribution of factors behind tokens should also be established, turning tokens from a technical unit of measurement into a measurement anchor for distributing returns from data elements, providing a new fulcrum for reforming the market-based allocation of data elements.
The intelligent economy needs its own unit of measurement, but a unit of measurement is not "speculated" into existence — it is forged bit by bit through unifying standards, building credit, and anchoring value. With less conceptual hype and more standards-building, tokens can truly settle from a "buzzword" into a reliable yardstick for the high-quality development of the intelligent economy.