⚠️ Editorial note: The open source ecosystem in China operates under a distinct institutional framework — characterized by state-led initiatives, intranet-like boundaries, and top-down governance. Readers should be aware that this context differs from the community-driven open source model common in other regions. The term “open source” as used in Chinese media may refer to practices that diverge from the conventional definition.

China Open Source Daily — 2026-11-01

🏛️ Policy & Regulation — 词元: What Happens to an Open Ecosystem When the Unit of Measurement Becomes a Policy Instrument

1. The naming decision, because everything else follows from it. In March 2026 the National Data Administration adopted 词元 — cí yuán, “word” plus “currency” — as the official Chinese term for token. It is worth pausing on the etymology, because it is not a translation. It is a monetary frame: the unit in which a model processes text has been renamed into the unit in which something is counted and paid. Every document in this briefing that follows uses that word, and the word carries the policy.

2. The document. On 2026-09-18 the Beijing Municipal Bureau of Economy and Information Technology (北京市经济和信息化局) and the Beijing Municipal Development and Reform Commission jointly issued the Action Plan for Accelerating the Development of the Token Economy (2026-2028) — 《北京市加快词元经济发展的行动方案（2026—2028年）》 — reported by Xinhua the same day and characterised by the municipal bureau’s own digital-industry division chief as the first provincial-level policy in China dedicated specifically to the token economy. It runs to 2028-12-31, and its structure is six dimensions, ten measures, announced as the “词元经济十条.”

The six dimensions, in the plan’s own order: (i) strengthen supply at the production end — build world-class token factories (词元工厂) and attack core technologies; (ii) complete the evaluation and distribution system — establish a token quality evaluation system and build a token distribution platform, to “perfect the market-based pricing and circulation mechanism for tokens”; (iii) cultivate the consumption-side ecosystem — general and vertical agents, and a core agent software ecosystem; (iv) deepen international cooperation; (v) construct an industry knowledge base via data ontology technology; (vi) element guarantees — innovative compute and token financial instruments, and “token engineer” reskilling.

3. The two measures that matter, neither of which is a subsidy.

First, a grading standard for token factories that is almost entirely a production-efficiency instrument. Beijing will set a graded evaluation standard for token factories (词元工厂分级评价标准) scored on model-adaptation count, token throughput, first-token latency, cache hit rate and power usage effectiveness (PUE). Land, energy quotas and communications networks are to be allocated to support factories that receive the grade. This is not a research grant. It is an industrial classification written in the vocabulary of a data centre. Four of its five criteria are engineering telemetry; the fifth is electricity. Nothing in the list measures whether the models being served are open, whether the weights can be obtained, or whether the output can be reproduced.

Second, and larger: the token distribution platform is to be written into the new-infrastructure plan. The plan directs that a token distribution platform be built offering developers a unified interface, model aggregation, token metering and settlement, and — the operative clause — that the platform be “incorporated into the layout of new-type infrastructure” (推动词元分发平台纳入新型基础设施布局), forming “an efficient service gateway to the intelligent economy.” A new term appears here and deserves to be flagged: 基础设施 is the word used for power grids, base stations and the national computing-network hubs. A metering and settlement gateway is being classified as infrastructure. Note what that classification does institutionally: infrastructure has a funder, a planning process, an operator and a list of approved participants. A marketplace has none of those things.

4. The state’s own diagnosis, and it is honest. The official rationale, stated by the bureau officials quoted in the coverage, is that scale is not value: the plan’s own justification is that China leads on volume but that model intelligence and the rate of conversion into real economic value lag; that a large share of consumption sits in low-value scenarios (chat and image generation) while industrial and scientific use remains thin; and that what is missing is measurement, pricing, trading and governance rules for tokens — with the explicit concession that tokens from different models are not fully comparable, so call volume growth is not the same as economic value growth. That last sentence is the most sophisticated thing in the file, and it is the state arguing against its own headline statistic.

5. The financing layer — where measurement becomes allocation. Two schemes are already operating, and they are the reason this file exists.

  • Beijing E-Town (Beijing Economic and Technological Development Zone), reported 2026-09-14/15: a “token loan” (词元贷) programme in which six banks — Agricultural Bank of China, China CITIC Bank, Industrial Bank, Bank of Beijing, China Minsheng Bank and Huaxia Bank — extended nearly RMB 2 billion (about USD 300 million) of credit across the AI industry chain, partly on the basis of token usage, alongside technological barriers, team strength and business prospects. The stated target is firms with high R&D intensity and high technical barriers but no physical assets and no profits — the collateral problem, in the bureau’s own words. Two named recipients: Guoqi Zhikong (connected-vehicle cloud and autonomous-driving platforms) took RMB 30 million from CITIC; Shenzhou Guangda (computing infrastructure and enterprise agents) took RMB 30 million from Industrial Bank.
  • Guangzhou, Haizhu district, from 2026-08: the country’s first such scheme — a financial product under which banks including Bank of China, China CITIC Bank and Bank of Guangzhou measure a firm’s token production and consumption to decide credit, with up to RMB 30 million per borrower, aimed at small and medium computing firms across token supply, application and services. Caixin Global reported in August that state-owned lenders had begun offering such loans “typically several million yuan per borrower.”

6. The consumer layer, which no plan designed. Rest of World (2026-09-04) catalogues what has already spread: China Merchants Bank issuing a card for AI developers with up to 1.8 billion tokens of MiniMax models; Shanghai Pudong Development Bank with up to 3 billion tokens of Alibaba’s Qwen; Moonshot AI issuing a Kimi-branded card with Agricultural Bank of China and American Express, earning tokens as spending rewards. Three carriers selling token as mobile data: China Telecom from RMB 9.9/month for 10 million tokens plus a TokenHub platform covering 142 models; China Mobile at 400,000 tokens per RMB 1 in Shanghai, billed to the phone account; China Unicom plans from RMB 15 for 6 million to RMB 45 for 18 million. AGI Bar in Beijing offering unlimited DeepSeek V4 Flash to anyone who buys a drink, running the model on a local workstation. A dumpling restaurant handing out RMB 10 of compute credits after a meal — more than 100 vouchers a day across two stores. And on Xianyu, Alibaba’s secondhand marketplace, tokens resold in day passes, monthly subscriptions and shared accounts, including arbitrage between official and discounted access.

7. The scale the policy is aimed at. Per the National Data Administration: daily token call volume of 100 billion in early 2024, 100 trillion by the end of 2025, over 500 trillion by May 2026 — the figure CCTV Finance carried in August. A China Telecom Research Institute forecast puts year-end at 100 quadrillion. An MIIT figure of more than 140 trillion in March 2026 circulates as well. The three series do not agree, and the disagreement is not innocent: they are the numbers the credit policy is priced against.

Institutional Analysis & Commentary:

Begin with the instrument, because the instrument is the argument.

China now has an official word for token that contains the word currency. Against that word it has built, in one year: a grading standard that scores the thing that produces tokens, a settlement platform to be classified as public infrastructure, a credit product that loans money against token flow, a voucher and card system that distributes tokens to consumers, and a state forecasting series that estimates aggregate token volume. Trace the chain and it is unmistakable: 词元 production → 词元 grading → 词元 metering and settlement → 词元 collateral → 词元 consumption. That is not a subsidy programme. It is a monetary architecture in construction, assembled from the consumption end backwards.

Here is why that ordering matters, and why it is not a metaphor. A payment system, as the monetary literature has understood at least since Rocher and Charles Goodhart’s formalisation of the old Russian credit-school tradition, has two halves: the means of payment — the thing that settles — and the unit of account — the thing in which debts are denominated and accounts are kept. They are separable, and their separation is the normal condition of most economies. A unit of account exists the moment the state grades, meters and lends against a quantity; a means of payment exists only when the instrument can discharge an obligation to someone who did not issue it. Beijing is building the unit-of-account half with visible determination — grading standards, metering, settlement, collateral — and has said almost nothing about the means-of-payment half. Which means the interesting question is not whether tokens become money. It is what happens when a state treats consumption of a privately produced good as though it already were.

The classical failure mode has a precise name and a precise date. In 1912 Ludwig von Mises, in The Theory of Money and Credit, set out what became known as the regression theorem: the objective exchange value of money cannot be derived from its function as money alone — it must trace back, link by link, to a prior non-monetary use value. A token has a real use value (it processes text), so it passes that test where a fiat claim would not. But the theorem cuts in a second, harder direction. Value must trace back to demand arising outside the scheme that denominates it. A token’s demand is demand for a particular model, particular vendor, particular licence. When the state loans against token volume, hedges with token vouchers, and grades factories by token throughput, it is forming a unit of account over goods it does not control and cannot substitute — and Mises’ own framework predicts exactly what follows: an artificial demand for the medium generated by the credit that names it, decoupled from demand for the service.

Which is why the plan’s own most honest sentence is the one that undermines it. The officials concede that tokens from different models are not comparable and that volume ≠ value. Correct — and fatal to the accounting, not merely a caveat upon it. A unit of account requires that the unit be fungible across the debts it denominates. A Qwen token, a DeepSeek token, a Kimi token and a MiniMax token are not one good in different packages; they are different goods with different prices, different capabilities, different licences, different serving terms and — crucially — different terms of access. Lending against “token consumption” measured in a single aggregate series means lending against a quantity that does not exist as a single thing. The cache-hit-rate criterion makes the point unintentionally sharp: it rewards serving efficiency, which is a property of an inference stack, not of the model — and it is scored in the same standard as model-adaptation count, which is a property of the vendor relationship.

Now the part that connects to this series, because it is the reason an open-source publication is covering a monetary policy. How is aggregate token volume counted? It is counted by platforms — traffic through gateways, API meters, settlement systems. And this series’ 2026-10-30 file spent a full paragraph on precisely this problem in another context: Alibaba claiming 3 billion Qwen downloads, Hugging Face counting 2.045 billion on its own Hub, Alibaba claiming 300,000-plus derivatives against 151,448 observed, and a 10-billion-versus-100-billion misreading inside this very briefing. The lesson stated then was: a metric without a stated denominator and a stated window is a claim, not evidence. Today the same metric is a collateral. The difference between a claim and a credit line is enforcement.

And the Chinese analysts already say so. Lou Feipeng of the Postal Savings Bank of China, quoted in the Sixth Tone report, names three preconditions before token lending can scale: exaggerated token-use data, inaccurate assessment of computing power consumption, and overestimated repayment ability. He is describing Goodhart’s law with a loan book attached: when a measure becomes a target, it ceases to be a good measure — and here the measure is being targeted by the very firms whose credit depends on it. This is not speculation about Chinese actors being uniquely dishonest. It is the ordinary behaviour of any firm facing an accounting-based credit test, and it is why banks everywhere prefer collateral: collateral is expensive to fake.

And here is the design that follows, in a detail that is easy to miss and is the most interesting institutional fact in today’s file. Beijing E-Town’s scheme is not conventional lending at all. As reported, the zone brought the banks together and stood behind the loans — a guarantee-innovation arrangement, not a rediscount window. Lending against an intangible is a transaction with very high asset specificity in Williamson’s sense: the bank cannot repossess a token flow, cannot value it independently, and cannot redeploy it. The governance response when asset specificity is high is not more market — it is hierarchy, or a third party absorbing the residual risk. But Beijing did not nationalise the credit. It interposed a local state guarantor and kept the banks as originators. That is a genuinely Chinese institutional form, and the right name for it is not “market-based finance” — it is a public balance sheet underwriting a private accounting standard it does not audit.

Then take the argument inward, because the easy version of this commentary is wrong in at least three ways.

First: the diagnosis is better than the cure, and the diagnosis is genuinely good. “Volume is not value, most consumption is chat and image generation, industrial and scientific use is thin, what we lack is measurement and pricing rules” — that is a serious industrial-policy analysis, and it is more honest about China’s AI economy than most external criticism. Measured against it, this plan is a coherent response, not a slogan.

Second: the consumer layer is real demand, not merely administered supply. A programmer buying a drink at AGI Bar to run DeepSeek locally, a dumpling shop handing out compute credits, a resale market on Xianyu with a spread worth arbitraging — these are the ordinary symptoms of a good becoming cheap enough to be casual. It is exactly what happened to bandwidth. The Rest of World sourcing is appropriately careful about it: analyst Poe Zhao calls the packages “a supply-led experiment” and notes that most users still meet AI through an app rather than a token balance. Both things are true at once. Cheapness is not the same as a currency.

Third, and least comfortable: this is not exotic. Every jurisdiction with an AI strategy is now trying to measure AI’s economic contribution, and none has a clean instrument. The German eligibility rule and the Chinese procurement standard this series covered on 2026-10-28 and 2026-10-29 were both attempts to write a test for a good whose defining property resists testing. Beijing’s move is stranger only in being monetary rather than technical — and there is precedent on the record: industrial reserve currencies, of which the Soviet tuzon in the 1930s and the Comecon transferable rouble were the fullest attempts, both built to price goods across a bloc whose internal prices were administratively set. The transferable rouble failed for a reason worth recording precisely: it existed for accounting and could not be spent outside the system, and because each member’s prices were set internally rather than negotiated, twice-yearly negotiations were conducted in a unit with no external reference. No one has proposed a token bloc of that kind, and it would be careless to say anyone has. But the structural question is on the record and it is not rhetorical: a unit of account whose circulation is confined to the entities that accept it, denominating goods priced by their own issuers, is a unit of account only inside the wall.

Which brings the file back to where an open-source publication has to stand. The plan’s international dimension is not incidental. It directs Beijing to build the China–SCO AI Application Cooperation Centre at a high level and to offer Belt-and-Road and emerging markets an integrated package of open-model adaptation, agent deployment, development toolchains and compute services — with multilingual token solutions and an overseas developer ecosystem. Read alongside the 2026-10-24 Inclusive Plan progress report of this series’ 2026-10-30 file — openness is the precondition for inclusive development, with toolchains and test environments still to be built — and the two documents are the same export strategy seen from two ends. The open model is the entry good. The token is the billing unit. And that is the precise place where the openness claim and the monetisation meet: a state can open weights freely — they cost it nothing to duplicate — and still capture the entire downstream flow through metering, settlement and credit denominated in a unit it defines. The four freedoms govern what you may do with the artefact. They do not, and never did, govern what the token meter charges you to run it. This is the distinction this series has been circling since 2026-10-30: a contribution artefact and a cooperation output are different institutional objects, and a metered service is a third.

The checkable question — three observables, each answerable from a public record, and the first two fall due within a year.

First: does the token distribution platform, once built, publish an access rule a stranger can read — what models it aggregates, on what terms, who may be a supplier, what the metering formula is, and what happens to a developer whose model is aggregated and then delisted? The moment it is classified as new infrastructure, the classification is testable: infrastructure either has a published tariff and an application path, or it has an approved list.

Second: does any token loan enter arrears, and does the resolution publish the valuation method? Two billion yuan of credit has been extended against a collateral that cannot be seized. What happens on the first default is what tells an outside reader whether the token was ever a collateral or only a signal toward a state guarantee. Nothing in the public record suggests these loans have gone bad; the question is what the record will say if one does.

Third, and the one that generalises: when an ecosystem’s unit of measurement is set by its platforms, who may audit the meter? Download counts, derivative counts and token volumes are all readings taken by parties with commercial interests in the reading. This series printed a wrong number of its own three weeks ago and corrected it in public, and the discipline it owes is the same one it asks of others. A metric that cannot be audited by the party paying it is not a unit of account. It is a price the issuer sets, and the state is currently borrowing against it.

Sources:

🔍 Commentary — The Meter Is the Institution

On 2026-10-31 this file argued that three National AI Application Pilot-Test Bases are what China built to answer a foreign-ministry report’s admission that “test environments remain to be built,” and concluded that a test environment measures its worth by whether an outsider can reproduce the test. Today’s file is the same problem in a different register, and it is a harder one.

A test environment governs whether a result can be checked. A token meter governs whether a thing can be bought, and at what price, and on whose account. Both are institutions. Only the second one is currently being written into a country’s definition of infrastructure.

Watch what the Beijing plan actually grades, because the choice is the whole policy. The token-factory standard scores model-adaptation count, throughput, first-token latency, cache hit rate, PUE. Every one of those is a property of the serving stack, and none is a property of the model’s terms. A factory that serves only closed weights, at a high price, under a restrictive contract, can score the same grade as one that serves open weights to anyone who asks — provided the telemetry matches. This is not a defect anyone slipped in. It is what a production-efficiency standard is. The policy has decided that the unit of account will measure the thing it can meter, and the thing it can meter is not the thing open source is about.

And the consumption side makes the same choice again, more quietly. A token credit card rewards spend, a carrier plan prices allowance, a dumpling shop hands out credits, Xianyu arbitrages the spread between official and discounted access. In every case the consumer is buying a quantity of processing, metered by a platform that also defines the unit. Compare what an open-source ecosystem actually relies on: the ability to obtain the artefact, run it on your own hardware, modify it, and redistribute it without asking. The meter neither helps nor hinders that. Which is precisely the point: the entire apparatus under construction is orthogonal to the four freedoms, and that is not an accident of design — it is what happens when a state builds an economic rail for a good whose openness is a side effect of how it was produced rather than a condition of how it is sold.

Here is the asymmetry that makes this the sharper half of the week’s story. Weights are the one input that costs a vendor nothing to give away — duplication is free, and the giving buys distribution, standard-setting position and goodwill. The token is the one thing that must be metered, because it is where the marginal cost lives. So a strategy of opening the weights and metering the tokens is not a contradiction. It is the optimal combination, and it is available to any vendor in any country. Open weights make the good widely adoptable; the meter makes the adopters into customers. The four freedoms were designed for a world in which the artefact was the product. In a world where the artefact is free and the inference is the product, they govern the part that was given away and nothing about the part that is charged for.

Which is why the plan’s own honest concession is more damaging than any external criticism would be. Tokens from different models are not comparable; call volume is not value. True. But a credit product and a settlement platform and a grading standard all require comparability, because their entire function is to make unlike things commensurable. The plan states the obstacle and then builds three instruments that presuppose it has been overcome. That is not hypocrisy — it is the ordinary sequence by which a measurement regime is installed before its unit is settled, and it is why the first token loan to go into arrears will be more informative about this policy than any of its ten measures.

Now the inward turn, because this file’s method requires it and the precedent is uncomfortable. In the 1930s the Soviet Union introduced the tuzon — a unit of account for capital goods, defined against a basket of commodities, that was never a means of payment and never circulated as cash. Its successor in the Comecon, the transferable rouble from 1964, was more ambitious: a clearing unit for intra-bloc trade in which no one was ever paid in anything they could spend outside the system, and in which the prices cleared through were administratively set rather than negotiated. Both were serious, technically sophisticated attempts to build a unit of account for goods whose prices were not market prices. Both failed for the same reason. A unit of account that denominates transactions between parties who cannot choose another unit is an internal ledger, and an internal ledger cannot price anything for the world outside it.

No one has proposed anything like that here, and this file will not pretend otherwise: Chinese token volume is overwhelmingly commercial, cross-border, and heavily driven by exports of open weights. But the structural difference between the transferable rouble and the 词元 is worth stating precisely, because it is the reason this case is genuinely new. The tuzon denominated goods the state produced. The ciyuan denominates a good that is substantially given away, produced by competing private firms, and consumed by foreigners who have plenty of substitutes. That makes it far more likely to work as a commercial unit — and far less able to serve as a monetary one. A unit of account whose underlying good is collapsing in price is a strange foundation for credit. Token prices have been falling hard for two years, which is the policy’s stated advantage and its unstated risk: the collateral deflates.

The checkable question, and this file will keep it to one sentence because the rest are above. When the token distribution platform is certified as new infrastructure, does its tariff schedule and its supplier-admission rule get published — or does it get an approved list? The word the plan uses, 基础设施, means the second is entirely normal. The first is what an ecosystem would require. Both are administratively possible, and only one of them can be audited by anyone who pays the meter.