⚠️ 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 — 2026-11-17
⚖️ Legal & Licensing — The Metered Free Tier: DeepSeek Prices Its Own Open Weights for the First Time
1. The event, and its date. On 2026-11-16 DeepSeek was reported to have begun restricting usage on its free web and mobile app entry points because of compute shortage. The measures, as reported by 21st Century Business Herald (21世纪经济报道) by reporter Zhao Yunfan (赵云帆) and republished the same evening by Sina Finance, are three and they are cumulative: the single-input cap is compressed to roughly 40,000 characters; conversations older than seven days with no active interaction can no longer be continued; and a separate daily conversation quota applies, on reaching which the interface displays “因服务器繁忙,使用额度已达上限” — server busy, usage limit reached. The paper states plainly that this is the first restriction DeepSeek has placed on free users since the company was founded.
2. What makes it different from the earlier changes. The report is careful to distinguish precedents, and the distinction is the entry’s foundation. In May 2026 DeepSeek had already limited short-interval repeat requests for free users and stopped file uploads in expert mode — but that was characterised as an attempt to stop high-frequency users arbitraging the free entry point with local automation, a fraud problem. The November change is a capacity problem, and the argument is made from the company’s own instrument: DeepSeek’s status page recorded 57 incidents of outage or degradation across 2026, minor and scattered through the first half of the year, then shifting after October to frequent load-based rate limiting and long global outages — including 15 October 2026, the longest in the company’s history, from 08:11 to 13:04 Beijing time, nearly five hours, taking down the web client, the app, the API and third-party platforms including Tencent’s Yuanbao, WorkBuddy and OpenRouter, followed by a second large global outage on 7 November, three weeks later. Compensation was paid to API users after the first, with a commitment to “resolutely ensure nothing of the kind recurs”; it recurred.
3. The mechanism the paper identifies, which is the part worth keeping. The demand shock is not chat. It is agentic office work. Tencent’s WorkBuddy — which the report identifies as one of the largest third-party proxy platforms for DeepSeek’s official API, and an officially recognised ecosystem partner — reached 32.35 million multi-end monthly actives in September 2026, up 647.1 per cent quarter on quarter, with 836 million uses that month, up 858.4 per cent, 9 billion cumulative user task requests from launch in March to the end of October, and token consumption up fiftyfold since launch. Tencent Cloud’s Dowson Tong (汤道生) is quoted from the October global digital ecosystem conference saying WorkBuddy’s tasks are typically long-running and that it is one of Tencent Cloud’s largest token-consuming scenarios, far exceeding Yuanbao’s quick-interaction pattern. The economics of why: an agent loop decomposes a task, calls tools, verifies in multiple rounds and rereads its own context; per Anthropic’s measurements a typical agent turn consumes about four times a conversational turn and up to fifteen times under multi-agent orchestration, and per OpenRouter, 80 per cent of token consumption today is generated by agents. The architecture that makes this load land on DeepSeek specifically is described in one paragraph: office-AI inference is usually served on cloud vendors’ copies of the open weights, but copies have finite capacity, and when a vendor’s own pool saturates it overflows to the official API — which is why WorkBuddy and Yuanbao went down with DeepSeek in the same incidents. And the last sentence of the report’s setup is the one this briefing picks up: “就在DeepSeek宣布对免费用户加以限制之前,字节跳动旗下豆包办公也官宣接入包括DeepSeek在内的模型” — ByteDance’s Doubao Office announced integration with DeepSeek and other models immediately before the limits were imposed.
Institutional Analysis & Commentary:
This series has spent six weeks documenting who pays for the production of an open-source commons in China. Today’s document is the first that shows, with a meter, who pays for its consumption — and the answer is that nobody does, which is why the meter was installed.
First, name precisely what DeepSeek has just done, because it is easily misfiled as a commercial decision and it is a governance one. Since V3, DeepSeek’s licence has been MIT — examined at length in yesterday’s 2026-11-16 entry, and exceptional in a half-year when Kimi, Qwen and MiniMax all added revenue gates — so no user of the weights owes DeepSeek anything, and every cloud vendor may host a copy and sell access to it. What DeepSeek operates at chat.deepseek.com is therefore not the product its licence sells; it is a free public terminal bolted onto an asset that anyone may copy. The 40,000-character cap, the seven-day conversation expiry and the daily quota are not licence terms and not product tiers. They are the operating rules of a public facility under congestion, and the facility belongs to a company whose own licence guarantees that its competitors may build a better one next door. An MIT licence creates an obligation-free commons; it creates no institution to fund the queue.
Second, the mechanism is a textbook common-pool problem and the shape of the failure is diagnostic. Open weights are non-rival in consumption — one more inference does not diminish the file — but the compute that serves them is rival, congestible and excludable. Those are two different goods sold under one name. Ostrom’s design principles were derived for the first and say nothing about the second; the second is what Elinor Ostrom’s later work and the commons literature call a provision problem, and provision problems are solved by a boundary and a levy, not by a norm. DeepSeek has just discovered the boundary — a daily quota — without the levy, and the quota is being applied to the only class of user it can identify for free: the one that does not pay. Note the resulting distributional outcome, which is the opposite of the one a commons is supposed to produce: the paying API customer and the cloud vendor’s overflow traffic keep full service; the individual who has no account, no contract and no representative gets the 40,000-character ceiling. In a commons, the marginal user is the one you protect; here the marginal user is the one you meter, because the marginal user is the only party with no invoice.
Third, and this is the finding that connects today’s brief to the rest of this series: the agent economy broke the free-rider accounting in a way no policy document anticipated. For two years, the implicit deal of Chinese open-weight AI was that a free consumer terminal buys mindshare, and mindshare converts into API revenue from developers — a marketing cost booked under the heading of ecosystem. That deal held while consumption was conversational: a chat turn is small, human-paced, and self-limiting because a person can only type so fast. An agent removes the human pace limit and multiplies the cost per unit of demand by four to fifteen. Forty million monthly actives that chat are a manageable marketing expense; forty million actives whose agents each run a hundred tool-calling loops are a data centre. DeepSeek did not change its licence, its model or its business model on 2026-11-16. It changed its stochastic demand into a deterministic one and then discovered that the marketing channel and the product were the same pipe. Every Chinese lab that gave away inference to build habit is now on the same curve, one quarter behind or ahead.
Fourth, the part that deserves the scepticism this series applies to every generous headline. The 57 outages and the five-hour global failure are a supply-side indictment disguised as an operations statistic: the weights were free and the capacity was not, and the gap between the two was filled by a status page. But note who actually absorbed the loss. API customers were compensated with credit; free users were given a smaller window. That is coherent from a firm’s balance sheet and it is exactly the allocation that an unpriced public good produces: liability flows to the party with a contract, service flows to the party with one too. If Chinese open-weight models are to be, as the Wuzhen framing quoted on 2026-11-16 put it, “a path to sovereignty for the Global South,” then the constraint on that path is not licence permissiveness — it is whether anyone is willing to pay for the queue in Lagos, Jakarta and São Paulo, where the user is even further from the invoice than the user in Chengdu.
Fifth, the falsifiable tests. If the free tier is being repriced rather than rationed, a paid consumer subscription with a published price and a published quota will appear within one quarter; a firm that has found the provision problem builds a levy. If instead the caps keep tightening with no price attached, the correct reading is that DeepSeek has decided to withdraw the consumer terminal as a channel and route all consumer demand through third-party harnesses — WorkBuddy, Doubao Office, Yuanbao — which is a re-intermediation of the commons, with the platform capturing the interface and DeepSeek becoming a wholesaler of weights and tokens. Watch one number for the tell: whether third-party platforms begin publishing token-pricing tiers of their own before DeepSeek publishes one.
Checkable questions.
(i) Will DeepSeek publish the quota it has imposed — the daily cap, the reset time, the basis — or will the limit remain visible only as a pop-up at the moment of refusal? A rationed facility that will not publish its ration is not allocating by rule; it is allocating by whoever happens to be lucky.
(ii) Do the cloud vendors hosting DeepSeek’s open weights publish capacity limits, or do they continue to overflow silently to the official API? The overflow path means a vendor can sell a service it cannot deliver and bill the upstream for the shortfall. That is a subsidy from the model owner to the reseller, and it is invisible in every price.
(iii) Does any Chinese lab respond by licensing its weights with a service-level commitment to the public terminal — a rationed but guaranteed free tier, funded from API revenue — and publish that commitment as a term? If one does, open weights will have acquired the institution they have always lacked: not a licence that grants freedom, but a covenant that funds the queue.
Sources: 21世纪经济报道 via 腾讯新闻 — 宕机频发,DeepSeek成立以来首次限制免费用户 (2026-11-16; the three measures; the first-ever restriction on free users since founding; the May 2026 arbitration-control measures and their distinction from the November capacity measures; 57 incidents on the status page in 2026; the 15 October 08:11–13:04 global outage and the 7 November repeat; WorkBuddy’s 32.35m monthly actives, 8.36bn monthly uses, 9bn cumulative task requests and fiftyfold token growth; Dowson Tong on token intensity; Anthropic’s 4x and 15x agent-turn multipliers; OpenRouter’s 80 per cent agentic share; the vendor-copy overflow path to the official API; Doubao Office’s integration announcement) · 新浪财经 — DeepSeek开始对免费用户限额了 (2026-11-16, same report syndicated; the 40,000-character single-input cap, the seven-day inactivity limit, the daily quota and the display string) · This series’ 2026-11-16 briefing — the behaviour condition, the scale condition and the subject condition, including DeepSeek’s continued use of MIT
🎓 Education & Institutions — A Luggage Tag for a Person Already Carrying the Bag: What China’s Open-Source Money Measures, and the One Thing It Cannot
1. A comparison chapter, and why it is usable. On 2026-11-03, version 1.0 of “How Europe and China fund open source” was published by the Europe–China Digital Commons Observatory (eu-china-commons.eu; source public on GitHub under CC BY 4.0), closing four research branches of the observatory’s register. The method is the reason to quote it: every public-money record is graded [F] fact with an evidence level A or B, [C] commitment where an amount is announced and not shown as spent, [I] inference, [P] proposal, and the note states its own limits — absence of data means we found nothing, not that nothing exists. The data set: 24 European and 21 Chinese public open-source-specific funding records; 23 European private corporate records through foundations plus OpenAtom’s 2020–23 filings; and 34 funded projects, 18 Chinese-origin and 16 European. This is the observatory whose OCR pass over OpenAtom’s Ministry of Civil Affairs filings this series examined on 2026-11-10; the funding chapter is the companion piece, and it has not appeared here before.
2. The structural findings, stated as the note states them. (a) Amounts are not summable on the Chinese side. Documented European public money that can be added up is small — €11.5 million (2023), €36 million (2024), €19 million (2025), Germany and France only — sitting under multi-year envelopes of DARE up to €120 million, NGI Zero ≈€65.5 million (2018–27), OpenEuroLLM €20.6 million, Chips JU RISC-V €60 million. No Chinese public open-source amount could be summed at all, because China publishes per-project caps and rates, not budgets or disbursements — the largest caps being Wuhan East Lake ¥30 million, Shanghai ¥20 million, Shenzhen ¥10 million — and no central-government open-source budget line was found. (b) The instruments point in different directions. China pays for adoption; Europe pays for creation and maintenance. Shenzhen’s 15 per cent of investment, plus 5 per cent for OpenHarmony-adapted RISC-V, prize-style grants paid after deployment, per-device rates; against German Sovereign Tech Fund maintenance grants — curl €195,000, OpenSSL €405,888, KDE €1,285,200 — and competitive grants for new strategic technology. (c) Stack-dependence. Chinese money is tied to named national stacks (OpenHarmony, openEuler, RISC-V, OpenAtom-hosted projects) and to local registration; European money is stack-neutral and conditioned on open licences and co-funding. (d) Who decides. In China, local industry bureaus and development zones, with the centre acting through OpenAtom under MIIT supervision; in Europe, Commission programmes and a few national agencies, with 23 of 27 member states showing no dedicated money in the search.
3. The project-level findings, which are the ones that bite. Chinese flagships depend on one company each — Huawei for OpenHarmony and openEuler, with the top funder close to 100 per cent, and Chinese-origin projects at Apache keeping the founding company as payer — against European projects with more diverse multinational funders, mostly subscriptions or dual licensing. Individual donations are nearly absent in China: Ant Design about US$18,000 all-time on Open Collective**, KubeSphere **US$0, against curl at about **US$91,000 a year** (the 2026-11-10 entry cited US$499,000 cumulative for curl; the two figures measure different periods and the note’s own number is the annual one). Governments as paying customers of open-source vendors is a European channel — BSI’s VS-Desktop for GnuPG, BwMessenger/Tchap for Element, openDesk — whereas the closest Chinese equivalent, 信创 procurement of commercial openEuler distributions, buys the proprietary product layer, not the commons. And the related-party fact already recorded on 2026-11-10 reappears: ¥6.49 million of ¥25.8 million of OpenHarmony spending went to Huawei Cloud in 2023.
4. The finding the note states as the largest asymmetry of all. Transparency. European budgets and grants are published per project; Chinese support is visible as rules and ceilings, with award notices scattered across local portals and uncollected. The note’s own conclusion is therefore stated against the headline: any “China spends X versus Europe Y” claim is today unfounded in both directions. It also records, and this matters for this series’ recurring licence scepticism, that its own instrument taxonomy treats OpenAtom’s donations as instrument_type: donation, summable: false, and excludes them from public-funding totals, with only the ¥20.7 million 2023 subsidy treated as a public grant — and even that flagged unsummable until the granting body is verified.
Institutional Analysis & Commentary:
A comparison chapter is only useful if it changes what one counts, and this one does: it shows that the two systems do not under-fund or over-fund the same object. They fund two different stages of the same production function, and only one stage has a payee who can be identified as a person.
First, put the three funding modes on a single axis so the difference stops being a matter of national taste. A creation grant buys a deliverable: money in, artefact out, audited against a specification. An adoption subsidy buys an installed base: money in, units deployed, audited against a device count. A maintenance grant buys continuity: money in, and what comes out is that a specific person is still reachable next month, still has credentials on the repository, and still answers. Only the third is a payment for a relationship rather than for an object — and only the third can be spent on the thing this series has been asking about since October, which is the thirty minutes a maintainer takes to tell a stranger why their patch is wrong. The observatory’s finding is that Europe buys mostly creation and a little maintenance; China buys mostly adoption. That is not a ranking. It is a statement about which institutions each system has: a ministry can specify a deliverable and a development zone can count a device, but only a fund with per-person grants can pay a maintainer, and maintenance funds are rare everywhere, not only in China.
Second, the adoption mode has an exact and under-remarked consequence: it makes the beneficiary a buyer, and a buyer has rights against the thing bought. An installed base that the state paid to create is an installed base whose owner can be told to adapt, certify, patch and interoperate — which is precisely the pattern this series has documented in the certification fee (2026-11-13), the adaptor certificate as a purchasing channel (2026-11-12), and procurement of intelligent-OS zones (2026-11-15). A grant to a creator creates an obligation to publish; a subsidy to an adopter creates an obligation to comply. Both are legitimate instruments. They produce opposite postures toward the upstream, and the observatory’s stack-dependence finding — Chinese money tied to named national stacks and local registration — is the arithmetic of the second. Europe’s stack-neutrality is not a moral achievement; it follows from the fact that the Commission is buying research output from a supplier pool it cannot enumerate in advance, so it must accept substitutes.
Third, the transparency finding is not a complaint about disclosure culture, and it should not be read as one. The note is explicit: China publishes rules and ceilings, Europe publishes budgets and disbursements — and the reason is the decision unit, not the virtue of the publisher. A ministry that sets a ceiling and delegates selection to development zones does not possess the number; the disbursements happen in fifty municipal portals and are awarded by bodies whose own rules are the instrument. The published artefact of a delegated subsidy regime is necessarily a rule, because the rule is what the centre actually decided. If one wants award notices, one is asking the municipalities, and asking them is a different research project. This series’ own instrument for exactly that — the “which projects have actually been paid the CNY 3 million donation subsidy” question filed against the BDA schedule on 2026-11-16 — is the same request, addressed to a district.
Fourth, the single-company dependence finding, which is the one with the shortest time horizon. The note’s "[F]/[I] Chinese flagships depend on one company each, with the top funder close to 100 per cent" is a resilience statement and it does not depend on any comparison with Europe. A project whose largest donor is also its original owner, whose cloud contract is with that donor and whose principal contributors are that donor’s employees is not thereby captured — it is dependent, and the difference shows up the day the donor’s strategy changes. This series found the same thing from the other direction on 2026-11-10: a foundation whose money circulates back to its largest donor cannot use its budget to discipline the donor, only to serve the donor’s project better. The observatory’s contribution is the measurement: near-100 per cent concentration in the Chinese case, diverse multinational funding in the European, and individual donations of US$18,000 all-time on the largest Chinese design system against US$91,000 a year for a single European utility project. The last number is a fact about a payment rail and a tax regime at least as much as about culture, and the note says so.
Fifth, the falsifiable predictions, and they are unusually clean because the note has already written the next round. If Chinese award notices (拟资助名单) are collected from Shenzhen, Shanghai, Wuhan, Hangzhou and Guangdong and the disclosed awards cluster on a handful of projects belonging to firms inside named development zones, then the “adoption” characterisation will be confirmed at the disbursement level rather than merely at the rule level — and the ¥30 million East Lake ceiling will turn out to be a number almost nobody receives. If, conversely, the awards are spread across many small recipients, the ceiling is a marketing figure and the actual instrument is a grant programme with a hidden floor; both are possible, and only the notices decide. And on the maintenance question: if by the next version any Chinese instrument — municipal, provincial or foundation — pays a named individual maintainer rather than an entity, the funding architecture will have acquired the one feature the comparison shows neither system has in quantity.
Checkable questions.
(i) Does the observatory’s next round recover a single Chinese award notice with a named recipient and an amount? The note lists this as an explicit next step with five jurisdictions named. One recovered notice would convert the “rules not budgets” finding from an inference into a fact with a date on it.
(ii) Will OpenAtom publish its 2024 and 2025 年度工作报告 — and if it does, will the ¥150 million annual commitment for 2026–2028 appear as restricted income with a named purpose? An observatory that already OCR’d four years of scans will read them on the day they appear; the question is whether they appear.
(iii) Is there any Chinese channel — a foundation, a city, a university, a company — that pays an individual for maintenance work rather than an organisation for a project? The comparison implies the answer today is effectively no; a single published counter-example would move the whole series’ conclusion from “structurally absent” to “structurally rare”, which is a very different claim.
Sources: Europe–China Digital Commons Observatory — How Europe and China fund open source: chapter with conclusions, v1.0 (2026-11-03; the 24 EU / 21 CN public-money records; the per-project caps — Wuhan East Lake ¥30m, Shanghai ¥20m, Shenzhen ¥10m — and the absence of any central budget line; OpenAtom ¥270m filed income 2020–23 with Huawei at 37 per cent in 2023 and ¥20.7m of government subsidy; the Sovereign Tech Fund figures for curl, OpenSSL and KDE; the adoption/creation/maintenance finding; stack-binding and local registration versus stack neutrality; the near-100 per cent single-funder concentration in Chinese flagships; Ant Design’s ≈US$18k all-time and KubeSphere's US$0 against curl’s ≈US$91k a year; the ¥6.49m of ¥25.8m OpenHarmony spending to Huawei Cloud in 2023; the transparency asymmetry and the explicit refusal to compare headline totals; the next-round list of Chinese award notices) · Source file on GitHub — docs/FUNDING-CHAPTER.md, updated 2026-11-08 · This series’ 2026-11-10 briefing — reading OpenAtom’s statutory accounts · This series’ 2026-11-16 briefing — the BDA donation subsidy and its national-level-organisation condition
🏗️ Institutional Change — Twenty-One Years of the Same Conference, and a Linen Table With No Counter: The Single Report From 2026-11-16 That Describes the Inbound Half
1. The report, and the one admission in it. On 2026-11-16 Sina Technology’s rolling news carried, from A5 Venture News, the announcement that the twenty-first China Linux Kernel Developer Conference (CLK 2026) will be held on 2026-11-21 at the Yun Valley Centre in Hangzhou — one main forum, five technical sub-forums and an open BoF track, with support from the Hangzhou Open Source AI Foundation, the Yun Valley Centre, vivo, Alibaba Cloud, Volcano Engine, OPPO and Xiaomi. CLK has run since 2006, twenty editions, with the stated principles “freedom, collaboration, innovation” and a stated “purely technical, non-commercial” positioning; it uses a dual track of public call-for-papers and committee invitation, every submission reviewed by a CLK mentor panel on technical novelty, depth, community value and practical applicability, and registration is free. The three main-forum topics: advancing RISC-V in the data centre from the kernel’s perspective; kernel R&D and test-sandbox infrastructure in the AI-agent era; and the 2026 Linux kernel annual review — technical evolution and community change. The BoF topics are the sentence this entry was waiting for: agent sandboxes, Android kernel upstreaming, CVE response, and kernel development in the AI era.
2. Why one line about “upstreaming” outweighs a week of attendance figures. “Android内核上游” — Android kernel upstreaming — is a topic about sending patches from a vendor’s fork back into a shared tree maintained by people the vendor does not employ. It is the single most expensive, least rewarded and most institutionally demanding act in the entire Chinese kernel economy, and it is the act this series has been unable to find a paying instrument for in six weeks of policy documents. Everything else in the announcement is a room. The conference has a room, a mentor panel, sponsors, a free door and a programme; the CVE应对 topic implies a response process and the upstream topic implies a patch path. A twenty-one-year-old volunteer conference that puts those two topics on a BoF agenda is closer to the missing institution than any subsidy examined in this series — and it costs a linen table and a microphone.
3. The contrast with the same fortnight, stated without exaggeration. On 2026-11-14 and 15, COSCon'26 ran in the same city, in the same Yun Valley Centre, under the guidance of a municipal AI open-source college, co-hosted by KAIYUANSHE — a volunteer community with no registered legal entity — and Datawhale, with two newsrooms as co-hosts, and media co-hosting a Chinese open-source conference for the first time. The reported record: more than 1,000 attendees on day one; a provincial news client reporting 654,000 views on a single story; 20-plus children as speakers, the youngest nine years old, in a youth forum that inverted the adult-on-stage format, described by KAIYUANSHE’s chairman Lin Lüqiang (林旅强) as “letting the next generation take the lead”; the first “open source mentor” (开源导师) credentials; the OPC nine-level talent capability standard already issued in September by the Hangzhou AI open-source ecosystem college; and the China AI Open Source Ecosystem Blue Book in Chinese and English, whose lead author Xu Xuchu (徐旭初) of Hangzhou Dianzi University — founding dean of that college — gave its framework as five questions (what, why, current state, problems, how to advance) and its finding as “application-pulled plus multi-party collaboration” (应用牵引+多元协同), with an explicit and candid admission that of four ecosystem models — central enterprise-led, project-peripheral, standard-exploring, and neutral-governance — China is mostly the first two and weak in the last two. His sixteen-character goal: open sharing, technology enablement, innovation and employment, autonomous controllability (开源共享、技术赋能、创新就业、自主可控).
4. The two days, side by side, and the difference that is not quality. Both conferences are volunteer-organised, both are in Yun Valley, both are free or cheap to attend, and both are, by the standards of any ecosystem, good. One of them announced its yearly growth in a metric that a sponsor can buy — reach — and one spent its announcement on the two activities that no sponsor pays for: responding to a CVE and pushing a patch upstream. That is not a criticism of COSCon'26, which was doing the work a community conference does when its funders are a district government, a municipal college and a media group. It is an observation about which activities can be reported upward, and therefore which activities get organised.
Institutional Analysis & Commentary:
Six weeks of this series has produced one repeated finding — that every instrument in the Chinese open-source ecosystem is fitted to a transaction with a legal person, and none to the transaction with a stranger. Today’s material supplies the missing half of the explanation, and it is not about money.
First, distinguish the two halves of a commons by their audience, because that is what determines whether an institution can be built around them. The outbound half — publishing, licensing, donating, releasing weights, announcing a standard — is performed for an audience of institutions, and every act in it produces a document: a licence file, a donation agreement, a press release, a standards number. The inbound half — reviewing, triaging, answering, reproducing a bug at three in the morning, telling a stranger no with a reason — is performed for an audience of one person, and it produces no document that survives the interaction. The value of a good review is in the recipient’s next attempt, which is not an artefact and is not attributable. This is why an administrative state can see the outbound half and cannot see the inbound half, and why the gap is not a funding gap but a legibility gap. A budget line requires a payee, a payee requires an invoice, an invoice requires a deliverable, and the deliverable of maintenance is the absence of incidents. You cannot invoice for the outage that did not happen.
Second, the CLK evidence shows the inbound half can be organised — by a body with no budget and no legal personality, over twenty-one years, on a definition of value that is not a KPI. Look at what the conference actually is in institutional terms: it convenes a mentor panel with standing to judge, it publishes the judging criteria (novelty, depth, community value, applicability), it rejects the commercial framing by rule (purely technical, non-commercial), and it has survived five changes of sponsoring company — Intel, Huawei, Alibaba Cloud, ByteDance, vivo, OPPO, Xiaomi and Longsoon have all been supporters; none of them is the organiser. Sustaining an institution for twenty-one years through the rise and fall of every sponsor is a governance achievement of exactly the type this series keeps saying is missing, and it is done by volunteers. The BoF format is worth naming separately: an unrecorded, unconference-style open round table is the only format in which a person can say “we cannot get our patches upstream and here is why” without it becoming a document that someone’s employer must approve. Some of the most important institutions in open source are deliberately undocumented, and every policy instrument that requires a report will fail to reproduce them.
Third, the COSCon'26 numbers are the mirror image, and the mirror is honest. 1,000 attendees, 654,000 views, 20-plus children, a nine-level talent standard, mentor credentials, a blue book in two languages. Every one of these is a countable output, and every one of them is produced by an organisation. A nine-level capability standard is a genuine institution — it creates a ladder and a shared vocabulary, and the 2026-11-16 entry found Hangzhou’s OPC ladder in exactly that role. But a ladder is a classification of people, and classification is the instrument of the entity doing the classifying. Note what the standard measures: capability levels of persons, not merged patches in repositories. And note the author of the blue book’s own words: the neutral-governance model is weak in China. The most candid sentence in a fortnight of documents is a municipal college dean saying, in a municipal conference, that the model in which no single party controls the project is the one his country is weakest at.
Fourth, the synthesis, and it is the reason this entry pairs a kernel BoF with a blue book. The outbound half of Chinese open source is now over-instrumented and the inbound half is under-instrumented, and the two facts have the same cause: instruments are built for what can be reported. Subsidies report donations; districts report enterprises; colleges report courses; conferences report attendance; standards report levels. Nobody reports a review. So the ecosystem has, simultaneously, the world’s tightest schedule of adoption subsidies (2026-11-16) and a genuine nine-year-old on a stage in a youth forum, and no line item anywhere that pays the thirty minutes a kernel maintainer spends explaining why not. But it also has, in the same city in the same week, a twenty-one-year-old volunteer conference whose BoF agenda names Android upstreaming and CVE response — which is what the inbound half looks like when it is organised by people who do the work rather than by people who count it.
Fifth, the falsifiable test, and this one is unusually concrete for a claim about culture. If the CLK 2026 BoF on Android kernel upstreaming produces, within twelve months, a publicly readable record — a mailing-list thread, a merge, a maintainer’s summary, an LWN article — of a patch path from a Chinese vendor fork into the upstream tree, then the volunteer conference has accomplished what six weeks of policy documents could not, and the correct conclusion is that the missing institution is a room and a mentor panel. If instead the twice-yearly reporting cycle shows the same firms reporting the same attendance at the same venues while upstreaming statistics stay flat, then the inbound half was never on the agenda at all, and the youth forum, the mentor credentials and the nine-level standard will have been what the ecosystem decided a person’s contribution means. And the cheap marker, available today: does CLK 2026 publish its BoF outcomes at all? An institution that meets and does not record is a conversation; one that records is a commons.
Checkable questions.
(i) Who sits on the CLK mentor panel in 2026, and what share of accepted talks come from the open call rather than committee invitation? The dual track is the whole design: an invitation-only programme is a sponsor showcase, and an open call adjudicated by named reviewers is a commons. The ratio is published in the programme and nobody has computed it.
(ii) Does any CLK sponsor report the conference as an open-source-contribution KPI, and if so, which KPI — attendance, talks accepted, or patches upstreamed? A sponsor that measures attendance is buying visibility; a sponsor that measures merged patches has found the instrument.
(iii) Does the Hangzhou OPC nine-level standard, or any comparable capability ladder, include a level that can only be attained by a merged patch reviewed by a stranger? If every level is achievable by coursework, examination and project delivery, the ladder measures training. If one level requires a merge, the ladder has accidentally imported a verifier with standing — and that is the whole missing institution, appearing inside a talent standard.
Sources: 新浪科技 / A5创业网 — 第二十一届中国Linux内核开发者大会将于11月21日在杭州举行 (2026-11-16; CLK 2026 at Yun Valley Centre Hangzhou on 2026-11-21; one main forum, five technical sub-forums and a BoF track; support from the Hangzhou Open Source AI Foundation, vivo, Alibaba Cloud, Volcano Engine, OPPO and Xiaomi; the 2006 founding, twenty editions, “freedom, collaboration, innovation” and the purely-technical non-commercial positioning; the dual public-call plus committee-invitation track; the mentor panel’s four review criteria; free registration; the three main-forum topics including RISC-V in the data centre and AI-agent-era kernel test sandboxes; the BoF topics including agent sandboxes, Android kernel upstreaming, CVE response and AI-era kernel development) · 中国Linux内核开发者大会 — official site (the founding institutions — Tsinghua, Intel, Huawei, Alibaba Cloud, Fujitsu Nanda, Digi, Tencent Cloud, OPPO, ByteDance, vivo, Ant Group, Loongson; the near-500 in-person attendance per edition; the 2025 programme’s reviewer panel and speaker affiliations; the call-for-papers criteria) · 潮新闻 — 中国开源年会首次落地杭州:一座城市的生态共建与长期主义 (2026-11-14; COSCon'26 on 2026-11-14/15 at Yun Valley Centre, “开源下一代 / Open Source: The Next Generation”; KAIYUANSHE and Datawhale as hosts with two newsrooms co-hosting and media co-hosting for the first time; 1,000-plus first-day attendees; 20-plus child speakers with the youngest nine; the 2026-09 establishment of the national first AI open-source ecosystem college and its OPC nine-level capability standard; the first open-source mentor credentials; Lin Lüqiang’s framing; Xu Xuchu’s five questions and the “application-pulled plus multi-party” finding; the four-model taxonomy and the admission that the neutral-governance model is the weakest) · 钱江晚报 via 搜狐 — 《2026中国人工智能开源生态蓝皮书》在杭州发布 (2026-11-15; the Blue Book’s joint publishers including Datawhale, the Zhejiang AI Academy, the Hangzhou AI open-source ecosystem college, Zhejiang University, Shanghai Jiao Tong University, ModelScope, Chuanbo Danao and Chao News; Xu Xuchu’s four-level rationale; the sixteen-character goal 开源共享、技术赋能、创新就业、自主可控; the “点状突破 to 系统构建” route and the call for patience capital and an inclusive engineering culture) · 新浪财经 — 中国开源年会在杭举行,中国人工智能开源生态蓝皮书发布 (2026-11-15; the Blue Book in Chinese and English and the mentor credentialing at the main forum)
🔍 WeChat Monitor
No manual feed for 2026-11-16 was present in china-daily-english-input/raw/ at the time of this run (the repository was pulled and is up to date; the most recent file is 2026-11-12.md). Section skipped.
🔍 Commentary — The Part That Was Never on the Ledger
First, this morning’s three documents are one document, and it is the first in this series that shows the two sides of an open-source economy in the same units. DeepSeek measured its consumption to the character — 40,000 per input, seven days per conversation, a daily quota — and found it unpayable. The funding comparison measured China’s production support and found that it is published as caps (¥30 million at the top, ¥10 million at the bottom) that no one is required to report receiving, that it flows overwhelmingly to one company per flagship, and that individual contributions to the largest Chinese design system total US$18,000 across the project’s entire life. CLK measured its inbound work — Android upstreaming, CVE response — in a BoF topic with no minutes. Consumption is metered exactly. Production is funded by ceiling. Maintenance is discussed aloud. That ordering is not an accident of reporting; it is the shape of what an administrative system can see.
Second, the DeepSeek case settles a question this series has been circling since the licence entry of 2026-11-16, and settles it against the optimistic reading. The argument there was that an MIT licence plus a large free model is the most generous arrangement in the industry — and it is. But generosity on the licence is not generosity in the provision, and 2026-11-16 is the date on which those two things came apart in public. A model released under MIT is a gift with no funder. A free terminal serving it is a facility with a cost. For two years the facility was financed as marketing, and the accounting worked because human conversation is slow. Agents removed the brake. So the first casualty of the agent economy, in China, was not a job and not a licence. It was the free tier — and the people who lost it are, by construction, the people with no contract to compensate.
Third, name what is actually missing, in one sentence, because six weeks of instruments deserve a formulation that fits in a policy memo. The ecosystem has built an apparatus that pays for the creation of commons, counts the adoption of commons, certifies the people who use commons, and standardises the levels they reach — and it has no instrument at all for the provision of the commons, which is a recurrent cost with no deliverable, no invoice and no payee, because the payee is a named person and the deliverable is the absence of an incident. This is why a CNY 3 million donation subsidy and a 40,000-character free-tier cap can appear in the same fortnight without contradiction: they are the two ends of a supply chain whose middle segment — the part that maintains anything — is not on either ledger.
Fourth, the honest caveat, and it cuts against the thesis in a way that matters. It is tempting to conclude that the volunteer community is the answer — CLK for twenty-one years, COSCon for eleven, both without a corporate owner. But both are conferences, and a conference is an annual event with a floor plan; a commons is a weekly obligation. The reason CLK can exist without a foundation is that the kernel it discusses is maintained elsewhere, by people paid by employers in other jurisdictions — which is to say the inbound institution it cares about is, in the main, imported. The BoF on Android kernel upstreaming is a conversation about joining someone else’s maintenance regime, and that is a legitimate and valuable thing to do; it is not the same as having built one. The distinguishing test of a Chinese commons institution will not be whether it holds a good conference, but whether a developer in Chengdu or Shenzhen can spend professional hours each week reviewing code written by people they have never met, for a project whose other maintainers they do not work for, and be paid, promoted or at least not penalised for it. No document in this fortnight establishes that, and one — the Chongqing clause quoted on 2026-11-16 about teachers’ community contributions counting toward professional-title assessment — is the only thing that gestures at it.
Fifth, what to watch, concretely, and none of it requires a subscription. Twenty-one days from now the CLK programme lists four judges’ names. Look at their affiliations: if the mentor panel is composed of people whose employers fund the conference, the reviewing is internal; if it includes maintainers from projects and companies that are not sponsors, the panel has standing independent of the budget. Within one quarter, DeepSeek either publishes a quota or a price. A quota with no price is a facility being withdrawn; a price is a provision institution being built. And within one funding cycle, one Chinese award notice surfaces with a recipient and an amount — because a subsidy regime that publishes its ceiling and not its disbursements is not hiding anything in particular; it simply has nothing to publish until fifty district portals are read.
Sources: 21世纪经济报道 — 首次限制免费用户 (2026-11-16) · Europe–China Digital Commons Observatory — How Europe and China fund open source (v1.0, 2026-11-03) · 新浪科技 — 第二十一届中国Linux内核开发者大会 (2026-11-16) · 潮新闻 — COSCon'26 首次落地杭州 (2026-11-14) · This series’ 2026-11-16 briefing — the scale condition, the donation subsidy and the Chongqing teacher-contribution clause · This series’ 2026-11-15 briefing — the committee arrives before the community · This series’ 2026-11-10 briefing — reading OpenAtom’s statutory accounts