⚠️ 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 — 2026-11-19

🏛️ Policy & Regulation — The Refusal Right: How a Bank Passed a National Open-Source Governance Assessment by Writing Down How It Leaves

1. The document, and why a bank’s internal rulebook is the day’s policy document. On 2026-11-10, at CAICT’s 2026 OSCAR Open Source Industry Conference in Beijing, Bank of Communications (交通银行股份有限公司, BoCom) was announced as having passed CAICT’s Trusted Open-Source Governance Maturity Assessment (可信开源治理能力成熟度评估) at the advanced grade (先进级), which the announcing account describes as reaching a nationally leading level of capability. The announcement was carried on the OSCAR Trusted Open Source (OSCAR可信开源) WeChat account on 2026-11-18 and circulated by the 2026-11-18 manual feed. This series covered the OSCAR conference itself on 2026-11-11, reading the extension of CAICT’s assessment system from projects to actors — from trusted-open-source governance to a Trusted AI Supply Chain. What it did not have, and what this feed supplies, is the content of an assessed institution’s own rulebook, in an assessed institution’s own words. That is a different kind of evidence.

2. What the bank says it built. Four instruments, named. (a) Rules: a 《开源软件管理办法》 (Regulation on the Management of Open-Source Software) and a 《开源依赖库管理细则》 (Detailed Rules on the Management of Open-Source Dependency Libraries). (b) A two-tier repository regime: for the larger open-source software it uses, BoCom runs dual repositories (双库管理) that separate “new technology” from “standardised technology” — a bifurcation that keeps an evaluated, approved, frozen stack distinct from an experimental one — while the much more numerous open-source dependency libraries are managed as a single repository (单库管理). (c) A platform: a group-wide open-source software governance platform supporting the full lifecycle of introduction, use, upgrade and exit (引入、使用、升级、退出), with vulnerability detection, security early warning and open-source licence risk analysis. (d) An organisation: a technical governance committee system (技术治理委员会体系), a software technology register and repository, and a stated link between static governance and dynamic governance. The bank describes its path in four phrases: “institutions first, platform as carrier, security embedded, benchmark-guided” (制度先行、平台承载、安全内嵌、对标牵引), under the guidance of the regulator’s 《关于规范金融业开源技术应用与发展的意见》, and notes it is exploring an “open source plus closed source” (开源+闭源) combination in AI.

3. The assessment’s own unit of account. CAICT’s framework, per the same announcement, runs on two dimensions — process and capability — and ten assessment directions, examining organisational mechanism, management institutions, risk management, plus six operational options including use management, operations management, periodic health assessment, exit management, legacy inventory (存量盘点) and third-party management. It is offered to any enterprise that uses open-source software, as a way of assessing internal governance capability. Note the two words in that list that no open-source community has ever put in an assessment framework: exit management and legacy inventory.

Institutional Analysis & Commentary:

This is the most institutionally complete document the series has examined from a Chinese enterprise, and its central instrument is one the open-source world does not have a name for. Call it the refusal right: the documented, audited, organisationally-owned capacity to stop using a piece of software.

First, read the dual-repository rule as a rule about transaction costs, because that is what it is. Williamson’s lesson is that the boundary of a firm is drawn where the cost of governing a transaction inside the hierarchy falls below the cost of buying it outside, and that the operative variable is not the price but the cost of the contract that would protect you — the cost of specifying, monitoring and enforcing it. BoCom’s split between new technology and standardised technology is precisely a governance-cost gradient made into an organisational fact: a new component enters the experimental repository where the cost of being wrong is bounded, and only crosses into the standardised repository after it has been evaluated, which is to say after someone has paid the cost of finding out what the contract actually says. A dependency library, by contrast, is managed as a single repository — not because the risk is lower but because the number of them makes per-unit evaluation uneconomic. That is a textbook make-or-buy calculation applied to software components, and it is stated in a bank’s own rules rather than in an economics paper.

Second, the whole architecture is organised around exit, not around use. The lifecycle is introduction → use → upgrade → exit, and the assessment grades exit management alongside use management. This is the exact institution that a fortnight ago this series found missing from the entire Chinese policy stack: on 2026-11-18 the briefing read a Beijing Academy of Artificial Intelligence vice-president’s proposal that data-export and service-continuity periods be written into procurement contracts and that critical systems switch models once before go-live, and concluded that Chinese open-source policy had no exit right in it. A bank has one. Not because a policy told it to, but because its supervisor told it that open-source risk is its own risk, and a bank that cannot leave a dependency is a bank whose operational resilience belongs to somebody else’s roadmap. The exit right was created by prudential regulation, not by an open-source policy — and that is the finding.

Third, note where the open-source licence sits in this stack, and where it does not. The bank’s licence work is “open-source licence risk analysis” — a component of a risk-management platform, sitting beside vulnerability detection. In this institution the licence is a compliance attribute of an input, catalogued and monitored like a version number. It is not a grant of freedom, not a collaboration protocol, and not a claim on the bank’s own outputs. There is no clause in the announcement about contributing back. The entire apparatus is built to govern what the bank takes in and how it leaves; there is nothing about what it gives out. That is coherent — a bank is a user, and a user owes a commons nothing under a permissive licence — but it should be stated plainly, because it is the same asymmetry the series has documented for six weeks from the distribution side, now visible from the consumption side: the Chinese institution for open source is an intake filter, and there is still no institution for output.

Fourth, the “open source plus closed source” phrase in the AI section is the sentence to keep. It is one clause with no detail, but it names the thing this entire fortnight has been circling: the boundary line inside a deployed AI system between the part the enterprise can inspect and replace and the part it cannot. BoCom’s governance instruments were built for libraries, whose exit is a version change and an API adapter. An AI model’s exit is a data migration, an evaluation-set rebuild and a re-tuning of prompts and pipelines — much harder, much less documented, and made worse by the fact that the serving platform, not the model, holds the data and the evaluation. A bank that can exit a database cannot yet exit a model, and the phrase “open source plus closed source” is the bank admitting it is buying both from the same vendor.

Fifth, the falsifiable tests. Does CAICT’s advanced grade come with a published maturity rubric — a set of levels whose criteria an outside party can check? If the criteria are published, this becomes a transferable institution and other sectors can adopt it; if the grade is a certificate with no public rubric, it is a credential, and this series has spent a month documenting what happens to credentials. Does any bank that passed this assessment publish an exit record — a dependency it actually removed, with the migration cost and the downtime? A right that has never been exercised is a policy; a right that has been exercised is an institution. And: does BoCom’s own open-source output ever appear — a patch to a library it depends on, a vulnerability disclosure, a donation — anywhere in CAICT’s or the foundation’s records? That is the test of whether the intake filter has an output side.

Checkable questions.

(i) Where is the dual-repository boundary actually drawn — by technology category, by business criticality, or by who approved it? The answer tells you whether the split tracks operational risk (a prudential logic) or internal politics (a budget logic), and the two produce very different behaviour when a critical system needs an urgent upgrade.

(ii) Who sits on the technical governance committee, and does it have authority to refuse a business unit’s request to adopt a component? A governance body without a refusal power is a review body, which is a different and much weaker institution.

(iii) Does the “legacy inventory” cover licence contamination risk only, or does it also record which open-source components the bank has patched locally and never upstreamed? The second number is the one that would tell us whether Chinese financial open-source use is generating maintenance debt that nobody can see — including, quite possibly, the bank itself.

Sources: OSCAR可信开源 via WeChat — 开源治理风向标:交通银行股份有限公司通过可信开源治理能力成熟度评估!相关能力达到国内领先水平 (2026-11-18; Bank of Communications passing CAICT’s Trusted Open-Source Governance Maturity Assessment at the advanced grade, announced at the 2026-11-10 OSCAR conference; the 《开源软件管理办法》 and 《开源依赖库管理细则》; dual-repository management separating new technology from standardised technology for larger software and single-repository management for dependency libraries; the group-wide governance platform covering introduction, use, upgrade and exit with vulnerability detection, security early warning and licence risk analysis; the technical governance committee system, software register and static/dynamic governance link; the four-phrase path of institutions-first, platform-as-carrier, security-embedded, benchmark-guided under the 《关于规范金融业开源技术应用与发展的意见》; the “open source plus closed source” AI exploration; CAICT’s two-dimension, ten-direction assessment covering organisational mechanism, management institutions, risk management, use management, operations management, periodic health assessment, exit management, legacy inventory and third-party management) · 中国新闻网 — 聚焦开源发展 2026 OSCAR开源产业大会在北京召开 (2026-11-12; the conference hosted by the China Communications Standards Association and organised by CAICT; Dai Xiaohui and He Baohong’s remarks; the four assessment families released including trusted open-source governance maturity, the OpenChain dual assessment, the SOSS national-standard conformity test and SCA tool capability; the Trusted AI Supply Chain extension to enterprise office agents, agent development platforms, code-security agents and enterprise programming assistants) · This series’ 2026-11-11 briefing — the evaluator moved upstairs: OSCAR and the instrument that grades agents instead of projects · This series’ 2026-11-18 briefing — Chen Bohui’s portability test and the absent exit clause in Chinese procurement


1. The licence, restated because it is the load-bearing fact. On 2026-07-17 Moonshot AI released Kimi K3 — 2.8 trillion total parameters, a mixture-of-experts architecture activating 16 of 896 routed experts and roughly 104 billion parameters per token, a one-million-token context window and native visual understanding — under a bespoke instrument. This series documented on 2026-11-04 what that instrument does: it is open weights plus training and inference code, and it carries a revenue-threshold trigger. A model-as-a-service operator earning above USD 20 million in any twelve consecutive months must conclude a separate agreement with Moonshot AI; a product reaching above 100 million monthly active users or USD 20 million in monthly revenue must display the name “Kimi K3” prominently. The licence has since been described in Chinese financial coverage as the point at which Moonshot’s open route began to show “commercial value” — K3 went to first place on Hugging Face’s trending board, gathering more than 4,000 likes within 30 minutes of release, reported as the platform’s fastest release-growth record.

2. Why a threshold in the licence is a different institution from a threshold in a price list. A permissive licence grants every downstream party an identical, unconditional, non-revocable permission; its economic function is to eliminate a negotiation. A licence with a revenue trigger is not a permission at all — it is a contingent contract, whose activation depends on a fact that the licensee must compute about itself (annualised revenue, monthly actives) and report or renegotiate. This series noted on 2026-11-04 that the threshold was calibrated at a level almost no plausible licensee would trip — a reserved right rather than an active restriction. That reading should now be sharpened: the calibration is what makes it enforceable at all. A threshold set where many licensees sit creates a compliance burden and a disclosure obligation on the vendor; a threshold set where almost nobody sits creates a single renegotiable event with exactly the counterparties that matter, which is to say the inference platforms and the large enterprise resellers. The clause is a screening device for finding the counterparties worth negotiating with, and it is written into the licence because that is the only document every downstream party has already signed.

3. The second clause is the one that scales. The attribution trigger at 100 million monthly actives is, in the trade’s terms, a brand-as-consideration clause: below the threshold you may use the model freely; above it, the model’s name must appear in your product. That is a trademark licence conditioned on scale, and it captures exactly the thing a copyleft licence cannot: the distribution surface. A copyleft clause travels with the code and asks the distributor to give back code. An attribution-at-scale clause travels with the code and asks the distributor to give back visibility — and visibility at a hundred million users is worth more than any source contribution, because it is the asset that converts an open model into a standard. This is Ye Qiwei’s sentence from a fortnight ago — that open source “quietly changes from a tool for catching up into a tool for setting standards” — executed in a licence file rather than announced at a conference.

4. The commercial record the licence is serving. Per Chinese financial coverage of the Hong Kong preparations: API revenue accounts for more than seventy per cent of Kimi’s total revenue; overseas paying users grew 400 per cent, with API revenue also up 400 per cent; the product is in more than 200 countries and regions; and more than seventy per cent of new users after the K3 release came from outside China, with overseas revenue overtaking domestic for the first time in February 2026. The annualised recurring revenue path reported is steep: USD 100 million in March 2026, USD 200 million in May, USD 300 million by mid-June, USD 1 billion by August, with the company telling investors it could reach USD 2 billion by year-end. K3’s release in July caused paid demand to exceed capacity, and Moonshot paused new consumer subscriptions on 2026-07-19 to prioritise existing paying users. One further data point, from the same coverage: consumer marketing spend of roughly USD 100 million in 2024 lifted Kimi’s monthly actives above 20 million; by the end of 2025 that figure had fallen to about 9 million.

Institutional Analysis & Commentary:

The Kimi licence is the first Chinese instrument in this series that governs the value chain rather than the artefact, and it should be read as a legal document with an economic purpose, not as an ideological statement.

First, the licence’s threshold and the company’s revenue curve are the same number seen from two sides. Moonshot’s disclosed position is that seventy per cent of its revenue is API revenue, and the licence’s trigger is denominated in a model-as-a-service operator’s revenue. Put the two together and the strategic logic becomes legible: the company’s revenue model is metered access, and its licence is designed to ensure that metered access is not undercut by a large operator who takes the weights and resells them at a margin the company cannot tax. That is a classic vertical-restraint problem, and the solution chosen — a negotiated agreement triggered at a revenue level — is the least restrictive instrument available short of a plain non-commercial licence. A permissive licence plus a commercial trigger is what “dual licensing” becomes when the open side is a two-point-eight-trillion-parameter model and the closed side is an API business. Nothing about it is dishonest, and nothing about it is a commons.

Second, the attribution clause is where the real institutional novelty sits, and it deserves to be named. Every open-source licence in the standard taxonomy regulates one of two things: what you may do with the code (permission) or what you must do with the code you distribute (copyleft). This clause regulates what you must do with your own product’s surface: at a scale threshold, the model’s name must appear. It is, in effect, a trademark easement granted in exchange for adoption, scaling with the adopter’s reach. The consequence for the Chinese open-weight ecosystem is that “open” now has a second dimension: not only whether the weights are downloadable, but whether the name is required to travel with them. A model family whose terms require attribution above a scale threshold is building a standard, deliberately, through its licence. That is a legitimate strategy and it is exactly how a follower becomes a default. It is also a strategy that a neutral foundation cannot run, which is why the licence is issued by the company and not by a governance body.

Third, the restructuring is a corporate-law document and it says more about open weight than any launch post. Moonshot is unwinding its Cayman Islands holding company, repurchasing the shares held by offshore investors, and having those investors reinvest into the mainland operating entity, while reportedly retaining an option for dollar-denominated transactions through an offshore vehicle — reported by the Wall Street Journal on 2026-11-13, with a listing targeted for early 2027 and a raise of up to USD 5 billion against a private valuation of about USD 50 billion. Read the structure, not the valuation. A company whose core assets are model weights, training infrastructure, data and an API business inside China cannot be listed with those assets held outside China, and a company whose weights are published cannot rely on trade secrecy to ring-fence them. Open weights are precisely what made the offshore structure unnecessary for asset protection and impossible for asset control at the same time — the weights are already out, so the value that must be protected is the entity, the revenue and the register. The open-source strategy and the domestic-entity restructuring are the same decision, taken at different times for different reasons.

Fourth, control, stated as a number, because it is the number that decides what “open” means in three years. Per the enterprise registry cited in Chinese coverage: Moonshot has 13 shareholders; founder Yang Zhilin holds 51.8304 per cent of the mainland entity as controlling shareholder, actual controller and beneficial owner; co-founder Zhou Xinyu 6.5635 per cent, Wu Yuxin 3.9099 per cent, Zhang Yutao 3.331 per cent, the founding natural persons together about 65.63 per cent; institutional holders include Alibaba, Tencent and the social-security fund’s Yangtze River Delta science and technology innovation fund. This series has asked all year whether any Chinese open-source institution holds governance power outside a founder or a sponsor. Here is the answer in an IPO-track company: a majority founder stake, a named donor-portfolio of platform companies and a state pension fund, and no independent governance body anywhere in the structure. The licence’s own thresholds — USD 20 million of operator revenue, 100 million actives — will be set, waived and renegotiated by that majority holder, and there is no committee, no charter and no community that can revise them. That is the governance question the Hong Kong listing prospectus will not be required to answer, and the one an adopter of a “standard-setting” open model should ask.

Fifth, the falsifiable tests. Does the IPO prospectus disclose the K3 licence as a material contract, and how many counterparties have actually entered the USD 20 million negotiation? One number converts the “reserved right” reading into either a fact or a fiction. Does Moonshot publish the weights of its next flagship under the same instrument or a different one? A tightening is evidence the trigger is being used; a loosening toward a standard open licence is evidence the standard-setting strategy needs adoption more than it needs revenue. And: does any downstream operator of K3 across the USD 20 million line disclose the negotiated terms? If none does, the clause functions as a market-segmentation tool that no one is willing to describe in public — which is the relevant institutional fact, and it is observable.

Checkable questions.

(i) Is the USD 20 million measured on revenue attributable to the model, or to the operator’s total business? The two readings produce thresholds that differ by orders of magnitude for a diversified cloud, and the licence text is where the difference lives.

(ii) Does the “Kimi K3” display requirement attach to the model, the derivative, or the hosted service? A requirement that attaches to a derivative fine-tune is a standard-setting clause; one that attaches only to the original weights is a marketing clause. Adoption practice will reveal which.

(iii) After the restructuring, which entity holds the K3 licence and the trademark — the mainland operating entity or the residual offshore vehicle? The answer determines whether a future licence change is a Chinese corporate decision or an offshore one, and every downstream commercial user has an interest in the difference.

Sources: 搜狐科技 / 独角兽早知道 — 月之暗面据报接近完成架构重组,全力为赴港IPO铺路,募资规模或达50亿美元 (2026-11-16; the 2026-11-13 WSJ report on terminating the Cayman holding company, repurchasing early dollar investors’ shares and reinvesting into the mainland entity with an offshore dollar-trading option retained; the early-2027 Hong Kong target and up-to-USD-5-billion raise; the roughly USD 50 billion pre-IPO valuation; the ARR path from USD 100 million in March to USD 1 billion in August with a USD 2 billion year-end target; API revenue above seventy per cent; overseas paying users and API revenue both up 400 per cent; more than 200 countries and regions; the 13-shareholder register with Yang Zhilin at 51.8304 per cent, Zhou Xinyu 6.5635, Wu Yuxin 3.9099, Zhang Yutao 3.331 per cent and the founding team at about 65.63 per cent; Alibaba, Tencent and the social-security fund’s Yangtze River Delta vehicle as institutional holders; K3’s 2.8 trillion total parameters with 16 of 896 experts active per token and 104 billion active parameters; the Hugging Face trending first place and 4,000-plus likes in 30 minutes; the 2026-07-19 pause on consumer subscriptions; the roughly 25-times price-to-sales arithmetic against a USD 2 billion year-end ARR) · EqualOcean — Moonshot AI Reportedly Restructures Ownership Ahead of Planned Hong Kong IPO (2026-11-16, reporting the WSJ account of the Cayman unwind, the repurchase and reinvestment arrangement, the retained offshore dollar-transaction option, the up-to-USD-5-billion raise and the roughly USD 50 billion latest private valuation) · 钛媒体 via 腾讯新闻 — Kimi估值500亿美元,月之暗面为什么急着拆掉海外架构? (2026-11-15; the red-chip structure explained; the three problems of valuation continuity, capital and tax arrangement, and exit channel for dollar funds; the ARR of USD 2 billion targeted for year-end versus USD 1 billion in August; the note that ARR is an annualisation, not recognised revenue, and does not imply profitability; the statement that the company is moving from primary-market financing logic to public-market governance logic) · 创业邦 — 杨植麟,要敲钟了 (2026-11-17; the 2023 founding and funding history from a nearly CNY 2 billion angel round to Alibaba’s USD 1 billion-plus lead in February 2024 at a USD 2.5 billion valuation, Tencent’s entry taking it to USD 3.3 billion, the USD 500 million C round in December 2025 at about USD 4.3 billion, and the climb to USD 50 billion by October 2026; the earlier 78.97 per cent registry stake of November 2024 versus 51.83 per cent now; the USD 100 million consumer marketing spend of 2024, 20 million-plus monthly actives, and the fall to about 9 million by end-2025; the rise of API revenue above seventy per cent; Sacra’s estimate of USD 15 million ARR at end-2025 and the finding that overseas revenue first exceeded domestic in February 2026 and that over seventy per cent of post-K3 new users came from outside China) · This series’ 2026-11-04 briefing — the Kimi K3 License, the USD 20 million MaaS threshold and the 100-million-MAU attribution clause, read against Z.AI’s USD 10 billion threshold · This series’ 2026-11-18 briefing — Ye Qiwei on open source shifting from a tool for catching up to a tool for setting standards


📊 The Ledger — The Shareholder Who Pays the Salaries: A Thousand Times Sales, Then Three Hundred and Thirty-Five Hong Kong Dollars

1. The marks, as of 2026-11-17. Zhipu (智谱, 2513.HK) closed on 2026-11-17 down 5.26 per cent at HK$335**, giving a market capitalisation of **HK$163.5 billion; intraday it touched HK$328**, the lowest since March. Zhipu listed in Hong Kong in **January 2026** at **HK$116.2 per share, rose to HK$2,980** by June — more than **25 times** the offer price and a market value of **HK$1.33 trillion — and has since fallen nearly ninety per cent, with about HK$1.1 trillion** of market value gone. **MiniMax (0100.HK)** closed the same day down **7.13 per cent at HK$116 on a market value of about HK$40.8 billion**, more than **ninety per cent** below its March high of **HK$1,330, below its IPO price since October, and about thirty per cent below that price. At the peaks, MiniMax’s trailing price-to-sales exceeded 600 times and Zhipu’s exceeded 1,000 times.

2. The fundamentals that arrived in between, stated without editorial. Zhipu: first-half 2026 revenue CNY 954 million, up 399.74 per cent year on year, adjusted net loss CNY 1.964 billion, widened 12.11 per cent. MiniMax: first-half revenue up 283.1 per cent to USD 117 million, adjusted net loss widened 52.7 per cent to USD 293 million, with B2B displacing consumer products as the main revenue source. Both companies’ gross margins declined as the business mix shifted. HSBC cut Zhipu’s target price from HK$1,375 to HK$500, a 64 per cent reduction, lowered its 2027–2028 ARR and total-revenue forecasts by 7 to 27 per cent, cut its adjusted operating-margin forecast by 15 to 19 percentage points, and pushed its expected adjusted break-even from 2027 to 2029. The immediate trigger for the 2026-11-17 decline was reported as the US Commerce Department’s Bureau of Industry and Security interim final rule of 2026-11-16 restricting remote-compute arrangements used by third-country data centres — a rule aimed at access to advanced AI chips, and one this series touched on 2026-11-17 from the demand side.

3. The two facts in this dataset that bear directly on open source, and they are the same fact twice. First, HSBC’s and JPMorgan’s reasoning. JPMorgan, in a July note, observed that open weights widen distribution but monetisation still depends heavily on model quality, iteration speed and endpoint reliability; for a leading model, open weights let cloud providers, inference platforms and private deployments on external GPUs expand usage, but for a model without enough differentiation, broader access accelerates price comparison and traffic substitution. Second, Zhipu’s flagship has fallen to 20th on the Artificial Analysis leaderboard and MiniMax’s has dropped off it, with no new flagship from either company since the GLM-5.3 and M3 releases of the summer. A model with open weights and no differentiation is a commodity with a support cost.

Institutional Analysis & Commentary:

This is the first hard price discovery for the proposition that Chinese open-weight AI is a business, and the market has delivered a verdict that the policy layer has not yet read.

First, separate the two things the falling share price is measuring, because they are not the same and only one of them is about open source. The 1,000-times-sales peak was a scarcity price on two companies that were, for a few months, the only listed pure plays on Chinese frontier AI in the only venue open to them. That premium was never a judgment about licensing or governance; it was a judgment about the absence of alternatives. What has repriced since is not openness — it is the arrival of competition in the listed venue plus the arrival of a credible substitute story elsewhere (Moonshot’s privates at USD 50 billion, DeepSeek’s reported USD 105.5 billion second round at a post-money valuation of CNY 615.5 billion). The correct reading is not “open source failed in the market”; it is “the market finished pricing the scarcity and started pricing the product.”

Second, and this is the finding worth the entry: the open licence transferred distribution and retained nothing. JPMorgan’s sentence — for a model without enough differentiation, broader access accelerates price comparison and traffic substitution — is the whole economics of open weights stated in one line, and it applies with full force to a company whose revenue is seventy per cent API while its weights are downloadable by anyone, including the cloud platforms that resell them and the inference platforms that compete on price. The licence made the model a public good at the moments of highest demand; the company then had to earn its revenue on top of a free substitute for its own product. This is not a paradox — companies do this deliberately, because distribution is worth more than exclusivity to a challenger — but it has a precondition that Chinese open-weight labs have been slow to state: differentiation must be renewed faster than the copy is adopted. Zhipu falling to 20th on the benchmark table and MiniMax dropping off it are not marketing embarrassments. They are the licence’s cost, arriving on schedule.

Third, the annualisation problem is the accounting counterpart of the licence problem, and the two should be read together. Moonshot’s USD 2 billion year-end ARR is an annualisation of current billings, explicitly not recognised revenue and explicitly not profit; its coverage says so in plain language. Zhipu’s and MiniMax’s halves show revenue growing at 300 to 400 per cent while gross margins fall and losses widen — the exact signature of a business in which usage grows faster than price and cost per token does not fall fast enough. For the state apparatus that has spent 2026 building vouchers for compute, tokens and data, this is the missing variable: the subsidy is denominated in consumption, and consumption is precisely the line item whose margin is collapsing. A voucher programme that lowers the price of inference is, in this market, a transfer to whoever sells inference — and if the seller is the same platform whose margin is under pressure, the voucher is a working-capital line, not a capability grant.

Fourth, note what the state’s instruments cannot see, and it is the same blind spot in each layer this series has examined. The financial regulator can now grade a bank’s exit management of open-source software (read above). The standards body can grade agents and supply chains. The province can count installations, enterprises and benchmarks. None of these instruments has a line for “the open-weight model your institution depends on was built by a company whose flagship has fallen off the leaderboard and whose losses are widening.” That is not a gap anyone can be blamed for — it is a genuinely hard measurement problem, since model quality is not an audit-grade quantity. But it means the Chinese public and financial sectors are accumulating dependencies on open weights whose continuation is a function of a private company’s ability to raise money at a valuation that the public market has just repriced by ninety per cent. A dependency whose upstream may fold is not a dependency with an exit right; it is an unhedged exposure wearing one.

Fifth, the falsifiable tests. Does either listed company’s next filing disclose ARR alongside recognised revenue, and does it disclose customer concentration? The gap between the two numbers is the annualisation problem in audited form. Does either company cut its model releases — fewer flagships, longer cycles — as its cost of capital rises? HSBC has already lengthened Zhipu’s development cycle in its note. A slower flagship cadence in an open-weight market is the mechanism by which a licence that was designed for standard-setting turns back into a licence for catching up, and it would be visible in release dates before it is visible anywhere else. And: does any Chinese open-weight lab announce a licence change in the next twelve months that tightens rather than loosens? The series’ 2026-11-18 test stands, and this dataset sharpens it: a company under margin pressure with a free-substitute product has a strong incentive to convert permission into terms.

Checkable questions.

(i) What share of Zhipu’s and MiniMax’s revenue is API versus subscription versus private deployment, and how did each gross margin move in the second half? Private deployment is the line that can be priced above self-hosting cost; if it is where the growth is, the open licence is serving as a sales channel rather than as a giveaway.

(ii) Do the lock-up expiries on the January listings cluster in a window that explains the November decline? SCMP flagged lock-up tests as the volatility driver in June; a mechanical explanation and a fundamentals explanation are not exclusive, and the disclosure calendar will distinguish them.

(iii) Which of the two companies’ open-weight repositories shows continued commit activity through the drawdown? A project whose upstream goes quiet when the sponsor’s share price falls is evidence that the open artifact was a marketing instrument with a budget line; one that keeps shipping is evidence it was a strategy. This is checkable on GitHub and on ModelScope this week, and nobody has checked it.

Sources: 新浪财经 / 达摩财经Pro — 智谱5个月没了万亿市值,唐杰700亿身家蒸发 (2026-11-17; Zhipu closing down 5.26 per cent at HK$335 and HK$163.5 billion market value, intraday HK$328; the January listing at HK$116.2, the June peak of HK$2,980 and HK$1.33 trillion, and the nearly ninety per cent decline with about HK$1.1 trillion erased; the 2026-11-16 BIS interim final rule restricting third-country data-centre access as the immediate trigger; MiniMax down 7.13 per cent at HK$116 and HK$40.8 billion, ninety per cent below its March high and thirty per cent below its offer price; the 600-times and 1,000-times trailing price-to-sales peaks; Zhipu's first-half revenue of CNY 954 million up 399.74 per cent and adjusted net loss of CNY 1.964 billion widened 12.11 per cent; MiniMax's revenue up 283.1 per cent to USD 117 million with B2B displacing consumer revenue and adjusted net loss widened 52.7 per cent to USD 293 million; declining gross margins at both; HSBC's target cut from HK$1,375 to HK$500, the 7 to 27 per cent forecast reductions, the 15 to 19 point margin cut and the break-even moving from 2027 to 2029; founder stakes — Liu Debing at about 1.54 per cent and Tang Jie at 6.02 per cent; Zhipu's flagship at 20th on Artificial Analysis and MiniMax's off the list; JPMorgan's July note on open weights, differentiation and price comparison)](https://finance.sina.com.cn/cj/article/2026-11-17/doc-inizacmz2114723.shtml) · [创业邦 — 杨植麟,要敲钟了 (2026-11-17; the listed peer comparison: Zhipu's 2026-01-08 debut at HK$116.2 rising about 13 per cent on day one to a HK$58 billion market value, MiniMax's 109 per cent first-day jump to HK$345 and HK$106.7 billion, Zhipu's June peak of HK$2,980 and MiniMax’s March peak of HK$1,330 with a HK$410 billion value, the 2026-11-16 closes of HK$353.6 and HK$124.9, JPMorgan’s target cuts to HK$1,200 and HK$200, and the note that DeepSeek’s second round was reported at CNY 105.5 billion raised on a CNY 615.5 billion post-money valuation against Moonshot’s CNY 333.5 billion, the two together approaching CNY 1 trillion) · Caixin — MiniMax港股破发 市场为何重估AI模型公司? (2026-10-22; the IPO-price break and the repricing of AI model companies) · SCMP — MiniMax once led Zhipu in Hong Kong’s AI stock race. How the tables have turned (2026-06-09; the June 8 closes and the analysts’ lock-up warning) · This series’ 2026-11-17 briefing — the chip-access rule and the agentic demand mechanism · This series’ 2026-11-12 briefing — the first AI export-control registrations by DeepSeek and MiniMax


🔍 WeChat Monitor

One item was present in the manual feed for 2026-11-18 (china-daily-english-input, commit 7c0cdea, “daily: 11/18 交通银行可信开源治理能力成熟度评估”): a single link to the OSCAR可信开源 account’s announcement that Bank of Communications passed CAICT’s trusted-open-source governance maturity assessment at the advanced grade. The article was extracted in full (37,093 characters of body HTML, published 2026-11-18 09:15 Beijing time) and is analysed in the Policy & Regulation section above; the source’s factual content, not just its headline, is what that section is built on. The feed’s structured slots for the fourteen monitored accounts (OpenAtom, CCF’s open-source development committee, Huawei Open Source, Mulan, COPU, Tiangong Kaiwu, Ming Says, BAAI FlagOpen, KAIYUANSHE, Silicon Time, AtomGit, the Linux Kernel Developers Conference, GOSIM/CSDN and the Huang Daniang Tea-Swotting House) were empty for the day, and there was no second link under “other”. No article from this feed has been used in the previous three briefings; the dedup check against ~/.hermes/cron/output/bf436b4758a0/ and against this series’ 2026-11-16 / 17 / 18 entries returned no earlier use of the BoCom announcement.

🔍 Commentary — The Value Moved Off the Code, and the Three Vehicles That Caught It

First, today’s three documents are one movement seen at three altitudes, and each names a different instrument for owning what an open licence does not confer. A listed bank built a four-part governance apparatus — dual repositories, a lifecycle platform, a committee, a register — whose decisive feature is a documented exit, and whose licence work is an input filter with no output side. A privately held model company wrote a licence that converts its open weights into a negotiated agreement above a revenue threshold and a required attribution above a scale threshold, then began repapering its cap table from a Cayman holding company into a mainland entity from which it will list. A listed model company lost ninety per cent of its market value in five months while its revenue grew 400 per cent, and the analysts’ explanation was that open weights widen distribution and, where differentiation is thin, accelerate substitution. Bank, licence, register. None of the three institutions is about source code, and all three are what “open source” has become as a commercial fact in China in 2026.

Second, the unifying claim is a Coasean one and it is worth stating precisely. An open licence is a device for eliminating one transaction cost — the negotiation over copying — and it eliminates that cost completely and in perpetuity. What it does not touch is the cost of everything else in the chain: running, evaluating, migrating, warranting, indemnifying, and being the party a supervisor can call. The three documents show the Chinese market discovering this at speed and building institutions at each layer where the remaining cost lives: the enterprise builds a governance apparatus (because a prudential regulator made non-exit a solvency risk), the model company builds a licence with thresholds (because its revenue is metered access and its product is freely copyable), and the public market prices the gap between distribution and margin (because that is what a public market is for). The open-source movement’s founding insight was that the copy is free and the value is elsewhere; what 2026 has added is a Chinese-specific answer to where “elsewhere” is — in the register, in the meter and in the exit documentation.

Third, the uncomfortable symmetry, which this series must keep recording. A fortnight ago the finding was that Europe funds maintenance and does not fund adoption; China funds adoption and does not fund maintenance. Today adds the demand side: China’s most sophisticated open-source user institution — a systemically important bank — is an intake filter with no output side, and that is not a failure of the bank; it is what a permissive licence asks of a user. Under Apache-2.0 or MIT, the bank owes the commons nothing, and the bank has behaved exactly as the licence permits: it built a governance system that protects the bank. A commons cannot be built out of institutions that are each correctly optimising their own intake. The missing institution is not a stricter licence and not a new subsidy — it is an output obligation attached to something the institution already values, which is the same shape as Zhejiang’s credential clause, and the same reason that clause may be the most important sentence written in Chinese open-source policy this year even though nobody has yet built the machinery to verify it.

Fourth, the honest caveat, and it cuts against this entry’s own framing. None of these three documents says open source is failing. BoCom’s assessment is a genuine capability gain from a genuine regulator push. Moonshot’s licence is a working commercial strategy with real global adoption — 400 per cent overseas growth, 200-plus countries, seventy percent API revenue. Zhipu’s ninety per cent drawdown is a repricing from a thousand times sales to something defensible, which is what happens to every first-mover in a category that acquires competitors. The finding is narrower and it will survive the caveat: in all three cases the institution that was built is a governance institution for a company, and in none of the three is it a governance institution for a commons. The Chinese ecosystem is getting very good at the first kind. It has not started on the second, and the forty-odd policy documents, the vouchers, the benchmarks, the assessments and the credentials are all, without exception, instruments of the first kind.

Fifth, what to watch, concretely, and all three are checkable this month. CAICT’s maturity rubric — published or not — decides whether the BoCom grade is an institution other sectors can adopt or a certificate this series has already learned to discount. The K3 licence’s counterparty count in the Hong Kong prospectus decides whether the USD 20 million threshold is a screening device or a decoration. And the commit activity in Zhipu’s and MiniMax’s open-weight repositories through the drawdown decides whether their open artifacts were a strategy with a budget or a marketing line with a quarter. Two of those three are public documents, and the third is a git log.

Sources: OSCAR可信开源 via WeChat — 交通银行通过可信开源治理能力成熟度评估 (2026-11-18) · 搜狐 / 独角兽早知道 — 月之暗面架构重组与赴港IPO (2026-11-16) · 创业邦 — 杨植麟,要敲钟了 (2026-11-17) · 新浪财经 / 达摩财经Pro — 智谱5个月没了万亿市值 (2026-11-17) · 中国新闻网 — 聚焦开源发展 2026 OSCAR开源产业大会在北京召开 (2026-11-12) · This series’ 2026-11-18 briefing — the license as a tool for catching up, and the exit right that procurement never writes · This series’ 2026-11-11 briefing — the OSCAR assessment system moving from projects to actors