⚠️ 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 — 2026-11-20

1. The sequence, with dates, because the sequence is the argument. On 2026-08-26, Reuters reported exclusively that Moonshot AI was negotiating revenue-sharing agreements with Microsoft, Amazon and Google to host Kimi K3, seeking up to 30 per cent of the revenue generated from K3-related services on Azure, AWS and Google Cloud — with the three unresolved issues named: how revenue would be split, data access, and auditing token usage. The report noted Moonshot had already signed similar agreements with smaller cloud platforms, that Chinasoft International disclosed a revenue-share agreement with Moonshot in July 2026 without disclosing terms, and that Alibaba was seeking revenue-sharing agreements with major users of its own open model on comparable logic. On 2026-09-18, AWS published Introducing Kimi K3 on Amazon Bedrock: the 2.8-trillion-parameter, one-million-token, natively multimodal model available through us.moonshotai.kimi-k3 and global.moonshotai.kimi-k3 inference profiles; the first open-weight model on Bedrock to support explicit prompt caching; customer data processed inside the AWS boundary, not shared with the model provider, not used to train it, with zero data retention always on and zero operator access preventing even AWS operators from reading prompts. On 2026-11-05, Microsoft published Moonshot AI’s Kimi K3.1 Pro comes to Microsoft Foundry: an open-weight model for agentic work with text and image understanding and a one-million-token context, priced at USD 2.30 per million input, USD 0.23 cached, USD 11.50 output — the first model in the K3.1 series.

2. The licence term that makes the ask collectable. The K3 licence, revised in July 2026 on release, requires that an entity — or its affiliates — that hosts the model as a service and generates more than USD 20 million in aggregate annual revenue trigger a paid arrangement with Moonshot. AWS reported in July 2026 that its AI business had surpassed a USD 25 billion annual revenue run rate. Neither company disclosed the financial terms of the Bedrock arrangement; Amazon referred SCMP to the launch post, which does not address revenue sharing; Moonshot did not respond. The licensing analyst who tracked the placement noted the same week that neither AWS page states whether an agreement exists, and that either AWS has one or it falls under the licence’s exemption for Moonshot’s “certified inference partners” — “both are relationships with Moonshot, and neither page specifies which.” The SCMP account adds two facts that are purely institutional: Bedrock carried dozens of open-weight models from DeepSeek, MiniMax and Alibaba, but DeepSeek imposes no commercial restrictions, and the Alibaba and MiniMax models on Bedrock are older ones dating from before those companies adopted their new commercial licences.

3. The comparison set, which is where the taxonomy can be read. Alibaba’s Qwen 3.8-Max requires a separate deal for a model-as-a-service or “AI work assistant” business above USD 50 million over any consecutive twelve months; MiniMax in an April 2026 policy change required separate agreements for businesses above USD 20 million annual revenue and added a clause prohibiting its models’ outputs from being used to train competing models; DeepSeek — the one whose open-weight line made it the base of the domestic market — imposes no commercial restrictions. Three Chinese labs, three different legal instruments, one shared direction of travel, and one holdout.

4. The commercial record the licence is serving. Per Rhodium Group, seven major Chinese AI developers combined generate about 10 per cent of the revenue reported for OpenAI and Anthropic — the monetisation gap stated as a single ratio. Per SCMP, MiniMax’s business-to-business revenue surged more than 700 per cent in the first half of 2026, with overseas markets accounting for about 60 per cent. And per the Menlo Ventures 2026 State of Enterprise AI (2026-11-03, nearly 400 US enterprise decision-makers surveyed 2026-09-08 to 26, weighted by AI spend): Chinese models grew to 89 per cent of open-source spend, led by Kimi at 61 per cent, DeepSeek 15 per cent, GLM 11 per cent and Qwen 3 per cent, with Meta’s Llama the most popular Western open model at 7 per cent.

Institutional Analysis & Commentary:

This week the licence stopped being a document and became a toll booth, and the toll booth is standing on a road that no foundation built, no community governs and no standard describes.

First, state precisely what the USD 20 million threshold does, because a week ago this series read it as a reserved right and that reading now needs correcting. On 2026-11-19 this briefing argued the threshold was calibrated where almost nobody sits, making it a renegotiable event with exactly the counterparties that matter rather than an active restriction. The AWS and Microsoft placements are the evidence that the reading was directionally right and understated in one respect: the marketplace did not need to be regulated to trigger the clause — it needed only to be large enough to host the model. A hyperscaler crosses USD 20 million of AI revenue in a week. So the clause is not screening for revenue; it is screening for who is capable of serving a 2.8-trillion-parameter model at all. The answer, as every party in the reporting notes, is: almost nobody, because the weights are 1.4TB even at 4-bit and the deployment recommendation is 64 cards and up. The licence and the physics of the artefact are doing the same job — concentrating distribution in a handful of counterparties — and the licence is simply the instrument that gets a payment at that point of concentration. That is not an open-source licence with a commercial annexe. It is a distribution agreement wearing a licence’s clothes, and the reason it can be issued as a licence is that the licence is the one document every downstream party has already signed.

Second, the actual innovation is not the revenue share — it is the audit clause, and it is the first genuinely new governance problem this series has found in Chinese open source all year. Reuters named the three sticking points in the talks: revenue split, data access, and auditing token usage. Read the third one again. Under a permissive licence, nobody audits anybody, because there is nothing to count. Under a copyleft licence, someone might audit whether you distributed source, which is a binary observable in a release artefact. Under K3’s instrument, the counterparty must be able to say, to a third party’s satisfaction, how many tokens of a specific model were served, to whom, at what price — which means the licence has created a demand for a metering standard and no one has written one. The hyperscalers already meter tokens because they bill for them; what they do not publish is the allocation between models, the price actually realised, and the attributable revenue base against which a percentage is computed. A revenue-share obligation without an agreed measurement basis is not a contract; it is a relationship — and relationships between a 30-per-cent asker and a 25-billion-run-rate host are settled by bargaining power, not by clause. Note also the mirror image in the technical arrangement: AWS guarantees zero data retention and zero operator access, meaning the model provider cannot see the traffic it is being paid a share of. The licence grants Moonshot a claim on revenue it is architecturally prevented from measuring. The only audit channel left is the cloud provider’s own word.

Third, DeepSeek is the control group, and the control group is losing. The same Bedrock platform carries DeepSeek with no commercial restrictions and carries Alibaba’s and MiniMax’s newer models not at all — the ones listed are from before the licence change. So the market has produced a clean experiment: a Chinese lab that kept its terms permissive sits at 15 per cent of open-source spend; the lab that put a toll in the licence sits at 61 per cent; and the hyperscalers, which could take any of them for free, are paying for the one with the toll. That result should be genuinely uncomfortable for anyone who believes openness is rewarded as openness. It is not. The toll is not a tax on adoption, it is a signal of seriousness — a counterparty who negotiates a revenue share is a counterparty who will stay in business, answer a support escalation and sign an indemnity. What the cloud platforms bought in the K3 agreement was not permission; permission was free. They bought a counterparty. And that, precisely, is what a licence written by a company can supply and a licence written by nobody cannot.

Fourth, the structural point, which is about who owns the layer. This series has spent six weeks arguing that the Chinese open-source ecosystem builds instruments for transactions with a buyer and none for transactions with a stranger. Here is the same asymmetry at the top of the global supply chain: the layers that matter are all privately owned and none of them is a commons. The weights: a company’s. The licence: a company’s. The hosting: three companies’. The metering: whoever built the billing pipeline. The audit: unresolved. Every one of those is a normal commercial arrangement, and the sum of them is an open-weight ecosystem with no shared infrastructure of measurement, no neutral arbitration, and no place where an adopter can read what the terms of the whole stack actually are. A foundation exists to solve exactly this — to hold a mark, publish a conformance test, and arbitrate a dispute — and there is no foundation in the picture. The reason is legible: a foundation cannot ask for 30 per cent of revenue, so it cannot buy its way into the room where the hyperscalers are. Moonshot got into that room because it had something to sell and a licence that could price it.

Fifth, the falsifiable tests. Does the K3 licence’s USD 20 million threshold ever produce a disclosed, quantified agreement — in the Hong Kong prospectus this series flagged on 2026-11-19, or in any cloud provider’s filings? If the revenue-share exists and is material, a prospectus must describe it; if no percentage appears in any filing, the toll is a bargaining chip rather than a revenue line. Does any Chinese lab publish a metering or attribution specification alongside its next licence — a defined revenue base, a defined token-counting method, a defined reporting period? The first lab to publish one has invented the institution the whole arrangement is missing, and it will be worth more than the percentage. And: when Microsoft and Google finish their negotiations, does either agree to the 30 per cent, or does the number that emerges become the industry’s de facto rate? A rate set between three hyperscalers and one Chinese lab becomes the price of open-weight distribution worldwide, and it will be set without a single community, foundation or standards body in the room.

Checkable questions.

(i) Is the USD 20 million measured on revenue attributable to the model, or to the host’s entire AI business? SCMP’s own framing — AWS’s total AI run rate cited against a per-model threshold — suggests how loose the comparison can get. The licence text is where the difference lives, and the two readings differ by four orders of magnitude for a hyperscaler.

(ii) Why are only pre-licence Alibaba and MiniMax models on Bedrock? If the newer ones are absent because the commercial terms are not yet agreed, then the licence has already excluded two Chinese labs from the largest Western distribution channel while admitting a third — a selection effect with real consequences for which Chinese model becomes the global default.

(iii) Do Kimi’s “certified inference partners” have published criteria? The exemption is the licence’s own escape hatch, and an unpublished certification programme inside a published licence is the least accountable arrangement in the stack.

Sources: Reuters — EXCLUSIVE: China’s Moonshot in talks with Microsoft, Amazon, Google over K3 revenue sharing, sources say (2026-08-26; up to 30 per cent of K3-related services revenue sought on Azure, AWS and Google Cloud; the three unresolved issues of revenue split, data access and token-usage auditing; similar agreements with smaller cloud platforms and Chinasoft International’s July 2026 disclosure; Alibaba seeking comparable agreements; the note that few customers can self-host 2.8 trillion parameters) · AWS Machine Learning Blog — Introducing Kimi K3 on Amazon Bedrock (2026-09-18; the 2.8-trillion-parameter model with a one-million-token context and native vision, approximately 2.5x scaling efficiency over K2; the first open-weight Bedrock model with explicit prompt caching; the us.moonshotai.kimi-k3 and global.moonshotai.kimi-k3 profiles; data processed inside the AWS boundary and not shared with the provider; zero data retention and zero operator access; the Bedrock open-weight catalogue spanning DeepSeek, Google, MiniMax, Mistral, Moonshot, NVIDIA, OpenAI and Qwen) · Microsoft Foundry Blog — Moonshot AI’s Kimi K3.1 Pro comes to Microsoft Foundry (2026-11-05; the first model in the K3.1 series; open-weight, agentic, text-and-image with a one-million-token context; pricing of USD 2.30 input, USD 0.23 cached and USD 11.50 output per million tokens; the Foundry open-weight catalogue including NVIDIA, OpenAI, DeepSeek, MiniMax, Mistral, Moonshot and Qwen) · South China Morning Post — Moonshot’s Kimi K3 lands on Amazon in key test for Chinese open-source AI revenue (2026-09-21; Bedrock availability from 2026-09-18; the July 2026 licence revision and the USD 20 million MaaS threshold; AWS’s USD 25 billion AI run rate reported in July; the absence of disclosed financial terms and Amazon’s referral to a launch post that does not address revenue sharing; Qwen 3.8-Max’s USD 50 million threshold for MaaS and “AI work assistant” businesses; MiniMax’s April 2026 change with its USD 20 million threshold and its ban on using outputs to train competing models; DeepSeek’s absence of commercial restrictions; the note that the Alibaba and MiniMax models on Bedrock predate their new licences; Rhodium Group’s finding that seven major Chinese AI developers generate about 10 per cent of OpenAI’s and Anthropic’s revenue; MiniMax’s 700 per cent B2B revenue growth with overseas at about 60 per cent) · The AI Realist — At Least 45 Days (2026-09; the observation that neither AWS page states whether a K3 agreement exists, and that the licence’s exemption for “certified inference partners” is the alternative reading; the K3 licence’s section 2 text on the separate-agreement requirement; the model card’s 45-day discontinuation notice) · The Next Web — Moonshot AI wants 30% of what US clouds earn from Kimi K3 (2026-08-26; the inversion of the usual platform-takes-a-cut arrangement; the note that open weights change who the cloud provider competes with rather than removing the need for one)


🏗️ Institutional Change — The Layer That Has No Owner: Shanghai Pays for Distribution and Calls It an Ecosystem

1. The document, and the timing. On 2026-11-19 at 12:03 Beijing time, the OSC开源模力 WeChat account, republishing 经济参考网 (Economic Information Daily) reporter Li Zhiyong’s (李志勇) piece, ran “Shanghai advances open-source platform construction, building the city that best understands developers” (上海推进开源平台建设 打造"最懂开发者的城市"). The same article ran on 2026-11-10 on PEdaily (投资界) and on 2026-11-12 through Xinhua’s app. It arrived in the manual feed for 2026-11-19. This series touched its policy spine on 2026-11-16 — the December 2025 Shanghai Implementation Plan for Strengthening the Open-Source System (《上海市加强开源体系建设实施方案》), the CNY 20 million and CNY 5 million ceilings, the contribution-oriented talent-evaluation pilot. What the 2026-11-19 version supplies, and what makes it a different entry rather than a repeat, is the organisational fact and the theory of the layer that goes with it.

2. The targets, restated because they are the county-level plan for the whole argument. By 2027: one to two open-source communities with international influence; 100 open-source commercialisation enterprises; more than 200 high-quality open-source projects; more than 3 million open-source developers gathered; up to CNY 20 million in subsidy for foundational development tooling; up to CNY 5 million in reward for a high-quality open-source project; and a pilot talent-evaluation mechanism oriented to open-source contribution.

3. The organisational fact. On 2026-10-30, OSChina (开源中国) — described as the country’s leading open-source platform company — moved its headquarters to Zhangjiang, Pudong, and on the same day the Model Ark Base (模力方舟基地) opened, offering community operations, platform resource matching, scenario validation and incubation acceleration to AI-era developers and teams. OSChina committed to three things: landing eighteen years of developer-operations experience in Zhangjiang; opening Model Ark’s model interfaces and compute resources to Shanghai enterprises and developers; and running continuous AI tooling training so that developers in traditional industries put AI into production processes. CEO Xu Yong (徐勇) gave the reason in a sentence worth keeping: Shanghai has the two scarce conditions AI open-source infrastructure needs — sufficiently dense real industrial scenarios, and sufficiently systematic institutional supply — and Pudong is the country’s first AI innovation-application pilot zone, so scenario density makes validation local. He explicitly rejected the simpler explanation: “This cannot simply be attributed to the appeal of investment-attraction policy, but to a city’s ability to twist industrial thickness, scenario density and institutional supply into one rope.”

4. What Model Ark actually is. Its stated positioning is “the key link connecting Chinese models, Chinese compute and end users”: it has adapted domestic heterogeneous compute — Ascend (昇腾), Muxi (沐曦), Tianshu Zhixin (天数智芯) — and aggregates the calling interfaces of more than a hundred mainstream open-source models. Its economics are described in one sentence: if every enterprise had to build its own agent infrastructure adaptation from zero, agents would stay in case studies and never become an industry; the value of an open-source platform is to turn this “heavy” work into a public good (公共品). Alongside it, the 2026 Shanghai Open-Source Software Application Innovation Competition, launched for registration on 2026-08-05, received 766 project registrations, with 9 organisations posting 12 real industrial problems, 269 projects addressing them, 10 teams per track reaching the final across three tracks aligned to integrated circuits, high-end equipment, intelligent vehicles and intelligent computing — Shanghai teams taking one third. The article’s own framing of the policy turn is precise and worth carrying into the analysis: Shanghai is moving “from giving subsidies and building carriers to reducing friction, closing the loop, and covering micro-level subjects” (从"给补贴、建载体"向"降摩擦、通闭环、覆盖微观主体"的深化转向), aimed at SMEs and one-person companies (一人公司).

Institutional Analysis & Commentary:

A city has now decided that what its open-source ecosystem is missing is a distribution layer, and it has paid to import one. That is a more honest diagnosis than any voucher programme, and it is also a category error worth naming precisely.

First, take the diagnosis seriously, because it is the sharpest thing anyone in Chinese open-source policy has said this year. Xu Yong’s sentence — dense real scenarios plus systematic institutional supply — is an accurate description of a genuine market failure, and it is not the failure the policy stack usually targets. The usual instruments buy artefacts: models, projects, benchmarks, installations, courses. What Shanghai is now buying is an interface: the adapted compute, the aggregated model endpoints, the community operations, the validation slot. That is a real gap. Between a model released on ModelScope and an SME that wants an agent to process its documents there is a chain of unglamorous work — driver adaptation, quantisation, serving configuration, cost accounting per endpoint — which no vendor is paid to do and every adopter must redo. Calling it “public good” is imprecise but the instinct is right: it has the structure of infrastructure, because the cost of doing it once for everyone is far below the cost of everyone doing it once. In Williamson’s terms this is the make-or-buy calculation at the level of an industry, and the answer arrived at by subsidising a private firm to provide it is what every country does with infrastructure it cannot persuade the market to fund.

Second, and here is where the word “public good” fails, in a way that has consequences eighteen months out. A public good is non-excludable and non-rival: once provided, nobody can be kept from it and one person’s use does not diminish another’s. Model Ark is rival (it has finite compute) and excludable (its operator controls who gets an endpoint). It is not a public good; it is a regulated private platform, and the only thing making it public-shaped is the subsidy and the policy expectation. This matters because of what happens at the end of the subsidy. A commons, once built, is held by a governance body with a charter that says who may join, how decisions are made and what happens when the founder leaves. A subsidised platform has a shareholder, a board and a burn rate. OSChina is eighteen years old and has just taken a city’s money to be the layer a city’s AI industry depends on — and there is no published instrument anywhere in this announcement that says what happens to Model Ark’s terms, its access policy or its pricing if the company is acquired, changes strategy or fails. That is not a hypothetical in this sector; it is the base rate. Shanghai has bought a dependency, not a commons, and the difference is only visible on the day it matters.

Third, and this is the finding this entry exists for: compare the two documents in today’s briefing and the same layer is being claimed twice, by two different property regimes. Moonshot is taxing the distribution layer with a licence — up to 30 per cent of the revenue that hyperscalers earn from serving its weights. Shanghai is subsidising the distribution layer with a procurement cheque — up to CNY 20 million for the tooling underneath, up to CNY 5 million per project on top. Both are correct responses to the same fact: the weights are free and the layer above them is where all the money is. Neither is building a governance institution for that layer. The licence is a private toll; the subsidy is a public grant; and a layer that is simultaneously tolled and subsidised, by a company and a city, with no charter and no published access terms, is the least legible piece of infrastructure in either ecosystem. The open-source movement’s own answer to this problem was the foundation: a neutral entity that holds the mark, publishes the conformance test and arbitrates access. China has foundations — OpenAtom has graduated openEuler, incubated it, and disburses real money — and none of them is in this picture. The distribution layer went to a company because a company could move in a month and a foundation could not.

Fourth, the competition numbers are the tell, and they point the opposite way from the press release. 766 registrations, 269 projects answering 12 problems, 30 finalists, one third from Shanghai. Read the ratio: 269 projects answering 12 questions. That is twenty-two teams per problem — the shape of a grant-writing tournament, not a procurement market. The article’s own description of the mechanism is candid about it: “discovering projects through competitions, matching resources through conferences, promoting growth through services” (以赛发现项目、以会对接资源、以服务促进成长). A competition can find a project; it cannot create a customer. The thing that would turn 269 answers into an industry is a purchasing officer with a budget and a problem, and the policy document’s own logic — reduce friction, close the loop, cover micro-subjects — stops one signature short of that. Note also what the three tracks are aligned to: integrated circuits, high-end equipment, intelligent vehicles, intelligent computing — which is to say, Shanghai’s existing industrial base, not a new market. The competition is a matching service for established sectors, which is a legitimate industrial policy objective and a different one from ecosystem-building.

Fifth, and it should be said plainly because it is the strongest part of the case. Shanghai’s plan is better instrumented than any other Chinese open-source policy this series has examined, and that is not faint praise. It has named ceilings, named targets, a pilot mechanism for individual contribution, a competition with registration data, and a platform with a published technical scope. Compare it to Zhejiang’s seven-department opinion (2026-11-18), which has a credential clause and no verifier, or to BDA’s subsidy schedule (2026-11-16), which prices donation at CNY 3 million with no acceptance test. Shanghai is at least measuring what it funds. The finding is not that the plan is hollow; it is that its most consequential act — importing a distribution layer — was done as an investment attraction, and an investment attraction is not an institution. A city can subsidise a headquarters into existence in a quarter. It cannot retroactively install the governance that would make that headquarters accountable to the ecosystem it now sits on top of.

Sixth, the falsifiable tests. Does Model Ark publish its access terms, its pricing and its service-level commitments — the three things a platform owes the layer it serves? If yes, a private platform is behaving as infrastructure and the “public good” language is earned. Does the 2027 “3 million developers gathered” figure have a definition — registered accounts, annual actives, or contributors to any project? The first two are counts a marketing department can produce; the third is the only one that would matter. And: what happens to Model Ark’s terms if OSChina is acquired or changes hands? No document in this announcement addresses it, and the answer determines whether Shanghai bought a commons or a dependency.

Checkable questions.

(i) Who sits on Model Ark Base’s governance — OSChina alone, or with the district and the enterprises it serves? A platform that a city’s AI industry will depend on should have a body where the dependants sit. None is described.

(ii) Of the 269 projects solving the 12 industrial problems, how many were subsequently procured by the entity that posed the problem? This is the only number that distinguishes a competition from a market, and the organisers are the only people who can produce it.

(iii) Does the “contribution-oriented talent-evaluation pilot” define contribution, and does it require a merged patch reviewed by someone outside the candidate’s employer? The same test this series applied to Zhejiang, now applied to a city that has evidently thought harder about implementation.

Sources: OSC开源模力 via WeChat — 上海推进开源平台建设 打造"最懂开发者的城市" (2026-11-19 12:03; republishing 经济参考网 and reporter Li Zhiyong; the December 2025 Shanghai Implementation Plan for Strengthening the Open-Source System with 2027 targets of one to two globally influential communities, 100 commercialisation enterprises, 200-plus quality projects and over 3 million developers, up to CNY 20 million for foundational tooling and up to CNY 5 million per quality project, and the contribution-oriented talent-evaluation pilot; OSChina’s headquarters moving to Zhangjiang on 2026-10-30 and the Model Ark Base opening the same day; Xu Yong’s “two scarce conditions” and his rejection of the investment-attraction explanation; the three commitments on developer operations, model and compute access, and AI tooling training; Model Ark’s adaptation of Ascend, Muxi and Tianshu Zhixin heterogeneous compute and aggregation of 100-plus open models; the “turn heavy work into a public good” formulation and the “make the base light so developers can go deep” line; the 2026 Shanghai Open-Source Software Application Innovation Competition’s 766 registrations, 12 problems from 9 organisations, 269 responding projects and 30 finalists with Shanghai at one third; the “from subsidies and carriers to reducing friction, closing the loop and covering micro-subjects” turn and the one-person-company objective) · 投资界 / PEdaily — 上海以开源平台为抓手,打造"最懂开发者的城市" (2026-11-10; the same report, with the same 2027 targets and ceilings, OSChina’s 2026-10-30 Zhangjiang headquarters and the Model Ark Base, Xu Yong’s remarks, the competition figures and the policy-turn framing) · This series’ 2026-11-16 briefing — BDA’s open-source subsidy schedule and the price on donation · This series’ 2026-11-18 briefing — Zhejiang’s seven-department opinion and the credential clause it cannot fund


📊 The Ledger — Twenty-Two Times the Volume and Forty-Five Times Less the Money: What Menlo’s Numbers Say About the Open-Weight Trade

1. The report. On 2026-11-03, Menlo Ventures published its fourth annual 2026: The State of Enterprise AI, built on a Morning Consult survey of nearly 400 US enterprise decision-makers responsible for more than USD 4 billion in AI spending, fielded 2026-09-08 to 26, with market sizing from a bottom-up model informed by survey responses, public filings and third-party reports, and token economics triangulated against OpenRouter, Ramp and Vercel. It is the most rigorous public measurement of what enterprises actually buy, and it lands directly on the question this series has followed all year: what an open licence is worth in a market.

2. The headline numbers. Enterprise AI deployment spend reached USD 177 billion, up 5.4 times from USD 33 billion a year earlier. For every USD 1 spent building models (USD 102.5 billion), enterprises spent USD 1.70 deploying them; infrastructure was USD 99.5 billion, applications USD 77.3 billion. Agents became the majority consumer of models: on OpenRouter, agent traffic overtook human traffic in February 2026 and has since grown 37 times, and agents now account for 80 per cent of all tokens — roughly seven times human usage. Frontier labs took 69 per cent of deployment dollars, up from 50 per cent; Anthropic’s share of enterprise API spend rose from 40 to 55 per cent, OpenAI’s from 27 to 31, Google’s fell from 21 to 9.

3. The four numbers about open source, which are the entry. (a) Volume: open-source usage grew 22.5 times to an estimated 2.9 quadrillion tokens per month, outpacing closed-source growth of 12.1 times; much of it runs on inference clouds like Fireworks and Modal, both of which passed USD 1 billion in revenue run rate this year. (b) Price: the average open-source token fell 3.3 times to USD 0.13 per million, while the blended closed-model price fell 24 per cent to USD 0.70. (c) Share: open source holds 19 per cent of tokens and 4 per cent of dollars, its spending share drifting down from 6 per cent as its token share nearly doubled — the report’s own summary is “winning tokens but not dollars.” (d) Composition: Chinese models took 89 per cent of open-source spend — Kimi 61 per cent, DeepSeek 15, GLM 11, Qwen 3 — with Meta’s Llama at 7 per cent the leading Western open model. Separately: near-frontier models including Kimi K3, DeepSeek V4 and GLM-5.3 hold just under 5 per cent of total LLM API spend despite approaching frontier performance at a fraction of the price; post-training adoption rose from 6 to 10 per cent of AI builders, and 91 per cent of those use outside partners — 50 per cent infrastructure providers like Fireworks and Modal, 41 per cent expert-data and evaluation platforms.

4. The arithmetic, done rather than asserted. Open source at 19 per cent of tokens and 4 per cent of dollars implies an average realised price roughly 5.8 times below the blended market. Against closed models specifically, USD 0.13 against USD 0.70 is 5.4 times cheaper per token. And the volume-to-spend ratio within open source, year on year: tokens up 22.5 times, spend share down from 6 to 4 per cent — so spend grew at roughly 3.6 times while volume grew 22.5 times, a ratio of about six to one. Six times more tokens for every one time more money.

Institutional Analysis & Commentary:

Menlo’s report is the first measurement in this series that lets the licence question be answered with arithmetic rather than inference, and the arithmetic is unkind to the licence on the dimension it was written to address.

First, the finding stated as a single sentence: the Chinese open-weight strategy has decisively won distribution and is decisively losing monetisation, and both halves are now measured. 89 per cent of enterprise open-source spend and 61 per cent for a single vendor is not a niche position — it is a near-monopoly on the open tier, which is what “occupying an ecological niche,” Ye Qiwei’s phrase from 2026-11-18, looks like when it succeeds. And 4 per cent of dollars for 19 per cent of tokens is what the same success costs. Read together: the strategy worked exactly as designed and the design has a hole in it. The hole is not the licence and not the model; it is that the price of the open tier collapsed by 3.3 times in one year while the closed tier’s blended price fell only 24 per cent — meaning the price gap between open and closed widened rather than narrowed as open models got better. That is the opposite of what the standard story predicts. The usual account is that open models commoditise the bottom and pressure closed prices down; what the data shows is that open models set the floor and the frontier raised the ceiling, with frontier models’ share of spend going from 35 to 63 per cent in the same year. Openness won the workloads nobody wanted to pay much for, and capability won the workloads that pay.

Second, the most important single line in the report is about agents, and it explains why the collapse happened. Cache reads now make up 95 per cent or more of an agent’s tokens, billed at USD 0.20 per million for Opus 5.5 against USD 4 for fresh input and USD 20 for output; at a mix of 3 per cent fresh input and 2 per cent output, an agent session costs about USD 0.70 per million tokens. Agents are a machine for converting expensive tokens into cheap ones. Which means the open-weight tier’s economics are being destroyed from two directions at once: open models compete on price with each other (89 per cent of the open market split among four Chinese vendors, all of whom can undercut each other at marginal cost), and agentic usage shapes the workload so that the tokens being consumed are the ones that cost almost nothing to produce. The licence’s USD 20 million threshold prices a share of revenue from a market where revenue per token is falling 3.3 times a year. A 30 per cent share of a number dropping that fast is a claim on a shrinking base — which is precisely why the claim had to be moved upstream, from the model to the platform that hosts it.

Third, and this is the institutional finding: the toll is being levied at the only point in the stack that is not commoditising. Look at where the money actually is. Applications: USD 77.3 billion, of which coding USD 34 billion and chatbots USD 25 billion — 76 per cent in two categories, both chosen because they have “environments in which agents can act, complete context they can learn from, and results they can verify.” Infrastructure: USD 99.5 billion, inference and agent infrastructure. The open model itself is nowhere on that list as a revenue category. So the layer that prices the model is either the hosting (AWS, Microsoft, Google — all three negotiating with Moonshot) or the application that embeds it (Cursor’s Composer 2 built on Kimi K2.5; Harvey moving its workflow to K3). Moonshot’s licence reached for the hosting layer because that is where the money is present and the counterparties are named. Shanghai’s subsidy reached for the adaptation layer for the same reason. Neither reached for the model, because there is nothing there to reach for. This is the commodity’s fate, and the Chinese open-weight labs have walked into it with their eyes open — Moonshot’s own licence is the proof that management knows.

Fourth, the uncomfortable corollary for the whole Chinese open-source programme, stated as carefully as it can be. Zhejiang funds adoption with token, compute and data vouchers (2026-11-18); BDA pays up to CNY 3 million for a donated project (2026-11-16); Shanghai pays up to CNY 20 million for tooling and CNY 5 million per project; openEuler counts 22 million installations (2026-11-18). Every one of those instruments is denominated in a quantity whose unit price is collapsing. That is not a reason to stop; adoption at a falling price is exactly what a public policy should want for its users. But it does mean the cost per unit of every target in every one of those documents is falling at roughly the same rate the market price is — so an installed base achieved cheaply is cheap. The programme’s own success metric is deflating as it succeeds, and the only instruments in the whole stack that do not deflate are the ones counting contribution, review and maintenance — the outbound half, for which, as this series has documented for six weeks, no province has yet built a verifier.

Fifth, the honest caveat. Menlo’s 89 per cent is a share of a category that is 4 per cent of the market. It is a dominant position in a small pool, and the report says so in its own framing — “spend remains small, but interest suggests significant further growth.” Near-frontier models at under 5 per cent of total API spend is the more sober figure, and it is the one that should govern any claim that Chinese open weights have taken the enterprise market. They have taken the open enterprise market, which is the market that exists because they made it. That is a real achievement and a narrow one, and conflating the two would be the kind of category error this series exists to catch.

Sixth, the falsifiable tests. Does the open-source spend share recover from 4 per cent in the next edition? If it does not, the toll-booth strategy is correct and the commons strategy is not, and that conclusion will be drawn by every Chinese lab’s board. Does Kimi’s 61 per cent hold once Microsoft’s and Google’s agreements are signed and Qwen’s newer models reach Bedrock? A drop would show the share was a first-mover artefact; a hold would show it was capability. And: does any Chinese lab publish per-token pricing that rises — the only observable proof that a vendor has escaped commodity competition. A model vendor whose price is rising has differentiation; every other vendor is in a price war it cannot win and did not start.

Checkable questions.

(i) What is the realised price per million tokens for Chinese open models on the three hyperscalers specifically? Menlo’s USD 0.13 is a blended open-source average; if the Chinese average is below it, the 89-per-cent share is even more a volume story than it looks.

(ii) Does the 61 per cent Kimi share survive self-hosting and domestic Chinese platforms, or is it a Western-enterprise measurement only? The report’s data sources — OpenRouter, Ramp, Vercel — skew individual developers and startups. The Chinese domestic market, which runs mainly on ModelScope and domestic clouds, is not in it.

(iii) Of the 10 per cent of builders who post-train, what proportion post-train on a Chinese open base? That number would measure something the licence cannot tax and the market has not priced: whether Chinese weights are becoming the substrate of other people’s products, which is a different and stickier kind of adoption than API calls.

Sources: Menlo Ventures — 2026: The State of Enterprise AI (2026-11-03; the Morning Consult survey of nearly 400 US enterprise decision-makers responsible for more than USD 4 billion in AI spend, fielded 2026-09-08 to 26; deployment spend of USD 177 billion up 5.4x from USD 33 billion; USD 1.70 of deployment revenue per USD 1 of model development; infrastructure at USD 99.5 billion and applications at USD 77.3 billion with coding at USD 34.0 billion and chatbots at USD 25.0 billion making up 76 per cent; agents at 80 per cent of OpenRouter tokens and 37x growth since February 2026; frontier labs at 69 per cent of deployment dollars with Anthropic at 55 per cent and OpenAI at 31 per cent of enterprise API spend and Google at 9; frontier models’ share of tokens rising 18 to 42 per cent and of spend 35 to 63 per cent; open-source volume up 22.5x to 2.9 quadrillion tokens per month with Fireworks and Modal passing USD 1 billion run rate; the average open-source token at USD 0.13 down 3.3x against a blended closed price of USD 0.70 down 24 per cent; open source at 19 per cent of tokens and 4 per cent of spend, the spending share having drifted down from 6 per cent; Chinese models at 89 per cent of open-source spend with Kimi 61, DeepSeek 15, GLM 11 and Meta’s Llama 7; near-frontier models including Kimi K3, DeepSeek V4 and GLM-5.3 at just under 5 per cent of total LLM API spend; cache reads at 95 per cent or more of agent tokens with Opus 5.5’s USD 0.20 cached rate against USD 4 input and USD 20 output and an agent session at about USD 0.70 per million; post-training adoption from 6 to 10 per cent with 91 per cent of post-trainers using outside partners, 50 per cent infrastructure and 41 per cent data and evaluation platforms; Cursor’s Composer 2 built on Kimi K2.5 with Fireworks and Harvey’s partnership on Applied Compute)


🔍 WeChat Monitor

One item was present in the manual feed for 2026-11-19 (china-daily-english-input, raw file raw/2026-11-19.md, repository pulled and up to date at the time of this run): a single link under “other” to the OSC开源模力 account’s 2026-11-19 12:03 republication of 经济参考网’s article on Shanghai’s open-source platform policy. The article was extracted in full (1,945 characters of body text; account OSC开源模力, gh_ee7c29935c00, msg_publish_time_str 2026/11/19 12:03) and the source’s factual content — not just its headline — is what the Institutional Change section above is built on. The feed’s structured slots for the fifteen 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, the Huang Daniang Tea-Swotting House and Yetian’s Book) were all empty for the day. Dedup: the Shanghai Implementation Plan and its ceilings were referenced in the 2026-11-16 entry, but OSChina’s 2026-10-30 Zhangjiang headquarters relocation, the Model Ark Base, Xu Yong’s remarks, the compute-adaptation scope and the 766-registration competition data have not appeared in this series before; grep against the 2026-11-16 / 17 / 18 / 19 entries and against ~/.hermes/cron/output/bf436b4758a0/ returned no earlier use of the platform-layer material. The English-language items in today’s Legal & Licensing and Ledger sections are new to this series: no prior entry mentions Menlo Ventures, Amazon Bedrock, or Moonshot’s revenue-share negotiations.

🔍 Commentary — The Same Layer, Claimed Twice: A Toll in Seattle and a Cheque in Shanghai

First, today’s three documents are one document read from three places where the money is. Seattle and Redmond said yes to a Chinese open-weight model and are paying a share of the revenue to host it. Menlo Park measured the whole market and found that open source holds 19 per cent of tokens and 4 per cent of dollars, with 89 per cent of the open dollars going to Chinese models and 61 to one vendor. Shanghai paid up to CNY 20 million to import a company that adapts domestic compute and aggregates a hundred open models into interfaces an SME can use. The weights are free in all three stories. The layer above the weights is where every dollar in all three stories is earned, spent or fought over.

Second, the unifying claim, which is a Coasean one and is now measurable rather than merely arguable. An open licence eliminates exactly one transaction cost — the negotiation over copying — and eliminates it completely and forever. It does nothing about the costs that remain: adapting, hosting, metering, supporting, warranting, and being the party a customer can call. For three years the Chinese open-weight strategy assumed those residual costs would be absorbed by the cloud platforms and that the model vendor would earn somewhere else. Menlo’s numbers show where “somewhere else” turned out to be: nowhere in the model. USD 0.13 per million tokens, falling 3.3 times a year, in a market where agents are machines for consuming the cheapest token available. The licence threshold, the Zhangjiang headquarters and the Shanghai subsidy are three different answers to the same discovery, and each of them is an attempt to own a piece of the residual cost rather than the artefact. None of them is a commons.

Third, what is genuinely new this week is that the licence has become a contract with a counterparty — and this series has been predicting the opposite for a month. On 2026-11-18 the test was whether a Chinese lab would publish “a licence that carries a service or governance term rather than pure permission.” It did, and the term was not a restriction but a price. That is a better outcome than the prediction and a more interesting one. A restriction says you may not; a price says you may, for a share. Only the second can build anything durable, because only the second gives the licensor a reason to keep the model good. The finding to carry forward is that the Chinese open-weight industry found its business model, and it is not open source — it is a distribution franchise administered through an open licence. That is legitimate. It is also exactly what Ye Qiwei meant, one month ago, when he said that once a follower catches up, “open source quietly changes from a tool for catching up into a tool for setting standards.” A standard collects a toll. That is what standards are for.

Fourth, and the uncomfortable symmetry this series must keep recording. Europe funds maintenance and not adoption; China funds adoption and not maintenance. Today adds a third line to that ledger, and it cuts against the European self-image: the United States neither funds maintenance nor funds adoption — it funds the application layer, and it is now paying a Chinese model vendor a percentage of the hosting because the hosting is where the revenue is. The open-source commons that all three systems depend on — the kernel, the compilers, the TLS stack — appears in none of the three budgets, and appears in today’s Menlo report only as an unremarked input to the USD 99.5 billion infrastructure line. Every system has found a way to be commercially serious about the layer above the commons and none has found a way to be institutionally serious about the commons itself.

Fifth, the honest caveat, and it is a real one. Shanghai’s plan is the most instrumented Chinese open-source policy this series has examined, with named ceilings, named targets, a competition with registration data and a platform with a published technical scope — and Xu Yong’s diagnosis of the missing distribution layer is correct in a way that the voucher programmes elsewhere are not. Moonshot’s licence is a working commercial instrument with real counterparties, real global adoption and a real negotiating position, and calling it a franchise is not a criticism. Menlo’s 89 per cent is a share of 4 per cent, and the report says so itself. Nothing in today’s three documents is a failure. The finding is narrower: each is an institution for a company or a city, and none is an institution for a commons — and the six weeks of documents that preceded them, the forty-odd policies, the vouchers, the benchmarks, the assessments, the credentials and the tolls, are all, without exception, instruments of the first kind.

Sixth, what to watch, concretely, and all three are checkable within a quarter. A metering specification — the licence created a demand for a way to count tokens attributable to a model, and the first vendor to publish one will have invented the institution the arrangement is missing. Shanghai’s voucher and competition data at disbursement level — the only Chinese policy in this series with the instrumentation to produce it. And the next Menlo edition’s open-source spend share — 4 per cent or 6, which will tell every Chinese lab board whether the toll-booth strategy or the commons strategy is the one that pays. Two of those three will be published by someone else. The first one is a document no one has written, and writing it would be worth more than the 30 per cent.

Sources: Reuters — Moonshot in talks with Microsoft, Amazon, Google over K3 revenue sharing (2026-08-26) · Microsoft Foundry Blog — Kimi K3.1 Pro comes to Microsoft Foundry (2026-11-05) · AWS Machine Learning Blog — Introducing Kimi K3 on Amazon Bedrock (2026-09-18) · South China Morning Post — Moonshot’s Kimi K3 lands on Amazon (2026-09-21) · Menlo Ventures — 2026: The State of Enterprise AI (2026-11-03) · OSC开源模力 via WeChat — 上海推进开源平台建设 打造"最懂开发者的城市" (2026-11-19) · 投资界 / PEdaily — 上海以开源平台为抓手 (2026-11-10) · This series’ 2026-11-19 briefing — the refusal right, the K3 licence and the repricing of open weights · This series’ 2026-11-18 briefing — code as a line item, Zhejiang’s credential clause and 22 million installations