Anthropic's framework provides a deconstructible sample, but the China version requires a complete rewrite across three dimensions—capital structure, compliance signals, and data sovereignty—and once rewritten, it may actually form a deeper moat than the US version.
Anthropic's three-layer structure (JV channel + pre-built agents + compliance endorsement) is transferable, but all three elements require local substitutes in China
The "Dimon equivalent" for China's compliance signal is not an entrepreneur—it's the president of ICBC/CCB or an actual deployment case from a regulatory body; harder to obtain, but once secured, the moat is deeper
Data localization regulations = foreign LLMs cannot process China's core financial/medical/government data = de facto monopoly for domestic LLMs in these verticals (a severely underestimated advantage)
China has an equivalent "resident engineer" culture, but it hasn't been systematically productized—this is where Anthropic's FDE model is truly worth learning from
The most dangerous trap: equating the "domestic substitution" narrative with a competitive barrier—policy dividends are a window, not a moat; the window is approximately 3–5 years
The real competitors for China's LLM enterprises are not each other, nor OpenAI/Claude—they are software companies like Wind, Jinzheng Technology, and Yonyou Network that have already built industry data pipelines. Whoever captures them first seizes the entry point to China's Bloomberg Terminal equivalent.
In scenarios like government/financial regulatory reporting, foreign competitors are effectively absent—this matters more than any parameter advantage.
Pick a vertical with "highly standardized formats + data localization requirements + heavy regulatory pressure," and complete the three-layer connection of industry JV + pre-built agents + industry endorsement—the first to do so becomes the Anthropic of the China market.
What Anthropic accomplished in 72 hours was not three independent things—it was a tightly interdependent three-layer system. Understanding this system is a prerequisite for transplanting it to the China market.
Key Insight: Strong dependencies exist between the three layers—without Layer 3 (compliance credit), enterprises won't procure; without Layer 2 (concrete products), the JV has nothing to sell; without Layer 1 (distribution pipeline), products can't reach decision-makers. The problem for China's LLM enterprises is often that all three layers are missing simultaneously, or they exist in isolation without being connected into a system.
CTO / CFO joint approval, decision cycle 2–6 months, technical evaluation carries high weight
Procurement committee + IT department + compliance department + (for SOEs) Party committee, decision cycle 6–18 months, qualifications, classified protection compliance, and domestic certification carry equally high weight
→ No matter how good the product is, missing Level 3 classified protection + domestic certification + SOE precedent can lose the deal
Dimon's endorsement = JPMorgan's legal team reviewed it = industry ticket of admission. A private-sector CEO's endorsement is sufficient
Requires use cases from institutions with regulatory lineage. Endorsement from ICBC/CCB president = China's Dimon equivalent; harder to obtain, but once secured, the lock-in effect is stronger (SOEs typically don't procure from competitors simultaneously)
→ Compliance signals are harder to obtain, but the resulting moat depth far exceeds the US version
Financial data can go on AWS/Azure private clouds; cross-border data flows are relatively free. Anthropic faces global competition from OpenAI, Gemini, etc.
Data Security Law + Measures for Security Assessment of Cross-border Data Transfer: core financial/medical/government data effectively cannot leave the country. Foreign LLMs cannot process this data
→ Absence of foreign models = de facto monopoly for domestic LLMs in core financial/government/medical scenarios. This is the competitive landscape Anthropic envies most: geographic borders are the moat.
US: Independent PE networks (Blackstone holds 2,700+ enterprises) + Accenture/Deloitte systems integration
China: No equivalent independent PE enterprise network. The SI backbone is Huawei / Yonyou / Kingdee / Glodon; the pipelines are Alibaba Cloud/Tencent Cloud/Huawei Cloud/Tianyi Cloud
→ China JV equivalent = Huawei Cloud + a central SOE fund / Yonyou + ICBC Strategic Investment Dept
US: Anthropic vs OpenAI vs Gemini, globalized competition, sensitive verticals also open
China: In core verticals (government/financial regulation/medical), foreign players are effectively absent; only domestic vendors compete
→ Competition is more focused, the pie more concentrated, and homogenization risk is also higher
SaaS penetration ~40%, mixed SME + large enterprise market, cloud-native is the default deployment. Anthropic's agents are delivered via API / cloud subscription with no on-prem needed; enterprises can self-serve
SaaS penetration ~10%–15%, procurement budgets highly concentrated in central SOEs, state-owned enterprises, and top private companies. Large enterprises universally require on-premises deployment (localization)—data must not leave their own data centers; this is a hard requirement, not a negotiating chip
→ The China version of pre-built agent SKUs must include three deployment forms: cloud / on-prem / domestic Xinchuang edition; pure cloud SaaS cannot enter core scenarios
The distribution barrier in the US is SIs (Accenture/Deloitte). In China, large enterprise workflows are deeply embedded in ERP / industry management software, creating an even harder barrier to bypass:
Covers 7 million enterprises, top choice for large manufacturing / SOE ERP, already has built-in YonGPT—Yonyou is growing AI capabilities from within ERP; it's a competitor, not just a distribution channel
Ultra-large enterprise ERP platform, deep partnership with Huawei Cloud, already has an AI agent capability framework. Existing data + workflow integration + customer trust: triple barrier stacked
Monopoly digital platform for the engineering and construction industry, currently AI-ifying engineering budgets / BIM scenarios. External LLMs can barely bypass it to access engineering digitization data
Cloud-native as default deployment, SaaS subscriptions cover SME to large enterprise, Anthropic agents can be delivered directly via API
Budgets concentrated in central SOEs / state-owned enterprises / top private companies; large enterprise procurement must also provide on-prem deployment options
Hard requirement for central SOEs / state-owned banks / critical infrastructure; data does not leave own data centers
Non-core applications on public cloud, classified processes on private cloud; flexible but high integration costs
Primarily mid-sized private enterprises, non-core business scenarios; corresponds to Anthropic's main target market
Using Anthropic's three-layer structure as coordinates, evaluating the current positions of major China players:
Highest composite score, but "SOEs don't trust Alibaba" is a real hidden barrier—especially evident in financial regulatory reporting and government scenarios. DingTalk's distribution pipeline is extremely strong among private enterprises, but cannot reach the core systems of state-owned financial institutions
Strongest compliance credit (★★★★★), but model capability is the weak point. Huawei Cloud's optimal strategy may not be to become an agent product company, but to be the "safest AI infrastructure provider"—a Layer 1 role, letting other LLMs run on Huawei Cloud
All large Chinese enterprises have their business processes locked inside ERP systems. Yonyou (YonGPT) and Kingdee are growing AI capabilities from within ERP—they already have data, integration, and procurement relationships. Large enterprise AI decisions will likely be "Yonyou's AI or buy separate AI," not choosing between LLM vendors. The optimal solution may not be to defeat Yonyou, but to become the LLM that Yonyou chooses to partner with.
Key move: Not "get SOEs to buy," but "make SOEs shareholders"—shareholder relationships bring channels, compliance credit, and data access rights
Anthropic chose finance because willingness to pay is strongest. China's equivalent logic: highly standardized formats + data localization requirements + heavy regulatory pressure = scenarios worth building pre-built agents for.
For SMEs, startups, mid-sized private companies. Subscription-based, rapid deployment, suitable for non-sensitive data scenarios
For central SOEs, state-owned banks, large manufacturing enterprises. Data stays in local data centers, Level 3 classified protection certified, deep integration with internal ERP/OA
For government, military, critical infrastructure. Runs on full stack of domestic chips (Kunpeng/Hygon) + domestic OS (Kylin/UOS) + domestic databases (Dameng/Renmin Jincang)
Anthropic's agents are pure cloud delivery—China's version must offer all three in parallel, otherwise it's excluded from 70% of large enterprise procurement
Become the AI engine for Yonyou/Kingdee/Glodon—model capabilities go outward, workflow + data + customer relationships stay with the ERP company. Fast, but weak bargaining power, long-term risk of being replaced
Build AI-native enterprise software, bypassing the ERP system from process design—replacing rather than integrating. Long cycle, capital-intensive, but if successful, you own the complete data flywheel and pricing power
→ The two paths are mutually exclusive—choosing A means signing strategic partnership agreements early to get data; choosing B means finding a vertical where ERP defensiveness is weak and breaking through first
This is the hardest and most valuable step. The equivalent conditions for China's "Dimon":
The framework formed by the Data Security Law (2021) + Measures for Security Assessment of Cross-border Data Transfer (2022) has effectively created:
Bloomberg Terminal's barrier is the trinity of data + workflow + network effects. China has a direct equivalent:
Wind: Full A-share data + bonds/futures/macro, covering virtually all Chinese financial institutions—any financial AI agent without Wind data access is inherently incomplete
All domestic LLMs simultaneously attacking finance + government, leading to price wars + client decision fatigue + vertical agents operating in silos with mutual incompatibility
Government procurement contracts are stable but have 6–18 month payment cycles, many requirements, sensitivity to policy changes, and no true commercial flywheel
If you believe "regulatory protection will last 10 years" and slow down productization, you'll be caught off guard when foreign competitors find workarounds
Releasing "new SOTA models" every quarter, but enterprise clients' agents still need 3 months of customization—building models, not products
These 72 hours from Anthropic gave China's LLM enterprises their first glimpse of a complete template—not the misread version of "model companies doing consulting," but the real three-layer structure of PE-led distribution + industry agent productization + compliance credit establishment.
The China version's rewrite is harder than the original—finding the "Chinese Dimon" requires years of cultivation, not a single press conference—but the moat after success is deeper than Anthropic's: data sovereignty + compliance lock-in + domestic certification, triple barriers stacked, with extremely high switching costs.
The window is real. The question is whether any company can connect the three layers into a system before the window closes.
From model companies to enterprise AI service platforms—tracking the strategic evolution and market landscape of Anthropic, OpenAI, and China's LLM enterprises
First published 2026-05-09