Anthropic's framework provides a deconstructible sample, but the China version requires a full rewrite across three dimensions—capital structure, compliance signals, and data sovereignty—and after rewriting, 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 need local substitutes in China
The "Dimon equivalent" for China's compliance signals 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 moat—policy dividends are a window, not a moat; the window is approximately 3–5 years
v2 Update (2026-09): Four months later, the judgments are accelerating validation—Zhipu has listed on the HKEX and initiated A-share STAR Market IPO tutoring; Yonyou and Kingdee, the two ERP giants, showed in their 2026 interim earnings that they have fully pivoted to self-built enterprise agents (Kingdee's "Lingji", Yonyou BIP 6), validating this article's judgment that "ERP is a hidden moat, not just a distribution channel"; Anthropic's own annualized revenue surged from ~$9B at end of 2025 to $30B in April 2026, showing this question is worth even more points than in May
The real competitors for China's LLM enterprises are not each other, nor OpenAI/Claude—they are Wind (万得), Kingsign Tech, Yonyou Network and other software companies that have already built industry data pipelines. Whoever captures them first captures the entry point to China's Bloomberg Terminal.
In scenarios like government affairs / financial regulatory reporting, foreign competitors are effectively absent—this matters more than any parameter lead.
Pick a vertical with "highly standardized formats + data localization + heavy regulatory pressure," and complete the industry JV + pre-built agents + industry endorsement three-layer connection—the first to complete becomes the Anthropic of the China market.
What Anthropic completed in 72 hours was not three independent events—it was a tightly coupled three-layer system. Understanding this system is a prerequisite for migrating it to the China market.
Key Insight: Strong dependencies exist between the three layers—without Layer 3 (compliance credit), enterprises won't buy; 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 each exists in isolation without connecting into a system.
CTO / CFO jointly approve, decision cycle 2–6 months, technical evaluation carries high weight
Procurement committee + IT dept + compliance dept + (for SOEs) Party committee, decision cycle 6–18 months, qualifications, classified protection compliance, domestic certification carry equally high weight
→ No matter how good the product, missing Level 3 classified protection + domestic certification + SOE precedent may lose the deal
Dimon endorsement = JPMorgan legal reviewed = industry admission ticket. Private-sector CEO endorsement suffices
Requires a use case from an institution with regulatory lineage. ICBC/CCB president endorsement = China's Dimon; harder to obtain, but once secured, lock-in is stronger (SOEs typically don't simultaneously purchase competing products)
→ Compliance signals are harder to obtain, but the resulting moat depth far exceeds the US version
Financial data can go on AWS/Azure private cloud; 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
→ Foreign models absent = de facto monopoly for domestic LLMs in core financial/government/medical scenarios. This is the competitive landscape Anthropic envies most: geographic boundaries are the moat.
US: Independent PE networks (Blackstone holds 2,700+ enterprises) + Accenture/Deloitte systems integration
China: No equivalent independent PE enterprise network. SI backbone is Huawei / Yonyou / Kingdee / Glodon; pipelines are Alibaba Cloud/Tencent Cloud/Huawei Cloud/Tianyi Cloud
→ China JV counterpart = Huawei Cloud + a central SOE fund / Yonyou + ICBC strategic investment arm
US: Anthropic vs OpenAI vs Gemini, globalized competition, sensitive verticals also open
China: Core verticals (government/financial regulation/medical) have foreign players effectively absent; only domestic vendors compete
→ Competition is more focused, the pie more concentrated, and homogenization risk higher
SaaS penetration ~40%, mixed SME + large enterprise market, cloud-native is default deployment. Anthropic's agents deliver via API / cloud subscription, no on-prem needed, enterprises can self-serve
SaaS penetration ~10%–15%, procurement budgets heavily concentrated in central SOEs, state-owned enterprises, top private enterprises. 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
→ China's pre-built agent SKUs must include: cloud version / on-prem version / domestic Xinchuang version—three deployment forms; pure cloud SaaS cannot enter core scenarios
The distribution barrier in the US is SIs (Accenture/Deloitte). Chinese large enterprises' workflows are deeply embedded in ERP / industry management software, forming a harder barrier to bypass:
Covers 7M enterprises, top choice for large manufacturing / SOE ERP, has built-in YonGPT—Yonyou is growing AI capabilities from inside ERP; it's a competitor, not just a distribution channel
Ultra-large enterprise ERP platform, deep partnership with Huawei Cloud, already has AI agent capability framework. Existing data + workflow integration + customer trust—triple barrier stacked
Monopoly platform for engineering/construction industry digitization, now AI-ifying engineering budget / BIM scenarios. External LLMs can barely bypass it to access engineering digital data
Cloud-native as default deployment, SaaS subscriptions cover SME to large enterprise, Anthropic agents deliver directly via API
Budgets concentrated in central SOEs / state-owned enterprises / top private enterprises; large enterprise procurement requires on-prem deployment option
Hard requirement for central SOEs / state-owned banks / critical infrastructure; data does not leave own data center
Non-core apps on public cloud, classified processes on private cloud; flexible but high integration cost
Primarily mid-sized private enterprises, non-core business scenarios; corresponds to Anthropic's main target market
Yonyou Network H1 2026 revenue RMB 3.727B (+4.1%), net loss narrowed to RMB 928.5M (vs. RMB 944.5M prior year), first-time separate disclosure of AI-related contract signings of RMB 906M, AI-related revenue of RMB 464M; launched Yonyou BIP 6 on August 7, 2026, releasing six AI-native products at once including YonWork (enterprise workspace super-agent) and YonCode (AI coding platform), with 65 vertical agents and 623 agent instances accumulated.
Kingdee International H1 2026 total revenue RMB 3.625B (+13.6%), AI-native product revenue RMB 296M, up 189% YoY; launched "Lingji" enterprise agent OS on May 20, 2026, signing 26 customers and deploying 40+ agents within three months, included in Gartner's "Enterprise Agent Capability" Innovation Insight report.
Using Anthropic's three-layer structure as coordinates, evaluating the current position of major Chinese players:
Highest composite score, but "SOEs don't trust Alibaba" is a real hidden barrier—especially visible in financial regulatory reporting and government scenarios. DingTalk's distribution pipeline is extremely strong among private enterprises, but cannot reach 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' business processes are already locked inside ERP systems. Yonyou (YonGPT) and Kingdee are growing AI capabilities from inside ERP—they already have data, integration, and procurement relationships. Large enterprises' AI decisions will likely be "Yonyou's AI or buy AI separately," not choosing among LLM vendors. The optimal solution may not be to defeat Yonyou, but to become the LLM that Yonyou chooses to partner with.
Zhipu, judged "weakest in distribution" in the matrix, was instead the first to break through the Layer 1 bottleneck—but not via Anthropic's PE JV path, rather via the capital markets: listed on HKEX Main Board on January 8, 2026 (stock code 02513.HK), becoming the world's first listed general-purpose LLM enterprise, with first-day market cap exceeding HK$50B and IPO oversubscription exceeding 1,000×; completed STAR Market IPO tutoring verification in June 2026, planning "A+H" dual listing with A-share fundraising of ~RMB 15B, RMB 12B earmarked for LLM R&D. Public market financing itself is a compliance signal—a prospectus reviewed by regulators and cornerstone investments from licensed institutions somewhat substitute the function of an "ICBC president endorsement." This path is faster than finding an SOE JV, but the moat is shallower: stock price and market cap face continuous public market scrutiny, unlike binding a single SOE which locks in customer relationships.
Key action: not "get an SOE to pay," but "make the SOE a shareholder"—shareholder relationships bring channels, compliance credit, and data access rights
Anthropic chose finance because willingness to pay is highest. China's equivalent logic: highly standardized formats + data localization + heavy regulatory pressure = scenarios worth building pre-built agents for.
For SMEs, startups, mid-sized private enterprises. Subscription-based, rapid onboarding, suitable for non-sensitive data scenarios
For central SOEs, state-owned banks, large manufacturers. Data stays in local data center, Level 3 classified protection certified, deep integration with internal ERP/OA
For government, military, critical infrastructure. Runs on domestic chips (Kunpeng/Hygon) + domestic OS (Kylin/UOS) + domestic DB (Dameng/Kingbase) full stack
Anthropic's agents are pure cloud delivery—China version must run all three in parallel, otherwise excluded from 70% of large enterprise procurement
Become the AI engine for Yonyou/Kingdee/Glodon—model capability external-facing, workflow + data + customer relationships belong to 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, full data flywheel and pricing power
→ The two paths are mutually exclusive—choosing A means signing a strategic cooperation agreement early to get data; choosing B means finding a vertical with weak ERP defenses and penetrating it 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) effectively creates:
Bloomberg Terminal's barrier is: data + workflow + network effects combined. China has a direct counterpart:
Wind (万得): A-share full 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 attack finance + government, leading to price wars + customer decision fatigue + vertical agents each going their own way with mutual incompatibility
Government procurement large orders are stable but payment cycles are 6–18 months, requirements are numerous, sensitive to policy changes, no real commercial flywheel
If believing "regulatory protection will last 10 years" and slowing productization pace, caught unprepared when foreign players bypass restrictions
Releasing "new SOTA model" every quarter, but enterprise agents still need 3 months of customization—building models, not products
Those 72 hours gave China's LLM enterprises their first glimpse of a complete template—not the misread version of "a model company doing consulting," but the real three-layer structure of PE-led distribution + industry agent productization + compliance credit establishment.
The China rewrite is harder than the original—finding the "Chinese Dimon" requires years of cultivation, not a single press conference—but the resulting moat 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.
This article was published in May 2026; four months later, a review shows: core judgments have not been overturned, and are instead accelerating simultaneously on three fronts—the original template itself is growing, China's "substitutes" are starting to appear, and faster than expected.
Anthropic's annualized run-rate revenue surged from ~$9B at end of 2025 to ~$30B in April 2026 (Bloomberg data), and the number of enterprise customers paying $1M+ doubled within months. In other words, those 72 hours in May were not a one-time PR stunt, but a growth curve being realized. This means "China's Anthropic" needs to benchmark not the Anthropic of the May press conference day, but a company whose revenue is still growing exponentially—the difficulty threshold is rising.
Zhipu, judged "weakest in distribution" in the §03 matrix, used HKEX + STAR Market dual listing to solve half the Layer 1 (distribution/capital signal) problem (see matrix update); Kingdee and Yonyou, the two ERP giants, didn't stop at "embedded AI" but went straight to building enterprise-grade agent operating systems (see Data 05 update); Wind didn't wait to be acquired, but opened its MCP data entry point (see §05 update); Huawei's partner ecosystem shows ~80% of licensed financial institutions nationwide have deployed financial agents (see §04 update). Looking at all four threads together, the China version of the three-layer structure is not a question of "whether it's possible," but "who can first assemble the three layers into a system"—the playbook proposed in §04 still holds, but there's even less time left for fence-sitters.
What Hasn't Changed: The legal framework behind the data sovereignty moat (§05) has not been amended—the various regulations CAC has issued since 2026 (Personal Information Cross-border Certification Measures, Compliance Guidelines for Promoting and Regulating Cross-border Data Flows in the Financial Sector, etc.) are refinements of supporting rules, not loosening of core restrictions. The "window of ~3–5 years" judgment in §05 currently has no evidence of narrowing early, but also no evidence of extension—the moat's "expiration date" assumption remains the variable in this article most in need of continuous tracking.
From model company to enterprise AI services platform—tracking the strategic evolution and market landscape of Anthropic, OpenAI, and China's LLM enterprises