Anthropic's business model is a product of the US cloud-native software market — a trinity of cloud delivery, standard agents, and PE distribution networks. Entering China means all three layers need to be completely redesigned: the deployment layer becomes on-premises, the product layer becomes deeply customized, and the distribution layer becomes a state-owned enterprise (SOE) trust chain. It's not impossible, but the transformation cost and timeline are severely underestimated.
The main competitor of a China-version Anthropic is not DeepSeek or Zhipu, but Yonyou and Kingdee. The data is in the ERP, the trust is in the ERP, and the decision budget is in the ERP. Before LLM companies can enter large enterprise customers, they must first answer a prerequisite question: embed into the ERP, or replace the ERP from the outside? The two paths have fundamentally different margins, sales motions, and capital structures — there is no middle ground.
To assess the portability of the Anthropic model, we must first establish software industry baselines for three markets — not market size, but structure: business models, deployment preferences, revenue composition. The gaps between the three markets are larger than most people realize.
Europe is the closest reference frame to China among the three: data sovereignty constraints (GDPR ≈ Data Security Law + PIPL), higher government procurement share, and stronger On-Prem preference — these three points share structural similarity with China. But China adds three unique layers on top: LLM filing requirements, Xinchuang domestic substitution mandates, and the political vetting chain in SOE procurement. European adaptation experience has reference value, but cannot be directly copied.
New deployments default to cloud; AWS/Azure hosting is the standard path. On-premises is the exception, seen only in government classified or special sectors (defense/nuclear)
Mittelstand ~40% still prefer On-Prem; GDPR requires data to stay within specific jurisdictions, driving "sovereign cloud" (Deutsche Telekom/OVHcloud) demand
On-premises is the default, not the exception. Central SOEs/state-owned banks/critical infrastructure data cannot leave local data centers — this is a hard compliance requirement, not a negotiating chip. 64% of large enterprise AI procurement requires on-premises deployment
→ Anthropic's API cloud delivery model fails directly in China's large enterprise market. Must provide an on-premises version that can be independently deployed in customer data centers, meaning: GPU procurement/leasing costs shift from cloud to vendor, fundamentally altering the gross margin structure.
"Fit to standard" culture: enterprises are willing to adjust internal processes for standard products. Salesforce and Workday's success is built on the philosophy of "configuration, not development"
German SMEs especially lean toward customization: the "Made in Germany" spirit extends to software, with lower willingness to standardize processes; SAP's rise itself served this need
SOE/central SOE projects can reach 80%+. Historically, due to rampant ERP piracy, Yonyou/Kingdee built competitive moats through extensive custom services; enterprise clients thus internalized the mental model of "software is a custom product"
Subscription-first: Salesforce seat fees $25–300/month, Anthropic Claude Enterprise API pay-per-token + fixed seat fee, revenue is predictable and scalable
Light implementation: GSI charges implementation fees, product vendor doesn't participate in implementation, margin structure is clear
Upgrade path: Auto-updates, invisible to customers, marginal cost approaches zero
Perpetual license dominant: Enterprises习惯one-time buyout; of Yonyou's FY2024 ¥9.15B revenue, subscription is only ~20%, the rest is license + implementation fees
Implementation is a profit center: Software license : implementation fee = 1:3 to 1:5 (US ~1:1), implementation teams are the core competitive advantage
Upgrade resistance: Heavy custom code means version upgrades require re-implementation, customers tend not to upgrade — creating massive "implementation debt"
→ Pure subscription / pure API consumption pricing logic requires a hybrid model in China's large enterprise market: subscription base + implementation services + custom development, meaning implementation delivery capability must be built
Anthropic's PE JV model works in the US because Blackstone/KKR hold 2,000+ portfolio companies; a JV collaboration simultaneously opens procurement channels to these enterprises. This structure does not exist in China.
Blackstone portfolio companies 2,700+; one JV agreement is equivalent to issuing entry tickets to 2,700 enterprises. GPM unaffected, distribution marginal cost extremely low
Capgemini/Sopra Steria are main distribution channels; Europe lacks equivalent PE networks, but GSI penetration is high, covering large enterprises through GSI partnerships is relatively feasible
No independent PE network; distribution core is the SOE trust chain: ICBC/China Mobile/State Grid procurement → other central SOEs reference → leading private enterprises follow. Building the trust chain takes 2–4 years
CTO/CIO-led, technical POC is standard process; Europe adds Data Protection Officer (DPO) and legal review, but overall path is clear. Dimon's endorsement = JPMorgan legal sign-off = industry entry ticket
Typical large SOE procurement path: Technical department research → IT committee review → Information security compliance review (classified protection, data security) → Procurement committee → (Party committee) study → Leadership approval. Each step can become a bottleneck
→ China-version "Dimon endorsement" = ICBC president's public endorsement; but far harder to obtain than in the US, and SOEs typically don't simultaneously purchase competing products — once secured, the lock-in effect is much stronger
Cloud-native is the default path; Anthropic agents delivered via API/cloud subscription, no on-premises needed, enterprises can self-onboard
Germany/France on the low side (~18–22%), Nordics higher (~35%); GDPR drives sovereign cloud demand rather than hindering SaaS overall
Budget concentrated in large enterprises, which generally require on-premises; SaaS growth mainly from mid-size private enterprises' non-core applications
| Metric | 🇺🇸 US (Current) | 🇨🇳 China (Adapted Estimate) |
|---|---|---|
| Delivery model | Cloud API / Claude.ai Teams | On-prem + Cloud API (hybrid) |
| Gross margin (GPM) | ~70–75% | ~30–40% (on-prem dilution) |
| Typical enterprise contract size | $500K–5M / year | ¥3M–15M / year (~$400K–2M) |
| Sales cycle | 3–6 months | 12–24 months |
| Implementation cost (vendor bears) | Low (GSI bears, vendor doesn't participate) | High (must build implementation team or deep SI partnership) |
| Pricing basis | Token consumption + seat subscription | Annual subscription + custom dev fee + implementation fee |
| Standard configuration share | ~80% (low customization) | ~30% (high customization) |
| Time to scale to 100 large enterprises | ~18–24 months | ~4–6 years |
Anthropic's enterprise AI business model consists of three interdependent layers. Understanding why each layer works in the US is prerequisite to precisely judging which layer will fail first in China.
Enterprises can directly access Claude via Anthropic API, no on-premises needed. Cloud hosting means Anthropic's compute costs are amortized by Google/Amazon infrastructure, GPM ~70%+. Token billing enables "pay for what you use," lowering procurement threshold
Cloud API is ineffective for 64% of large enterprises requiring on-premises deployment. Even accepting a hybrid model, LLM service filing requires in-country services to use filed models. Foreign models face compliance obstacles providing services within China; Anthropic itself cannot directly enter core scenarios via API
US financial industry report formats (SEC 10-K/10-Q), compliance processes (SOX/Dodd-Frank) are highly standardized. 10 financial agents can cover 80% of JPMorgan's common scenarios without customization. GSI handles the remaining 20% adaptation
Chinese financial industry report formats are fragmented (A-share/HKEX/CBRC/CSRC/PBOC formats are mutually inconsistent), OA approval workflows are highly individualized, supply chain data standards lack unification. Standard agents' "out-of-the-box" capability fails before entering the meeting room — every enterprise will say "our situation is somewhat special"
A JV agreement with Blackstone et al. simultaneously opens procurement channels to 2,700+ portfolio companies. Dimon's public endorsement = JPMorgan legal/compliance has reviewed = other financial CTOs can reference this precedent to reduce their own decision risk
Chinese PE/VC don't hold operating enterprises' procurement rights, can't "bundle" channels. Trust signals come from different sources: SOE CEO's statement > private CEO's endorsement. Building an SOE trust chain requires not JV equity structures, but real deployment cases + government procurement records — this takes 2–4 years to accumulate
→ Layer 3 is the hardest to transform of the three layers. Layer 1 and Layer 2 can be solved through technical investment; Layer 3 requires time and political capital, neither of which can be accelerated by spending money.
The following three scenarios are thought experiments in structural feasibility, not predictions of Anthropic's actual business decisions — scores only answer "if attempted, can the commercial structure work," and do not address political-level possibilities such as export controls and national security reviews, which almost preclude Scenarios A and B from actually occurring.
Assume Anthropic keeps its current cloud API model + pre-built agents + PE JV structure and enters the Chinese market without any modification — where does it fail first structurally?
Actual probability of occurrence: ≈0 — export controls and LLM service filing requirements would eliminate this path before the structural test, no need to wait for market validation
Cloud API cannot enter on-premises deployment scenarios; foreign model filing obstacles; data cannot leave borders
US financial agent formats don't match Chinese reporting standards; no native Chinese workflow support
PE JV has no equivalent structure in China; no SOE trust chain; compliance credit is zero
Conclusion: Can only serve a small number of foreign enterprise China branches and export-oriented leading private enterprises; cannot enter government/finance/SOE core scenarios, which account for 60%+ of China's AI spending
Assume Anthropic or an equivalent international AI company is willing to establish an independent legal entity in China, jointly developing an on-premises version, rewriting Chinese agents, and building SOE JV structures with domestic partners — corresponding to Google/Microsoft's past China strategies — how far can it go structurally?
Actual probability of occurrence: Low. 55/100 is only the pure commercial structure score; the real barrier is the premise itself — whether a US frontier model company would accept the accompanying export control scrutiny and geopolitical friction, which is harder to overcome than Layer 3 trust chain building, and is outside the structural scope of this article
Theoretical timeline (assuming decision to proceed): earliest 3 years to enter non-core central SOE scenarios, 5–7 years to enter core finance/government scenarios; capital requirement: ~¥1–3 billion buildout investment
Chinese AI vendors replicate Anthropic's three-layer structure, but fill each layer with domestic resources: domestic LLMs + Chinese enterprise agents + SOE/cloud vendor distribution networks. This is the most likely path to succeed.
DeepSeek / Kimi / Qwen / ERNIE · Filed · Support on-premises · Xinchuang adaptation feasible
Rewrite A-share report agents / SOE OA agents / supply chain agents; embed Yonyou/Kingdee data interfaces
Central SOE strategic investment + Alibaba Cloud/Huawei Cloud pipeline + Yonyou/Kingdee ISV channel; saves 2–3 years of trust building vs foreign vendors
Conclusion: No single company currently possesses all three layers simultaneously. The "fastest winner" of the Anthropic model in China may be a new JV entity, not an existing single vendor. This is also the scenario this article considers closest to the realistic trajectory — not because Anthropic would do this, but because Chinese vendors don't need to cross the political threshold in Scenarios A/B; the outcome depends only on who first assembles Layer 2 and Layer 3
The following eight changes are each score-reducing if missing. The first four are technical and product-level, the latter four are commercial and compliance-level — the latter are harder, because you can't trade engineering capability for time. This checklist is reverse-engineered from the structural requirements of Scenarios A/B/C thought experiments, and applies equally to Scenario B's foreign-led entrant and Scenario C's Chinese vendors; but since Scenarios A/B have near-zero real-world probability, the ones most likely to fulfill this checklist item by item are probably Scenario C's Chinese vendors, not Anthropic itself.
Must simultaneously provide three deployment forms: ① Public cloud version (mid-size private enterprises, non-core applications) ② On-premises version (large enterprises/finance/SOEs) ③ Xinchuang version (SOEs/government, running on Feiteng/Kunpeng CPU + Kylin OS). The three SKUs are not simple repackaging of the same codebase — the Xinchuang version requires adaptation for non-x86 architectures, engineering effort ~30–50% of building a new product
Shift from pure subscription/consumption model to three-part pricing: ① Annual platform subscription base (recurring revenue source) ② Custom development fee (priced per module, not per person-day) ③ Implementation service fee. Key: custom development fees must be productized (with fixed SKUs), otherwise you devolve into a zero-margin project company rather than a scalable AI platform
Not translating US agents, but natively rewriting: A-share annual report analysis agent (connecting Wind data), SOE OA approval agent (connecting DingTalk/WeCom workflows), supply chain agent (connecting Yonyou/SAP data fields), bank anti-fraud/intelligent customer service agent (connecting PBOC data standards). Each agent's underlying data format follows Chinese standards, not translated US standards
Provide a low-code customization framework rather than fixed templates: customer IT or implementation partners can drag-and-drop assemble on the framework without vendor engineers participating in every project. This is the key to shifting "customization debt" from vendor to customer/channel. Reference Salesforce Flow / ServiceNow's no-code configuration layer logic
Four compliance certifications are all indispensable: ① LLM service filing (CAC "Interim Measures for the Management of Generative AI Services", prerequisite for providing generative AI services) ② Classified Protection Level 3 certification (minimum threshold for entering finance/government scenarios) ③ PIPL/DSL compliant data processing agreements ④ Xinchuang compatibility certification (joint testing with specific hardware vendors). Filing cycle ~3–6 months, classified protection certification ~6–12 months, start early
China-version "PE JV" is a triangular partnership: ① Central SOE industry fund (provides compliance credit + SOE customer procurement endorsement, minority stake) ② Top cloud vendor (Huawei Cloud/Alibaba Cloud, provides compute infrastructure + cloud marketplace distribution) ③ ERP vendor strategic partnership (Yonyou/Kingdee, data interfaces + existing customer reach). Missing any side of the triangle significantly slows entry into core scenarios
The only viable trust chain building sequence: Government scenario reference customer (regardless of revenue size, first produce a publicly referenceable case) → policy banks / state-owned banks → central SOEs → top city commercial banks / insurance → leading private enterprises → mid-size enterprises. There is no shortcut that bypasses "government first." Anthropic's Dimon path works in the US because US private financial institutions are highly independent; China needs government nodes as trust anchors
Two things from Europe are worth directly borrowing: ① Sovereign cloud architecture (GDPR-compliant data localization implementation can directly correspond to DSL/PIPL requirements, reducing redundant development) ② German SI ecosystem model (Germany handles high customization demand by strengthening SI partnerships rather than product standardization, which is closer to China's demand structure). Not worth borrowing: Europe's AI Act framework — the EU's risk classification logic is fundamentally different from China's filing system logic, and cannot be directly mapped
CAC may require AI agents to be filed separately (rather than only filing the underlying model); compliance costs multiply, and requirements may change with policy cycles, creating retroactive adaptation costs
If Xinchuang requirements extend from government/SOEs to the entire financial industry or critical infrastructure, the hardware ecosystem available for AI deployment will narrow significantly, and NVIDIA GPU-based on-premises solutions may become invalid
Each on-premises project requires the customer or vendor to procure GPU clusters (¥millions to ¥tens of millions scale). As on-premises project count grows, GPU procurement/maintenance costs become an explicit liability, while cloud API model's economies of scale cannot transfer
Yonyou YonGPT and Kingdee AI are growing AI capabilities from inside the ERP. Despite Yonyou's FY2024 ¥2.06B loss, it controls data access for 7M enterprises — external LLM vendors must answer "why not just use the ERP's built-in AI"
Update (2026-09): Yonyou's 2025 annual report net loss narrowed to -¥1.389B, 2026 half-year report further narrowed to -¥928M with AI-related contract signings of ¥906M; Kingdee International has turned profitable in FY2025 (net profit ~¥93M), cloud services revenue share 82.5%, subscription revenue ¥3.556B (+20.9%), and launched AI-native entry "Xiao K" and ~20 industry intelligent agents — ERP vendors' AI transformation is materializing faster than the May assessment
Highly customized projects mean each customer is a separate code branch; version upgrades require re-implementation. When custom customers exceed 50, maintenance costs may exceed new business revenue, creating a "growth trap"
Hybrid talent who understand both LLMs and enterprise ERP/OA systems, and can communicate with SOE customers, is extremely scarce. Unlike the ERP era, AI implementation requires greater technical depth, but the existing SI ecosystem hasn't yet built this talent reserve
The Anthropic model works because it highly aligns with the structure of the US software market: cloud delivery corresponds to cloud-native deployment defaults, standard agents correspond to 25% customization rates, PE JV corresponds to private PE enterprise networks. There is no friction between the layers — each layer paves the way for the next.
The six structural differences in China's software market cause this alignment to disappear entirely. It's not that one layer doesn't match — all three layers don't match, and each layer's mismatch has a different cause that can't be solved with the same approach.
This doesn't mean China has no market — the geographic moat formed by data sovereignty actually creates a domain that domestic AI service providers can occupy but US competitors can never touch. But entering this domain requires rebuilding the three-layer architecture using Chinese industry logic, not translating Anthropic's three layers.
European experience provides partial reference — sovereign cloud architecture, SI-led customization handling, government-procurement-first trust chains — but China's regulatory density and SOE-dominated landscape mean European cases can only be referenced, not copied.
Final judgment: China doesn't lack model capability (Layer 1). What's missing is a composite entity that can simultaneously do Layer 2 (Chinese agents) and Layer 3 (distribution trust) well. Whoever assembles these two layers first becomes China's Anthropic — but it won't be called Anthropic, and it won't be led by a US company.