> ## Content Index > Fetch the complete content index at: https://blog.aperion.ai/llms.txt > Use this file to discover other available public pages before exploring further. # One Global AI Policy Plane, Including China - URL: https://blog.aperion.ai/smartflow-china-architecture/ - Published: 2026-10-26T21:05:58.000Z - Updated: 2026-10-26T21:05:58.000Z - Description: How a multinational gives its China teams governed AI on in-country models, under the same identity and policy as the rest of the company, with prompts, responses and records kept in China. - Author: Craig Alberino - Tags: AI Regulation, Enterprise Strategy, AI Governance *Architecture overview. Regulatory references verified as of October 26, 2026.* ## The problem: China is a telemetry black hole Multinationals operating in China face a structural gap. Employees there cannot use the Western frontier models the rest of the company runs on, so they turn to in-country services built on Qwen, DeepSeek, Kimi and similar models. Those services are capable, but they sit outside the company's governance perimeter: no policy enforcement, no caching, no usage records, no audit trail. Global IT cannot see who is sending what to which model. It cannot show compliance in either direction, with China's own data and AI rules or with the company's global standards. ![Two paths for the same request. Today, China-based employees go straight to in-country AI services with no policy, cache or record. Through an in-country gateway, the same requests carry corporate identity, pass policy and cache, and leave a record that stays in China.](https://storage.ghost.io/c/56/dd/56dd2e81-b5ad-4845-bc3a-a53d4e7c40f8/content/images/2026/10/aperion-china-two-paths-2026-10-26.webp) Today's direct path, and the same requests through an in-country gateway. ## The architecture: an in-country gateway under one global policy plane Smartflow is designed to run as a Kubernetes-native gateway inside China, on the company's own in-country infrastructure or in a Chinese cloud region. China-based employees sign in with the company's existing identity provider, and their requests go to the in-country models the company has approved. Prompts, responses and audit records are processed and stored in China. Policy definitions are written once and travel to every region. Enforcement runs inside each one. ![One global policy plane above two regional gateways. Policy definitions travel down to both. The gateway in mainland China routes to approved in-country models and models the company hosts itself, and keeps prompts, responses, audit records and provider credentials in China. A minimized rollup crosses the boundary only where the transfer is approved.](https://storage.ghost.io/c/56/dd/56dd2e81-b5ad-4845-bc3a-a53d4e7c40f8/content/images/2026/10/aperion-china-policy-plane-2026-10-26.webp) Policy definitions travel. Enforcement, governed traffic and evidence stay inside each region. #### Identity: named people, not shared keys Smartflow authenticates users through the company's existing OIDC identity provider, such as Okta or Microsoft Entra ID. Group and role claims from the identity provider drive team assignment and policy, so the corporate identity that opens the laptop also governs the AI request. Sensitive routes can require a verified human principal before a request proceeds. #### Routing: in-country models, chosen by policy The gateway connects to in-country commercial model APIs and to open-weight models the company hosts itself in China. The supported provider list is in the documentation. Model choice is a policy decision made per request: China-based users reach the endpoints the company's counsel has approved, and traffic crosses the boundary only where policy allows it. #### Policy: fail-closed by default Policy is enforced on the request path, before traffic leaves the gateway: approved models and routes per team, data-sensitivity rules, controls on tool calls, and a fail-closed default for anything not explicitly permitted. The same definitions apply globally, with regional overlays for local rules. #### Records, cost and cache Every request produces a usage, cost and audit record, written inside China, which closes the black hole. Semantic caching serves repeated requests from the in-country gateway, which matters where in-country model capacity is constrained. Up to 80% provider-side cost reduction on cache-eligible traffic.\* \* Observed across design-partner deployments over 90-day windows. Results are workload-dependent and should be measured on the customer's own traffic during a scoped POC. ## Compliance posture Verified as of October 26, 2026. | Concern | How the architecture addresses it | | ------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | **China data rules** (Personal Information Protection Law, Data Security Law, Cybersecurity Law) | In this design, prompts, responses and audit records are processed and stored in China. Where the deployment supports it and the company has approved the transfer, global IT receives a minimized operational rollup: control results, policy versions, attestations, evidence hashes and aggregate cost by team. Prompts and responses stay in China. Routing China-based users to in-country endpoints keeps the inference path inside China. | | **Generative AI services** (CAC interim measures) | A generative AI service offered to the public in mainland China needs a model that has completed CAC filing, and the service itself is registered with the local cyberspace authority. Use by the company's own employees generally falls outside that regime. Per-team approved-model lists keep public-facing routes on filed models. | | **Global corporate policy** | One policy plane, one identity source and one audit format worldwide. Regional overlays express local requirements without forking governance. | | **Export controls** | Smartflow is commercial software, delivered without hardware. APERION provides Smartflow's export classification on request. Which users may reach which models is a policy the company's counsel sets and the gateway enforces on every request. | | **Vendor lock-in** | The gateway is provider-neutral. Adding or swapping an approved model is a configuration change. | ## What a deployment looks like **Footprint:** Docker or Kubernetes in the company's China infrastructure. No prompt, response or cached artifact traverses APERION infrastructure. **Identity:** point the gateway at the existing Okta or Entra ID tenant, and group claims map to teams and policies. **Model catalog:** publish the in-country set the company's counsel approves through one standard chat-completions endpoint. **Records:** usage, cost and audit records stay in China, and the minimized rollup reaches the global dashboard where the transfer is approved, with per-team cost attribution from the first day. For the same pattern applied to a guest record that crosses three jurisdictions in one week, see [Governed AI for multinational hotel groups](https://aperion.ai/blog/governed-ai-multinational-hotel-operations/). *Regulatory summaries are general and current as of October 26, 2026\. They are not legal advice. Confirm requirements with counsel in each jurisdiction.*