One Global AI Policy Plane, Including China
How multinationals can give their China-based teams first-class AI access on approved in-country models while keeping the same identity, policy, telemetry, and cost governance they enforce everywhere else in the world.
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 reliable policy enforcement, no caching, no usage telemetry, no audit trail. Global IT cannot see who is sending what to which model, and cannot prove compliance in either direction, with China's own AI and data regulations or with the company's global standards.
The architecture: an in-country gateway under a global policy plane
Smartflow deploys as a Kubernetes-native gateway inside China, in the company's own in-country infrastructure or a Chinese cloud region. Chinese users authenticate with the company's existing identity provider and are routed to approved Chinese models. Prompt and response content stays in-country. What leaves is governance metadata: who used AI, under which policy, at what cost, with what outcome. Global IT sees one consistent picture across every region, and every region enforces the same rules.
Identity: real humans, not shared keys
Smartflow authenticates users through the company's existing OIDC identity provider, including Okta, Microsoft Entra ID, Auth0, and Google Workspace. Group and role claims from the IdP drive team assignment and policy, so the same corporate identity that opens the laptop governs the AI request. Verified-identity enforcement can require a known human principal before sensitive routes are allowed.
Routing: Chinese models, natively
The gateway speaks natively to Alibaba Qwen via DashScope, DeepSeek, and Moonshot Kimi, alongside self-hosted models served in-country through vLLM, Ollama, and compatible runtimes. Model choice is a policy decision made per request: Chinese users reach approved Chinese endpoints, and traffic never silently crosses the data boundary.
Policy: deterministic, fail-closed
Policies are enforced on the request path before traffic leaves the gateway: allowed models and routes per team, data-sensitivity rules, guardrails on tool calls, and deterministic deny when a request is not explicitly permitted. The same policy definitions apply globally, with regional overlays for local regulation.
Telemetry, cost, and cache
Every request produces usage, cost, and audit records, closing the black hole. Semantic caching serves repeated traffic locally, cutting token spend (up to ~80% in production deployments) and keeping response times fast even when in-country model capacity is constrained.
Validated against shipping code, not slides
The claims above were verified directly against the current Smartflow release binary: native endpoint support for DashScope (dashscope.aliyuncs.com) and DeepSeek (api.deepseek.com), Moonshot support with its private-VPC endpoint override (MOONSHOT_API_BASE, default api.moonshot.cn), OIDC single sign-on with Okta and Entra ID group-to-role sync, per-request policy enforcement with fail-closed default, and per-request usage, cost, and audit records.
Compliance posture
| Concern | How the architecture addresses it |
|---|---|
| China data rules (PIPL, generative-AI measures) | Prompt and response content is processed and stored in-country. Only governance metadata leaves. Approved-model routing keeps Chinese personal data on Chinese endpoints. |
| Global corporate policy | One policy plane, one identity source, one audit format worldwide. Regional overlays express local requirements without forking governance. |
| Export restrictions | Smartflow is commercial gateway software developed by APERION. It ships no model weights and no controlled hardware, so the U.S. export restrictions that apply to advanced AI chips and frontier model weights do not apply. Chinese model endpoints are supported natively in the standard product. |
| Vendor lock-in | The gateway is provider-neutral across 30+ model providers. Swapping or adding an approved model is configuration, not a rebuild. |
What a deployment looks like
Footprint: Docker/Kubernetes in the company's China infrastructure, with no SaaS dependency. Identity: point the gateway at the existing Okta or Entra ID tenant; group claims map to teams and policies automatically. Model catalog: approve the in-country set (Qwen, DeepSeek, Kimi, local vLLM-served models) and publish it through one OpenAI-compatible endpoint. Telemetry: usage, cost, and audit data flows to the global dashboard while content stays in-region, with per-team cost attribution from day one. The result is simple to state: China stops being an exception. The same employee identity, policy, telemetry, and cost controls, everywhere the company operates.