JD.com blocked six external AI platforms in late March — IT Home first broke the story — cutting employee access to Doubao, Qwen, DeepSeek, ChatGPT, Gemini, and Grok. Meituan imposed approval gates on Alibaba’s Qwen the same week. Both gave data-security justifications. The concern is simpler: every prompt routed through a rival’s model carries proprietary transaction data with it.

When an employee queries a rival’s model with purchase volumes, delivery routes, or pricing signals, the prompt is the data export. No breach is required. The information arrives by invitation, packaged as natural language. An operations manager asking Qwen to optimise delivery zones hands Alibaba a map of the company’s logistics network. That banned list is a competitive census.

A data-security policy that requires senior-management approval for Alibaba’s Qwen while exempting ByteDance’s Doubao is not a data-security policy.

JD.com routes blocked employees to JoyAI, its in-house model. Meituan steers teams toward LongCat, the 560-billion-parameter model it open-sourced last September. The asymmetry is the strategy: open source is the offence; blocking rivals is the defence.

The fear has precedent. Anthropic disclosed in February that three Chinese AI labs had routed over sixteen million conversations through Claude via 24,000 fraudulent accounts, extracting step-by-step reasoning to train competing systems. The technique is distillation.

Alibaba’s Qwen processed over 200 million one-sentence transactions during the Spring Festival campaign. When a Meituan employee queries Qwen about pricing, Alibaba’s model absorbs something about Meituan’s operations. JD.com and Meituan drew the obvious conclusion: the rival model is the rival pipeline.

The platform war used to be about who delivered faster. It is now about whose model learns from whose receipts.