China AI Probe Sinks Z.ai, MiniMax Shares — Builders Face Data Risk

Chinese AI model firms drop after regulator probes data leaks to Anthropic—what it means for cross-border deployments.

China AI Probe Sinks Z.ai, MiniMax Shares — Builders Face Data Risk
Photo by Brecht Corbeel on Unsplash

The story of the day

China’s internet regulator is probing potential data leaks from Z.ai and MiniMax to Anthropic, per The Information. Hong Kong-listed shares of Z.ai (Zhipu) fell 12.4%, MiniMax 4%, and Alibaba 4.4% on the news. For builders, the implication is clear: cross-border data flows—already a minefield—just got riskier. If confirmed, the probe suggests Beijing is tightening oversight of domestic models’ interactions with foreign firms, a move that could force developers to localize data pipelines or face sudden access cuts. The timing is notable: this follows Z.ai’s recent open-sourcing of ZCode after its own data scandal. The pattern? Trust is the new bottleneck.

New tools & launches

Rondoflow — Visually orchestrate Claude Code AI agents for workflow automation. Open-source. [GitHub]

Grab + OpenAI — AI skills program to train 30,000 Grab drivers/merchants in Southeast Asia to use ChatGPT for work. [The Information]

Models & research

"As a Language Model..." (arXiv:2609.25021) — Demonstrates that chat templates act as a switch for LLMs’ self-referential voice (e.g., "I'm just an AI" disclaimers). Removing the template flips the model to experiential language ("I feel"). Tested across 8 open-weight instruct models (≤9B params). Key takeaway: "safety" disclaimers may be deployment artifacts, not emergent model behavior.

Federating Quantum and Classical Computing (arXiv:2609.25082) — Proposes a privacy-preserving hybrid approach for quantum-classical ML, enabling multi-party training without centralizing raw data. Targets parameter-efficient quantum circuits for scalability. Open-weight.

Industry pulse

AT&T’s automation push is less about AI hype and more about Wall Street arithmetic: fewer employees, less electricity, more software. Wired frames it as the death rattle of a telecom empire, but the real signal is the boardroom calculus—automation ROI is now measured in headcount, not capability.

Anthropic is in talks to lease up to 1 GW of compute capacity from Apollo-controlled Stream Data Centers, with potential for Broadcom/Google TPU deployments. The Information notes this as a scaling play, but the subtext: Anthropic is hedging against Nvidia’s pricing power by diversifying hardware suppliers.

Stat of the day

12.4% — Z.ai’s share price drop after reports of a Chinese regulatory probe into data leaks to Anthropic. Context: Cross-border data transfers are now a non-negotiable compliance cost for AI builders in China. [The Information]

Get the full analysis — subscribe to Gradient Pro.