China’s Cyberspace Administration is investigating DeepSeek and Moonshot following Anthropic’s unproven allegation of covert API routing โ and the structural logic of the dispute is more informative than either the allegation or the inquiry.
Mandatory disclosure: This article was drafted with a Claude model made by Anthropic โ which is a named party to the underlying dispute. No wrongdoing by any company has been established. This is not legal advice. Readers should weigh that context accordingly throughout.
The Information reported on September 22, 2026 โ with Bloomberg and others carrying the story โ that China’s Cyberspace Administration has opened an investigation into DeepSeek and Moonshot. The inquiry follows an allegation Anthropic made in its own September threat report: that both companies, along with several others, covertly routed user conversations through Claude and used its responses to distil its capabilities. That allegation is Anthropic’s and has not been confirmed by DeepSeek, Moonshot, or the Cyberspace Administration.
As reported, officials interviewed executives and employees at the companies’ offices to assess the severity of any potential breaches. The inquiry began with representatives of seven companies summoned, then narrowed to these two. The criteria for that narrowing have not been published. Regulators are reported to be examining whether sensitive Chinese police, military and state-linked data reached a United States AI system through the alleged routing โ which defines the scope of an inquiry, not a finding that any such data moved.
Anthropic’s report also named Alibaba, Xiaomi, Zhipu, SenseTime, and MiniMax. None of the allegations against any of the named companies are established facts. No confirmation, denial, or statement has been published by any of the seven companies or the Cyberspace Administration as of the date of this article, and the absence of such statements is not characterised here as meaning anything.
The key insight: One alleged data flow has produced two complaints pointing in opposite directions โ Anthropic’s concern is about what was taken out of its system; the regulator’s reported concern is about what Chinese data left its jurisdiction and reached a US system. Those are two distinct potential harms at opposite ends of the same pipe. That structure โ not the allegation itself โ is what turned a commercial claim in one country into a regulatory matter in another within a fortnight.
The Structural Read
The Permission Layer framework โ one of the Business Engineer tools for reading how regulatory power shapes which AI systems actually operate and at what scale โ maps this dispute cleanly. The framework holds that in AI, capability does not determine market access; permission does. And permission is increasingly being exercised at the data layer, not the product layer.
What the reported investigation makes structurally visible is that a single alleged technical event โ API calls routed through a foreign model โ can simultaneously trigger the permission mechanisms of two different jurisdictions, in two different directions, under two different standards of proof. That is not an anomaly. That is the ordinary architecture of a cross-border data dispute when the pipe crosses a regulatory boundary.
The same fact pattern is now three different matters. As a commercial dispute, it would turn on terms of service, proved to a contractual standard. As a data-security question, it turns on what category of information moved, assessed against a regulatory standard. As an investigation, it presently requires no proof at all โ an inquiry is a process for finding out, not a conclusion about what happened. Each framing is real and each demands a different analytical lens, but they are not interchangeable, and treating them as equivalent produces bad reads.
Permission Layer โ Structural Principle
The Regulator Establishes a Question, Not an Answer
A regulator opening an inquiry establishes that the regulator considered the question worth asking. That is a statement about the regulator’s threshold for action โ not about the companies under examination. The hazard specific to this news cycle is that readers encounter all three framings in the same week and treat the existence of an investigation as evidence for the original allegation. It is not. And the narrowing from seven companies to two is not evidence against the two that remain, nor exoneration of the five that are no longer the focus. It is a procedural step whose reasoning has not been published.
On the Volume Figures โ A Counting Problem
Anthropic’s report alleged Moonshot sent over 23 million exchanges to Claude across May through July 2026, and that DeepSeek sent 12.1 million in a fourteen-day window in July 2026. Those figures cover different periods โ the Moonshot figure spans roughly six times the duration of the DeepSeek figure. They are not comparable with each other, and presenting them adjacently invites an inference neither figure supports. No rate, ratio, or per-day derivation is computed here, because the published figures do not license that arithmetic. A number does not become comparable because the party publishing it would benefit if it were treated that way.
Three Implications
IMPLICATION 1 โ THE PERMISSION LAYER IS NOW BIDIRECTIONAL
When an AI company makes an allegation about a cross-border data flow, it activates regulatory permission mechanisms in both jurisdictions simultaneously โ and potentially in opposite directions. The company that raises the complaint is not insulated from scrutiny; the act of surfacing the data flow draws attention to the flow itself, not only to the alleged wrongdoer. This is a structural feature of cross-border data disputes, not a peculiarity of this one.
IMPLICATION 2 โ STANDARDS OF PROOF DO NOT TRAVEL ACROSS FRAMINGS
The commercial, data-security, and investigatory framings each carry their own evidentiary standard, and none of them borrows credibility from the others. A reader who exits this news cycle believing that an open investigation strengthens the original allegation โ or that a commercial complaint validates a regulatory inquiry โ has made a logical error that the coverage structure encourages but the underlying facts do not support. Each framing must be assessed on its own terms.
IMPLICATION 3 โ PROCEDURAL NARROWING IS INFORMATION ABOUT PROCESS, NOT ABOUT GUILT OR INNOCENCE
The narrowing from seven companies to two indicates that the inquiry is discriminating between cases rather than treating a broad allegation as uniformly credible. That is the informative part. But it carries no verdict in either direction: it is not evidence against DeepSeek and Moonshot, and it is not exoneration of the five companies no longer in focus. Reading a conclusion into a procedural filter โ in any direction, about any of the seven โ invents information that does not exist.
Business Engineer Framework
The Permission Layer
In AI markets, capability does not determine access โ permission does. The Permission Layer framework maps how regulatory power shapes which AI systems operate, at what scale, and across which borders. This dispute is a case study in what happens when a single data flow crosses a jurisdictional boundary and activates permission mechanisms in two directions at once. The Map of AI places the Permission Layer as the governance architecture above the infrastructure and model stack โ and this story is a live example of why that layer increasingly determines competitive outcomes.
What the reported investigation establishes โ and what it does not โ are two different things, and keeping them separate is the minimum the story requires: an allegation by one company, an inquiry opened by one regulator, no findings, no penalties, no confirmations, and a structural logic that explains exactly why a commercial claim in one jurisdiction became a regulatory question in another within a fortnight, without that sequence saying anything at all about whether the original allegation is true.
Sources: The Information, September 22, 2026; Bloomberg (carrying the same report); Anthropic September 2026 threat report (underlying allegation, unverified). Structural analysis is original to FourWeekMBA / Business Engineer. This article was drafted with a Claude model made by Anthropic, a party to the underlying dispute. No wrongdoing by any company has been established. This is not legal advice.
No wrongdoing has been established by anyone, and this is not legal advice. The matter described above is an investigation that has produced no findings, arising from an allegation made by Anthropic that remains unproven. No confirmation from the Cyberspace Administration of China, DeepSeek or Moonshot is established here, and nothing above treats that absence as meaningful. The reported examination of whether police, military or state-linked data moved describes the scope of an inquiry, not a finding that any such data did move. Anthropic’s alleged volume figures cover different and non-comparable periods, and nothing above derives any rate, ratio or per-day comparison from them. The narrowing of the inquiry is neither evidence against the companies still under examination nor exoneration of those no longer its focus. Nothing above takes any position on any government, on export controls, or on relations between any countries.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
