The Spinoffs

This installment is about what happens when the arrow reverses.

For a long time, the corridor was a one-way pipeline. China imported leadership credentials from American graduate schools, imported scientists from American laboratories, imported engineering talent from American joint-institute partnerships, and ran covert reconnaissance against the foreign entities its state most needed to surveil. The product, in all five cases, was capacity returning home to Beijing.

That phase is ending. Tsinghua University’s laboratories, principally one called the Knowledge Engineering Group, have now spawned a generation of Chinese artificial-intelligence companies that are the rival, on at least some technical benchmarks, of the American ones. The corridor has stopped only absorbing. It now exports: models, talent, and on at least one prominent occasion, a company that the United States Department of Commerce has formally accused of advancing the People’s Liberation Army.

The Knowledge Engineering Group

Inside Tsinghua University’s Department of Computer Science and Technology sits a research laboratory called the Knowledge Engineering Group, abbreviated KEG. By the standard description, it is an academic research group focused on knowledge graphs, natural language processing, and social-network analysis. By a less standard description, it is one of the most prolific commercial-spinoff laboratories in the entire Chinese science establishment.

The KEG laboratory was, for years, directed by Professor Tang Jie, who later co-founded the most consequential spinoff. The current directorship is held by Professor Li Juanzi, who was Tang’s closest collaborator and co-founder. Together they trained generations of Chinese artificial-intelligence researchers, including several of the founders of the rival Chinese AI startups now competing on the same frontier.¹

The most important academic credential Professor Tang holds is, for our purposes here, that he was the doctoral advisor of a man named Yang Zhilin — a former Tsinghua undergraduate who went on to a PhD at Carnegie Mellon University, then founded a separate AI company called Moonshot. We will return to him in a moment.²

What the KEG laboratory has produced, beyond academic papers and trained PhDs, is the corporate seed corn of the Chinese large-language-model industry. The KEG lab incubated at least three named companies that have, by 2026, become significant commercial entities: Zhipu AI, DeepLang, and through the academic-lineage thread of its alumni, the founding of Moonshot AI.³ Of the six companies that financial analysts and the Chinese technology press collectively call the Six Tigers — the leading large-model startups in China — two come from this single laboratory at Tsinghua.

Zhipu AI

Zhipu was incorporated on June 11, 2019. The founders were Tang Jie and Li Juanzi. The address was the Tsinghua University Science Park in Beijing’s Zhongguancun district — the same science park, you may recall from the Interlude, that houses Tsinghua Holdings Zijing Education, the corporate cutout that operates American university partnerships without triggering disclosure requirements. Different floor, same building.⁴

For the first eighteen months, the company struggled. The local administrative commission of Zhongguancun Science Park provided rent-free office space to keep the lights on. The product was knowledge graphs — useful technology, modest business. Then, in 2020, the founders made the call that defined the company. They committed to building a Chinese-language large language model from scratch, on the premise that the existing English-trained models were inadequate for the linguistic structure of Chinese.

They named the architecture GLM, for General Language Model. In August 2022, four months before the public release of ChatGPT, Zhipu released GLM-130B, a one-hundred-and-thirty-billion-parameter bilingual model, as open source — from a GitHub organization called THUDM, the abbreviation of Tsinghua University Department of Mining... actually, Tsinghua University Data Mining, which is the KEG laboratory’s public-facing technical handle. In 2023 came ChatGLM, the conversational variant. In 2024, GLM-4 and GLM-4-Plus. By the end of 2024, Zhipu had eight hundred employees, a multi-billion-dollar valuation, and the description that recurred in every Chinese-press profile: China’s OpenAI.⁵

The corporate name was Beijing Zhipu Huazhang Technology Company. Its commercial brand outside China became Z.ai after a July 2025 rebrand. Its investors include the four largest Chinese internet companies — Alibaba, Tencent, and the rest of the so-called “Four Dragons” — plus state-backed funds.⁵

On the technical level, Zhipu is what a healthy company that came out of a healthy university laboratory should look like. The papers are public. The benchmarks are competitive. The open-source release of GLM-130B was, in the open-source AI community, treated as a serious contribution. None of what follows takes any of that away.

What follows is what the United States Department of Commerce decided about the company on January 15, 2025.

The Entity List Designation

On January 15, 2025 — five days before the inauguration of the second Trump administration, in the final regulatory actions of the Biden Bureau of Industry and Security — the Department of Commerce placed Zhipu AI and nine of its subsidiaries on the Entity List.⁶

The Federal Register notice, published the following day, is the kind of document that rewards reading in its original language. Volume 90 of the Federal Register, page 4617, the section captioned “Addition of Entities to and Revision of Entry on the Entity List.” The relevant paragraph is short enough to quote in full:

The ERC determined to add Beijing Zhipu Huazhang Technology Co., Ltd.; Beijing Lingxin Intelligent Technology Co., Ltd.; Beijing Yuanyin Intelligent Technology Co., Ltd.; Beijing Zhipu Future Technology Co., Ltd.; Beijing Zhipu Linghang Technology Co., Ltd.; Beijing Zhipu Qingyan Technology Co., Ltd.; Hangzhou Zhipu Huazhang Technology Co., Ltd.; Nanjing Zhihu Information Technology Co., Ltd.; Shanghai Zhipu Huanyu Technology Co., Ltd.; and Shenzhen Zhipu Future Technology Co., Ltd., to the Entity List, under the destination of China. These additions are being made because these entities advance the People’s Republic of China’s military modernization through the development and integration of advanced artificial intelligence research. — Federal Register 90 FR 4617, January 16, 2025

The legal basis cited is Section 744.11 of the Export Administration Regulations, the section that permits the Commerce Department to add entities whose activities are “contrary to the national security and foreign policy interests of the United States.” The standard is not criminal conviction. It is administrative finding by the Entity List End-User Review Committee, which combines representatives from the Departments of Commerce, State, Defense, Energy, and the Treasury.

Read what the United States government actually said. These entities advance the People’s Republic of China’s military modernization through the development and integration of advanced artificial intelligence research. Not might advance. Not could be misused to advance. Advance. Present tense. Indicative mood.

And what makes the Zhipu listing different from prior Entity List designations involving Chinese AI is the category. Zhipu was, on the public record, the first Chinese large language model company placed on the Entity List. The earlier listings — SenseTime, Megvii, Yitu, Hikvision, Dahua, Iflytek, Meiya Pico, all designated in October 2019 — were facial-recognition and surveillance-camera companies, designated specifically for their role in the Xinjiang surveillance apparatus.⁷ Zhipu is something new. It is a frontier-model company that the United States government has assessed is integrating its research into Chinese military modernization.

What Zhipu Said Back

Within hours of the designation, Zhipu issued a public response. The English-language statement, posted on the company’s social channels and reproduced widely in the Chinese tech press, included the following sentence, which is the sentence that matters for this installment:

Zhipu originated from the achievements of Tsinghua University. As one of China’s earliest large model companies, we began developing the GLM pretraining framework in 2020. — Beijing Zhipu Huazhang Technology Co., Ltd., January 15, 2025

The company did not contest the genealogy. It led with it. Zhipu’s response to being named by the United States government as an entity advancing the People’s Liberation Army was, in its first paragraph, a public statement of its Tsinghua origin. That is not the rhetoric of a company trying to distance itself from its university lineage. That is the rhetoric of a company that understands its university lineage is its primary credential.⁸

The same statement said the designation “lacks a factual basis” and that the company “strongly opposed” it. That is the part of the response the headlines carried. The genealogy disclosure was, for our purposes here, the more interesting part.

The Fleet

Zhipu, however, is not the only ship the harbor has launched. The Knowledge Engineering Group has been productive in a way that complicates any simple “Zhipu is one company” reading of the spinoff story.

In March 2023, four men sat down in Beijing to incorporate a different artificial-intelligence company. Their names were Yang Zhilin, Zhang Yutao, Zhou Xinyu, and Wu Yuxin. All four were Tsinghua alumni. Yang — whose Tsinghua doctoral advisor had been Tang Jie, the KEG laboratory’s long-time director and Zhipu co-founder — held a Carnegie Mellon PhD and had been a researcher at both Google and Meta. The company they founded was called Moonshot, and the Chinese name, Yue Zhi An Mian, translates to Dark Side of the Moon. The product is a chatbot named Kimi.⁹

Moonshot is not on the Entity List. Moonshot is not, on the documentary public record I have been able to find, accused by the United States government of advancing Chinese military modernization. The point is something different and harder to dismiss. The point is that the same Tsinghua research laboratory whose director founded the company the United States named as a military-modernization vector trained the doctoral student who founded a separate frontier-AI company at sufficient commercial scale to raise three and a half billion dollars at a twenty-billion-dollar valuation.¹⁰

The KEG laboratory is not a one-shot fluke. It is a recurring industrial process. The same lab. The same advisor lineage. Multiple commercial spinoffs, each at frontier-model scale, each backed by the dominant Chinese internet companies, each operating in a national-security-relevant technology category. This is what a flagship-university-as-strategic-asset looks like in practice, and it is what the architectural map of Parts I through V was, all along, going to deliver. The corridor has reached its commercial output stage.

One laboratory. Three companies. Two of them at the frontier of a technology the United States government has placed under export control on national-security grounds.

The Public Listing

And then, in January 2026, came the move that should have ended any remaining ambiguity. Zhipu — the company on the United States Entity List for advancing the People’s Liberation Army — took itself public.

The legal entity name on the listing application was Knowledge Atlas Technology Joint Stock Company, Limited. The trading venue was the Hong Kong Stock Exchange. The ticker symbol was 2513. The transaction was, on the public record, the first initial public offering by a Chinese large-language-model foundation-model company in history.¹¹

Read that twice. The company that the United States Department of Commerce, in the Federal Register of January 16, 2025, accused by name of advancing Chinese military modernization through artificial-intelligence research, listed publicly on a major international stock exchange less than twelve months later. International institutional investors, including the many funds whose portfolios pass through MSCI indexes and Hong Kong-listed exchange-traded products, can today buy shares of Knowledge Atlas Technology in a routine brokerage transaction.

The Entity List restricts exports of American technology to the listed entity. It does not, on its own, restrict American capital flowing into the listed entity through the public markets. That is a different regulatory authority entirely — the Treasury Department’s outbound-investment review process under the 2023 executive order, which is itself, under the second Trump administration’s America First Trade Policy memorandum, under continuing review.¹² The gap between the two authorities is wide, and the Knowledge Atlas listing has been the largest single demonstration of that gap to date.

Restrict the technology going out. Leave the capital coming in. The corridor stopped being only an importer of American capacity at some point in the last five years. It is now, at least at the level of the foundation-model startups, an exporter of Chinese securities into American investor portfolios. The same name is on both ends of the trade.

What the Trade Floor Looks Like Now

The political context, as of the publication of this installment, is in motion in a way that earlier installments of this series did not have to navigate.

The Zhipu Entity List designation was a Biden-administration action in its final week. The picture has changed materially every quarter since.

The Zhipu listing remains in force. The technology controls associated with it remain in force.

The architecture around it — the rules that determine which subsidiaries are also restricted, which capital flows are reviewed, which licenses are granted — is now a moving picture rather than a static legal regime.¹³¹⁴

A reader of this series in May of 2026 should not assume that the regulatory environment will look the same in November.

The Shape of the Corridor, After Six

Five installments of architecture. One installment of output. The architectural section described what the corridor is built to do. This installment describes what it now actually does, in the present-tense indicative mood of commercial transactions and Federal Register designations.

One institution. One machine. The same university that produces the Chinese state’s political leadership and hosts its weapons laboratories and absorbs its returning scientists and anchors its joint institutes and operates its academic-network backbone is also, as of January 2025, the genealogical origin of the company the United States government has named as a People’s Liberation Army modernization vector.

The same military-civil fusion architecture documented in Part II now operates in plain commercial sight. A Tsinghua research laboratory spawns frontier artificial-intelligence companies that the United States government itself has assessed are integrating their work into People’s Liberation Army modernization. The doctrine the second installment described as state policy is the doctrine the sixth installment finds running as a publicly listed company.

The same five lanes of architecture from the earlier installments are not separate stories. They are the upstream infrastructure of an industrial-scale exporter that has, in the months since Part V went to readers, raised its initial public offering on the Hong Kong Stock Exchange, watched its commercial peer Moonshot triple in valuation, and continued to publish frontier-model research at a cadence that competes seriously with the American laboratories.

Two installments remain. Part Seven turns to the statutory architecture of the Chinese state that makes this commercial output possible in the form it takes — the laws, the party-state administrative organs, the National Intelligence Law of 2017 and the Military-Civil Fusion Strategy of 2014 and the Data Security Law of 2021 that make every Chinese AI company a partner-of-last-resort to the Chinese state, whether it wishes to be or not. Part Eight is the reckoning — the installment that names the American institutional decisions still in front of us, including the question of how the United States holds its own boards accountable for relationships with an institution that has now been documented across seven prior installments to be acting as it does.

The corridor is now exporting. The remaining question is who, at the receiving end, is still pretending it is not.