Summary

China’s investability hinges less on formal rules than on a hierarchy in which Xi Jinping’s signals cascade through regulators that businesses almost never challenge. Zhang’s compact formulation—“signal is policy, and policy is signal”—explains why bureaucratic discretion can coexist with sudden, system-wide turns. For equity holders, the same execution capacity that builds infrastructure quickly also creates unusually concentrated policy risk.
The 2020 Ant Group shock was less a personality feud than a collision between private financial scale and Communist Party control. Ant sought a $320 billion valuation—above JPMorgan at the time—while resisting classification as a financial institution; its IPO was stopped “at the eleventh hour,” followed by an 18-month crackdown. Zhang says the campaign wiped trillions from Chinese tech valuations, drove investors away, and damaged “animal spirits,” even if the systemic-risk concerns were legitimate.
Beijing’s recent embrace of entrepreneurs is a rotation toward strategic hard tech, not permission to rebuild Ant-style financial empires. Xi’s table featured leaders from AI, EVs, clean energy, robotics, and Huawei; consumer platforms remain useful for employment, but semiconductors are the critical bottleneck. The policy signal is sectoral: self-sufficiency and manufacturing capacity outrank platform expansion.
DeepSeek supports Zhang’s “one to ten” thesis: China may lag at basic breakthroughs yet excel at adoption, cost compression, and diffusion. It started a domestic AI price war in mid-2024 and was dubbed the “Pinduoduo of AI” months before the West noticed; she presents it as a largely independent private-sector success, not evidence of a state-run Manhattan Project. Erik notes that Jeff Ding’s research reaches the opposite conclusion on China’s diffusion strength.
Zhang expects export controls to accelerate, not prevent, a Chinese semiconductor breakthrough by aligning private incentives with state resources. With 1.4 billion people, returning engineers, and restricted access to advanced foreign technology, she predicts “another DeepSeek moment” within a few years and thinks the tech war could then end. Erik remains unconvinced: his stated view is that sustained controls have a good chance of mattering.
DeepSeek also exposed a US incentive problem: upstream AI monopolies earn from compute scarcity, while efficient downstream models threaten that economics. Zhang points to Nvidia’s above-60% operating margin, four dominant cloud providers, and loss-making model labs to argue that concentration “blinded [the US] to the direction of innovation.” Her claim is not that America lacks talent, but that its market structure rewards scaling demand more than lowering cost.
Neither country is governing frontier AI as though catastrophe is a live risk, and Zhang doubts cooperation will become serious before a contained disaster forces attention. Her “high-wire” pattern is late top-level intervention after harms become undeniable; the United States is simultaneously relaxing oversight. She proposes an IPCC-like scientific body to share capability evidence, arguing that even a 1% or 0.1% disaster probability justifies insurance.

Deep dive

1. China’s visible strengths coexist with deep economic pessimism

Zhang’s research combines an insider’s background—born, educated, and briefly practicing in China—with seven years at the University of Hong Kong, industry relationships, court-case research, and secondary evidence. Her epistemic warning is categorical: “It is impossible to truly understand China without being there,” because “you don’t get it from an interview.”
China understands the United States better than the reverse, in Zhang’s view: Chinese audiences consume extensive US news, while China’s opaque bureaucracy, censorship, and information controls obstruct outsiders. The knowledge is still biased—state media emphasizes bad news abroad—but the information asymmetry remains substantial.
Her macro balance is deliberately uncomfortable: China is “not ten feet tall,” its economy is at “rock bottom,” and young people have a very pessimistic outlook. Yet she contrasts California’s incomplete high-speed rail effort with a San Francisco–Los Angeles line she believes China could build in one or two years: “China speed” remains a real execution advantage.
Communist Party legitimacy depends heavily on performance—prosperity and social stability—rather than elections. Zhang says many ordinary Chinese accept ubiquitous cameras because they feel safer, while zero-COVID initially drove very high approval as China resumed production before other countries. Sentiment reversed after Omicron, when the rest of the world reopened and China maintained severe lockdowns.

2. Xi sits atop a pyramid governed by signals

Zhang’s dynamic-pyramid model has four levels: top leaders, regulators and bureaucrats, regulated firms, and platform participants such as merchants, delivery workers, and consumers. Similar actors exist elsewhere, but their relative power and accountability differ radically.
Nathan asks whether “top leadership” really means roughly 24 Politburo members rather than a broader elite. Zhang compresses it further: “You can just think about it as 1 person: the president,” because officials ultimately recognize one “big boss,” even though Xi must delegate implementation throughout 31 provincial entities and numerous central agencies.
Bureaucrats possess expertise and discretion, but continuously watch for changing priorities from above. Zhang’s formulation captures the mechanism: “Signal is policy, and policy is signal.” Agencies then frame their preferred enforcement actions around those signals, mixing top-down alignment with their own bureaucratic incentives.
Firms sit below agencies because they almost never confront the state directly; platform users sit lower still because they bargain weakly against platforms and face censorship when organizing complaints. Regulators answer primarily upward, not to the affected public.

3. Government dominates in China, courts mediate in Europe, and business leads in America

The clearest Chinese power test is corporate behavior. Zhang notes that Elon Musk “behaved like a sheep” in China, courting officials rather than sending an agency the kind of insulting emoji he might use in America: “He understands in China you play by the Chinese rules.”
Europe occupies the middle of Zhang’s spectrum. US technology companies routinely challenge European Commission decisions in Luxembourg; courts often defer to regulators, but firms sometimes win, and European officials accept litigation as a normal feature of the rule of law rather than disloyalty.
America is the opposite extreme: years of hearings produced no comprehensive federal privacy law or broad legislation disciplining domestic tech giants, while major measures targeted TikTok or sensitive data transfers involving countries such as China. Watching technology CEOs surround the president at inauguration, Zhang asks whether the US has developed its own “crony capitalism.”

4. Censorship doubles as a legitimacy-production system

Chinese tech firms and the government invest heavily in content moderation. Social-media platforms receive timely instructions from the Cyberspace Administration of China about permissible content and alignment with Communist Party values. The machinery’s internal operation remains opaque even as it forms a core compliance function.
Complaints from consumers and workers are normally muted, but authorities can selectively amplify them when public anger helps legitimize enforcement. Media organizations, propaganda departments, and regulators then operate “in tandem,” turning a technical intervention into a publicly intelligible moral narrative.
Didi proceeded with a US IPO despite CAC pressure to delay while data-security concerns were assessed. Two days after listing, CAC launched a cybersecurity review and Didi’s shares fell more than 10%. Zhang calls the legal fit dubious: procurement-oriented review rules were repurposed to address information submitted during a US listing.
Media reports then depicted Didi as a traitor capable of revealing sensitive traffic patterns around ministries, while its president was accused of selling Chinese secrets. Around Ant’s aborted IPO, the same system portrayed Jack Ma as a greedy businessman and Ant as a company poised to exploit consumers—claims that made intervention appear protective rather than arbitrary.

5. China’s tech champions won brutal contests inside a protected market

The Great Firewall created barriers against foreign social-media and search businesses: Google exited, while Bing retained only about 1% of Chinese search. That shelter reduced external competition for firms such as Baidu, but Zhang still believes domestic champions would have emerged without it.
Nathan tests the distinction directly: Alibaba, Tencent, and their peers benefited from excluding US competitors, but Chinese rivals still competed for each category “fair and square.” Zhang strongly agrees—the domestic market itself was exceptionally competitive rather than populated by state-designated winners.
Her innovation taxonomy separates foundational invention from commercialization. China is weaker at “zero to one”: DeepSeek did not discover transformers or neural-network architecture. It is much stronger from “one to ten”—adopting a breakthrough, improving efficiency, lowering cost, and diffusing it to consumers.
DeepSeek began a domestic AI price war around mid-2024, slashing prices until larger players followed; Chinese observers called it the “Pinduoduo of AI.” The West noticed only eight or nine months later. Zhang sees the same cost-and-delivery machine in Temu, which she found roughly 30% cheaper than Amazon, and in Shein’s expansion abroad.

6. Ant Group turned regulatory neglect into an 18-month crackdown

From the early 2010s until late 2020, consumer technology enjoyed a decade-long honeymoon with light-touch regulation. Zhang argues that neglect allowed excessive concentration and unresolved harms; after Ant’s IPO crisis, the pendulum swung into an intense crackdown lasting roughly 18 months, through early 2022.
Jack Ma’s critical speech contributed to the confrontation, but Zhang treats it as secondary to tensions between Ant and financial regulators. Ant had expanded from payments into other financial services while insisting it was not a financial institution, even changing its name from Ant Financial to Ant Group and presenting itself as “tech-fin,” not fintech.
Its proposed $320 billion valuation—higher than JPMorgan’s at the time—made regulators see a bubble and potential systemic risk. Zhang also emphasizes political control: Ma had controlled more than 50% of the voting rights in Ant, leaving one private entrepreneur with extraordinary influence over what had become the world’s largest fintech business.
The state wanted both intervention against potential systemic risk and to “remind them who the big boss was.” But the crackdown, Zhang says, wiped trillions from technology valuations, injected little new competition, drove private investors away, and sabotaged entrepreneurial confidence. She also corrects Nathan’s shorthand: Jack Ma was not under house arrest; she had seen him recently.

7. Beijing’s thaw is reserved for industries that strengthen the state

The recent entrepreneur symposium did not signal renewed enthusiasm for lightly regulated private finance. Zhang reads Xi’s guest list as the message: the DeepSeek founder, Huawei’s leader, and entrepreneurs in AI, EVs, clean energy, and robotics occupied the strategic center.
Consumer platforms may continue selling games and services because China “desperately” needs employment for young people. But Zhang distinguishes job support from transformation: hard technology, especially semiconductors, is expected to create durable competitive advantage and reduce dependence on foreign suppliers.
Xi’s larger project is to move China from an “old and dirty economy” and property bubble toward advanced manufacturing and clean-energy exports. Zhang cites a Bloomberg assessment that most Made in China 2025 targets were achieved; US pressure has only intensified Beijing’s conviction that China must become self-sufficient.

8. DeepSeek’s breakthrough came from private “animal spirits”

On Nathan’s question about top-level clearance or government monitors, Zhang offers no evidence of prior sign-off. Her answer is that DeepSeek developed products “quite independently,” although every Chinese business must follow state rules and a player this important now needs close communication with officials.
Zhang’s deliberately harsh assessment is that “anything the Chinese government touches, it’s not doing that well.” State-directed semiconductor investment produced corruption scandals, inefficiency, and limited progress; the government still relies on Huawei, itself a private firm, because real innovative capacity comes from entrepreneurial “animal spirits.”
DeepSeek was a maverick rather than an established national champion: a quantitative hedge-fund business, constrained in parts of its trading operation by regulation, redirected resources into AI training and became an overnight success. Alibaba, Tencent, other major platforms, and AI startups also improved rapidly, even if DeepSeek led.
Open sourcing served several purposes in Zhang’s account: it provided a global public good, increased DeepSeek’s credibility, and gave Chinese models a chance to shape industry standards as China gained a reputation as the second-most-important AI superpower. She characterizes those intentions as benign while acknowledging that Beijing will seek to guide the technology toward national interests.

9. Export controls may align the forces Washington wants to separate

Hardware remains China’s chief AI bottleneck, but Zhang thinks stricter US controls could be self-defeating. She cites calls such as Dario Amodei’s to “double down” with smarter chip controls, arguing that this school discounts how restrictions change Chinese private-sector incentives.
Once advanced foreign technology becomes inaccessible, companies must innovate domestically, cooperate with government, and search for substitutes. Export controls therefore create an alignment between public objectives and private survival.
Zhang stresses China’s 1.4 billion people and large engineering base, including engineers who returned after US anti-espionage investigations under the China Initiative. She quotes Huawei founder Ren Zhengfei’s message to Xi: “Even if we lose everything else, as long as we have people, we can innovate.”
Her conditional forecast is bold: within a few years, a semiconductor breakthrough will produce another “holy cow” DeepSeek moment, after which the tech war ends because the largest technological obstacle has fallen. Erik explicitly reserves judgment—he thinks sustained controls may still matter and notes that Jeff Ding treats diffusion as a Chinese weakness, not a strength.

10. US–China hostility has become an institutional feedback loop

Zhang separates demand from supply. The demand for tougher policy came from Chinese imports, lost US manufacturing jobs, trade imbalance, and China’s geopolitical rise challenging US primacy. But institutions and leaders determined how that discontent became policy.
Trump supplied the pivotal turn after 2016, especially from 2018: Section 301 tariffs, capital restrictions, sanctions against firms such as Huawei and ZTE, the proposed TikTok ban, and the China Initiative’s investigations of ethnic-Chinese scientists. His rhetoric—calling Chinese students spies and COVID the “China virus”—also enlarged the constituency for confrontation.
Biden inherited nearly all of that architecture and intensified export controls. Once established, the agencies and restrictions developed “long-lasting impacts”; even Trump’s shift from TikTok antagonist to its “biggest savior” illustrates the volatility of a system built around executive discretion.
In trying to contain China, Zhang argues, America has become more China-like: concentrated executive power, abrupt policy swings, and fragile outcomes with unintended consequences. Her broader judgment cuts both ways: China’s biggest threat is “China messing itself up,” while America’s is damage to its own democracy and governance.

11. DeepSeek exposed an American market-structure failure

Zhang turns Nathan’s question back on Silicon Valley: if US companies had greater resources, technological leadership, and knowledge of Chinese efficiency gains, why did none pursue the same low-cost trajectory? “It was a deliberate business choice,” not evidence that American engineers lacked capability.
Her answer starts upstream: Nvidia is a virtual monopoly with an operating margin above 60%, while four large cloud providers dominate infrastructure and invest downstream. Model companies, including OpenAI, were characterized as loss-making. “The money is all upstream,” where scarcity and expanding compute demand support the profit pool.
Those firms therefore benefit from projecting enormous future demand for chips and cloud resources, not from proving that comparable models require much less compute. DeepSeek’s lesson, in Zhang’s framing, is that US concentration “blinded us to the direction of innovation”—a failure of incentives rather than science.

12. Serious AI cooperation may require a disaster first

Asked whether Xi might privately be an AI “doomer,” Zhang’s honest answer is: “I have no idea what President Xi is thinking.” Her inference is only behavioral—China and the United States are pressing development “at full speed,” which would be irrational if their leaders believed harms clearly outweighed benefits.
China’s dynamic pyramid is particularly risky for AI because problems often receive top-level attention only after becoming severe; Zhang lists pandemic control, technology regulation, and the property crisis as recurring examples. Meanwhile, the United States is moving toward the opposite regulatory extreme despite remaining the frontier leader.
Her bleak assessment of track-two dialogue is that “what the world needs is a disaster”—hopefully one contained enough not to “kill us all,” but serious enough to make dangerous capabilities politically undeniable. Nuclear rivals negotiated because they understood nuclear destruction; leaders still do not understand what AI systems may become capable of.
Zhang’s practical proposal is an independent international scientific institution modeled on the IPCC, sharing evidence about AI capabilities and updating governments on emerging hazards. Since “we don’t understand AI,” even a 1% or 0.1% chance of loss of control warrants modest precaution: “That’s the whole point of buying insurance.”

Tau Language: The Software Synthesis Future [Sponsored] - Ohad Asor Joseph Henrich — Humans defeated smarter species with cultural evolution