Xi Jinping’s paranoid approach to AGI, debt crisis, & Politburo politics — Victor Shih

Summary

• China will keep funding AI regardless of its debt hole, but with a kill switch built in. Shih’s read of Ding Xuexiang’s Davos speech: “We need to invest in AI, but we can’t go all out in investing in it without knowing what the brakes are” — for the Party, “developing the brakes is just as important as developing the AI itself.” He is “all but certain” a team inside DeepSeek’s headquarters can “pull the plug,” because such teams exist in every major Chinese internet company.
• Total government debt is pushing 200% of GDP — 60-70% central plus an estimated 120-140% local — and the ~10 trillion RMB local-debt swap authorized late last year is “an accounting exercise” that swaps higher- for lower-yielding paper without reducing the stock. Yet teachers go unpaid four or five months a year while AI, defense, and tech keep getting funded: “the financial system is geared toward fulfilling the priorities of Xi Jinping.”
• The socialist banking system maximizes output, not profit — 1% deposit rates and $20k/year capital controls leave savings in state banks that help finance strategic sectors on bureaucrats’ sign-off. The famous winners — BYD, CATL, DJI, Xiaomi, High-Flyer — were mainly privately funded; state-financed successes like Huawei sit amid “maybe even over a dozen failed cases” per success, including MIIT semiconductor officials jailed for approving bogus projects.
• Succession is the sharpest tail risk in the system. FX reserves of $3T are “less than 5% of the money supply” — if people rationally reallocate just 10% of their assets abroad, “the foreign exchange reserve is gone,” forcing ~20% rates and mass bankruptcy. Financial repression only holds while bureaucrats fear Beijing; if the command structure lapses “even for a week or two, I think we would have a financial crisis, at least.”
• Taiwan: no fixed trigger. Xi has been in power 12 years without invading, so his desire for unification “is not so strong that he would take a very risky gamble” — the invasion threshold is conditional, and Ukraine “likely increased the threshold” by proving confident leaders get bad intelligence.
• Bearish 2040 call: China at only 1x-1.2x the US economy, versus consensus dominance — Shih doesn’t “quite believe in these PPP calculations” and puts real income today at ~70% of the US. No big consumption stimulus is in sight at these debt levels, and Dwarkesh argued that export-led growth eventually “has to conquer Europe… take North America. That’s going to be really tough.”
• DeepSeek is the best tool for reading the CCP — Shih finds it clearly weighted toward policy documents and high-level meetings, likely because High-Flyer trained it to extract policy alpha for Chinese-market trading. He still won’t install it on his phone.

Deep dive

1. China looks decentralized on paper — the money says otherwise

• Dwarkesh’s opening puzzle: local+provincial governments account for ~85% of spending in China vs. 50% in the US — is the authoritarian state actually decentralized? Shih’s answer: it was, from the mid-1970s to mid-1990s, and that fiscal decentralization drove “a very good period of mainly private sector–driven growth.”
• The turn came with the 1994 tax centralization — Beijing, fearing a Soviet-style breakup, grabbed the VAT and eventually all major tax categories, reimbursing provinces conditionally: “If you do this thing, then I’ll give you a little bit of money.” Localities regained autonomy via land sales from 2000-2020, but the center “basically killed the land market in 2022.” Today, “localities in China are highly dependent on the central government.”

2. Real degrees, real experts — but Party preservation trumps everything

• The Politburo’s STEM credentials are mostly genuine — Ma Xingrui from Tsinghua, military-industrialists from top programs — but Shih’s caveat: technical training doesn’t confer governing skill, and STEM-tracked officials “may never learn anything about supply and demand.” The one universal requirement, ratcheted up recently, is government ideology (“Situation and Policy” class).
• His resolution of the competence paradox (Dwarkesh: why scrub airport runways during Zero-COVID?): expertise flows upward through real channels, but the Party’s instinct is worse than self-preservation — it’s power preservation. “Some people die, whatever. We have slower growth for a couple of years. But if we get to preserve power, we will go ahead and preserve power.”
• The vaccine episode is the cleanest evidence: Shih has a paper showing leaders knew mRNA vaccines were better but protected domestic pharma — “Okay, we believe you, but we still want to help our domestic pharmaceutical industry” — while propaganda denigrated Western vaccines. Experts are respected and consulted constantly, then overridden for political reasons; the Paxlovid non-purchase might’ve been a political decision too.

3. Every decision funnels into one man — “the whole thing became Xi Jinping”

• Leading Small Groups killed debate structurally: when the Politburo Standing Committee decided things, members held notionally equal rank and could overturn the Secretary General; in LSGs, everyone below Xi is outranked and “in no scenario will they debate a particular policy with him.” Once decisions moved into LSGs, “the whole thing… became Xi Jinping” — briefing books, policy options, his gut call becoming law.
• The workload is Stalin-like and Shih confirms the parallel: meetings “almost every day, like 270 days out of the year.” Politburo study sessions are “somewhat real” — speakers face an unstaged Q&A (“It’s very scary”) and Xi’s post-lecture remarks become policy.
• Shih’s honest assessment of Xi’s own ability: from internal Party speeches, “he has a really good political nose” for controlling Party and military; on economics “some advisor gives him some talking points”; on tech he cares intensely but only through the lens of beating the US — “he wants to win.” Still, “comparing him to American leaders, especially today, I think he’s probably better prepared.”

4. Ding Xuexiang: the mystery man holding the AGI brakes

• Thus far, the Chinese AI effort appears to sit under Ding Xuexiang — Standing Committee member, head of the Central Science and Technology Commission and of the office running Xi’s cybersecurity commission since 2022. His background: metallurgical forging from “not the MIT of China but maybe closer to the IIT of China,” then decades in Shanghai.
• The trust is a genuine puzzle: Ding worked directly under Xi for only one year in Shanghai, yet was pulled to Beijing in 2013 to run the flows of information across Xi’s desk. Shih’s best guess — he fed Xi intelligence on Jiang Zemin’s Shanghai stronghold — but “what else he’s done for the big boss to earn his trust is a big mystery.”
• Ding’s doctrine, from Davos: “We have to develop the brakes at the same time.” The Party fears an actor — outside or inside the Party — using AGI to usurp its power, so designated humans in every agency and company must be able to “put their foot on the brake.” The trigger Shih names: AGI generating Falun Gong-related content faster than censors can contain it. Despite the forging degree, Shih rates Ding “definitely in the top quartile” of Politburo members you’d want running AI, “if not the top two or three.”

5. How an AGI mobilization would actually run — and why the personnel matter

• Dwarkesh’s setup: computer-use agents might arrive next year, most white-collar work — potentially 40% of the economy — within five years. What institutional signals to watch? Shih doesn’t expect a new AGI leading group — AI stays under the cybersecurity LSG, because “it’s the security aspect of it that’s the most important.” So far the response to DeepSeek has been pure acceleration: Xi met Liang Wenfeng, every big tech company was pushed to adopt it fast.
• But full mobilization (all Huawei GPUs to DeepSeek, land, power) crosses into NDRC territory under Li Qiang or He Lifeng — which could itself justify a dedicated AGI group if Xi won’t let Ding share power. The foreign layer: Gulf-state cloud capacity needs a customer willing to pay billions — China. Dwarkesh’s label: “Belt and Road v2.”
• Why leadership matters: some trusted figures are demonstrably bad — He Lifeng, now chief US trade negotiator, “is well known for starting and perpetuating the largest real estate bubble the world has ever seen in Tianjin,” a “New Manhattan” of empty office towers. And Shih flags a structural drag: content paranoia means even well-aligned AI will get humans “double- and triple-checking the content” — “they’re just so afraid of subversive content getting out.”
• Shih’s own tooling tip: DeepSeek beats Grok and Baidu for tracking CCP policy — it surfaces high-quality policy documents and meetings where Chinese social-media-trained models return chatter, almost certainly because High-Flyer trained it to mine policy for trading alpha. “I don’t dare install it on my phone.”

6. Xi’s rise was the Stalin playbook, run patiently for 30 years

• From the Cultural Revolution, two lessons were possible; Xi drew the darker one: “Just don’t be on the losing side. In any political struggle, make sure you’re on the winning side. Because if you’re on the winning side, then you can do terrible things to your enemies.”
• The evidence of strategic patience: in Fujian in the late 1980s-90s Xi spent “an inordinate amount of time” cultivating military officers — built them a dormitory, joined an anti-aircraft regiment — and 30 years later many of those people were generals he later promoted (and, lately, has begun purging). Like Stalin, he stayed low-key, took peripheral postings while other princelings fought over Beijing, then formed serial coalitions — convincing Hu Jintao to help purge security czar Zhou Yongkang, as Stalin did Trotsky — until absolute power. The Trotsky analogue: Bo Xilai, the flamboyant rival, now imprisoned.

7. The debt mountain was built by kickbacks, not KPIs

• The numbers: central debt 60-70% of GDP is only “sort of” low because Beijing pushes obligations down while authorizing local issuance — Shih estimates local debt at 120-140% of GDP, total pushing 200%. The debt bought high-speed rail and, lately, industrial policy: AI science parks and local seed funds doing venture deals with borrowed money.
• Why returns collapsed: infrastructure was genuinely productive in the 1980s-90s, but “once you’ve built the first high-speed rail between Beijing and Shanghai… there’s very rapidly diminishing returns” — especially with populations shrinking in northeastern and southwestern cities.
• Shih’s contrarian mechanism for why officials kept building: not GDP-for-promotion but rent-seeking. A $1B light-rail project yields ~$100M in contractor kickbacks, of which $50M goes to your superior to buy promotion. James Kung’s work shows selling land cheaply to princelings statistically raises promotion odds.
• Dwarkesh’s Robert Moses riff, which Shih buys: China’s corrupt equilibrium makes every faction want the project finished, while America’s regulatory stakeholders profit by prolonging — “I’m going to get a little consulting fee to slow this down by five years.” In China the party secretary “can cut through all the red tape… as long as he can benefit himself somehow.”

8. Socialist finance: maximize output, never mind profit

• Shih’s deepest frame: “socialism only cares about output… capitalism wants to maximize profit.” Chinese companies are profit-seeking, but “because the financial system is socialist, they’re forced into socialist-like behavior” — a state bank can finance a strategic-sector firm that “never, ever” makes money, so long as it produces the thing.
• The funding base is financial repression as a tax on savers: 1% deposit rates, $20k/year outflow limits, leaving money in banks that finance industrial policy. The lived result of decades of 5%+ growth: migrant workers barely getting by, rising homelessness, abysmal elderly care — “the part about truly helping people live a good life, they’re not making as much progress as they should.”
• On rebalancing, Shih rejects Dwarkesh’s fix of simply ending repression: deposits are concentrated in the top 10% of households, so higher rates help only net-savers; the median household will “just save like crazy in anticipation of getting sick” without better welfare. His actual solution — cut defense (“China already has the largest navy in the world. Why does it need a navy that’s even bigger?”) and many industrial subsidies, fund welfare — is one Beijing “will never, ever in a million years” adopt while it prioritizes US competition and supply-chain dominance.

9. Taiwan: a moving threshold, not a countdown

• Shih’s revealed-preference argument: Xi sincerely wants unification but “if he wanted it so badly that he’d do whatever it takes… he would’ve done it already” — 12 years in power without invading. Dwarkesh’s pushback: the stockpiling of oil and grain, clean-tech self-sufficiency, and amphibious buildup look like preparing to cross a threshold that couldn’t have been crossed in 2015.
• Shih says the threshold is conditional, not fixed: Ukraine “likely increased the threshold” — Putin was confident, got bad intelligence, and failed, a possibility Xi “cannot discount.” Other exogenous factors could lower it.

10. No succession plan — and the transition is priced as a financial event

• The mechanism Shih spells out: repression works because the SAFE bureaucrat approving a fake billion-dollar Rolex invoice fears “I’ll be in jail in a week.” Remove the fear — “there’s nobody in Beijing… this guy promises me $100 million” — and outflows explode. With reserves under 5% of money supply, a rational 10% reallocation abroad exhausts them, forcing ~20% rates and mass bankruptcy. “If there’s any sense that there’s no command structure in Beijing, even for a week or two, I think we would have a financial crisis, at least.”
• Naming a successor “is going to die for sure, in a horrible way” (Stalin, Mao); the workable model is a transitional figure who can’t become number one — Mao’s Jiang Qing, and now possibly Xi’s wife or increasingly high-profile daughter, barred by Party sexism and lack of official posts. But unlike 1976’s Long March cohort, today’s Politburo lacks mutual trust — by Xi’s design — so this transition will be “more ruthless, more brutal, and potentially more disruptive.” One stabilizer: military-industrialist statists who, like post-Stalin Soviet vested interests, back whoever subsidizes them.
• The closing forecast, notably bearish: absent AI, China in 2040 is 1x-1.2x the US economy — Shih doesn’t believe PPP adjustments and puts real income today at ~70%. Debt leaves no big consumption stimulus in sight, leaving supply-side doubling-down into a world that will push back: “is the world really going to switch to 100% buying BYD?”