How Beijing Governs AI, Diaspora Networks, and Retirement

Opening

Welcome to today's China report.

This week is about how Beijing governs under three pressures: scarce computing power, political work beyond its borders, and a population growing older. All three reports appeared within one week in August twenty twenty-six. Only three reports qualified, fewer than usual, and the artificial-intelligence and retirement pieces sit at the edge of domestic politics; they earn their place because they show the Party-state managing technological and demographic pressure, not because they make claims about elite politics.

So this will be a shorter crop, taken one report at a time. The common question is practical: when does a stated priority become a working system?

Jamestown on open-weight AI

Sunny Cheung and Shijie Wang, writing for the Jamestown Foundation, say Beijing is preparing tiered governance for open-weight artificial-intelligence models. Basic models would remain freely available. Frontier models would face security review and export controls. The most sensitive weights would stay inside China or circulate only within a controlled group.

Their explanation starts with constraint. One Chinese founder says China has about five percent of United States computing power and trails by two years. Jamestown also cites a United States government commission’s estimate that four big American technology firms spent at least three hundred and fifty billion dollars on artificial intelligence in twenty twenty-five, against less than forty billion dollars for Chinese peers. If Beijing can’t win by piling up chips, open weights let developers worldwide improve Chinese models and give domestic chipmakers something fixed to optimise around. Government, financial and military users can also run them offline, beyond a foreign provider’s switch.

The evidence for that strategy is mixed but real: official plans and speeches, legal scholarship, company executives, one author interview with a senior machine-learning engineer, model uptake figures, and a Reuters report that the commerce ministry discussed advanced-model limits with Alibaba, ByteDance and Zhipu.

But tiered governance rests on less. The three-tier scheme comes from a Nankai University scholar at a Supreme People’s Court roundtable. Other scholars endorse classification, and state television popularised it. No rule has been issued. Jamestown says legislation is likely, but gives no date. So the forecast is hard to catch out. A real test would be a binding process that treats basic, frontier and national-security models differently. Continued frontier releases without it would cut against the claim.

There’s another leap. Chinese models are spreading, but the report doesn’t show that users then adopt Chinese technical standards, the geopolitical payoff it claims. Downloads, derivatives and Singapore building on Qwen establish uptake, not rule-setting. The report needs named standards, adopters and decisions linking model choice to governance choice. Until then, tiered regulation is a plausible direction, not a demonstrated system.

Jamestown on overseas Chinese affairs

Arran Hope’s Jamestown report reads two July conferences as the opening of a more coordinated phase in overseas Chinese affairs. One brought more than four hundred association leaders from one hundred and forty countries and regions to Beijing. The other was the first national overseas-affairs work conference since twenty seventeen, and the first since the twenty eighteen reforms put overseas Chinese affairs offices under the Communist Party’s United Front Work Department.

Senior sponsorship signals priority: Politburo Standing Committee member Wang Huning and United Front chief Li Ganjie attended both events, carrying Xi Jinping’s instructions. Party rules call the work political guidance. Conference coverage ties it to business, talent, propaganda, scientific self-reliance and Taiwan unification. Discussions about weak grassroots staffing, standardising associations, attracting younger people and offering commercial incentives supply the machinery. Hope argues that political direction from Beijing will become more organised action abroad.

The sources reveal intent but not effect. They’re overwhelmingly Party regulations, official-media readouts, institutional commentary and remarks from invited association leaders. They tell us what Beijing wants. They can’t show whether associations comply after delegates fly home, how many people they reach, or whether staffing and money rise.

Hope predicts aggressive expansion in the coming months and years. But there’s no clock, budget or baseline. A testable version would be growth in United Front staffing, association coordination, funded programmes and pro-Beijing political activity after July twenty twenty-six.

Here’s the report-specific problem. Its executive summary says Beijing seeks control over sixty million citizens outside China. The body calls them overseas Chinese, a Party count mixing citizens with foreign citizens of Chinese descent. That isn’t a small wording slip. Beijing’s effort depends on blurring exactly that boundary, while the report risks repeating it.

I checked this against Singapore’s Institute of Southeast Asian Studies, now the ISEAS–Yusof Ishak Institute. In March twenty twenty-five, historian Mara Yue Du found that Xi’s language does blur those categories, but political responses run from cooperation to rejection and recruitment has had limited success. That doesn’t make the conferences harmless. It shows why four hundred selected leaders, including a French reunification campaigner, can’t stand in for sixty million people. Jamestown proves an expanded Party bid for influence, not expanded control. Host-country activity records, association finances and interviews with members rather than invited leaders would settle the difference.

RAND on retirement reform

The RAND Corporation’s Sichen Shawn Chao asks what China’s gradual rise in statutory retirement ages changes by twenty thirty-five. This is a labor and demographic study, not elite politics. It says the reform creates a large temporary legal buffer, but doesn’t solve aging or guarantee actual workers.

Men move from sixty to sixty-three. Women on the old white-collar track move from fifty-five to fifty-eight, and those on the old blue-collar track from fifty to fifty-five. RAND simulates the official month-by-month phase-in against the old rules while holding the population constant. In its preferred case, the statutory working-age population rises by five and a half million between twenty twenty-four and twenty thirty-five; under the old rules it falls by forty-two point eight million. The gap is forty-eight point three million people. It peaks near sixty-seven million in twenty forty-three, then erodes.

The model uses United Nations population projections by age and sex, United Nations-published Chinese census occupation tables from twenty ten and twenty twenty, the legal schedule, and Penn World Table growth data. Seven scenarios put the twenty thirty-five effect between forty-five and fifty-five million. That consistency makes the mechanical headcount finding strong.

But the key assumption is also the politics of the policy. China doesn’t publish how women divide between retirement tracks. RAND maps occupations onto them, assigns one quarter of service and sales workers to the higher-age white-collar track, and carries forward half the occupational upgrading observed across just two censuses. Women provide fifty-four percent of the twenty thirty-five buffer. Blue-collar women on the old-fifty track alone provide thirty-eight percent.

RAND predicts the statutory window, not employment. Its generous upper bound says even if every retained person worked, annual growth through twenty thirty-five would be only about zero point three percentage points higher than under the old rules. A twenty thirty-five statutory count outside its range of forty-five to fifty-five million would falsify the model, though the unpublished track split makes that hard to observe cleanly.

It doesn’t model participation, employer demand, informal work, pension cash flow or whether artificial intelligence complements these jobs or removes them. The title invokes artificial-intelligence-driven change, but the model rightly refuses to forecast it.

The weakness is the meaning of “buying time.” For old-fifty-track women, a later statutory age may delay benefits without adding paid work. And their large modeled share depends on an occupational proxy, not observed retirement classification. I looked for a comparable conversion from statutory headcount to workers at the Korea Institute for International Economic Policy and couldn’t find one. What would decide this is cohort data separating the female tracks, then tracking employment, hours, health, benefit claims and social-insurance payments. Until then, forty-eight million is a sound legal-demographic estimate and a poor description of usable labor.

What to watch

The quickest test is whether Beijing turns the artificial-intelligence discussion into a binding tiered rule, with named thresholds and actual reviews. Then watch whether July’s overseas-affairs conferences produce measurable staffing, money and activity rather than more conferences. And as the first retirement cohorts move through the new ages, employment and contribution records will tell us whether RAND’s window is productive time or merely later eligibility.