Opening 0:00
Welcome to today's China report. Today it's technology competition, chokepoints, export controls, and what happens if artificial general intelligence actually arrives. Only three reports qualified this week, fewer than usual, and a quiet month in the literature is worth saying out loud. Two of the three landed on the same day, with a U.S.-China summit approaching and rare-earth controls still the live test of whether export-control leverage runs in both directions.
CFR on China's supply-chain chokepoints 0:36
First, "Dismantling China's Supply-Chain Weapon," published on the twenty-first of September by the Council on Foreign Relations, written by Joris Teer, who runs the economic security portfolio at the EU Institute for Security Studies.
The argument is that rare earths are a distraction, not because they don't matter, but because they're one chokepoint among many. For seventeen of the thirty-four materials the European Union calls critical, Teer says China holds at least a seventy percent share of global mining or refining. Then add batteries, industrial gases, semiconductor wafers, legacy chips. And he says the Busan truce is a façade: controls postponed to November twenty twenty-six, not withdrawn, with exports still running below pre-control levels.
Here's the mechanism, and it's the good part. Subsidising Western supply can't work alone, because Chinese state firms can flood the market and undercut whatever you build. So you need protected demand, aligned allied tariffs, local-content rules in public procurement, an ex-China market new producers can actually sell into, plus a deterrent threat to Beijing's own market and technology access.
What's it resting on? Synthesis, mostly. The International Energy Agency's estimate of around three and a half trillion dollars a year of economic activity exposed if China cut permanent magnets to America and Europe, and the IEA again for the sixty billion dollars of investment needed by twenty thirty-five. Otherwise it's Financial Times and Bloomberg reporting and off-the-record conversations with European industry. No original dataset. The forecast is crisp, though: if the truce lapses in November, controls hit ten more rare earths, batteries and nine further materials, and diversification won't have closed the gap before twenty thirty-five.
Now the challenge. Teer is right that demand guarantees are the missing half of Western policy, and he names it clearly. But "supply-side approaches will fail" is carrying more than his evidence supports, so I went and checked. The Observer Research Foundation in New Delhi, Indian, close to Delhi's line, which is rather the point, published a paper in April this year putting Japan's rare-earth dependence on China at about ninety percent in twenty ten and forty-eight and a half percent by twenty eighteen, heading below half by twenty twenty-five. That's supply-side policy halving exposure in eight years. Teer's reply would be that Japan still got squeezed to near zero this year, and fair enough. But then the report owes us a number for what safe looks like, and never gives one. And his remedy needs exactly the allied cohesion his own Busan section shows breaking, he files the full version under "if transatlantic relations improve."
CSIS on the AI governance race 3:54
Second, from the Center for Strategic and International Studies, published the same day: "The State of AI Global Governance and Its Implications for the U.S.-China Summit," by Christopher Collins and Aalok Mehta.
There's an institutional race on, they argue, and Washington has entered it without a product. China offers developing states a bundle, a regulatory model built for state control, open model weights anyone can download, training places, infrastructure, and a seat at a new World AI Cooperation Organisation, WAICO, signed by twenty-nine countries in Shanghai in July. The American answer, Pax Silica, grew from eleven signatories to twenty-four, but it's about supply chains, not governance. And the United States has no domestic framework worth exporting.
The chain is short: if the contest is over who writes the rules and who joins, rather than what the rules say, then having nothing to export is a structural loss. So compete on capacity-building instead, they reach for Eisenhower's Atoms for Peace.
The evidence is almost all announcements and news reporting. Reuters for the WAICO count, official releases for Pax Silica. One genuinely strong number comes from a twenty twenty-four UN advisory body: a hundred and eighteen countries belonged to none of the significant AI governance initiatives, and only seven, all rich, belonged to all of them. The claim that five of ten Chinese frontier models released this year published safety evaluations is sourced to an International Bar Association explainer, with no sample frame and no definition of what counts as an evaluation.
And the headline finding, that middle powers won't grant either side exclusivity, rests on Kazakhstan. The report says so itself: Kazakhstan is the only confirmed overlap, Pax Silica in June, WAICO in July. Singapore and Indonesia are gestures at a pattern, not the pattern.
So I went looking for something sturdier. The ISEAS-Yusof Ishak Institute in Singapore fielded its State of Southeast Asia survey in January and February, two thousand and eight respondents across all eleven ASEAN states. Forced to choose, fifty-two percent picked China and forty-eight percent the United States, reversing the year before, and fifty-five percent said what they really want is an ASEAN resilient enough not to have to choose. That carries the hedging story better than Kazakhstan does. It also shows the hedge has a tilt, and the tilt moved. What I couldn't find anywhere was a regional survey asking these countries about AI governance specifically. ISEAS doesn't ask. So the central claim about what middle powers want from AI institutions has no polling under it at all.
Predictions: one hard date, the bilateral AI dialogue on the twenty-fourth of September, plus the UN Global Dialogue's second session in May twenty twenty-seven. The rest is clockless, if an OECD member joins WAICO, expect a domino effect. No threshold, no deadline.
CNAS on AGI and superpower competition 7:30
Third, from the Center for a New American Security, published the twenty-seventh of August: "Superpowers and AGI," by Jacob Stokes.
This one stipulates its world. Assume artificial general intelligence, AGI, is close or already here, then ask what it actually buys you. Stokes's answer: not automatic dominance. He traces five channels capability has to pass through to become power, economic growth, the information environment, military capability, misalignment, and domestic politics, and finds friction in each. Growth hits physical constraints. Information control stays partial, because information still moves on paper and by word of mouth. Military advantage lands on top of an existing balance. And he sets out four phases: both racing, one approaching, one holding it, both holding it.
On China he offers four readings, Beijing isn't chasing AGI; its firms are, without direction; it's taking a different route through embodied AI and robotics; or there's a secret programme, and concludes it's a messy mix of all four.
Evidence: a hundred and thirty-nine endnotes, mostly news reporting, lab statements and other think tanks. The sharpest number is Epoch AI's, capital spending at five big cloud firms has quadrupled since GPT-4 came out in March twenty twenty-three, growing roughly seventy-two percent a year and approaching half a trillion dollars in twenty twenty-five. There's a Pew poll of twenty-five countries showing people trust their own governments to regulate AI ahead of either Washington or Beijing. On China itself, it's English-language state media and Western reporting.
The forecast beat is nearly empty, and he says so: timelines run from one to three years out to the mid-twenty-forties, and Stokes writes AGI may never arrive at all. The only firm date in the paper is Xi's twenty twenty-seven Taiwan target, not his.
Here's my problem. The paper assumes AGI lands at roughly sixty to seventy percent of what bullish technologists say is possible. That figure is the author's own, it isn't defended, and the conclusion leans on it hard, because every friction he names, physical limits, partial information control, existing force balances, is precisely what a genuinely general intelligence is claimed to dissolve. Turn the dial to a hundred and the finding may well flip. So the headline isn't a result about AGI. It's a result about a parameter he picked. Add a China assessment that holds four incompatible readings at once, and a framework with no clock on it, and you have a report that can't be caught out. That isn't the same as being right.
What to watch 10:47
The nearest thing on the clock is November, when the Busan truce lapses. If China's wider controls, batteries, ten more rare earths, take effect, the CFR reading holds; if exports normalise, its "façade" verdict was too dark. Then January, and whether American defence contractors get their delay on Chinese magnets, which is a cleaner test of diversification than any market-share figure. And the two September reports quietly pull against each other: CFR's remedy assumes allies can be lined up behind a single tariff wall, while CSIS's whole finding is that partners are refusing to line up behind anything. Different countries, same instinct.