China's AI Strategy and the Middle-Power Squeeze

Opening

Welcome to today's China report. Today it's artificial intelligence policy, and what it does to everyone who isn't Washington or Beijing. Two reports, landing two days apart in early September. And let me say this plainly up front: the second one, from Brookings, isn't squarely about China, China is one of two poles in it, not the subject. It earns its place because its actual question, how third countries respond to Chinese and American AI dominance, is the question the first report raises and then declines to answer. Also worth noting: only two reports cleared the bar this week. That's fewer than usual, and a quiet month in the literature is itself a fact about the literature.

Bruegel on China's AI strategy

First, Bruegel in Brussels, published the tenth of September, by Alicia García-Herrero and Théo Storella.

Their claim is that China's AI posture only looks contradictory. It protects the domestic market while pushing its models out into the world. It chases self-reliance while promoting international governance frameworks. Bruegel says that's not incoherence, it's one state-coordinated strategy across the whole stack, aimed at leading on innovation, on standards, and on adoption. Three motives underneath it: economic, because AI is the answer to an exhausted investment model and a shrinking workforce; sovereignty, because US export controls exposed how dependent China was; and military, which they call the least publicly acknowledged and arguably the most significant.

The mechanism is tiered state support. Heaviest money where capital intensity is highest, which means chips. Lighter enabling measures for models. And a protected home market for the platforms, so they accumulate users and data, earn revenue, and fund the model development themselves. Then you publish the weights openly, which lowers barriers at home and puts Beijing at the centre of global AI development at the same time. The asymmetry is the whole point, Chinese models travel outward, foreign firms don't get the same access coming in.

What does it rest on? Funding flows, mostly. The third phase of the national chip fund, forty-seven and a half billion dollars, sourced to Reuters. A national AI fund of eight point two billion, sourced to a Chinese news outlet. Shenzhen compute vouchers capped at five hundred million renminbi a year, from a municipal bureau notice. Huawei taking seventy percent of SMIC's advanced capacity, and that one's an estimate from MERICS in Berlin, not an official number. But look at the shape of the reference list. A lot of it is Chinese state planning documents. The "AI plus" State Council document, the education ministry's action plan, the fifteenth five-year plan. Those are excellent evidence of what Beijing intends. They are not evidence that it's working. A report whose headline claim is coordination, resting substantially on the coordinating documents themselves, is a reading of a strategy rather than a measurement of one.

The forecast is thin, and I want to be fair about why. There's no date for chip self-sufficiency, no date for when Chinese models reach any particular share of anything. The one hard number with a year attached is American: about seven hundred and fifty billion dollars of US hyperscaler capital spending in 2026, roughly zero point six percent of global GDP, from S&P Global Ratings. Which means the biggest figure in a paper about Chinese state coordination belongs to five American companies.

The policy mapping here is good and I'd trust it. Two problems, though. They call the military objective arguably the most significant, and then the evidence for it is a PLA Daily piece from 2022 and a Georgetown CSET paper from 2021. The limb they weight most heavily is the one their own references support least. Second, the diffusion story. I went looking for an independent count of Chinese models actually being taken up outside China. The closest I found was an ISEAS Perspective out of Singapore, published in April by Wang Zheng, on Beijing's AI outreach to Southeast Asia. What it documents is training programmes and forums, Guangxi province alone put more than two hundred and sixty ASEAN officials through AI programmes by December 2025. That's real outreach. It isn't adoption, and ISEAS doesn't claim it is. If the diffusion half of Bruegel's thesis is right, somebody should be able to produce usage shares by country. I couldn't find one.

Brookings on middle-power AI agency

Now Brookings, the eighth of September, Douglas Rediker.

His claim is that middle powers, Europe, Japan, South Korea, risk political subordination through AI dependence, and that the answer is neither full sovereignty nor autarky. He calls it modular sovereignty: pool the things that need scale, training runs, foundation research, compute, and keep national control over applications and deployment. And he says the realistic arena isn't chatbots, it's physical and industrial AI, in places where the legal identity of whoever operates the system is itself a requirement. Hospitals, grids, ports, rail, defence.

The chain runs like this. Dependence on foreign models and cloud gives you formal sovereignty and no practical control. Competing on price or scale against China is a losing strategy. But regulated, security-sensitive deployment is a segment where jurisdiction beats cost, so build there, collectively, with binding procurement to create the demand.

The dependence-to-subordination link is asserted, not demonstrated. What he offers instead is two incidents: the 2025 licensing requirement on Nvidia's H20 chips, later reversed, and the June 2026 episode where Anthropic's nationality verification under export controls led it to suspend two models for everyone rather than sort users. Those are illustrations. There's no case study of dependence actually producing political subordination, and no model.

The numbers are all borrowed. Fifty-four percent of worldwide industrial-robot installations in 2024 and a stock of roughly two million, from the International Federation of Robotics. Two hundred and eighty-six billion dollars of US private AI investment in 2025 against China's twelve point four billion, from the Stanford AI Index. And that pair is where it gets interesting, because those two numbers can't both mean what he wants. If Chinese private AI investment really is a twenty-third of America's, then the Stanford measure isn't capturing whatever produced the robot dominance he's worried about. He's stacking a figure that undercounts China against a figure that shows China winning, and treating both as load-bearing.

His predictions are dated, which I appreciate. Japan's Noetra initiative, forty-four companies, a planned cluster of about twenty-seven and a half thousand Nvidia Rubin chips, targets an omni-modal model around fiscal 2028 and a robotics model around 2030. European AI gigafactories operational around mid-2028, off a July tender worth about ten billion euros against more than seven hundred billion in US corporate capex this year alone.

Here's my real challenge, and it comes from inside one of his own cases. In July, KIEP, the Korea Institute for International Economic Policy, which is funded by the Korean government, published a piece by Hyunjin Lee arguing almost the opposite prescription. Not collective capacity: interoperability. Korea as the adapter between two decoupling stacks, holding the high-bandwidth memory bottleneck so both sides need it. Lee cites an in-house patent-citation study finding that a half-decoupling of US semiconductor and Chinese image-processing technology would cost Korea about one point nine percent in innovation against one point seven for the United States, though KIEP publishes no methodology for that, so take the decimal point lightly. The direction is the point. Seoul's own institute reads the same squeeze Rediker reads and concludes that being indispensable to both beats building a third pole.

What to watch

The nearest thing that resolves is European: those gigafactories are meant to be running by mid-2028, and if the tender slips again, Rediker's financing chapter is the only part of his argument left standing. Watch also whether anyone publishes country-level usage shares for Chinese models, that single dataset would settle the weaker half of Bruegel's thesis in either direction. And note that these two agree Europe is behind while sharing almost no sources: one read Beijing's planning documents, the other read robot counts and investment tallies. Agreement arrived at separately is the kind worth something.