AI is expected to be a central issue at the Xi-Trump summit later this September, but the scope for concessions on either side seems much less certain. US rhetoric around AI safety is often viewed sceptically in China as a means of holding back homegrown companies, and given the recent exchange of barbs between the US and China over AI distillation, the scope for agreement over the terms of market competition appears even narrower.
Huang Ping, a scholar at CUHK-Shenzhen, nevertheless lays out why it would be in China’s interests to seek a much broader AI agreement to secure its access to global markets and influence over standards. His reasoning is that whatever disagreements other countries may have with Washington, they tend to share its fundamental concerns about industrial competition from China. Recent events at the G20, where China alone opposed provisions on addressing “non-market policies” in global trade, amply illustrate the US’s ability to pull together a united front on such issues.
Faced with the possibility of the US achieving something similar in global AI markets, he argues China’s best strategy is to “advance through concessions” (以退为进). In practice, this would mean provisionally agreeing a global “division of labour” with the US: China would provide open-source models, cheap inference, industrial applications and smart devices to global customers, while the US could maintain primacy in markets for cloud computing, high-end chips and frontier models.
This would seem a big compromise for China to make. However, Huang argues that the prospect of direct Sino-US confrontation in global AI markets is even less favourable to China. Faced with a potential bifurcation of the global AI stack, most countries, anxious to avoid the swallowing up of their industrial value chains by Chinese companies, would likely choose the US as the partner posing less of a threat to their domestic industries.
Huang’s strategic logic is consistent with China’s promotion of open-source AI as an alternative to proprietary ecosystems from the US, using non-monopolistic principles as a wedge to alter the terms of competition. Yet China’s AI ambitions evidently go beyond exporting low-cost open models. While much overseas use of Chinese models currently runs on US cloud servers, commentary has recently argued that a focus on “token exports” should in the future mean routing more overseas inference through domestic data centres.
Huang has earlier cautioned against China muscling in on the cloud infrastructure segment of AI markets. The ambition to “do everything itself” across entire value chains, he argues, is exactly what risks uniting other countries against Chinese AI.
— James Farquharson
Key Points
The support of the other 19 participating G20 members for language on addressing “non-market policies and practices”, which China alone opposed, shows that Washington is still capable of building a broad consensus around concerns over Chinese industrial competition.
By contrast, although many countries are enthusiastic about Chinese open-source models, this is mainly because they want affordable technology and managed reliance on multiple suppliers.
As such, most countries would currently favour Washington if forced into a mutually exclusive choice between China and the US, mainly because cooperation with the US seems more compatible with preserving their domestic industries.
By contrast, a widely perceived issue with China’s industrial model is that it tends to swallow up entire industrial chains and hollow out local industries, meaning that few countries would accept Chinese AI leadership.
The key to maintaining global AI market access is negotiation with Washington, which has the power to either ease the international acceptance of Chinese technology or consolidate a global technological system that excludes Chinese companies.
Chinese Whispers with Cindy Yu, columnist and contributing editor at The Times and The Sunday Times, is a podcast on Chinese history, politics and society. It goes high and low, bringing you conversations on all the underrated, quirky but important topics beyond the usual headlines. Listen here.
China and the US share core interests in expanding AI markets, AI safety and avoiding a disorderly end to the AI investment bubble, while their conflicts of interest relate mainly to the distribution of global revenues and leadership.
A provisional and non-exclusive division of labour could allow the US to leverage its strengths in frontier research, chips and cloud computing, while China could concentrate on serving markets in open-source models, industrial applications and smart devices.
Such an arrangement would be provisional and permit continued technological progress for China, which would only have to relinquish the urge to compete in every segment of the global market.
Chinese companies that expand overseas should allow local partners to control their own data and share benefits from application development, employment and revenues, rather than seeking to control the entire industrial chain.
To build international trust, secure market participation and gain influence over global standards, China should accept jointly negotiated safeguards and audits as limited concessions for broader access.
The Scholar
Name: Huang Ping (黄平)
Year of Birth: Not publicly disclosed
Position: Associate Professor and Assistant Dean (Student Affairs), School of Public Policy, The Chinese University of Hong Kong, Shenzhen; Deputy Director (Development), The Institute for International Affairs, Qianhai
Research Focus: Sustainability Transitions, Innovation Studies, Economic Geography, Energy Transitions
Education: Dual BS degrees in Management and Economics, Harbin Engineering University, China (2010); PhD in Management Science, Harbin Engineering University, China (2015), jointly trained with Utrecht University, Netherlands
Experience Abroad: Postdoc, The Bartlett School of Architecture, University College London, UK; Postdoc, The Fletcher School of Law and Diplomacy, Tufts University, US; Research Associate, The Urban Institute, University of Sheffield, UK
SINO-US AI TALKS VIEWED THROUGH THE LENS OF THE G20: BREAKING THE DEADLOCK WILL REQUIRE ADVANCING THROUGH CONCESSIONS
By Huang Ping (黄平)
Published by Greater Bay Area Review on 9 September 2026
Thank you to Huang Ping for his permission to share this article
Lightly edited machine translation
Illustration by ChatGPT
I. Introduction: Lessons from the G20 Meeting
Two consecutive G20 meetings took place in North Carolina, the United States, from 31 August to 2 September.
The first was the second meeting of finance ministers and central bank governors. Participants recognised that investment in artificial intelligence, computing capacity and digital infrastructure could raise productivity. They also stressed the need to address financial risks and use AI to strengthen cyber resilience. However, the meeting ultimately failed to produce a joint statement. Instead, the United States, which holds the rotating presidency, issued a “Chair’s Statement”. The disagreements concerned global imbalances, so-called “non-market policies”, trade surpluses and sovereign debt, rather than AI. The text published by the US Treasury specifically noted China’s opposition to paragraphs 4, 10, 11 and 13. These four paragraphs primarily addressed supply chains for energy, food, fertilisers and critical minerals. They also covered eliminating “non-market policies and practices” and restructuring sovereign debt. The provisions on AI investment, innovation and risk management fell outside these four paragraphs. [Note: China disagreed with the proposed memorandum’s wording on eliminating “non-market policies” in global trade; the other 19 countries were all in favour.]
The G20 innovation ministers’ meeting that immediately followed produced an even more noteworthy outcome. Participants agreed on a statement covering six areas: innovation policy, public services, technical talent, AI intellectual property, technical standards and investment across industrial supply chains. They also adopted the “Carolina Principles”, proposed under US leadership. The statement advocated greater investment in basic research, faster commercialisation of technology and priority use of existing sectoral rules to address AI risks. It also called for respect for each country’s right to formulate its own governance policies. The White House described this as a moment of “historic unity”.
Taken together, these meetings sent a signal that runs counter to the instincts of many people in China:
The United States under Trump has actually not been isolated by its friends [众叛亲离] to the extent imagined, and nor is China’s consensus with the rest of the world as solid as assumed.
Countries around the world do indeed worry about US technological monopolies and want to secure their own AI sovereignty. Yet they also worry about China’s formidable technological diffusion capacity, its industrial scale and its production strengths throughout the supply chain. They do not want to be tied to the United States, but neither do they want to be displaced by China. If forced to choose between the two systems, most countries are presently still more likely to lean towards the United States.
China and the United States have a packed negotiating schedule over the next four months. In mid-September, the two countries will hold their first high-level talks on AI safety; their heads of state are scheduled to meet in Washington on 24 September; and the APEC leaders’ meeting will take place in Shenzhen on 18–19 November, followed by the G20 summit in Miami on 14–15 December to round things off. AI will be the central theme throughout.
The difficulties faced by China are real, but there is also cause for hope. The key to breaking the deadlock may lie in recognising the genuine shared interests between China and the United States, rather than continuing to seek an ‘anti-US alliance’. Instead of asking the world to choose China over the United States, China could use limited concessions, a functional division of labour and shared benefits to integrate its AI into more countries’ industrial systems.
II. Support for Open Source Does Not Mean China Alignment
The US position at this G20 innovation ministers’ meeting was clear. It advocated reducing regulatory barriers, expanding basic research and private investment, and accelerating AI’s journey from the laboratory to the market. Existing sectoral rules should govern AI wherever possible, with dedicated rules introduced only for new risks that conventional regulations cannot cover.
Other countries did not entirely agree. European Commission Executive Vice-President Virkkunen stressed that Europe would continue implementing the AI Act to ensure that AI remained safe, transparent and compliant with European law. Andrew Bailey, Governor of the Bank of England and Chair of the Financial Stability Board, issued a fresh warning about risks to the global financial system. Rapidly evolving frontier AI models were fundamentally reshaping cybersecurity risks and had become a key factor threatening global financial stability. Countries urgently needed to coordinate regulation and implement standards for controllable deployment. India, meanwhile, supported innovation while emphasising domestic models, India-specific benchmarks and the development of sovereign capabilities.
But these are mainly differences of degree, rather than direction.
The United States, the EU, the UK, India and other G20 members all recognise AI as a general-purpose technology that raises productivity and drives economic growth. They all support greater infrastructure investment, its wider use in public services, the development of technical talent, the establishment of international standards and the adoption of measures to address genuine risks. Their dispute concerns whether regulation should come sooner or later, and whether it should be stricter or looser. It is not about whether AI should be developed and made widely available.
In other words, the United States may not have persuaded every country to accept its model, but it did successfully establish ‘innovation, growth, private investment and industrial applications’ as the meeting’s overarching framework. The gap between the United States and other countries is not as wide as imagined.
China, too, shares considerable common ground with countries around the world.
The G20 innovation ministers’ statement emphasised that countries should uphold national sovereignty in governing emerging technologies, promoting AI applications in healthcare, education, transport and public administration, while fostering competition through open, fair and transparent standards. These provisions are compatible with China’s longstanding positions on prioritising development, putting people first, sharing benefits inclusively and respecting each country’s independent choices.
The problem is that agreement on principles does not automatically translate into industrial choices.
Many countries endorse China’s call to make AI’s benefits widely accessible and are perfectly willing to use Chinese open-source models. But they will not necessarily want to build all their computing capacity, data systems, cloud platforms and critical applications on Chinese technological foundations. Supporting openness does not mean accepting Chinese leadership. Nor does supporting multilateral governance mean being willing to become part of China’s AI system.
The consensus reached at this G20 innovation ministers’ meeting shows that China can find common ground with the United States and other countries on AI governance. Yet disagreements at the finance ministers’ meeting over global imbalances, so-called “non-market policies” and “persistent trade surpluses” reveal another reality. When AI technology combines with China’s comprehensive manufacturing base, supply chains and export capabilities, other countries are not just going to be worried about security risks, but also about the survival of their industries. The support of every participating member except China for the US Treasury’s Chair’s Statement demonstrates how widely these economic anxieties are shared.
China’s consensus with other countries is therefore real, but its limits are equally clear. They welcome cheaper, more open technology from China, but they do not want China to enter every segment [of the market], from models, computing capacity and equipment to end-user applications, and ultimately displace their domestic industries.
The two G20 meetings appeared to offer contradictory answers [to the question of a consensus between China and the rest of the world].
On AI governance, China, the United States and other countries can reach a consensus. On trade, industry and global imbalances, however, China appears relatively isolated.
In fact, there is no contradiction.
Countries around the world want AI to keep developing, to share in the gains from higher productivity and industrial upgrading, and to prevent cyberattacks, financial risks and technological misuse. This is a functional consensus. But consensus turns to conflict the moment when attention shifts to the distributional questions of who supplies the chips, models and cloud services, who gains markets, data and profits, and whose companies get driven out.
China, the United States and other countries agree on “making the pie bigger”, but fundamentally disagree over who should make it, how it should be divided and who should set the rules.
This is precisely the difficulty facing China–US AI negotiations, but also a possible opening. Talks can begin with functional consensus. Conflicts over distribution must then be addressed through a division of labour, reciprocity and incremental compromise.
III. Why Global Markets Would Choose the US Over China
In the AI era, China and the United States are the only countries that genuinely possess capabilities across the full technology stack.
For third countries, “manageable dependence” means avoiding complete reliance on any single country, supplier or technological approach, rather than seeking autonomy in everything from chips to applications.
Australia is discussing how to secure reliable access to frontier AI through technology agreements, turning dependence into something that can be managed through institutional arrangements. It is not seeking to break away from the United States and build an entire AI industry independently. Brazil, meanwhile, is trying to use both Chinese and US technologies, maintaining digital sovereignty through diversified supply rather than choosing one to the exclusion of the other. India emphasises sovereign AI and domestic models while also treating US companies and technologies as important resources for cooperation.
These countries do not really want to choose China or the United States. They want to choose themselves.
But technological sovereignty is not just a matter of a political slogan: without chips, cloud computing, capital, models and talent, even the strongest desire for autonomy can amount only to “managing dependence”. Consequently, if an independent path is not viable and a country must choose a principal technological system, most countries are presently still more likely to favour the United States.
In the past, we have often attributed other countries’ misgivings about China to ideological prejudice or US political pressure. These factors certainly exist, but they cannot explain away everything.
For many countries, the deeper anxiety about China stems from fear of the sheer scale of our economy and industrial base.
Chinese AI’s expansion overseas has never been about just exporting AI models. Behind the model lie servers, communications equipment, power solutions, cloud platforms, smart devices and complete supply chains. Chinese companies have exceptional capabilities in controlling costs, delivering engineering projects and replicating solutions at scale. Once they enter a market, they are likely to go beyond simply supplying a technology and instead reshape the entire industrial chain.
This can be attractive to countries that lack industrial capabilities, but it may threaten those trying to nurture domestic industries. Their concern is not that Chinese technology will fail to work, but that it will work too well, cost too little and expand too quickly, ultimately depriving local companies of opportunities to grow.
The sharp disagreements at this G20 finance ministers’ meeting over so-called “overcapacity”, “non-market policies” and persistent trade surpluses were a concentrated expression of this anxiety about scale. The United States sought to frame China’s export capabilities as a problem of global imbalances. Every other participating member except China accepted the relevant wording.
One point is in need of clarification here. It isn’t that the United States operates on a smaller scale than China. Its control over chips, cloud platforms, foundation models, capital and global standards is equally formidable. The difference lies in how the two countries’ scale manifests itself.
The United States primarily controls high-value-added segments through chips, cloud services, software, finance and standards, and many partners can still expect to retain their own markets for local infrastructure, sector-specific applications and services. [They fear that] China, by contrast, may end up supplying equipment, models, construction and end-user products simultaneously, competing more directly with local manufacturing and digital industries. This perception may not be entirely objective, but it has a tangible influence on countries’ choices.
IV. The Importance of a US-China Deal on AI
This leads us to a counter-intuitive judgement: the most important means of advancing the globalisation of Chinese AI may be the United States itself, rather than its enemies.
The reasons are straightforward.
First, only the United States and China possess capabilities across the full AI stack, and only they can genuinely negotiate bargains over the global AI industry’s governance and division of labour. Other countries can choose, balance and hedge, but they struggle to shape global rules independently.
Second, despite their intense rivalry, China and the United States have substantial and quantifiable shared interests. US companies need China’s market, manufacturing capabilities, opportunities for practical applications and engineering talent. Chinese companies need US chips, capital, developer ecosystems, international networks and influence over rules. The more interdependent the two sides are, the clearer the scope for negotiation becomes.
Third, countries experiencing trade or political friction with the United States do not automatically become China’s technology partners. Canada is a case in point. It repeatedly stresses AI sovereignty and reduced dependence on the United States, yet its government, healthcare, finance and critical infrastructure remain heavily dependent on cloud and AI systems supplied by US companies such as Microsoft, Amazon and Google. Disagreements with Trump do not mean Canada is willing to shift its underlying AI infrastructure towards China.
China therefore cannot simply apply the logic that “the enemy of my enemy is my friend”. What determines the scope for Chinese AI’s global expansion is who is willing to admit Chinese technology into their domestic industrial ecosystem, rather than who happens to be on bad terms with Washington at the moment.
The United States has the greatest ability to influence those countries’ choices. If China and the United States can establish a minimal framework for coexistence in AI, the political resistance facing Chinese AI in global markets could fall significantly. If they move towards complete decoupling, however, Washington will be able to draw increasing numbers of countries into a technological system that excludes China.
This is why China–US AI negotiations may matter more than China’s negotiations with any third party.
V. Seeking Complementarity: A Division of Global Markets
China and the United States have major conflicts of interest in AI. Viewed globally, however, they also share more interests than might be imagined.
The first point of agreement is that AI should continue to develop and its applications should expand. Both countries regard AI as a general-purpose technology and a central engine of future growth. Both need more compute capacity, energy, practical applications and users to expand the global market.
The second is preventing AI risks from spiralling out of control. However intense their rivalry becomes, neither country wants AI used for large-scale cyberattacks, biological threats or attacks on critical infrastructure, or to cause misjudgements in nuclear command systems.
The third is preventing a twin fallout from an AI investment bubble [bursting] and disorderly investment in infrastructure from hitting the global economy simultaneously. Both countries need continued AI investment, but neither can afford a sudden popping of the bubble.
The real disagreement between China and the United States therefore concerns who leads, who secures markets and who sets standards, rather than whether to develop AI—a question of distributing benefits rather than [pursuing fundamentally] different objectives.
We must recognise the hard reality that, at least for now, the United States will not accept China standing on fully equal terms with it at the frontier of AI technology.
It is unrealistic to expect the United States to voluntarily surrender its technological lead, lift all chip restrictions or accept China’s wholesale replacement of the US ecosystem. Nor should these be negotiating objectives at this stage.
Yet, its unwillingness to accept technological equality does not mean that the United States will be unwilling to accept an industrial division of labour.
What could emerge at this stage should be a provisional and non-exclusive functional division of labour that could evolve and allow countries to move into more advanced roles, rather than a fixed hierarchy. The United States would continue drawing on its strengths in frontier research, top-tier chips, cloud computing and global capital. China would draw on its strengths in open-source models, low-cost inference, industrial applications, smart devices and deployment at scale. Third countries would retain control over local data and capabilities in local languages, sector-specific applications, operations and governance.
Such a division of labour would not consign China permanently to the lower end or require it to forsake technological autonomy; it could temporarily reduce direct confrontation by shifting current dynamics towards a more complementary market relationship. China could continue to catch up in core technologies without seeking to capture every segment of the global market.
This would ease both US anxiety about being displaced and third countries’ fear of China’s scale.
VI. “Advancing Through Concessions” in Two Main Areas
“Advancing through concessions” [以退为进] certainly does not mean abandoning independent research and development or accepting permanent US dominance. Still less does it mean trading core interests for short-term market access.
What needs to be curbed is the urge to exert control at two levels.
First, in business, “taking the entire industrial chain overseas” should no longer be taken to mean that “Chinese companies handle every part of it”. Chinese companies can provide models, equipment and engineering capabilities. But they should leave data, application development, operations, jobs and a share of profits to local partners. Partner countries should gain businesses, tax revenues and talent from the development of Chinese AI, alongside cheaper products.
Second, on rules, China should be willing to accept mutually agreed mechanisms for risk assessment, intellectual property protection, model provenance, data security and third-party audits, thereby earning market trust through greater transparency. Verifiable commitments could thereby secure a reduction in the United States’ expansion of sanctions and exercise of extraterritorial jurisdiction.
China should relinquish the idea that it must “do everything itself” across global industrial chains, seeking instead access to global markets, a seat at the table in setting standards and a lasting industrial presence.
Taking a step back is not about retreating [退一步，不是为了后退]. It is about enabling Chinese AI to become truly integrated into the world.
VII. Conclusion: Limited Concessions for Market Participation
The two recent G20 meetings have taught China a sobering lesson.
The United States is not as isolated as imagined. Even under the Trump administration, its chips, capital, cloud platforms, alliance system and rule-making capabilities allow it to turn its AI policy preferences into a broad international consensus.
Nor will the world embrace China as readily as imagined. Open-source technology, low prices and complete supply chains are highly attractive. Yet China’s industrial scale also leaves other countries wondering how much room their own companies will have left to grow once they adopt Chinese technology.
Consequently, the greatest obstacle to Chinese AI’s global expansion may be the combined strength of its technology, industry and supply chains, rather than any technological weakness. Together, they are so strong that potential partners fear they will ultimately become nothing more than markets.
This is also where the true value of China–US AI negotiations will lie.
China and the United States are the fiercest competitors, yet they are also best placed to reach arrangements with global influence. Both want the AI market to keep expanding, both need to contain major security risks and both require stable standards and predictable conditions for their industries. The real difficulty in negotiations lies in the divvying up of markets, benefits and global leadership.
China’s way forward is neither to wait for more countries to take its side nor to replicate a complete Chinese system in every overseas market. Rather, it is to exchange limited concessions for broader participation, reduce external fears through a functional division of labour, and secure a lasting presence by sharing benefits.
True “advancement” lies not in squeezing the US out of the global AI system, but in ensuring that China becomes a part of any global AI system that cannot be bypassed.
READ MORE
China’s AI Path and the Needham Question: From 1 to 10, Not 0 to 1
As both nations invest ever more significant resources into artificial intelligence, the divergent trajectories between China and the US have become a key issue—with many pointing to a philosophical—even quasi-religious—divide between the rapid adoption of practical AI applications in the former and the more abstract pursuit of AGI in the latter. Huang Ping, an up-and-coming associate professor at the Chinese University of Hong Kong, agrees with these analyses. He notes that despite state-funded investment drives in basic research, China’s innovation system remains relatively weak at original breakthroughs (moving from “0 to 1”) but excels at scaling and commercialising technologies (moving from “1 to 10”).
China’s AI Strategy: Encircling the Cities From the Countryside — by Di Dongsheng
Today’s edition opens with an introduction by my former colleague Rebecca Arcesati, Lead Analyst in the Science, Technology and Innovation programme at the Mercator Institute for China Studies (MERICS). In particular, her research examines China’s AI and data strategies, EU–China innovation relations and US–China tech competition. Much of her work can be found