As Chinese large AI models have delivered one global shock after another, alarm in American political and technology circles has deepened into something close to panic.
In fact, the campaign to hem in China's high‑technology industries began as early as the first Trump administration, the placement of Huawei on the Entity List being the emblematic case. Over the near decade that followed, the pressure has only intensified: the Biden administration erected its “small yard, high fence”; the second Trump administration has reportedly weighed banning Chinese large models outright.
Cai Cuihong, Deputy Director of the Center for Global AI Innovative Governance and Professor at the Center for American Studies, Fudan University, told Guancha.cn that although the three administrations have differed in emphasis, they share one premise: that China is the most important strategic competitor the United States faces in artificial intelligence.
Should Washington escalate its pressure on Chinese large models, Cai argues, it would trigger a significant counter‑effect, and one that Washington itself most fears: the transformation of a catch‑up dynamic organized around American closed‑source frontier models into a contest organized around a global open‑source model ecosystem.
It must be acknowledged that Chinese large models still trail the most advanced American systems. The challenges China must confront in seeking a further breakthrough, Cai notes, fall into three areas: core technologies and innovation inputs; application and industrialization capacity; and ecosystem and internationalization capacity.
The following is a transcript of the conversation.
Guancha.cn: China‑U.S. competition has until now centered mainly on trade, manufacturing and chips. With the rise of large AI models, a further dimension has been added. Over the past decade, what are the similarities and differences among the three administrations in their AI policy toward China?
Cai Cuihong: The three administrations display fairly marked shifts in their AI policy toward China, and these may be understood as three stages.
The first Trump term can be summarized as the securitization of technology competition, with AI treated as merely one among several strategic technologies. Under the Biden administration this was upgraded into the institutionalization of restrictions on critical technologies, with semiconductor controls explicitly directed at the formation of frontier AI capability. In the second Trump term, competition has been extended to the level of ecosystems and to a global scale, with AI folded into the overall framework of technology diplomacy.
A decade ago, following Trump's first election, the American AI Initiative was launched in 2019. It focused on increasing R&D funding, opening federal data and computing resources, advancing technical standards and developing talent, with the aim of preserving overall American technological leadership. Large models did not yet exist, so the policy was not directed specifically at China. In technology competition with China, the representative measures of that period included adding Huawei and other firms to the Entity List, tightening investment screening and export licensing, and progressively extending the extraterritorial reach of U.S. export controls.
Restrictions in the first Trump term were largely point‑by‑point, aimed at specific firms and specific transactions. Wherever a company was judged to pose a security risk, or a transaction was thought liable to transfer critical technology to China, administrative instruments such as the Entity List, investment screening and export licensing were brought to bear. This reflected a shift in perception: China's technological development was no longer a question merely of trade imbalances or market competition, but one to be placed within a national security framework.
Under the Biden administration, restrictions on individual firms were upgraded into institutional arrangements targeting the formation of frontier AI capability in China. The export controls on advanced computing and semiconductor manufacturing items introduced in 2022 are one example; in 2024, high bandwidth memory, additional semiconductor manufacturing equipment and software tools were brought within the scope of the controls.
Whereas the first Trump term asked chiefly who might obtain a given technology, the Biden administration attended more closely to what determines the ceiling on capability. It therefore identified control points along the chain by which AI capability is formed, encompassing chip design, wafer fabrication, software tools, high bandwidth memory and model training. The object was no longer any single product but the entire process by which China's frontier AI capability is generated.
The Biden administration customarily summarized this approach as “small yard, high fence”: control a small number of critical technologies bearing directly on national security, and place a very high regulatory barrier around them. The difficulty is that as chips, equipment, software and high bandwidth memory were successively brought within the perimeter, the actual boundary of the “small yard” kept expanding.
In the second Trump term a further change has occurred, in my view: a shift from restricting China's access to critical technologies toward contesting dominance over the global AI ecosystem.
Although Trump rescinded certain Biden‑era AI executive orders, this in no sense represents a relaxation of technology competition with China. America's AI Action Plan of 2025 states explicitly that the full American AI technology stack, comprising hardware, models, software, applications and standards, is to be exported to allies and partners. The United States seeks not only to build AI capability at home but to secure worldwide adoption of its AI systems, compute hardware and standards.
Compared with the Biden administration's emphasis on institutionalized, rule‑based policy instruments, the second Trump term allows considerably wider executive discretion, and its policies display a more pronounced commercial and political transactionalism.
In sum: the first Trump term sought to prevent particular Chinese firms from obtaining American technology; the Biden administration concentrated on controlling the critical nodes on which the formation of frontier AI capability in China depends; and the second Trump term is further concerned with securing global market adoption of an American‑led AI technology stack, standards and governance rules.
Notwithstanding these differences, the continuities are equally striking. All three administrations have held that China is the most important strategic competitor the United States faces in AI, and that AI‑related advanced compute and semiconductor manufacturing capacity carry a high order of national security and strategic significance. From the Biden administration onward, moreover, allied coordination and the technology choices of third countries have become an important component of the AI competition with China.
Guancha.cn: The release of Kimi K3 immediately drew intense attention in the United States, with some reports suggesting that the Trump administration is considering restricting or even banning Chinese large AI models. If the White House imposes further restrictions, might the pressure in fact make China's AI industry stronger?
Cai Cuihong: Whether to ban Chinese large AI models remains at the stage of policy discussion, but arrangements amounting to a de facto ban could well emerge. Among the measures under discussion are adding large‑model companies to the Entity List, or restricting American firms from hosting Chinese models. Washington need not announce a comprehensive ban publicly; through procurement restrictions, the Entity List, compliance liability, security advisories and pressure from public opinion, it can raise the risks and costs to American firms of using Chinese large models, thereby producing market exclusion in practice.
That Kimi K3 attracted such attention in the United States indicates that the gap between Chinese large models and American frontier models is narrowing, and that China is able to combine low cost, open weights and competence on complex tasks. It also demonstrates that the restrictions on advanced chips and compute imposed over the past several years have failed to prevent Chinese models from approaching the global frontier. That said, Kimi K3's own technical report acknowledges plainly that its overall capability still lags the strongest American closed‑source models.
Continued American pressure will not necessarily make Chinese AI capability stronger. A forcing effect is possible, but it is not automatic. The likely sequence involves higher costs in the short term, an impetus toward substitution over the medium term, and a possible reshaping of the technology ecosystem over the long term; only after a protracted process might a step change in capability be achieved.
Whether external pressure is ultimately converted into a leap in capability depends on whether capability can be built simultaneously across frontier technology, basic research, industrial application and the global ecosystem. If the response amounts only to duplicative construction or low‑grade substitution, external pressure may instead entrench the technology gap.
What Washington most fears, I believe, is a significant counter‑effect of further restrictions on Chinese models: the conversion of a catch‑up dynamic organized around American closed‑source frontier models into a contest organized around a global open‑source model ecosystem.
Broadly speaking, the China‑U.S. AI competition is undergoing a change of paradigm, from one governed by scarcity to one governed by diffusion. Competition is expanding from a contest over the ceiling of capability alone to a simultaneous contest over the speed at which capability diffuses, the scale of its application and the reach of its ecosystem.
Guancha.cn: American media often portray a society that is fearful of AI, even rejecting it, and many reports have followed the wave of protests against AI data center construction across a number of states. In your observation, is that an accurate picture?
Cai Cuihong: The characteristic condition of American society at present can be captured in three words: use, anxiety and distrust. It is not a rejection of AI. Americans are adopting it rapidly while worrying about the social consequences, and while lacking sufficient trust in either corporations or government.
These concerns arise at three levels as the scope of AI use widens. The first is employment and income security. The second is privacy, false content and algorithmic reliability. The third is a deeper set of governance questions: who holds the data, who bears responsibility when something goes wrong, and whether government and business can be trusted.
It should also be noted that in the current American political environment, the language of the “China threat” and of “competition with China” retains considerable mobilizing power. In the dispute over a data center in Utah, for instance, local opposition groups were at one point linked to China, an allegation later conceded to be without evidence. What lay behind it was the use of a national security narrative to obscure a conflict of local interests.
Local protests against data center construction do not amount to opposition to AI. What residents actually care about are highly concrete questions. Is decision‑making on the project transparent? Will the community obtain any real economic benefit?
They are also concerned about how costs are distributed between themselves and large technology capital. Will the project crowd out local water and power resources? Will it drive up utility bills? Will it generate noise or environmental pollution? Will it cause substantial job losses?
Protesting against a data center is therefore not the same as opposing the development of AI. Residents may well support American AI development, but that does not mean they will unconditionally accept a facility that consumes local electricity, water and land, enjoys tax concessions, and whose returns accrue chiefly to technology firms based elsewhere.
Guancha.cn: AI policy touches a great many stakeholders, so disagreement is inevitable. Could you describe the divisions between the two parties, among government departments, between government and Silicon Valley, and among firms themselves?
Cai Cuihong: None of these four relationships consist of camps with clear boundaries and fixed positions. The two parties are not simply opposed; the Treasury is not necessarily moderate, nor the Defense Department necessarily hawkish. Each specific question has to be examined on its own terms.
Take the two parties first. The evident consensus between them is that China is America's principal strategic competitor in AI and that the United States must preserve its competitive advantage. They differ over how the competition is to be won, and over what regulatory and social costs they are willing to bear in winning it.
The Republicans, particularly in the second Trump term, emphasize reducing ex ante regulation of American AI firms, accelerating the construction of data centers and of energy, grid and chip infrastructure, and avoiding conflicting regulatory rules across states. Their greater worry is that regulation arriving too early and weighing too heavily will slow the pace of innovation among American firms.
The Democrats attend to a different set of questions. Are workers being displaced? Will algorithms produce discrimination? Do consumers have recourse? Does training data infringe copyright? Is personal information protected? Are large companies acquiring excessive market power? Their greater worry is that if AI systems lack transparency and appropriate safeguards, the public backlash provoked by these social risks will ultimately impair American competitiveness over the long run.
The core difference, then, is this: Republicans fear that premature regulation will cause the United States to lose to China, while Democrats fear that the absence of regulation will produce social disorder that in the end undermines innovation itself.
Among government departments, a simple division into hawks and doves is likewise inadequate. It is more accurate to say that different departments have different priorities.
The Defense Department and the intelligence agencies focus on low‑probability, high‑consequence national security risks, on worst cases and on the upper bound of capability. Their concern is whether a given technology would enhance China's military modernization. Even where a technology carries substantial commercial value, they are therefore inclined to consider restrictions in advance.
The Treasury attends to capital and financial instruments. Would the American network of capital and financial services assist China in developing sensitive technologies? If financial sanctions were imposed, could they be executed with precision without inflicting collateral damage on a large volume of ordinary commercial activity?
The Commerce Department sits at the intersection of national security and industrial competitiveness. It must consider on the one hand whether a technology would strengthen Chinese AI or military capability, and on the other whether excessive restriction would cost American firms their global markets and accelerate domestic substitution in China.
The State Department and the diplomatic apparatus ask whether allies such as the Netherlands, Japan and South Korea are willing to follow, and whether the United States can offer third countries a sufficiently attractive alternative.
Between government and Silicon Valley the relationship is one of mutual dependence and mutual constraint. Government today relies increasingly on private firms for technology, including models, chips, cloud platforms and technical personnel. Firms in turn rely on the government for energy and infrastructure permitting, research funding, public procurement, export licenses and protection in international markets.
The relationship is symbiotic, yet the conflicts within it are plain.
The government asks whether China will distill American large models, whether chip exports will strengthen Chinese capability, and whether large models carry cyber, security or military risks.
Firms ask whether they can obtain sufficient water, power and land from the government, whether the rules are stable, and whether state laws will conflict with one another. They are also concerned about the boundaries of regulatory liability: when a model errs, does responsibility attach to the model developer, the deploying enterprise or the end user?
Among Silicon Valley firms, the divisions turn chiefly on business model and market position.
Chip companies, of which Nvidia is representative, worry that if Chinese firms cannot buy American chips, China will accelerate domestic substitution and American firms will forfeit revenue, market share and influence over standards. They therefore favor controlling only the most advanced products while permitting continued export of chips of relatively lower performance.
Firms such as Anthropic, which place greater weight on frontier model safety, hold that short‑term commercial revenue must not take precedence over preserving American strategic advantage in advanced AI and compute. They have accordingly supported strict export controls on advanced semiconductors in public, and have emphasized preventing the diffusion of frontier model capability to competitors through distillation and similar means.
Platform companies such as Meta, which favor an open‑weight approach, place greater emphasis on technological diffusion and an open ecosystem. Zuckerberg recently argued that banning Chinese large models is not an effective route to American competitive advantage, and that the United States would do better to address its own systemic capability bottlenecks in areas such as infrastructure.
Startups, together with the venture capital behind them, worry that complex model registration, safety evaluation and legal liability requirements will in the end be affordable only to large companies. They therefore tend to argue that regulation should concentrate on clearly defined high‑risk uses rather than imposing identical ex ante requirements on all foundation models.
That said, a very stable consensus underlies all of these positions: that China is the most important state‑level AI competitor the United States faces. The domestic American debate is therefore not about whether to compete, but about how to compete, where the limits lie, and what price is worth paying.
Guancha.cn: It is sometimes said that American AI concentrates on frontier development while China concentrates on application scenarios. How do you view these two choices?
Cai Cuihong: There is something to the observation, but to read it as meaning that America cares only about technology and China only about application would be an excessively binary framing.
A more accurate formulation is that the American model of AI development is driven principally by raising the ceiling of capability, building global platforms and securing high returns on capital, whereas the Chinese model places greater emphasis on joining technological breakthroughs to the industrial system, to application scenarios and to national development objectives.
One important difference concerns the drivers of innovation. The United States pursues a model of frontier breakthrough driven jointly by large technology companies, venture capital and leading research institutions. Possessing world‑leading technology firms, a venture capital system, university research institutions and cloud computing platforms, it finds it easier to concentrate resources on frontier models and general‑purpose technological breakthroughs, and then to diffuse them globally through its platforms.
Private investment in AI in the United States is very high. According to Stanford's AI Index Report 2026, American private AI investment in 2025 was roughly twenty‑three times that of China. This figure counts private investment only, and therefore understates Chinese government guidance funds, infrastructure and other public outlays; it should not be read as the gap in total AI investment between the two countries.
China's distinguishing feature is that it seeks not merely to build models but to bring AI into real settings such as manufacturing, automobiles, logistics, healthcare, education and public services. Behind this lie conditions relatively particular to China: a complete industrial system, an immense market and an abundance of application scenarios. The Chinese government also possesses considerable organizational and coordinating capacity in industrial planning, infrastructure construction and the promotion of applications, which makes it easier to drive rapid technological iteration through a large volume of real‑world scenarios.
It bears particular emphasis, however, that valuing application does not mean abandoning frontier innovation. The two paths are in fact converging. The United States is now making up ground on data centers, energy grids and real applications in manufacturing, while China is further strengthening foundational capabilities in frontier models, basic research, chips and system software.
Guancha.cn: Competition will surely dominate the future of China‑U.S. relations in AI, but there must also be room for cooperation. In which areas do you think the two countries can cooperate?
Cai Cuihong: Cooperation is indeed necessary, but expectations should not be set too high. Technologies bearing on core model capability, training data, advanced chips and military uses are unlikely to become focal points of cooperation in the near term.
The areas genuinely suited to cooperation generally share several features: the risks involved spill across borders; no single country can address them alone; and cooperation does not require either side to open its most sensitive technologies. Four directions are therefore worth considering: cooperation on strategic security, on risk governance, on public goods, and on global governance and capacity building.
The most urgent and necessary area is preventing AI from aggravating military miscalculation and strategic loss of control. AI is being used ever more widely in military decision‑making, intelligence analysis, early warning and target identification. It is especially important to affirm that human control must be maintained over the decision to use nuclear weapons.
Limited cooperation is also possible on frontier model risk assessment and safety testing. Without touching commercial secrets or core technologies, the two sides could seek common evaluation methods: how should a model be assessed; does it contain software vulnerabilities or malicious code; does it lower the threshold for biological and chemical weapons; and do agents deviate from human objectives in the course of carrying out tasks?
Incident notification and emergency communication mechanisms are likewise needed in cybersecurity and biosecurity. Basic behavioral boundaries could be discussed, such as refraining from using AI to attack hospitals, power grids, nuclear facilities and other critical civilian infrastructure. Sensitive biological data is not suitable for sharing, but the two sides could cooperate in establishing evaluation standards for high‑risk biological capabilities.
A mechanism for notification of model accidents, technical failures and major security incidents could also be established. Where a serious model vulnerability or major technical failure is discovered, the necessary channels for contact and notification should exist.
Technical cooperation is also possible on synthetic content, deepfakes and content provenance authentication, including work toward technical interoperability in the labelling of synthetic content and toward standards for digital watermarking and content provenance.
Public health, climate and disaster early warning are the areas in which the value of cooperation can most readily be explained to the public. They possess evident global public goods characteristics, their benefits cross national borders, and cooperation does not necessarily require the exchange of the most sensitive model weights or training data.
Beyond this, the two countries should undertake a minimum of coordination on global AI governance and on capacity building in developing countries. Within the United Nations and other multilateral mechanisms, they could work together to support developing countries' participation in international AI governance and to narrow digital divides in compute and data capability.
Guancha.cn: The prominent American AI scholar Gary Marcus recently wrote that Chinese AI models have very nearly caught up with the most advanced American systems, that the United States cannot win this AI race, and that Washington should stop treating AI as a zero‑sum game and explore instead a path of international cooperation and public goods. How do you assess his argument?
Cai Cuihong: He is a well‑known AI scholar in the United States and a longstanding critic of the prevailing large language model paradigm. His views, however, represent neither the American AI industry nor the position of the U.S. government and strategic community.
He holds that Chinese models are rapidly approaching the American frontier and that American technological leadership will be difficult to sustain over time. He therefore judges that the two countries may enter a condition of close parity in large models in which the lead changes hands repeatedly, and that the United States will find it hard to establish a permanent and overwhelming technological advantage.
When he first commented on DeepSeek, he made the point that this did not mean China had won the AI race, only that the two countries were converging rapidly.
More precisely, he argues that the American mode of competition is itself flawed and cannot produce a lasting technological monopoly. If the United States continues to pour in resources, capital and compute merely to contest a transient lead in models, only for competitors to follow rapidly, the result may be an arms race of high cost, low profit and high risk. Rather than continuing to expand such zero‑sum competition, he argues, part of its logic should be redirected toward international cooperation, so that AI serves global public goods in medicine, science and similar fields.
His assessment carries an important implication: AI competition cannot be understood as a single contest that is settled once and for all. Space for cooperation should be preserved in domains marked by cross‑border risk and global public goods characteristics. This does not mean, however, that cooperation can substitute for competition in AI, still less that core technologies can be opened without limit. The reason is straightforward: while the capabilities of large models diffuse readily, chips, compute, data, platforms, industrial ecosystems and military applications of AI retain a pronounced scarcity and a strong strategic character.
The greater significance of Marcus's argument is thus that it prompts us to reconsider what winning the AI competition means: to move from the pursuit of a short‑term lead in models toward a concern with sustained innovative capacity and overall ecosystem advantage.
Guancha.cn: What constraints does the United States face in developing AI?
Cai Cuihong: The most prominent bottlenecks lie in infrastructure, critical supply chains, talent and commercial returns.
The first is infrastructure such as power generation and the grid. The United States is attempting to address this, but it is not as simple as building a few plants; the process is highly complex. Models and chips are refreshed rapidly, whereas major energy infrastructure often takes years to build, and this mismatch of speeds is itself a bottleneck.
The second is the supply chain. The United States holds a clear advantage in advanced AI chip design, but strength in design does not mean the entire supply chain is held onshore. Nvidia can design world‑leading AI chips, yet its most advanced products depend principally on TSMC for fabrication, and the chain further involves high bandwidth memory, advanced packaging and other transnational links. China's advantages in rare earths and in the processing of certain critical minerals also constitute a potential constraint on the American AI supply chain.
Talent warrants attention as well. Current uncertainty in American immigration and visa policy affects the expectations of international talent about remaining in the country over the long term. Should this trend persist, it will weaken America's capacity to renew its talent pool.
Finally there is the question of commercial returns. Whether the very large sums flowing into AI can be converted into stable corporate profits and broad productivity gains remains to be seen. If capital investment continues to grow at high speed while real applications and commercial returns fail to keep pace, this too will constrain American AI development.
Guancha.cn: By comparison, what difficulties does China face?
Cai Cuihong: The constraints on China differ from those on the United States. They may be grouped roughly into three areas: core technologies and innovation inputs; application and industrialization capacity; and ecosystem and internationalization capacity.
First, there are shortfalls in core technologies and innovation inputs. China's most conspicuous bottleneck is of course advanced chips. This encompasses not only chip design but advanced‑node manufacturing, high bandwidth memory, advanced packaging, and the system software and development ecosystem that accompany them.
Beyond the technological base, data and talent are critical innovation inputs. China possesses vast data and an abundance of application scenarios, yet in industry, healthcare, scientific research and embodied intelligence there remain problems of inconsistent data standards, insufficient specialist annotation and difficulty in circulating data across institutions. As for talent, high‑end expertise in chip architecture, foundational algorithms, basic research, and at the intersection of AI with biology, materials and manufacturing, requires long accumulation and cannot be assembled quickly.
Second is how to convert an advantage in application into an advantage in industrialization. China's evident strengths are a complete industrial system, an abundance of scenarios and rapid deployment. But a large number of applications does not equate to high‑quality applications; in specialist domains such as industry, healthcare and finance, the problem of moving from workable to stable and reliable has still to be solved. Genuine industrialization depends further on whether productivity can be raised on a sustained basis and whether application models can be replicated and scaled. Efficiency in the allocation and use of compute and other resources must also improve, so as to avoid duplicative construction and mismatches between supply and demand.
Third is ecosystem and internationalization capacity. China can now produce globally competitive large models and holds a clear advantage in open‑source models. But it must move beyond releasing and open‑sourcing models to building an ecosystem as a whole: not only widening model use but developing a stronger developer ecosystem, industrial applications and international influence.
Entry into international markets brings a further question of trust. Overseas users look not only at model performance but at data security, cybersecurity and intellectual property, and these concerns may be amplified by geopolitical competition.
As open‑source models enter international markets, the relationship between openness and protection must also be managed well: tiered openness calibrated to different technologies, for example, together with improved licensing, patents and records of technological provenance, so as to strike a better balance between expanding the open‑source ecosystem and protecting core achievements.
Guancha.cn: The near‑duopoly of China and the United States in AI has caused concern among a number of middle powers, France and Japan in particular, which hope to achieve breakthroughs of their own and avoid being constrained by the two leaders. From China's perspective, how should it cooperate with these countries?
Cai Cuihong: Although frontier large models do exhibit a fairly pronounced pattern of China‑U.S. competition, the two countries have not monopolized the AI system as a whole. Besides France and Japan, Germany, South Korea and the Netherlands hold advantages of their own at many points in the AI chain, including industrial software, robotics, semiconductor materials, equipment, regulation and international standards. What these countries genuinely fear is that, as models, compute and platforms grow ever more concentrated, their own room for technological choice and their industrial autonomy will be squeezed.
China's aim, therefore, should not be to pull these countries onto its own side, but to help them acquire greater choice in technology and stronger bargaining power, so that China becomes one option by which they reduce dependence on a single source. This is not a matter of substituting dependence on China for dependence on the United States.
China can bring its supply‑side advantages to bear. Its open‑source models are relatively inexpensive to use and deploy, and comparatively easy to localize to local requirements. Priorities can be set country by country according to each partner's strengths and needs. With countries possessing a strong technological and industrial base, cooperation can take a more reciprocal form through joint research and development, technical standards, supply chains and market access. With countries whose markets are large and whose demand for applications is strong, cooperation can center more on localized models, compute, talent and industrial applications. Multilateral platforms such as the World Artificial Intelligence Cooperation Organization can also be used to draw more countries into discussions of AI rules and standards, strengthening the participation of middle powers in global AI governance. Whatever the form of cooperation, the aspirations of these countries to digital sovereignty and technological autonomy must be respected, so that mutually beneficial cooperation advances a more plural global AI landscape.
Original Link：https://mp.weixin.qq.com/s/bAM6qsFlfFOz2HomidLIeQ?scene=25&sessionid=#wechat_redirect
