China Has Not Won the AI Race. But America’s Strategy Is Failing

Kimi K3 does not prove that China has overtaken the United States. It demonstrates something more troubling for Washington: withholding advanced chips has slowed China without stopping it. As America commits extraordinary sums to increasingly expensive proprietary systems, Beijing is offering open models, industrial partnerships and cheaper access to countries that may never be able to afford the American frontier.

Within days of each other, two events revealed the diverging shape of the artificial intelligence contest. In Beijing, Moonshot AI unveiled Kimi K3, a model containing 2.8 trillion parameters, developed in a country that has spent years being denied the most advanced American processors. In Ohio, Nvidia was reported to be considering a financial guarantee of approximately $250 billion to help OpenAI occupy a computing complex whose total cost could exceed $500 billion.

One development came from a Chinese company forced to work around a shortage of frontier hardware. The other contemplated one of the largest private financial commitments in industrial history. Together, they pose a question more important than which model currently leads a particular benchmark. Is the future of artificial intelligence going to belong to the country that builds the most powerful system, or to the country that makes useful intelligence affordable and adaptable enough for the rest of the world?

The United States still possesses the stronger technological position. It controls the leading AI processors, the most important cloud platforms, the deepest pools of capital and many of the best research institutions. Kimi K3 has not overturned those advantages. But it has exposed the weakness in Washington’s central assumption: that China could be kept permanently behind by restricting its access to advanced chips.

The model Washington was not expecting

Moonshot AI announced Kimi K3 on 16 July. The company says the model has 2.8 trillion total parameters, native visual capability and a context window of more than one million tokens. It is intended for extended coding, research, reasoning and complex office work.

The enormous parameter count requires qualification. Kimi K3 uses a sparse mixture of experts architecture. It contains 896 specialist components, but selects only 16 of them for each token it processes. It does not activate all 2.8 trillion parameters whenever it answers a question. That efficiency is part of its significance. China is not merely attempting to reproduce the American approach with inferior hardware. Its laboratories are trying to extract more useful intelligence from the computing power available to them.

Independent assessments place Kimi K3 among the leading models, particularly in coding and agentic tasks, but not indisputably above every American rival. It is also relatively slow and, by Chinese standards, expensive. Its published price of $3 per million fresh input tokens and $15 per million output tokens is far above that of several Chinese competitors. Calling it simultaneously the cheapest and most powerful model in the world would be unsustainable.

The stronger claim is narrower. A Chinese laboratory has produced a frontier class model despite controls intended to prevent precisely that outcome. Moonshot has also promised to release the model’s weights, allowing developers to adapt and deploy it without remaining wholly dependent on the company’s own service. Open weight does not mean fully open source: Moonshot is not necessarily disclosing its complete training data, code and development process. But it offers users considerably more control than a closed model available only through a corporate interface.

Demand reportedly became so intense after the launch that Moonshot temporarily suspended new subscriptions. That was evidence of enthusiasm, but also of weakness. China can create a large and capable model; serving it at scale remains constrained by limited computing capacity.

Washington’s accusation

The American response moved quickly from technical scepticism to allegations of industrial theft. Anthropic says Moonshot, DeepSeek and MiniMax conducted large scale attempts to extract capabilities from Claude, generating more than 16 million exchanges through approximately 24,000 fraudulent accounts.

Michael Kratsios, the White House science and technology policy director, went further. He alleged that Moonshot distilled Anthropic’s Fable model, operated a sophisticated system to avoid detection and obtained access to restricted Nvidia GB300 servers through Thailand. Treasury Secretary Scott Bessent said American watermarks had been discovered in Chinese models and warned that sanctions and Entity List designations could follow.

These are serious allegations, but they remain allegations. Distillation, in which one model learns from the output of another, is a normal technique used throughout the industry. The legal and ethical dispute concerns whether Moonshot obtained protected output through deception, violated access restrictions or misappropriated proprietary capabilities. American officials have not yet published the full technical evidence. Independent specialists have also questioned whether Fable, released shortly before Kimi K3, could have been the principal source of the Chinese model’s abilities.

Moonshot says its advances arose from original architectural changes. Beijing has accused Washington of “AI hegemonism” and threatened countermeasures. If the American evidence is eventually substantiated, Washington will have a serious intellectual property case. If sanctions arrive before that evidence is disclosed, China will argue that intellectual property has become a pretext for suppressing a successful competitor.

The $250 billion gamble

Even if every American allegation were proved, it would not settle the larger strategic question. Knowledge crosses borders more easily than physical processors. China has been forced to economise on computing power while the United States is committing itself to a scale of expenditure without historical precedent.

Nvidia is reportedly discussing a guarantee of roughly $250 billion to help OpenAI lease a proposed 10 gigawatt data centre complex being developed by SoftBank’s SB Energy in southern Ohio. The complete project, including the processors installed inside it, could cost more than $500 billion. Nvidia may also become involved in financing those processors.

The arrangement has not been completed, and it would not simply be a $250 billion loan from Nvidia to OpenAI. Nvidia would provide a financial backstop for lease and construction obligations. Nevertheless, the circularity is difficult to ignore. The world’s dominant AI chip supplier would be helping its customer obtain the infrastructure and financing required to purchase and operate more of the supplier’s chips.

This resembles vendor financing seen during earlier technological booms. The supplier supports the customer, the customer buys the supplier’s products and the resulting sales appear to justify further investment. Nvidia’s shares fell after the proposal was reported, reflecting investor anxiety about the risk moving from OpenAI’s balance sheet to Nvidia’s.

There is nothing inherently improper about such an arrangement. Nvidia is immensely profitable, demand for computing power remains strong and the Ohio complex could support valuable services for decades. The unresolved question is whether AI revenue will grow quickly enough to support liabilities of this magnitude. America is constructing the most capital intensive technological system in history before the commercial return on that system has been conclusively demonstrated.

Two theories of technological power

The American strategy rests upon frontier processors, enormous data centres, proprietary models, cloud subscriptions and the protection of intellectual property. Its strength is the integration of capital, hardware, software, universities and global corporate networks. Its weakness is cost.

China is increasingly competing through a different combination: open weight models, architectural efficiency, state supported industrial capacity, rapid domestic substitution and lower cost distribution. Its strength is diffusion. Its weakness is continued dependence on foreign semiconductor equipment, imported materials and hardware that still trails Nvidia’s best systems.

The division is not absolute. Meta, Nvidia, Microsoft and parts of the American developer community support open models. Anthropic and national security officials want tighter control. Nor is every Chinese system genuinely open. Releasing trained weights is not the same as disclosing the data and process by which a model was created.

The real contest is therefore not simply closed America against open China. It is a struggle over whether advanced intelligence remains a costly service controlled by a small number of corporations, or becomes an adaptable industrial commodity distributed across thousands of companies and governments.

The physical war beneath the software

Artificial intelligence is routinely described as if it existed in a weightless digital realm. In reality, it depends upon mines, refineries, fabrication plants, electricity grids, cooling systems and international shipping routes.

China dominates important stages of rare earth separation, permanent magnet production and electronics manufacturing. It can delay licences and restrict selected materials upon which American, Japanese and European industries depend. Exports of some controlled heavy rare earths to particular destinations have approached zero.

But claims that all Chinese rare earth exports have collapsed are false. China continued to export thousands of tonnes of rare earth magnets in June. Beijing is applying pressure selectively, not closing every outlet.

China has vulnerabilities of its own. It remains dependent upon advanced chipmaking machinery, electronic design software, extreme ultraviolet lithography and elements of the international semiconductor supply chain. Its temporary prohibition on helium exports is particularly revealing. China produces little helium and imports most of what it uses. The measure is principally defensive, intended to preserve domestic supplies during disruption. It may tighten the world market, but it is not evidence that Beijing commands helium as it commands rare earth processing.

Each side therefore possesses the ability to impose serious costs on the other. America controls critical technologies at the top of the semiconductor chain. China controls much of the industrial foundation beneath it. Neither possesses a complete and independent AI supply chain.

Apple meets the cost of economic nationalism

The contradiction inside the American strategy is visible in the dispute between Apple and Micron. Apple is reportedly seeking permission to use memory chips from the Chinese companies CXMT and YMTC in products sold outside the United States. It wants to contain costs and defend its position against Huawei, Xiaomi and other Asian manufacturers.

Micron, America’s principal memory producer, wants the administration to prevent the arrangement. It argues that Chinese competitors benefit from state support and could eventually destroy the remaining American memory industry.

Both companies have a rational case. Allowing Apple to buy Chinese memory could help one American company remain competitive while strengthening the Chinese semiconductor industry Washington is attempting to contain. Refusing permission would protect Micron while forcing Apple to absorb higher costs or pass them to consumers.

This is why reindustrialisation cannot be achieved by presidential instruction alone. Producing advanced technology in the United States remains more expensive than producing it in much of Asia. National security may justify paying that premium, but the cost must ultimately be carried by companies, consumers, taxpayers or the state.

Has China won?

No. America retains the most advanced processors, powerful cloud companies, deep financial markets and leading research institutions. China still lacks unrestricted access to the best lithography equipment. Kimi K3 does not lead every benchmark, and its early capacity problems reveal the continuing hardware constraint.

But the belief that export controls would freeze Chinese progress has been badly damaged. The restrictions strengthened the political importance of Huawei and domestic chipmakers. They encouraged Chinese laboratories to improve efficiency and gave Beijing a compelling reason to distribute open weight systems that reduce dependence on American services.

Human capital further complicates the picture. China educates an increasing share of the world’s leading AI researchers. MacroPolo estimated that 38 per cent of top researchers were educated in China by 2024. Yet American institutions employed approximately 59 per cent of elite researchers. China has become a principal training ground for AI talent while America remains its most attractive destination.

The United States has imposed genuine costs and probably delayed China’s hardware development. It has not prevented Chinese laboratories from approaching the frontier. Containment has bought time, but Washington has not yet shown how that time will be converted into a durable victory.

The countries beyond the frontier

Most countries will never build a 10 gigawatt computing complex. Their businesses cannot spend hundreds of billions of dollars on infrastructure, and many governments cannot afford indefinite dependence on the most expensive American models. Their choice will not be determined solely by which system leads a benchmark by a few percentage points.

China understands this. At the World Artificial Intelligence Conference it promoted a new international organisation involving 29 countries, promised 5,000 training places over five years and proposed AI application centres connected with ASEAN, the Arab League, the African Union, CELAC, BRICS and the Shanghai Cooperation Organisation. Beijing is offering models, training, infrastructure and development partnerships as elements of a single diplomatic programme.

The proposition is attractive: countries that cannot build frontier models can adapt Chinese systems to local languages, public services, agriculture, medicine and industry. Open weights may offer greater autonomy than permanent dependence on an American corporate interface.

America has a strong counterargument. Its systems may offer greater reliability, better security, clearer commercial accountability and deeper integration with international business. Microsoft chief executive Satya Nadella argues that institutional trust may outweigh the appeal of cheaper Chinese models. Governments will also consider censorship, surveillance, data security and the political influence accompanying Chinese infrastructure.

The contest is therefore between two offers. China is betting on affordability, adaptation and development partnerships. America is betting on technological superiority, institutional trust and the depth of its corporate ecosystem.

The dollar will not disappear

China’s growing technology exports may also encourage greater use of the renminbi. Approximately 30 per cent of Chinese goods trade is now settled in its own currency. Beijing can combine equipment, finance and technical assistance in contracts that avoid the dollar and reduce exposure to American sanctions.

But this does not mean that Chinese chips will destroy dollar supremacy. The dollar accounts for about 57 per cent of allocated global foreign exchange reserves. The renminbi remains below 2 per cent. China maintains capital controls, its currency is not fully convertible and its financial markets cannot provide an equivalent to the depth and liquidity of the American Treasury market.

Technology exports could contribute to a more plural international payments system. They cannot by themselves create a reserve currency trusted on the same scale as the dollar. Here, as in artificial intelligence, the transcript’s prediction of immediate Chinese victory confuses a shift in direction with the completion of a revolution.

A divided American establishment

Washington’s final difficulty is that the United States does not possess one coherent AI interest. Anthropic wants protection against alleged distillation. Nvidia wants access to global markets and warns that excessive restrictions accelerate Chinese substitution. Apple wants cheaper Chinese memory. Micron wants Chinese competitors excluded. Meta supports open weights. Security officials want tighter control of both processors and models.

These are not minor disagreements. They reflect a conflict between national security, corporate profitability, technological openness and consumer prices. America has not yet decided whether victory means possessing the best model, controlling the world’s computing infrastructure, preserving domestic manufacturing or ensuring that global developers continue to build upon American technology.

The United States may continue to own the technological frontier. It has every advantage required to remain there. But China does not need to surpass every American model in every test to transform the market. It needs systems that are capable enough, affordable enough and open enough to be adopted by countries and companies that cannot or will not pay for America’s frontier.

Kimi K3 has not proved that China has won the artificial intelligence race. It has demonstrated that America cannot win merely by withholding chips. The race will be decided not only by who creates the most intelligent machine, but by who can turn intelligence into an industrial system the rest of the world can afford.