China Isn’t Winning the AI Race. It’s Changing the Rules of the Game

Only a year ago, the prevailing view was that America’s leadership in artificial intelligence appeared firmly established. OpenAI, Anthropic and Google defined the frontier of large language models, while most competitors remained focused on narrowing the performance gap with the leading US labs.

Then the competitive landscape began to shift.

Rather than competing solely to build the most capable AI models, Chinese companies increasingly focused on lowering costs, releasing open-weight models and expanding developer access. In doing so, they have begun changing the economics of artificial intelligence itself.

The latest evidence is difficult to ignore. Over the past year, Chinese AI companies have released a series of increasingly capable large language models—from Moonshot AI’s Kimi K3 to DeepSeek’s R1, Z.AI’s GLM and Alibaba’s Qwen family. More striking than their benchmark performance is the way they are being adopted. Businesses, including some in the United States, are beginning to deploy these models because they offer a compelling combination of high performance, significantly lower inference costs and, in many cases, open-weight access that allows organisations to run and customise them on their own infrastructure.

The emergence of these models raises a much larger question than whether China has built another ChatGPT competitor. Has China discovered a fundamentally different path to AI leadership—one based not on producing the single most powerful model, but on making advanced AI cheaper, more accessible and easier for the world to adopt?

This shift is about far more than another chatbot launch. China’s AI strategy is gaining momentum not because it has overtaken the United States in frontier research, but because it is challenging three pillars of American AI leadership simultaneously: the cost of intelligence, the global talent pipeline and the future of open AI ecosystems. Understanding those three forces is essential to understanding the next phase of the global AI race.

Pillar One: China is Making AI Cheap Enough To Change the Market

For much of the generative AI boom, the industry assumed that building the most capable model would inevitably create the most valuable business. The race was measured in benchmark scores, reasoning ability and ever-larger training runs. American AI labs led that race, investing billions of dollars to push the frontier of artificial intelligence.

Chinese AI companies appear to have reached a different conclusion.

Rather than competing solely to build the world’s most powerful model, they are increasingly competing to make advanced AI dramatically cheaper and more accessible. That distinction may prove to be one of the most important developments in the AI industry.

The release of Moonshot AI’s Kimi K3, alongside DeepSeek’s R1, Z.AI’s GLM series and Alibaba’s latest Qwen models, illustrates this shift. While each model differs in architecture and capability, together they represent a broader strategy: deliver performance that is close enough to frontier models while reducing the cost of deploying artificial intelligence. Instead of asking whether a model is the absolute best, many businesses are beginning to ask a more practical question: Is it good enough for our workloads at a fraction of the cost?

That question is increasingly influencing enterprise AI decisions.

San Francisco-based AI assistant platform Lindy AI publicly explained why it switched from OpenAI to DeepSeek for parts of its operations. According to its chief executive, Flo Crivello, the move saved the company millions of dollars while improving performance across several core use cases. Whether or not every organisation reaches the same conclusion, the example highlights an important reality: for most businesses, the economic value of an AI model matters just as much as its benchmark performance.

This is changing how enterprises evaluate AI.

A large bank, retailer or software company does not simply compare benchmark scores on reasoning tests. It measures the total cost of ownership, including inference costs, infrastructure requirements, latency, reliability and the ability to customise a model for specific business needs. Once a model reaches a sufficiently high level of performance, the commercial advantage often shifts towards whichever option delivers the best value rather than the highest score.

That is precisely where open-weight models are changing the conversation.

Unlike proprietary AI systems that are accessed exclusively through an API, open-weight models allow organisations to download the model parameters, fine-tune them using their own data and deploy them on infrastructure they control. This gives companies greater flexibility over costs, performance, data governance and vendor dependence. Instead of paying continuously for every API request, businesses can increasingly optimise AI workloads on their own hardware or through third-party cloud providers.

The implications extend well beyond lower operating costs.

Open-weight models encourage experimentation. Developers can modify them, researchers can improve them and enterprises can integrate them into specialised workflows without waiting for a model provider to release new features. Every successful deployment strengthens the surrounding ecosystem, attracting more developers, more optimisation tools and more enterprise adoption. In technology markets, ecosystems often become competitive advantages in their own right.

This is not the first time the technology industry has witnessed such a shift.

Linux became the operating system underpinning much of the world’s cloud infrastructure despite not being owned by a single commercial company. Android grew into the dominant mobile operating system by enabling a broad ecosystem of manufacturers and developers. In both cases, openness accelerated adoption and created network effects that proprietary competitors struggled to match.

Chinese AI companies appear to be applying a similar logic to artificial intelligence. Rather than relying solely on premium subscription revenue, they are helping establish an ecosystem in which advanced AI becomes cheaper to deploy, easier to customise and more widely available. If that strategy succeeds, the centre of competition may gradually move away from who builds the single smartest model towards who enables the largest number of organisations to use AI effectively.

For American AI leaders, this presents a different kind of competitive challenge. Maintaining leadership will require more than producing the next frontier model. It will also require demonstrating that proprietary systems can continue to justify their premium pricing in a market where increasingly capable alternatives are becoming both cheaper and more accessible.

History suggests that industries are rarely transformed by the company that builds the most advanced technology alone. More often, they are reshaped by those that make that technology affordable enough for everyone else to use.

Pilar Two: Open Models Are Becoming China’s Soft Power

Technology companies have long understood that the most valuable product is not always the one that generates the highest immediate revenue. In many cases, the greatest advantage comes from building the platform that everyone else chooses to build upon.

This distinction is becoming increasingly important in artificial intelligence.

Most leading American AI companies, including OpenAI and Anthropic, primarily distribute their models through proprietary APIs. Developers gain access to powerful capabilities, but the models remain under the control of the companies that created them. Every application built on these models depends on their pricing, infrastructure, usage policies and future product decisions.

Chinese AI companies are pursuing a noticeably different strategy.

By releasing open-weight models such as Kimi K3, DeepSeek R1, Qwen and GLM, they are not simply offering another alternative to proprietary AI services. They are giving developers, researchers and enterprises the ability to download, fine-tune and deploy these models on infrastructure they control. In doing so, they are encouraging adoption rather than dependence.

That difference may appear subtle, but it has profound strategic implications.

Throughout the history of computing, open technologies have often become the foundation on which entire industries are built. Linux powers much of the world’s cloud infrastructure despite not being controlled by a single technology company. Android became the world’s most widely used mobile operating system because manufacturers could adopt and customise it without building an operating system from scratch. TensorFlow and PyTorch accelerated advances in machine learning by giving researchers and developers common frameworks on which to build and share new ideas.

Artificial intelligence may now be entering a similar phase.

Every developer who fine-tunes an open model, every startup that builds products on top of it and every enterprise that integrates it into its operations contributes to a growing ecosystem. As more organisations adopt the same family of models, a wider network of optimisation tools, research, documentation and specialised applications begins to emerge. The value no longer lies solely in the model itself, but in the ecosystem that grows around it.

This is where China may be pursuing a form of technological soft power.

Soft power is often associated with culture. Hollywood films, South Korean K-pop and Japanese consumer electronics have all shaped global influence by becoming part of everyday life far beyond their national borders. Technology can create a similar form of influence. When developers around the world build on a common platform, they help establish its standards, expand its ecosystem and reinforce its position within the global technology landscape.

If Chinese open-weight models become the default foundation for thousands of AI applications, China gains something more valuable than short-term licensing revenue. It gains developer mindshare. Universities teach the models. Startups optimise for them. Enterprises invest in tooling around them. Researchers publish improvements built upon them. Each of these decisions strengthens the ecosystem and makes it more attractive for the next wave of users.

For American AI companies, this presents a strategic challenge that extends beyond model performance. A proprietary model may generate higher revenue per customer, but an open ecosystem can spread more rapidly, attract a larger developer community and establish itself as an industry standard. History has repeatedly shown that platforms with broad adoption often become more influential than products with the highest technical specifications.

The next phase of the AI race may therefore be determined by more than who builds the most intelligent model. It may be decided by who builds the ecosystem that the rest of the world chooses to build upon.

Pillar Three: The Return of China’s AI Talent

In 2019, Yang Zhilin faced a career decision that many aspiring AI researchers would have considered obvious.

Having completed his PhD at Carnegie Mellon University in just four years—roughly two years faster than the typical programme—he had opportunities that most researchers spend an entire career pursuing. Apple wanted to hire him. He had already interned at Google and Meta. Postdoctoral positions at Stanford and the Massachusetts Institute of Technology were available to him. By every conventional measure, Silicon Valley offered the most attractive path.

He chose Beijing instead.

Yang returned to China and founded Moonshot AI, naming the company after Pink Floyd’s The Dark Side of the Moon. His former PhD supervisor, Russ Salakhutdinov, later recalled Yang’s reasoning: if he did not at least try building his own company, he would regret it for the rest of his life.

At the time, the decision appeared unconventional. Today, it looks increasingly significant.

Last week, Moonshot AI released Kimi K3, a frontier large language model that has drawn comparisons with leading systems from OpenAI and Anthropic. Even Elon Musk described the release as “impressive work”. Yang’s story is no longer simply about one entrepreneur returning home. It illustrates a broader shift in where some of the world’s most talented AI researchers believe they can build globally competitive companies.

It would be tempting to explain this solely through American immigration policy. After all, Silicon Valley has long depended on attracting the world’s brightest scientists and engineers. But Yang’s story suggests the reality is more nuanced. According to Salakhutdinov, Yang had multiple opportunities to remain in the United States. Apple even explored employing him through its Beijing office, allowing him to stay closer to home while working for an American technology giant. His decision was not driven by a lack of opportunity in the United States. It reflected a belief that China had become a compelling place to build an AI company.

That perception appears to be shared by an increasing number of founders and researchers.

Several Chinese AI entrepreneurs have argued that building a startup in China offers advantages that extend beyond familiarity with language and culture. Professional networks are often stronger, recruitment can be faster and the domestic pool of AI engineers has expanded dramatically over the past decade. Some founders also point to a more unified regulatory framework and the ability to deploy AI products more quickly across large domestic markets.

None of this suggests that building an AI company in China is easy. Chinese startups continue to face restrictions on access to advanced semiconductors, growing geopolitical tensions, the possibility of overseas bans and a fundraising environment that remains less mature than that of the United States. Yet some entrepreneurs believe those disadvantages are outweighed by the opportunity to develop and commercialise AI at extraordinary speed.

The talent pipeline itself is also changing.

For decades, China relied heavily on researchers who studied or worked abroad before bringing their expertise home. Today, that dependence is beginning to diminish. Research by Stanford’s Hoover Institution into the team behind DeepSeek’s core AI models found that more than half of the researchers had never studied or worked outside China. Among those who had experience in the United States, the majority eventually returned to China.

At the same time, the flow of experienced researchers back to China appears to be strengthening. A 2024 Stanford survey of more than 1,300 US-based scientists of Chinese descent found that many no longer felt fully secure in the American academic environment. Nearly three-quarters reported feeling unsafe as researchers, while many cited concerns about anti-Asian discrimination and violence. Although these concerns represent only one factor influencing career decisions, they illustrate that talent migration is shaped by social and political conditions as well as economic opportunity.

The larger story is not that America is suddenly losing all of its AI talent. The United States remains home to many of the world’s leading universities, research laboratories and technology companies. Rather, China has become a credible alternative destination for researchers who once viewed Silicon Valley as the only place to build ambitious AI companies.

That shift has profound implications.

Artificial intelligence is ultimately a talent-driven industry. Breakthrough models are not created by governments or venture capital alone; they are built by researchers, engineers and entrepreneurs. If China can educate more of its own AI talent, persuade internationally trained researchers to return and provide an environment where they believe they can build globally competitive companies, its innovation capacity will become increasingly self-sustaining.

For decades, China’s technological rise depended largely on absorbing knowledge from abroad. Today, it is doing something more significant. It is retaining world-class talent, cultivating a new generation of researchers at home and increasingly creating the innovations that others are trying to match

Part IV: Why Open AI Has Become a National Security Debate

China’s rapid progress in open-weight artificial intelligence has triggered a debate that extends far beyond technology. It has become a question of national security.

In Washington, concerns about Chinese AI are increasingly framed around surveillance, intellectual property theft, cyber espionage and the possibility that critical infrastructure could become dependent on technologies developed by strategic competitors. Reports suggest that the White House has explored measures ranging from restricting the spread of Chinese open-weight models to imposing sanctions on some of them. The underlying concern is straightforward: if artificial intelligence becomes as fundamental to the digital economy as operating systems or cloud computing, should a geopolitical rival be allowed to provide that foundation?

These concerns deserve serious consideration.

However, understanding the risks requires distinguishing between how an AI model is created and how it is deployed. Those two issues are often treated as if they are identical, but from a technical perspective they are not.

The distinction begins with model weights.

Model weights are the numerical parameters learned during training that enable an artificial intelligence model to generate text, write code or perform reasoning tasks. When a company releases an open-weight model, those weights can be downloaded, fine-tuned and deployed by anyone with sufficient computing resources. Unlike a proprietary AI service, the model does not have to remain connected to the company that originally developed it.

That leads to a second distinction: hosting.

If an American enterprise downloads an open-weight Chinese model and runs it entirely within its own data centres or on a trusted domestic cloud provider, its data does not automatically flow back to China. Customer information, internal documents and business workloads can remain inside infrastructure controlled by the organisation itself. From a data governance perspective, this creates a very different risk profile from using a proprietary AI service hosted and operated overseas.

Inference is equally important.

Inference is the process of running a trained model to generate responses. When inference occurs on infrastructure controlled by the organisation using the model, the organisation—not the original model developer—typically controls how data is processed, stored and secured. In practical terms, this means that the location where the model is executed often matters more for data security than the country in which the model was originally trained.

Open-weight models also allow fine-tuning.
Organisations can adapt a model using their own datasets, improve performance for specialised tasks and modify safety or content restrictions to meet local regulatory requirements. This flexibility explains why open-weight models have attracted significant interest from developers and enterprises. It also means that two deployments of the same model may behave very differently depending on how they have been customised.This distinction also complicates another common assumption—that Chinese AI models are inherently subject to Chinese government censorship.

When accessed through services hosted inside China, some models may refuse to answer politically sensitive questions in accordance with domestic regulations. However, when the same open-weight models are downloaded and hosted elsewhere, developers can often modify or remove those restrictions during deployment or fine-tuning. The behaviour of an open-weight model is therefore influenced not only by who created it, but also by who operates it and under what policies.

None of this eliminates legitimate security concerns.

Chinese companies operate within a legal environment that includes the National Intelligence Law, under which organisations may be required to cooperate with state intelligence authorities. That reality understandably influences how governments assess strategic risk. At the same time, the relationship between companies and the state is not always as straightforward as it is often portrayed. Earlier this year, reports emerged that government inspectors attempting to obtain access to data from the e-commerce platform Pinduoduo encountered resistance from company employees, highlighting that legal obligations do not necessarily translate into unrestricted access in every situation.

The larger challenge for policymakers extends beyond the nationality of any individual model.

Artificial intelligence is rapidly becoming a powerful tool for cyber operations. Security researchers are already using open-weight models to analyse sophisticated cyber attacks. Following a recent AI-agent security breach, Hugging Face analysed the incident using an open Chinese model without any data leaving its own systems. The episode demonstrated both the capability of modern open-weight AI and the importance of secure deployment practices.

This shifts the debate away from a simple question of whether Chinese AI models can be trusted. A more useful question is whether governments and organisations have developed the technical safeguards needed to deploy any advanced AI model securely, regardless of where it originated.

Ultimately, the greatest security challenge may not be that China has built powerful open-weight models. It may be that governments are still approaching AI primarily as a geopolitical contest while the technology itself is becoming increasingly decentralised. In a world where model weights can be downloaded, modified and run almost anywhere, security will depend less on the passport of the developer and more on who controls the infrastructure, the data and the deployment environment.

The Bigger Question: Can America Sustain Its AI Leadership?

Taken individually, none of these developments suggests that the United States has lost its leadership in artificial intelligence. American companies continue to build many of the world’s most capable foundation models. They dominate private AI investment, possess unmatched hyperscale cloud infrastructure and remain home to the world’s leading AI laboratories, universities and semiconductor companies.

The more important question is whether those advantages are sufficient to preserve long-term leadership in an industry that is evolving far beyond frontier model performance.

For decades, America’s technological dominance rested on a remarkably successful formula. It attracted the world’s brightest researchers, connected them with abundant venture capital, allowed them to build companies at global scale and benefited from a regulatory and financial ecosystem that rewarded innovation. Silicon Valley became not only a place where technology was invented, but also where it was commercialised.

That formula is now facing pressure from several directions.

Immigration has long been one of America’s greatest competitive advantages. Many of the entrepreneurs and researchers behind today’s AI revolution were born outside the United States. Yet, as the story of Yang Zhilin illustrates, attracting exceptional talent is no longer enough if that talent increasingly chooses to build companies elsewhere. The American immigration system remains largely designed around employment rather than entrepreneurship, making it easier to hire skilled workers than to encourage them to found the next generation of globally significant technology companies.

At the same time, China’s domestic innovation ecosystem has matured considerably. It no longer depends exclusively on researchers trained overseas, nor does it rely solely on copying technologies developed elsewhere. Its universities are producing more AI researchers, its technology companies are investing aggressively in frontier research and an increasing number of founders believe they can build globally competitive businesses without leaving China.
Competition is also shifting from proprietary products to open ecosystems.

American AI companies have largely pursued business models built around proprietary APIs and premium subscriptions. Chinese companies, by contrast, are investing heavily in open-weight models that encourage broad adoption by developers and enterprises. Neither approach is inherently superior, but they compete on different dimensions. One maximises revenue from individual customers, while the other seeks to maximise participation across an entire ecosystem. History shows that technological leadership often belongs not only to those who invent the best products, but also to those who establish the platforms on which others build.
Capital remains one of America’s strongest advantages. Its venture capital ecosystem, public markets and institutional investors continue to provide levels of funding that few countries can match. Frontier AI research requires enormous financial resources, and the United States remains exceptionally well positioned to finance companies pursuing increasingly expensive training runs and advanced computing infrastructure.

Yet capital alone cannot guarantee leadership.

Artificial intelligence is becoming an infrastructure industry as much as a software industry. Training and deploying frontier models now requires massive data centres, advanced semiconductor supply chains, reliable electricity networks and long-term investment in energy generation. The race is increasingly being shaped by who can build the physical infrastructure required to support the next generation of AI systems.

Governments are also playing a more prominent role.

Artificial intelligence is no longer viewed solely as a commercial technology. It is increasingly regarded as a strategic national capability with implications for defence, intelligence, cybersecurity and economic competitiveness. This has intensified public investment, export controls and industrial policy on both sides of the Pacific. Success will depend not only on the strength of private companies but also on how effectively governments support innovation while managing security risks.

These developments suggest that the global AI race is becoming more distributed than many expected.

The United States continues to lead in frontier research, venture capital, advanced semiconductor design and many of the technologies that underpin modern artificial intelligence. China, meanwhile, is strengthening its position through a combination of lower-cost models, open ecosystems, expanding domestic talent and rapid commercial deployment. Leadership is no longer defined by a single metric, nor is it likely to belong permanently to one country.

The central question, therefore, is not whether America is losing the AI race. It is whether the next era of artificial intelligence will still revolve around a single centre of innovation, or whether leadership is evolving into a more competitive, multipolar ecosystem in which different countries shape different parts of the AI value chain.

If that is the future, then America’s greatest challenge is not simply building the next frontier model. It is ensuring that it remains the world’s most attractive place to discover, finance, deploy and scale the technologies that define the age of artificial intelligence.

What Investors and Business Leaders Should Watch

The developments discussed throughout this analysis are not merely geopolitical stories. They have the potential to reshape the economics of enterprise software, cloud computing and artificial intelligence over the coming decade. Whether China’s strategy ultimately succeeds or not, it is already forcing businesses, investors and policymakers to reconsider long-held assumptions about how AI will be developed, deployed and monetised.

For businesses, the most immediate implication is cost.

If Chinese open-weight models continue to narrow the performance gap with the leading proprietary systems while remaining significantly cheaper to deploy, enterprise AI economics could change rapidly. Many organisations do not require the absolute best model available; they require a model that is reliable, secure and cost-effective for their specific workloads. As more capable open-weight models become available, CIOs and technology leaders may gain greater bargaining power when negotiating with proprietary AI providers. Competition could place sustained pricing pressure on premium AI services while making advanced AI affordable for a much broader range of businesses.

Lower costs could also accelerate AI adoption across industries that have so far struggled to justify large-scale deployments. Manufacturers, retailers, logistics companies, financial institutions and public-sector organisations may find that open-weight models offer a more attractive balance between performance, flexibility and operating costs. If AI becomes substantially cheaper to deploy, the next phase of adoption may be driven less by technological breakthroughs and more by improved business economics.

For investors, however, the story extends beyond lower prices.

The emergence of increasingly capable Chinese models suggests that future competitive advantage may depend less on who builds the single most advanced model and more on who controls the surrounding ecosystem. Companies that provide AI infrastructure, cloud services, specialised semiconductors, data-centre capacity, cybersecurity, developer tools and enterprise integration may benefit regardless of which foundation model ultimately dominates the market. As AI matures, value creation is likely to spread across the broader technology stack rather than remain concentrated within a handful of frontier model developers.

At the same time, geopolitical risk is becoming an investment consideration in its own right. Governments are paying closer attention to the origin of AI technologies, particularly when they are deployed in critical infrastructure or sensitive industries. Future regulations, export controls and procurement rules could significantly influence which models enterprises are willing—or permitted—to adopt. Investors should therefore evaluate not only technological capability but also regulatory resilience and geopolitical exposure.

Over the next 12 to 24 months, four developments deserve particularly close attention:
1. Enterprise adoption of Chinese open-weight models.
Will more American and multinational companies follow early adopters by deploying models such as Kimi K3, DeepSeek, Qwen or GLM in production environments? Broader adoption would indicate that cost and performance are outweighing geopolitical concerns for many commercial use cases.

2. The economics of frontier AI.
Will competition continue to drive down inference costs and subscription pricing for advanced AI models? If Chinese open-weight models place sustained pressure on proprietary providers, pricing could become one of the industry’s most important competitive battlegrounds.

3. Global AI talent flows.
Will more internationally trained Chinese researchers choose to return home to establish companies and lead research teams? A sustained increase in return migration would strengthen China’s long-term innovation capacity and reinforce the talent ecosystem underpinning its AI industry.

4. Government policy and market access.
Will the United States and its allies introduce stricter rules governing the deployment of foreign-developed AI models in government agencies, critical infrastructure or regulated industries? Such policies could shape enterprise adoption just as much as technological capability.

Ultimately, the next chapter of the AI race may not be decided solely by who builds the most intelligent model. It may be determined by who can lower the cost of intelligence, attract the world’s best talent, build the strongest developer ecosystem and navigate an increasingly complex geopolitical landscape. Those are the signals that business leaders and investors should be watching most closely.

Conclusion

It is still too early to conclude that China has overtaken the United States in artificial intelligence.

America continues to lead in many of the areas that matter most: frontier research, advanced semiconductor design, venture capital, world-class universities and many of the companies defining the cutting edge of AI. Those advantages remain substantial, and they should not be underestimated.

Yet focusing solely on who builds the most capable model risks missing the more important transformation taking place.

China appears to be pursuing a different strategy. Rather than competing only at the frontier of AI research, it is lowering the cost of intelligence, expanding access through open-weight models, strengthening its domestic talent pipeline and encouraging the growth of a global developer ecosystem. That strategy does not necessarily replace America’s strengths, but it does challenge the assumptions on which much of its AI leadership has been built.

History suggests that technological revolutions are rarely won by innovation alone.

The companies that invented the personal computer did not dominate every stage of the computing industry. The pioneers of the internet did not capture all of its economic value. Open-source software reshaped enterprise computing not because it was always technically superior, but because it changed the economics of adoption. Time and again, industries have been transformed not only by those who created breakthrough technologies, but by those who made them accessible, affordable and indispensable.

Artificial intelligence may now be approaching a similar turning point.

The next phase of competition is unlikely to be defined solely by benchmark scores or the release of the next flagship model. It will increasingly be shaped by who attracts the world’s best researchers, who builds the most trusted AI infrastructure, who creates the strongest developer ecosystem and, perhaps most importantly, who makes advanced AI affordable enough to become a utility rather than a luxury.

This is why the question is no longer whether China is winning the AI race.

The more important question is whether China is changing the rules of the race itself.

If the future of AI is determined not only by intelligence but also by cost, openness, talent and adoption, then leadership will no longer belong exclusively to the country with the smartest model. It will belong to those who can make artificial intelligence the foundation upon which the rest of the world chooses to build.

That is the competition now unfolding—and its outcome will shape not only the future of technology, but also the balance of economic and strategic power in the decades ahead.