This article by Farrell Gregory was published at Palladium Magazine on October 6, 2026.
Become a PALLADIUM member today! You will receive our beautiful quarterly print edition and help support our non-profit publication.
Moonshot AI, one of China’s largest artificial intelligence labs, unveiled its latest model Kimi K3 the day before Xi Jinping’s speech at the World AI conference (WAIC) in July 2026. It may have been pure coincidence, but the sequence of events struck a few in the U.S. as a definitive moment when something about the future of Chinese machine intelligence was revealed. “We should adhere to the principle of openness,” Xi said, and “seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing.”While not a change in the established state perspective, many in the West believed that China would eventually have to move away from their open-source approach to keep up with the rising cost of rivaling frontier capabilities. The speech firmly established that China was not planning on changing anything anytime soon.
Encouraged by billions of dollars worth of compute vouchers and other support for small firms, Chinese labs and entrepreneurs have for the last few years pursued a strategy involving a multitude of labs releasing capable open-weight models, utilizing both highly innovative methods of development and distillation from American frontier models. Below these “foundation” models, there is a constantly shifting substrate of smaller firms that specialize intelligence for particular applications.
Though public statements from Chinese lab leads and the capabilities of their models are visible from America, an enormous delta exists between American rhetoric and the reality of Chinese AI. Only in discrete moments can the hazy fog of expectation be punctured by a jolt of undeniable reality.
For many in the English-speaking world, there has been a consistent expectation that once some threshold of capability is surpassed, China will act to restrain open-weight development. Whether this theory is justified by concerns for social stability, national security, or hypothetical existential risks, they believe there will be a “Mythos moment” in China: a sudden shift in AI policy once open models develop dangerous potential. This theory was presented in highly influential papers such as “Situational Awareness,” written by the German investor and former OpenAI employee Leopold Aschenbrenner, which in the summer of 2024 correctly predicted that confidence in an enduring American lead in AI was misplaced. Once China wakes up to the strategic implications of superintelligence, the essay argues, the previously dissonant mechanisms of Chinese firms and the Party will harmonize and organize the whole Chinese nation in pursuit of the transformative technology.
This argument might seem hard to disagree with, for as capabilities grow rapidly in the age of recursive self-improvement, the machine intelligence ecosystem will undergo an algae bloom. Such tremendous and varied capabilities, the thinking goes, will prompt a rapid, dramatic transformation in Chinese AI governance. Yet, even as Chinese and American model capabilities continue to improve, the Chinese state continues to defy such predictions.
Open-Weights Under State Oversight
The history of Chinese AI governance can be read in state documents filled with slogans and measured statements. Public pronouncements and comments on AI reveal a consistent focus on practical application and cautious advancement rather than the pursuit of transformative superintelligence. In 2017, China released its Next Generation Artificial Intelligence Development Plan, which set the ambitious goal of becoming a world leader in AI by 2030. Importantly, this was framed in terms of industrial competitiveness and economic output, not the pursuit of artificial general intelligence.
Then, at a 2018 Politburo collective study session, Xi Jinping stated that China must march “in the front ranks when it comes to theoretical research” and occupy “the high ground in critical and AI core technologies.” The following years were comparatively mute. Finally, in 2025, at the first collective study session to focus on AI in seven years, Xi’s statement was focused on policy support for technical self-sufficiency, a sharp contrast to the U.S.-China Economic and Security Review Commission’s call a year earlier for a “Manhattan Project-like program dedicated to racing and acquiring an Artificial General Intelligence (AGI) capability.”
America’s 2025 AI Action Plan then went on to model development as a winner-take-all race between America and China on the basis of LLM scaling, stating that “whoever has the largest AI ecosystem will set global AI standards and reap broad economic and military benefits.” A recent strategy document released by the White House Office of Science and Technology Policy mentions artificial general intelligence as a transformative technology, which “could present fundamentally new opportunities and threats.” Combined with the widespread perception that AGI is imminent for better or worse, the American AI strategy can reasonably be summed up as an LLM-scaling race towards superintelligence.
This does not seem to be how the Chinese see it. The Chinese AI+ Plan doesn’t mention competition at all, nor does it even mention America. Its focus is on diffusion, standards-setting, and making AI a value-add for other existing parts of the economy. Current Chinese AI policy involves decentralized distribution of benefits allocated by provinces and municipalities, supporting an open-weight ecosystem, and a multiplicity of regional labs. We should not necessarily take these documents at face value, but we should question why China’s AI narrative seems to ignore the “race to dominance” framework.
Today, the Chinese state retains a multitude of regulatory tools to intervene in model development and deployment. Generative AI models are subject to twin pre-deployment registration and review regimes. These include continual model testing and evaluation for security by the relevant provincial and then central Cyberspace Administration of China (CAC). The role of the CAC here is primarily licensing for algorithms, though they can also take direct action and force tech companies to maintain compliance or adjust their models. A recent campaign of theirs prompted Alibaba, Zhipu, MiniMax, and DeepSeek to take actions to combat malicious code generation and other cyber threats. Much of the reporting on China’s cybercrime enforcement and state capacity comes from state media, and thus should be read with a healthy dose of skepticism.
These regulatory tools, considered to be part of the wider cybersecurity state that exists in the PRC, help explain some of the present facts. As model capabilities remain comparable in America and China, the latter may, for the reasons outlined above, be able to sustain wide domestic access to highly-capable open-weight models.
If Xi Jinping’s latest speech at the WAIC is to be believed, Chinese AI governance will remain the same for the near future—open-weight models with formidable capabilities, distributed widely, and all under human control. No nationalization of the labs, and no Manhattan Project in Inner Mongolia, at least not yet. So how did American technologists misread China? And might China diverge even further from Western expectations?
The Power and Limitations of Bay Area Rationalism
If there is a constant historical flaw of U.S. foreign policy, it is assuming that other people, leaders of nations, and their citizens think the same way we do. Nowhere has that dynamic been more destructive than in U.S. policy towards China. As outlined in sociologist Richard Madsen’s book China and the American Dream, from the early American missionaries—in whose estimation China was always on the precipice of converting—to the ecstatic marketeers who expected political liberalization to follow economic liberalization, American thinkers have long held that China is primed to think and respond to material conditions as we do, if only given a bit of time and faith.
As a result of such thinking, policies that created a trade imbalance and supply chain overreliance remained in place for decades, fortified by assumptions of developmental symmetry with China that did not exist. Even as the American perspective on U.S.-China relations shifted from convergence to competition in the 2010s, we never stopped projecting our fundamental assumptions regarding economics, governance, and technology onto the Chinese. While such mistaken policy over the previous decades was always costly, America’s relative strength and prosperity masked the true impact of misconstruing China again and again. But today, with China generally recognized as a peer competitor in all forms of innovation, the consequences of making the same mistakes in Chinese policy will be immediately evident, and perhaps more irreversible.
Published two years before the inflection point it describes, the viral forecasting scenario “AI 2027,” authored by the AI Futures Project, commanded attention from Silicon Valley to the White House. It accurately predicted the exponential pace of the machine intelligence buildout and the suffusion of autonomous agents into the digital realm. Beyond technical predictions, the authors also anticipated a radical shift in Chinese governance of technology.
According to the scenario, the General Secretary decides in mid-2026 that China cannot afford to continue lagging behind America in AI capabilities, leading to the nationalization of Chinese frontier research. It starts as information sharing, and then becomes a formalized collective lab within a year. The researchers, compute, and supporting infrastructure are all conglomerated and orchestrated by the central government as part of the Party’s all-out attempt to remain competitive in AI and overcome its compute deficit. Yet as the authors of “AI 2027” themselves admit, this has not happened yet. Too strict an adherence to a particular vein of “rationality” may explain why American forecasters are still getting China wrong.
Bay Area rationalism spread across an influential group of blogs such as Overcoming Bias, LessWrong, and Slate Star Codex—the author of which was listed among the authors of “AI 2027.” In the 2000s, these writers were already concerned about the development and potential influence of artificial intelligence, more than a decade before the release of GPT-3. Thought experiments on AI risk, such as Nick Bostrom’s paperclip maximizer and orthogonality thesis, spread widely among early rationalists. In San Francisco, the language, expectations, and methods of rationalism have suffused much of the discourse both online and offline for years. As a result, the rationalist perspective is greatly overrepresented both among those working at major AI labs as well as those employees’ extended social circles. As AI becomes an increasingly mainstream issue—the subject of Congressional letters and daytime TV panels—more and more people are using rationalist terminology and intellectual framing, almost certainly without knowing where these ideas originated.
Rationalism is primarily concerned with epistemics, or the examination of how humans come to know and believe certain things. Instinctual thought is full of errors and dead ends, which rationalists try to overcome through clear, analytical reasoning in pursuit of being “less wrong.” This way of thinking lends itself to a highly structured and quantifiable mode of analysis. Thus, it is unsurprising that the rationalist community was a prime source of talent for the Effective Altruist movement, which strives for efficient philanthropy and personal optimization to maximize their perception of what is good. The idea of expressing expectations as probabilities, such as having a “p(doom)” that reflects one’s belief in the likelihood of AI-induced human extinction, is downstream of Bay Area rationalism. This outlook, that would be unusual in any other environment, tends to be the cultural default in Berkeley group houses, American AI labs, safety evaluators, and forecasting teams.
There is a particular aspect of rationalist thinking that seems primarily responsible for incorrectly forecasting Chinese reactions to machine intelligence. Rationalists have an expectation that different people and polities will have similar utility functions for the same technology once it has reached sufficient scale. Rational actors observing shared evidence will experience similar responses, so the thinking goes. They believe that for predictions cast forward over years, the best thing to do is assume mutual rationality and deemphasize the details that make potential responses to transformative technology distinct. This is a natural consequence of coarse-grained analysis, which requires deemphasizing details in order to identify larger trends and patterns.
In this view, someone who cites particular governing dynamics, personalities, or other contingent factors might be accused of missing the forest for the trees. It is certainly difficult to predict how Chinese researchers and political leadership might look at the same capabilities that we do and choose an entirely different course of action. But assuming a shared rational outlook encourages, if not requires, a blindness to the ways in which certain domestic and cultural factors may cause Chinese people to thoughtfully and sincerely see the technology in a different light. To ignore this, even willfully and with the best of intentions, is still a mistake.
In a recent quarterly timeline update, the authors from the AI Futures Project briefly address this, stating that “we probably would have heard by now if the CCP had consolidated the various Chinese AI projects…in general it seems that ‘China Wakes Up’ has not yet happened.” While acknowledging incorrect forecasting is admirable, more must be done to update underlying assumptions, not just timelines they produce.
As evidence demonstrates, we cannot expect China to converge on Western expectations anytime soon. Without understanding Chinese perspectives on machine intelligence, the CPC will continue to surprise American policymakers and intellectuals. One of America’s greatest strategic mistakes over the last fifty years was basing America’s economic and trade policy on the idea that greater economic liberalization in China would result in political liberalization. The result was the deindustrialization of America and the creation of countless supply chain reliances that are still used as leverage against the U.S. today. In the early stages of the transformation that machine intelligence promises, further policy based on a misapprehension of China would be even more devastating.
The Other Side of the Pacific
In Cybernetics for the 21st Century, philosopher Yuk Hui distinguishes the technological differences between the U.S. and China. Hui divides the characteristics of any technology into “tendencies” and “facts.” Technical tendencies are the universal principles that govern how a given technology develops and operates. Technical facts are the contingent factors—the multitude of particular ways of being for any given technology. While the terminology may seem counterintuitive, as the names of each concept could easily be swapped, “tendencies” in this context represent attractor basins; such as the cross-cultural evolution of tools like arrows and wheels. Whereas “facts” are incidental features that vary from one circumstance to another due to different initial conditions.
Applying this framework to machine intelligence, it can be said that the core, underlying technology of neural networks and deep learning is a technical tendency, shared from Silicon Valley to Beijing. It is the application and interpretation of that technology, made evident during the process of development and diffusion, where the technical facts emerge. One’s degree of concern about existential risk is a technical fact, as are fears of labor displacement and human-directed cyber attacks. The technology itself is the tendency, while the broader theoretical expectations are contingent cultural facts.
It seems that purely rationalist predictions of Chinese AI overestimate the technical tendencies and underestimate the technical facts. And what facts they are! How particular technology and its use in China is becoming. Look at the robotic Olympics in Beijing, where humanoids gallop unnaturally and break sprinting world records, occasionally bursting into sparks and shrapnel. Look at the constant churn of talent and share prices, the unceasing reconstruction, reinvention, and derivation of generations of models in the open-weight ecosystem, with mechanical patrilineage stretching back to Moonshot, DeepSeek, and perhaps even the labs in San Francisco.
It could be that this is just a matter of distinguishing between timelines, and Chinese AI governance will dramatically change in response to novel publicly-available capabilities. As models recursively self-improve and continual learning becomes a closer reality, it seems folly to expect the current paradigm to hold. But within the distance between that future and the present, there exists such a wide range of possibilities. Even if you think it is a foregone conclusion that China will “wake up” to the American perspective on AI, does it not matter tremendously whether that occurs in two months or two years? And when they do, why would anyone be certain that their reaction will mirror America’s?
The most immediate implication of the continued availability of highly-capable open-weight Chinese models is that it cuts into the moat that American frontier labs currently have, alongside the profits and expected value that this moat would ensure. That value, diminished by the availability of Chinese models, is what enables labs to command the capital they do, whether for hiring researchers, developing future models, securing compute and building infrastructure, or serving current inference. Additionally, infrastructure providers—from traditional cloud services to neoclouds and the manifold firms involved in the American AI buildout—have similar capital access predicated on those expectations.
This is fundamentally a question of where value accrues given model usage. Industry analyst reporting from earlier this year indicated that Anthropic’s margins on inference rose from 38% to 70%, growing alongside revenue. As American labs continue to lead at the frontier, highly complex and specialized inference will likely remain increasingly profitable and a continuing advantage. But how much model usage will this cede to commoditized and lower-complexity competitors? As Anthropic prepares its public offering for later this year and OpenAI is set to follow in 2027, the share of capital going towards inference and other financial indicators will transition from company leaks to a matter of public disclosure.
Dean Ball, Head of Strategic Futures at OpenAI, has stated that open-weight is decelerationist, which, constrained to the narrow sense of the AI buildout, seems possible. Imagine tremendous revenue and enormous profits only at the edge; price competition decreases the immediate value that frontier models will accrue to labs. There is a plausible world where machine intelligence becomes too cheap to meter for most usage, too commoditized, yet central to everything. This is hypothetical, but less and less so every day. And for every day that goes by, while Chinese labs maintain open-weight and low-cost models, American frontier labs are increasingly forced to compete on price if they want to remain widely used.
Chinese labs seem to be actively working to monetize the present paradigm, signaling that they do not anticipate a near-term change in governance. They are finding ways to remain open-weight yet capture value. As it released the weights of Kimi K3, Moonshot AI also updated the terms of its licensing agreement to require non-Moonshot API users who use the “model as a service” to renegotiate commercial usage if the user’s business generates over $20 million in revenue per year. The license also requires businesses with over 100 million monthly users or meeting the revenue requirement to display Kimi K3 branding. Further, after lowering V4 costs by seventy-five percent earlier this year, DeepSeek now seems likely to substantially raise prices for API usage.
The most candid view from a Chinese lab can be found in an investor call held by DeepSeek founder Liang Wenfeng. The May 20th call was leaked in July, disclosing sensitive information and disrupting that year’s second round of fundraising for DeepSeek. Liang, whose statements to the media are limited, especially when compared with American lab leads, was forthcoming about his expectations within this (supposedly) closed environment. The transcript is worth reading in whole, but a few quotes stand out. Regarding pricing, Liang stated that “We [DeepSeek] set our prices to earn only a reasonable profit, not to maximise income.” In this private setting, he predicted that DeepSeek’s frontier models would continue to have open weights. Focusing on OpenAI and its efforts to establish an enduring lead in frontier AI, he argued that “The U.S. will face challenges, and in the future it may also face challenges from China, because Chinese players are willing to take less in return for providing the service.” Putting it more directly, he said “Those who take more will be beaten by those who take less.”
The Chinese labs seem to need a much lower profit, level of perceived value, and subsequent capex to thrive. According to recent analysis, Chinese firms are set to spend about one-tenth the value that American firms will for the AI buildout. Alibaba, the Chinese tech behemoth, is issuing $10 billion in equity to invest in AI. DeepSeek recently raised $7.4 billion at a $50 billion valuation ahead of a possible 2027 IPO on the Shanghai stock exchange, and soon after began another raise at a $74 billion valuation. Moonshot AI is raising capital at a similar scale. In comparison, Anthropic and OpenAI were both valued at nearly a trillion dollars each when they raised earlier this year.
OpenAI has already cut token costs for GPT-5.6 Luna by eighty percent and Terra by twenty percent, which could be an early sign of price pressures forcing labs to further cut costs to compete. If the Chinese government eventually moves away from widely available, open-weight models, much of that value floods back to American frontier labs. If this were to happen in the next two months, the dynamic would be far less consequential and devastating to American labs compared to withering from a year or more of this price compression.
These are the technical facts, contingent on behavior, outcomes, and still more facts. The particularities of Chinese government, society, economics, and more—these are the essential elements for understanding why China has, and likely will always, continue to look at the same technical tendencies of machine intelligence and behave very differently than America.
Divergent Sinomodernity
Chinese thinkers tend to understand this better than their American counterparts. The latter expect that their experience and way of thinking are the standard of modernity, and at times, the only way. The idea that hurtled farthest out of the 18th century Enlightenment was that all men were capable of the same reason. Western civilization was comprised of universalists then, and is still now—an unyielding inheritance. The rationalists are inheritors of that universalism in their own way, placing their faith in a shared utility function. Comparatively, Chinese technologists seem much more accustomed to a timeless truth of the world known as contingency. The material conditions of any civilization are always a consequence of very particular historical circumstances.
In his 2018 book AI Superpowers, Taiwanese computer scientist Kai-Fu Lee applied this premise to AI’s role in automation, calling it Moravec’s revenge. The Austrian-born Canadian computer scientist Hans Moravec’s original formulation was that it would be far easier for machine intelligence to mimic the intellectual abilities of an adult than the physical capabilities of even a toddler. Lee’s book contradicted the conventional wisdom of the time, which believed that AI and automation would be worse for Chinese employment given the much higher proportion of people working in manufacturing. He instead predicted that labor-market adjustments caused by AI would hit America first. Lee argued that because physical intelligence would be slower to develop and harder to deploy than AI that could automate administrative and service work, America would be impacted first and more widely. As he put it, “it’s far easier to build AI algorithms than to build intelligent robots.”
Let’s say a major cause of the government response to AI would be due to the amount of human labor it automates in the short term without creating a suitable number of new jobs. Assume, also, that service and knowledge-intensive jobs are more quickly substituted by digital machine intelligence than manual labor. On this benchmark alone, there is reason to believe that American professional and office workers could face the threat of automation before more advanced physical machine intelligence displaces manual Chinese workers. As of 2023, 23% of Chinese workers are engaged in primary agriculture and related fields, while another 29% work in manufacturing and industry. In comparison, as of 2025, 43% of the American workforce is in management and professional occupations, 18% is in sales and office jobs, and 16% is in the service industry. Such asymmetries between the economic, social, and political realities in China and America inform how the “technical facts” of AI development and diffusion will determine different approaches to governance.
The details, so seemingly insignificant, are the totality. They determine the timeline. Returning to “Situational Awareness:” “Every month of [America’s capability advantage over China] will matter for safety too. We face the greatest risks if we are locked in a tight race, democratic allies and authoritarian competitors each racing through the already precarious intelligence explosion at breakneck pace—forced to throw any caution by the wayside, fearing the other getting superintelligence first.” Every month of close competition and price compression matters too.
At this point of discontinuity between the predictions and present, those outside of China should be more curious, circumspect, and imaginative than ever. Thinkers must scour over the technical facts, ravenous for a shred of insight that will tell us something of the world to come. If the story of the coming century is the agglomeration of talent, capital, and production towards the only two countries capable of utilizing them at scale, then getting the contours of the competition correct is all the more important. Right now, there is an entire parallel civilization of perspectives that American thinkers are not taking into account.
Divergent Sinomodernity, that form of contemporary society born in China, emerged slowly, then all at once. Those in the West often fail to appreciate how much of contemporary life is shaped by the particular circumstances that it too emerged from. Roman political structures, Greek philosophy, Arabic numerals, Christian faith, English common law, German idealism, the technology that grew out of an Industrial Revolution in Manchester and quickly enveloped the world; all of these incremental instances led us to where we are today. What was new— advances in technology, forms of government, and economic systems—was synonymous with civilization as embodied first in European capitals and then American steel towers. It was a form of life. At the time, the only one on that frontier. This is no longer the case.
The core truth here is that modernity, its norms and expectations, is contingent. Because of the dominance of European civilization and its inheritors for the last several hundred years, we have forgotten this. However, as modern China is clearly demonstrating, contingencies will continue to become ever more decisive. In the fifty years since the end of the Cultural Revolution, the rapid modernization of China has been compressed into the span of a single lifetime. As China becomes more prosperous and more capable of developing and deploying frontier technologies, there is no reason to expect that future innovations will not take on a particularly Chinese character.
Mid-September 2026 was a clear moment of cultural contrast. In the U.S. it began with the resignation of Jacob Coxon, an Anthropic researcher who very publicly warned of the potential catastrophic risk as a result of further development. In short order, the heads of Anthropic, OpenAI, Google DeepMind, and xAI each endorsed the idea of pacing the frontier, the proposal to slow down model development that had been earlier circulating as an open letter among the AI community.
That same day, September 12th, China’s Minister of State Security, Chen Yixin, published an article that detailed his own thoughts on the opportunities and risks inherent in the development of machine intelligence. After outlining a sequence of opportunities that essentially repeat the party lines about practical applications and international cooperation, Chen turned to the risks. He outlines how AI is being used to spread hostile and false information online, along with the new types of cyberattacks, data leaks, governance complications, and military applications it creates. He also mentions how “monopoly exacerbates the imbalance” due to how “relevant countries” are taking advantage of their developmental lead by “implementing technology controls, monopolizing industry standards, and establishing closed-source ecosystems.”
However, throughout the text, Chen does not describe anything resembling the job shortage or safety concerns gripping the popular press and lawmakers in America. One potential explanation for this is that despite whatever Chen or other security officials in China might think, they do not have the latitude to preempt the party line on such a consequential issue. That may be true, but it’s a nonfalsifiable proposition. The inclination to ignore what people in China say, whether lab leads or government officials, in favor of what we believe to be their true beliefs (which conveniently mirror our own), is not a wise one.
It would be extremely valuable for Americans living in China, or immersed in Chinese media and discourse, to try to find those societal contingencies masquerading as technological determinism. It is after all those that will uniquely shape the Chinese governance of machine intelligence. Looking at the same underlying technology and extrapolated trendlines, how and why will China react differently? Perhaps they are uninterested in pacing the frontier because they see America as already forcing them to pace via chip export controls. Maybe the relative lack of concern regarding human-directed and autonomous cyber incidents is because of the already existing cybersecurity state. Why establish an ecosystem of independent evaluators when the CAC can unilaterally order model changes with the authority of the party-state?
These are just hypotheses. The future of machine intelligence, which is the future of the world, can be glimpsed only dimly through the fog of uncertainty. Human faculties of reason, even as they are challenged by mechanical reason, remain as central to our understanding of the contemporary world as they were during the Enlightenment. However, the rationalist adoption of this frame, and the coarse-grained analysis they contribute, clearly has its own limits. They are starkly demonstrated by the difference between the projections of Chinese behavior and the irrefutable truth that can be seen before us today. A certain level of openness not yet commonly found will be necessary to anticipate Sinomodernity.
The Chinese, too, may need to be more imaginative. In Xi Jinping’s WAIC speech, he expressed a wish that “this fine steed of AI gallops with both speed and stability.” Perhaps instead he should prepare for the future of machine intelligence approaching as a herd of wild horses beyond counting or knowing—life run riot—graceful and glorious to behold; the thunderous clamor of hooves heralding the transformation of the Earth.
Farrell Gregory is a research fellow at the Foundation for American Innovation. You can follow him at @efarrellgregory.
PALLADIUM is the first magazine of the third millennium. We are non-partisan, non-profit, and on a mission to discover what the future of governance and society will look like.
Become a PALLADIUM member today! You will receive our beautiful quarterly print edition and help support our mission.
We thank Pardon, for underwriting our work in 2026 with their generous gift.