Deepseek is not the Chinese OpenAI

The thoughts underlying this thesis have been with me some time, and have been slowly been crystallising more and more clearly as time goes on and I realise that the rest of the world is seemingly diverging more and more with my own thoughts on the subject. This is not a fully formed proposal, and I will not attempt any detailed plan of action. As always, it is instead an attempt to try and formulate my own thinking on the subject, as well as to offer the reader a perspective on the Chinese economic philosophy that is heavily overlooked in the west. While informed by reading scholarly material and following the ongoing discussion about China, it is prompted mainly by my visit there and getting an idea of how the country works. That idea may be incorrect, and so feel free to e-mail me your thoughts on the subject.

In short, I am of the opinion that American and China are in a sort of dialectical movement with each other, with America (and Europe) wishing to copy the industrial strategy that it sees as having been key to China’s modernisation, and China simultaneously trying to create a hyper-competitive environment that fosters innovation. I think that the Chinese approach will succeed, and the western one will not. America and Europe risks losing the dynamism and economic resilience that free-market economics offers, while China encroaches on the gap that we gained during the great divergence.

China is grossly misunderstood in the west. I do not claim to be a Sinologist, but merely in my brief visits to the mainland I have gained much insight that I do not believe most policymakers have. First of all, the PRC is not the Soviet Union. This might seem obvious at first, but it is the viewpoint that I hear between the lines of much literature on how to deal with China. Instead of vast defence spending, it is industrial policy. The opponent is a vast authoritarian system that efficiently the production of a continent to those areas where it most serves the national interest. This straw man is not a reality. The authoritarian government of the PRC is its greatest weakness, just as it was for the USSR, and for the Russian Republic that succeeded it. But while both the Soviet and Chinese communists ruled over oppressive regimes, the similarity mostly ends there.

The ideological system of “Socialism with Chinese Characteristics” (中国特色社会主义, SCC) as straightforwardly viewed from a Marxist point of view, is an attempt to create a system of pure capitalism as quickly as possible. A broader, more holistic, and more radical form of Lenin’s New Economic Policy where the rural Soviet Union and PRC would both allow capitalist enterprise to grow, as long as it did so under the auspices of the communist party. This is precisely the path that societies follow according to orthodox Marxist theory (Feudalism -> Capitalism -> Socialism -> Communism). In the west, Socialism with Chinese Characteristics is instead commonly understood to be some sort of attempt by the CCP to maintain a veneer of communist ideology while in practice becoming a free-market dictatorship. This is also not the case. The CCP is communist through and through, and SCC is a genuine attempt to achieve it. When spending time on the mainland, one is struck by the sheer presence of the party, of the praise for workers, of and of what seems to be authentic belief.

On the ground, there is a feeling of division between state control and private enterprise. Roads are meticulously cleaned and well-maintained, while at the same time lined by buildings almost falling into disrepair. The state wishes that certain things, like the power grid, should be fully controlled, but in other sectors the private sector should be able to flourish freely. This schizophrenic reality is that of SCC, with certain industries — primarily related to infrastructure — fully nationalised and others in an almost Laissez-Faire environment.

To try and achieve socialism (and eventually communism) China must unleash the free-market completely. The “weariness” of big tech that the CCP is sometimes reported to hold is not just grounded in a fear of large corporate interests clashing against the party, but is also an attempt to genuinely increase competition and increase the freedom of the market. Another example behind this might paradoxically be the intervention that is most talked about, that being the production of Electric Vehicles (EVs). Chinese EV subsides are not oriented toward building “national champions” like the Export-oriented approaches undertaken across the Yellow Sea in the Republic of Korea. Instead it has tried to foster a breadth of EV producers that are now aggressively undercutting each other, becoming internationally competitive in the process.

In the beginning of AI 2027, a short story of rapid AI development¹, the leading Chinese AI company (collectively called DeepCent, a portmanteau of Deepseek and Tencent) becomes the leader of a nationalised and consolidated “collective”. Researchers and new integrated circuits (ICs) are sent to a secure centralised facility where AI development and strategy are carried out. This is a very accurate view of how American and western strategists view China, as a singular entity that makes unified decisions that are fully carried out. But this is perhaps the exact opposite of reality. Instead the rapid industrialisation of China has been due to thousands of decisions made by comparatively low-level provincial officials and below, all competing to look good in the eyes of higher-ups by growing GDP². There has been no large change in CCP governance that would suggest a different approach when it comes to AI development.

The authors of AI 2027 agree that there is no clear leading Chinese company, and in a footnote they remark that:

We consider DeepSeek, Tencent, Alibaba, and others to have strong AGI projects in China. To avoid singling out a specific one, our scenario will follow a fictional “DeepCent”.

But it is precisely this multifaceted competition that makes Chinese AI development different from its American counterpart. The competition between AI companies, all with extremely limiting compute resources, forces Chinese firms to compete using talent — something they will likely have more of in the foreseeable future due to what will probably be decreased brain drain to the US.

The “hawks in the CCP” warning about AGI do not exist, at least in the same way they do in America. Instead, the Chinese leadership is committed to using its dominance in manufacturing and infrastructure development to expand in the field of “Embodied AI” (具身人工智能). This approach requires numerous companies to make tailor-made AI solutions for different industries and tasks, and therefore motivates the opposite strategy from the centralised one that AI 2027 predicts. An example of this would be ZTE’s ongoing application merging drone use, 5G connectivity and AI to monitor things like the maintenance requirements for infrastructure.

Xi Jinping himself, the paramount leader of the PRC, seems quite keen on AI, having ordered a 2025 “study session” (集体学习) on the subject for the top CCP leadership. But he is old-fashioned, having gotten into Tsingua University (China’s top university) as a worker-peasant-soldier student during the Cultural Revolution. He has a greater interest in “hard” development like infrastructure; even in Made in China 2025, the CCP plan to expand into higher value-add manufacturing, it is specifically manufacturing that is the focus, not the service economy or software developments — areas where AI makes the biggest amount of sense. Chinese industrialisation through state-led infrastructure initiatives has been incredibly successful, and the CCP is much more likely to stick to what they know works (even if it is with new fancy tools).

His study notes from the aforementioned study session does mention a need to “concentrate forces” (集中力量) on high-end ICs and “foundational” software, but he also mentions creating an “enterprise-led industry-academia-research-user collaborative innovation system”, a far cry from a secret power plant/data center under state control. It also mentions such things like “Highlighting applied directions” (突出应用导向) and “Helping traditional industry reform and upgrade” (助力传统产业改造升级). These things are in-line with the ideas of embodied AI more than they are a centralisation of compute.

Deepseek then is not a state-sponsored champion, but one of many companies experimenting with AI development. This will likely accelerate AI development, and force AI firms adopt shorter release cycles and lower prices. The huge size of the Chinese market also allows for increased experimentation and development, experimentation that Western firms will not have access to. R1’s free software³ nature helps this further, by helping deploy AI in novel ways across the Chinese economy faster than in America, also increasing innovation. There is no large consolidation of the AI industry in the PRC.

American governance is in a unique place on the spectrum. Contrary to the PRC, there are enough checks and balances for the executive arm to feel “restrained” by them, but it is still empowered enough to project a significant amount of power. This seems like a worst-case scenario to me — not enough power to take the seat as an “enlightened despot”, but still enough to break things if you aren’t careful. This is the issue at the heart of American AI strategy. The bureaucracy does not have the capacity to grasp the development of AI, but it feels it necessary to intervene regardless.

When Deepseek’s R1 model was first publicly released, it was widely reported as a “Sputnik moment”, as a wakeup-call to the American government that it no longer held a decisive technological advantage over its strategic rival. But this is an over-exaggeration. In 1957, with the launch of Sputnik, the Soviets attained a technological capability that the Americans did not posses at all, not merely a level of parity. But one might argue that what Deepseek showed was Chinese capability to do something, but to do it at a fraction of the cost of American alternatives. But this is not exactly the case either, as Semianalysis recently showed, Deepseek’s low price-per-token has not resulted in companies flocking to Deepseek, but instead a slow decline in users, even as the overall market has grown. This is because they have merely chosen to make certain trade-offs to lower prices due to their compute-restrained nature, rather than any underlying technical improvement.

This compute-restrained environment that Chinese firms find themselves in is of course artificially created by American export-controls. As is being increasingly argued, perhaps most prominently by Nvidia CEO Jensen Huang, these export-controls might actually be decreasing American power relative to China, by effectively tariffing American ICs in the Chinese market, allowing domestic AI-chip manufacturers to catch up with Nvidia, TSMC, and the non-Chinese semiconductor manufacturing industry.

But what has not been discussed to the same degree is the risk that not only does this environment hurt the long-term prospects for the western-aligned IC industry, but it might also hurt American software as well. As Deepseek alarmists claim, China has largely caught up in the development of LLM, or are at least very close behind. But they have done so in a much less capital- and compute-rich environment. As Sutton writes in The Bitter Lesson, most of the gains in AI research have been due to increasing amounts of compute, and developments have rather been in how to make use of more and more computational resources, rather than in how to make use of it. Chinese development seems to have changed this trend, if only very slightly. If we then extrapolate Chinese AI so that it has the compute resources of OpenAI, would it not overtake OpenAI? There is a risk that, like the Sardaukar and Fremen from Frank Herbert’s Dune, we are helping Chinese AI develop in a more hostile environment, while at the same time supporting the development of a Chinese IC industry to support it in the long-term.

Similarly, American strategy, lead both by private-sector and public sector investment, risks “putting too many eggs in one basket”, that basket being OpenAI. 500 billion USD, the amount of money to be invested in the Stargate Project, is a lot of money, likely more than China is spending, so is it not important to get as much development as possible out of it? If China is able to us its compute more efficiently than the US and instead uses its money to bolster domestic manufacturing further, especially in the field of integrated circuits, it could undermine US national security. China’s vibrant free-market economy is a tremendous resource, exceeded only by that of the United States. It would be a shame to destroy that lead.

The US and Europe must resist the impulse to bet big on national champions, as (perhaps paradoxically to many) Deepseek shows. High technology advancements should instead be facilitated through an increase in basic research, ability for academia and free enterprise to intermingle, and deep capital markets that can bet big on new, innovative ideas.

China is a free-market economy, and it is committed to using the strengths inherent in that system to its national advantage. The PRC leadership will try and use AI to upgrade preëxisting industry, and will not engage in a rapid “race to AGI”, beyond continuing attempts to become a significant producer of integrated circuits. Its large domestic market allows it to rapidly test new innovations, but so does the vast American and (to some extent) European markets. While it the CCP is increasingly grasping after interventionism, it still is trying to remain committed to free enterprise.

The US and Europe must resist the impulse to bet big on national champions, as (perhaps paradoxically to many) Deepseek shows. High technology advancements should instead be facilitated through an increase in basic research, ability for academia and free enterprise to intermingle, and deep capital markets that can bet big on new, innovative ideas.