ChinaTech #1 Command and Innovate: China's Centralized Science Strategy
This new segment by Shobhankita Reddy is your go-to newsletter for updates and perspectives on China’s tech ecosystem. The debut edition explores China's centralized science policy under the CSTC, the implications of Deepseek's open-source AI breakthroughs, and the challenges of US export controls in the global tech race.
Early last week, a report on the National Science and Technology Work Conference caught my eye. Presided over by Vice-Premier Ding Xuexiang, who is the director of the Central Science and Technology Commission, the conference had several officials in attendance, including the presidents of the Chinese Academy of Sciences and the Chinese Academy of Engineering, officials from the Ministry of Finance and education as well as the National Development and Reform Commission, and the Central Commission for Discipline Inspection. The conference also had an admiral of the PLA navy and an official from the Central Commission for Military-Civil Fusion Development.
This is a note-worthy meeting, not just because of the presence of officials from different government departments and institutions but also because this was the first work conference since the commission was set up in March 2023. The CSTC is believed to be meeting quarterly, with its first meeting being held in July 2023, but there are still no details of its actual composition. It appears to be composed of both full-time and standing committee members.
The fact that the recent work conference was headed by Ding Xuexiang, a member of the Politburo Standing Committee, and co-chaired by science minister Yin Hejun is an important signal for party-state relations and the CSTC’s significant bureaucratic clout. As quoted in the report - ‘the meeting showed the new commission had become “more mature” in its operations, “breaking the barriers” as it boosts collaboration across various levels of government departments and sectors.’
In March 2023, China’s Ministry of Science and Technology (MOST) was restructured with some institutions (China Rural Technology Development Center and China BioTechnology Development Center, for example) being transferred to other ministries and commissions.
To centralize a unified leadership for science policy and funding along with direct CCP oversight, the new CSTC was formed as a superseding body, with the Ministry of Science and Technology being delegated much of the administrative and executive work. This is in line with Xi Jinping’s style of tightening Party control and weakening the hold of the professional bureaucracy. The choice of Ding Xuexiang is interesting too. Being a trained engineer who previously spent time as a materials science researcher and administrator before moving to politics as a deputy director of the Shanghai Municipal Science and Technology Commission in 1999, he brings the right blend of scientific management skills to the commission. Ding is also believed to be extremely close to and trusted by Xi, marking the importance of the commission to China’s national goals.
Another example of MOST being downgraded is the increased role of the CSTC General Office, housed within MOST, that manages the commission's day-to-day operations including approving funding, coordinating with regional S&T commissions, conducting periodic reviews regarding the implementation of projects, etc.
A key challenge such central commissions face is the difficulty in coordinating across different regional and central clusters and driving a singular mission of priorities across them all. It remains to be seen how the CSTC would tackle this.
As reported in November 2023 - ‘the Central Science and Technology Commission issued the "Overall Plan for the Construction of Regional Science and Technology Innovation System", proposing to improve the layout of the regional science and technology innovation system, optimize and upgrade the three international science and technology innovation centers in Beijing, Shanghai, and the Greater Bay Area, and accelerate the construction of three regional science and technology innovation centers in Chengdu, Chongqing, Wuhan, and Xi'an.’
Another challenge the commission faces is to actualize Xi’s “walk with two legs” slogan - that scientists must have one leg in exploratory research and the other in national priorities, that even exploratory research must be conducted with the goals of the techno-security state in mind; and the commercialization of this research into products for the market.
A lot is still unknown about this commission - its composition, scope of work and even the goals and policies it aims to further such as the Medium & Long-Range Plan for National S&T Development (2021-2035) which has still not been disclosed.
The recent report on the conference indicates that CSTC is playing an important coordinating role bringing together central ministries, local governments, academia and research institutions. Its planning mandate also seems to extend to all domains of science and tech policy, from finance to industrial development and talent cultivation. This is a signal that the commission is making strides in its goals, vision-setting and institutional collaboration.
Tracking Tech News
A month ago, Deepseek, a research-first private arm of a Chinese hedge fund, caught global attention for its open-source v3 model that it claimed was trained for a “joke of a budget” of $6 million and using “only 2048 GPU clusters for two months”. The third iteration in a set of AI models, v3 showcased a higher performance on mathematical and coding tasks compared to existing models. Additionally, it was superlative in its ability to parse academic texts in the Chinese language, owing to its higher context length capabilities. Deepseek’s previous two AI models were also offered at an extremely low cost, forcing Bytedance and Alibaba to cut the prices of their models as well.
Now Deepseek is back in the news — this time for an R1 model. Both v3 and R1 are fully open-source and have generous permissive licenses to use, copy, distribute and commercialize other software built over them. Both v3 and R1 models employ a similar technical architecture of Mixture-of-Experts models, where a bunch of specialist smaller models work together, and depending on the query asked, only a small number of parameters are hit, improving inference times as well as accuracy. Deepseek also claimed that R1 was 90-95% cheaper to build than its OpenAI counterpart.
A few questions come to mind.
Firstly, is this a point in favour of the open-source AI camp compared to the closed-source camp? Champions of open-source AI have long been harking about its long term benefits for transparency, innovation capabilities and lower cost, as compared to the monopolizing of technology by a few Big Tech companies. Deepseek’s decision to open-source their models also appears to be a move in cultivating street credibility, meant to catch the attention of developers world-wide.
Secondly, if a chinese company can independently and at a breakneck speed catch up to the capabilities of firms such as Open AI and Anthropic that have billions of dollars in funding, what real technological moat do these firms have?
Thirdly, what does this say about the effectiveness of US export controls on chips to China? While the accuracy of the reported compute costs and GPU utilization remains a debate among analysts, it is evident that ChinaTech should not be underestimated.
If the budget for training the models is as being reported, it would indicate the ineffectiveness of export controls in curtailing the transfer of critical know-how and skills relevant to the development of strategic technology.
In the scenario that the claims are potentially exaggerated and China still has access to superior computing power, the difficulties in cutting access to a select adversary in a world with complex global supply chains with trade partnerships that are coherent mainly from an economic viewpoint but not necessarily a security or strategic lens come to light. What might the US then do?