China: how Beijing plans to become a global leader in AI

China can no longer be con­sidered a second-tier com­pet­it­or in tech­no­lo­gic­al innov­a­tion. Since the mid-2010s, it has demon­strated its abil­ity to pos­i­tion itself as a lead­ing play­er. With regards to high per­form­ance in the glob­al race for AI, whilst not over­tak­ing the United States, it has clearly dis­tin­guished itself. Look­ing at the num­ber of Chinese pat­ents filed in this field; the pace has clearly accel­er­ated since 2019. Accord­ing to the World Intel­lec­tu­al Prop­erty Organ­isa­tion, Beijing filed nearly 30,000 pat­ents related to gen­er­at­ive AI between 2014 and 2024.

At the turn of the 2020s, the Chinese tech giant Hua­wei made a break­through that has Amer­ic­an com­pet­it­ors fear­ing the launch of Chinese chips just as power­ful as their own products. Today, accord­ing to obser­va­tions by Pierre Sel, a PhD stu­dent at the Uni­ver­sity of Vienna, “it’s open-source AI mod­els are of high­er qual­ity than those developed in the United States.”

In Janu­ary 2025, the Chinese com­pany Deep­Seek sur­prised the world by unveil­ing Deep­Seek-R1, a high-per­form­ing open-source AI mod­el, trained at a frac­tion of the cost of OpenAI, thanks in par­tic­u­lar to engin­eer­ing work that enabled developers to max­im­ise the use of the chips at their dis­pos­al. Liang Wen­feng, its founder with a back­ground in quant­it­at­ive fin­ance, has out­per­formed US tech, whilst OpenAI had until then seemed unbeatable.

Even more sur­pris­ing is that he achieves this per­form­ance at reduced costs and without access to the most effi­cient chips. Pierre Sel qual­i­fies this suc­cess by not­ing that “access to advanced chips is not a cut off. The start-up has cer­tainly been able to bene­fit from the Malay­si­an black mar­ket”¹ and is said to have “trained” its mod­el using Amer­ic­an mod­els. As for the actu­al cost, whilst Deep­Seek puts it at $6 mil­lion, “this fig­ure excludes costs related to the devel­op­ment cycle and the deploy­ment of infra­struc­ture²,” the research­er cla­ri­fies. The sud­den emer­gence of this mod­el was sub­sequently exploited extens­ively by both Beijing and Wash­ing­ton, against a back­drop of polit­ic­al and com­mer­cial objectives.

The Chinese Com­mun­ist Party (CCP) places AI at the heart of its eco­nom­ic frame­work and, more spe­cific­ally, advoc­ates for the wide­spread deploy­ment of gen­er­at­ive AI. The report by the China Inter­net Net­work Inform­a­tion Centre (2025) pro­motes the “grow­ing adop­tion” of these tech­no­lo­gies and notes that the use of “smart assist­ants” is already com­mon­place in many sec­tors, includ­ing industry, con­sumer goods, health­care, dis­aster pre­dic­tion and edu­ca­tion. Offi­cial state­ments that should be treated with cau­tion, as Pierre Sel points out, “ser­i­ous doubts remain regard­ing the qual­ity of this adoption”.

The gap between the two coun­tries lies in com­put­ing power, which is determ­ined by the chips used.

Accord­ing to offi­cial data pub­lished by Beijing, the uptake of these tech­no­lo­gies has already reached an advanced stage: “more than 78% of the com­pan­ies sur­veyed use some form of AI, ran­ging from cus­tom­er ser­vice to logist­ics.” On the user side, fig­ures released in Feb­ru­ary 2026 indic­ate 600 mil­lion users of gen­er­at­ive AI, out of a total of 1.125 bil­lion inter­net users. The plat­forms are also repor­ted to have pro­cessed over 600 bil­lion quer­ies per day³.

This massive uptake, com­bined with wide­spread use, requires con­sid­er­able com­put­ing power, a digit­al archi­tec­ture and infra­struc­ture com­men­sur­ate with the Party’s stated ambi­tions. “In terms of devel­op­ment cap­ab­il­it­ies, there is no doubt that China pos­sesses tech­nic­al and human resources (and, to a less­er extent, fin­an­cial resources) com­par­able to those of the United States. The gap between the two coun­tries lies in com­put­ing power, which is determ­ined by the chips used,” explains Pierre Sel.

There is no doubt that the United States dom­in­ates the semi­con­duct­or mar­ket, with their tech giants such as NVIDIA, Qual­comm and AMD. In 2024–2025, over 50% of glob­al semi­con­duct­or rev­en­ue was gen­er­ated in the United States. On the oth­er side of the Pacific, almost a quarter of the chips sold world­wide in 2024 were pur­chased by Chinese elec­tron­ics man­u­fac­tur­ers, accord­ing to estim­ates by the Semi­con­duct­or Industry Association.

In Janu­ary 2025, shortly before leav­ing the White House, Joe Biden intro­duced new reg­u­la­tions: the Frame­work for Arti­fi­cial Intel­li­gence Dif­fu­sion. This legis­la­tion aimed to restrict access to chips for Chinese buy­ers. A year and a pres­id­ent later, a policy shift took place. Don­ald Trump relaxed export con­trols by author­ising NVIDIA to sell the H200 chips, one of the most power­ful avail­able to Beijing. In return, a 25% tax must be levied by the United States.

Aware of this stra­tegic weak­ness, the Chinese gov­ern­ment is seek­ing to regain its sov­er­eignty, not­ably by strongly encour­aging man­u­fac­tur­ers to devel­op loc­al altern­at­ives. Moreover, accord­ing to Pierre Sel, “com­pan­ies can choose to use a data centre loc­ated abroad (often in Malay­sia or Singa­pore) to train AI mod­els. These data centres are not sub­ject to US restric­tions and can pro­cure NVIDIA chips freely”⁴.

Accord­ing to Chris McGuire, a mem­ber of the Coun­cil on For­eign Rela­tions think tank and advoc­ate of export con­trols, the gap between the two powers is widen­ing. Based on the highest estim­ates, Hua­wei would account for only 5% of NVIDIA’s glob­al com­put­ing power in 2025, 4% in 2026 and 2% in 2027⁵. Oth­ers point to some pro­gress: 20 to 30% of the equip­ment used in Chinese chip man­u­fac­tur­ing is expec­ted to be pro­duced in China by 2025, com­pared with 10% in 2022⁶.

Des­pite China’s over­all lag in this race, Pierre Sel warns against jump­ing to con­clu­sions. He par­tic­u­larly high­lights the ambi­gu­ity of the play­ers involved, who have mul­tiple interests in exag­ger­at­ing the lead of some or down­play­ing their own. Fur­ther­more, the cri­ter­ia used to give one coun­try an “advant­age” over anoth­er are vague. Thus, “while China may not pos­sess the most power­ful chips or mod­els, it is its open-source mod­els that are the most down­loaded⁷.” Down­load­ing does not neces­sar­ily imply usage; nev­er­the­less, this data con­tin­ues to cause con­cern among US stakeholders.

Alicia Piveteau