The PRC's Compute Push Goes Global

Executive Summary:

Physical deployment is only one form of Chinese compute going global. Chinese analysts distinguish it from the more mature model of “services crossing the border” (服务出境), in which computing hardware remains in the PRC while domestic data centers process AI workloads for foreign customers and export the resulting tokens and services (Zhong Lun, July 10). Even when hardware remains physically in the PRC, compute exported-as-a-service still represents a decision to devote computational resources to foreign rather than domestic demand. Computing capacity is inherently rivalrous: capacity devoted to processing foreign workloads cannot simultaneously be used for domestic workloads.

The model of cross-border compute exports has already developed into a wider industry push. At the 2026 World Artificial Intelligence Conference (WAIC), one exhibitor established a dedicated English-language “Token going global” (Token出海) section that demonstrated Chinese computing power packaged into standardized tokens services and exported overseas, attracting considerable attention from foreign attendees, particularly from the Middle East (Jiemian News, July 19; China Brief, July 22). In 2025, the China Academy of Information and Communications Technology (CAICT), the think-tank under the PRC’s Ministry of Industry and Information Technology (MIIT), collaborated with industry partners to launch a dedicated “Computing-Power Token Going Global Ecosystem Initiative” (算力词元 (Token) 出海生态计划). Its aim is to move the sector beyond experimental projects towards “standardized, efficient, and sustainable” (规范、高效、可持续) overseas expansion (CWW, July 1). Its organizers argued that doing so could help hedge against external risks and expand services trading (Open Data Center Committee, July 1).

Importantly, Chinese firms are actively laying the groundwork for the physical deployment of advanced domestic AI hardware abroad. Huawei has drawn up plans to export Ascend-series chips, including the Ascend 950DT and 950PR AI accelerator, to South Korea and has signed distributor agreements with local partners like Hansol PNS and SK Shieldus (Sina Finance, July 7). The company has also unveiled the Atlas 960E SuperPoD as part of its effort to address large-scale AI infrastructure demand overseas, while other Chinese companies have announced plans to deploy a “domestic computing power cluster” (出海国产算力集群) in Saudi Arabia (SenseTime, July 20). Similar intent is emerging vis-à-vis Southeast Asia, with reports that Malaysia is considering adopting Huawei’s 910C Ascend in its sovereign AI ambitions (Sina Technology, September 8).

Chinese elites have endorsed a wider version of this push. Wang Zhongyuan (王中原) of Fudan University has argued that the PRC should assist countries in the Global South by constructing local computing infrastructure, jointly building data centers, and exporting “computing services” (算力服务) (CASS, July 16). Song Guoyou (宋国友), deputy director of Fudan University’s Center for American Studies, is even more explicit, calling for Chinese firms to deploy “computing clusters” (算力集群) across different regions of the Global South (National Office for Philosophy and Social Sciences, July 16).

A growing number of Chinese industry and policy elites have explicitly described internationalization as the next stage for the domestic compute industry. Li Wei (栗蔚), deputy director of CAICT’s Cloud Computing and Digitalization Research Institute, has called for accelerating the overseas expansion of the domestic computing-power industry and integrating it into the “global computing-power industrial system” (全球算力产业体系) (People’s Tribune, July 20). Wang Chao (王超), executive director of CAS Star, the deep-tech venture capital firm founded by researchers from the Chinese Academy of Sciences, has similarly argued that domestic compute will not only serve domestic demand but also for overseas expansion (Beijing Daily, September 23).

Chinese companies have described similar ambitions. Huawei has actively touted Ascend and its Kunpeng computing platform as having created “a new choice for the world” (为世界构建了新的选择) (Huawei Tech, February 2024). iFlytek, meanwhile, has partnered with Huawei Ascend to “cooperate deeply on overseas large-model deployments” (在海外主权大模型上深度合作) and “plant flags and seize positions around the world” (在全球插红旗、抢阵地) (iFlytek, April 29). Other firms describe computing power going global as the “industry’s second growth curve” (Runze Intelligence, August 28) and argue that overseas expansion “is no longer optional, but imperative” (Sina Finance, July 29).

Southeast Asia has become a particularly important potential market. One CAICT report constructed a market-entry evaluation framework for overseas expansion and identified opportunities for Chinese computing-power providers in emerging markets, especially in Southeast Asia, where they note surging demand (CWW, May 13). One Chinese AI infrastructure company similarly concludes that compute in Southeast Asia will exceed 3GW in 2026 and 6GW in 2028, asserting that Chinese firms should leverage their accumulated technological capabilities to “rapidly enter overseas computing-power center markets” (快速切入海外算力中心市场) and “capitalize on the supply gap” (把握 … 供给缺口) in Southeast Asia’s high-end IDC market (Runze Intelligence, August 28). Meanwhile, Sugon, another major Chinese computing-infrastructure provider, has established an ASEAN business headquarters in Nanning, Guangxi. By providing a Chinese tech stack comprising Sugon’s servers and its high-end Hygon DCU compute, PRC firms can provide a “complete solution” for Southeast Asian countries (QQ News, September 21).

Chinese hardware providers often achieve pricing advantage by undercutting foreign competitors. One Chinese estimate concludes that for equivalent computing services yielding similar inference performance, Chinese compute providers can charge just one-fifth to one-seventh as much as comparable overseas services. The price advantage is derived partly from lower electricity costs in the PRC and vertical integration of large Chinese telecommunications operators, most of which own both computing and network infrastructure (Southern Daily, June 5).

The PRC’s overseas push is happening despite domestic compute constraints. Chinese sources depict a landscape in which shortages are concentrated in select types of high-performance hardware, while the country’s aggregate computing infrastructure expands rapidly.

Yang Siwei (杨思维) and Liu Rui (刘睿) of CAICT have written that the PRC’s domestic computing buildout has “laid a solid foundation for the computing-power industry to ‘go global’” (为算力产业“出海”奠定坚实基础) (Information and Communications Technology Policy, March 2025). By the end of 2023, the PRC already possessed more than 230 exaFLOPS of computing capacity, accounting for over 25 percent of the global total. Meanwhile, intelligent-computing capacity grew 70 percent, accounting for 30 percent of the PRC’s overall computational scale. By June 2024, more than 250 intelligent-computing centers were operating or under construction nationwide (Information and Communications Technology Policy, March 2025). Li Wei points to further growth since, claiming that the PRC has developed sufficient computing capacity to support 70 percent of domestic parallel-training demand for trillion-parameter models (People’s Tribune, July 20). Consequently, Yang and Liu call for Chinese firms to start building “comprehensive computing-power service ecosystems” (全面算力服务生态) in emerging country markets (Information and Communications Technology Policy, March 2025).

Importantly, Chinese firms do not need to physically ship domestic AI chips abroad to internationalize domestic computing power. Developers overseas can, for example, access Ascend computing power from the PRC through PyTorch, a software framework for developing and running AI models that reduces friction when accessing hardware (Beijing Daily, September 23).

The PRC has already experimented with this model at scale. In April 2026, the Shantou Overseas Chinese Pilot Zone in Guangdong completed the PRC’s first end-to-end commercial model for “Token going global,” which included consolidating the first phase of an international Token factory with 2,000 petaFLOPS of computing capacity (TYM, September 3). This represented a deliberate decision by Chinese authorities to add compute capacity with the express intent of serving the overseas market.

Overseas users have since accessed token services produced in Shantou, which Chinese reporting describes as the first time Chinese computing resources were exported in the form of “high value-added digital services” (高附加值的数字服务) (CWW, April 23). Zhu Jun (朱军), director of the Basic Theory Research Center at Tsinghua’s Institute for AI, described the “international Token-factory model” (工厂的建设模式) as a practical path for Chinese produced computing power and large models to expand into overseas markets (TYM, September 3). The economics of the model are increasingly plausible as the hardware itself remains in the PRC. In Shantou, network latency to Singapore reportedly fell to just 32.7 milliseconds, allowing overseas users to submit requests to models running on Chinese infrastructure and receive outputs with low network delays (Southern Daily, June 5).

Chinese researchers have also explored architectures that reduce dependence on the most high-performance accelerators. Zheng Weimin (郑纬民), a professor at Tsinghua University and former president of the China Computer Federation, has described an inference architecture in which a 10,000-accelerator cluster might employ 9,000 domestically produced accelerators for prefill and 1,000 higher performance accelerators for generation, raising the domestic share to 90 percent. [1]  Zheng remarked that the resulting high-quality tokens could serve both domestic industries and go global, adding that the PRC had “sufficient computing capacity” (算力充足) to achieve this in addition to a clear cost advantage (Ningbo Computer Federation, June 10).

Chinese inference services explain one part of the story, yet their viability makes the planned export of advanced Chinese accelerators even more puzzling. If Chinese companies can serve overseas customers while leaving their hardware within the PRC, there is less reason for them to physically export valuable hardware to foreign markets.

One answer is that overseas demand can feed back into the development of the PRC’s domestic technology ecosystem. Guo Liang (郭亮), Chief Engineer at CAICT’s Cloud Computing and Big Data Research Institute, has observed that emerging markets in Southeast Asia, the Middle East, Africa, and Latin America offer market conditions that have created ample room for the PRC’s computing services to expand overseas. More importantly, he argues that transitioning from building infrastructure toward exporting capabilities and services can promote the “iterative upgrading” (迭代升级) of the PRC’s “computing-power industrial chain” (算力产业链) and drive the coordinated, concurrent uplifting of domestic AI computing clusters, large models, and AI chips (Xinhua, July 24).

Liu Xiaoge (刘小鸽), a deputy director of research at the Development Research Center of the State Council, has similarly argued that Token exports can create a powerful “demand-pull effect” (需求牵引效应) that helps mitigate technological chokepoints affecting the PRC’s domestic AI hardware stack. Supplying domestically produced chips for large-scale application scenarios, he argues, will create opportunities for frequent iteration that can narrow gaps with the international frontier “in both manufacturing processes and ecosystems” (制程与生态) (State Council, May 8).

The overseas ambitions of Chinese firms do not demonstrate that the PRC has overcome its computing constraints, nor does it mean that Chinese companies can challenge Nvidia globally at a competitive scale. Chinese legal analysts have pointed out that Chinese firms targeting overseas markets should prioritize cross-border service exports from infrastructure located in the PRC before deploying physical hardware abroad (Zhong Lun, July 10). Huawei’s own leadership has been forthcoming about these constraints. Rotating chairman Eric Xu recently acknowledged that domestic shortages prevent a “fully-fledged” rollout of Huawei Ascend abroad (Reuters, September 17).

This qualification matters. Huawei is constrained from international expansion on a “fully-fledged” scale, but this does not mean that internationalization has stopped altogether. Chinese companies and policy elites do not appear to view compute scarcity as a reason to postpone international expansion before compute parity is achieved. As time progresses, and as the PRC consolidates increasingly novel algorithmic efficiencies, its ability and willingness to export compute abroad will only grow.

Daniel Fu is the Stonex PhD Scholar in International Relations at the London School of Economics and Political Science (LSE). His research examines how U.S. export controls shape Chinese policy and firm behavior, with a broader focus on international political economy, artificial intelligence, and China’s science and technology strategy.

This article appeared originally at The Jamestown Foundation's China Brief. 

[1] Prefill processes the input prompt while generation produces output tokens, with the two stages having different compute requirements. This allows them to run on different types of accelerators.