On July 7, 2026, the Chinese artificial intelligence company DeepSeek announced that it is developing its own AI chip.[i] The chip in question is designed not for training new models, but for the “inference” phase where already trained models generate responses to user queries. Two developments in the third week of October clarified the significance of this move:It was reported that DeepSeek is collaborating with Huawei on the design of next-generation Ascend chips, and around the same time, the US Department of Homeland Security (DHS) signed a decision excluding DeepSeek product from the information systems and procurement processes of federal civilian agencies.[ii] When these three developments are read together, it would be incorrect to view DeepSeek’s chip initiative merely as a company’s supplier diversification. Essentially, this can be interpreted as a shift in the US-China technology rivalry from the axis of chip export controls to an integrated struggle of “technology stacks” encompassing models, hardware, and software.
The backdrop to DeepSeek’s shift toward the hardware sector is US export restrictions. The company’s model, which shook the markets in January 2025, was trained using Nvidia H800 chips, the sale of which to China was banned by Washington in late 2023; following the bans, the company’s reliance on Huawei increased. During this period, Huawei held nearly half of China’s approximately $50 billion AI chip market.[iii] This picture demonstrates that US restrictions directed demand toward domestic suppliers rather than halting China’s capacity. However, the resulting structure, from DeepSeek’s perspective, is the replacement of one dependency with another. The company’s chip move can be read as a quest to manage this second dependency.
The focus on an inference chip is also no coincidence. As AI applications become more widespread, the sector’s computational load is shifting from model training to running models; this area can be sustained with cheaper and less energy-consuming chips compared to general-purpose graphics processing units. Considering that DeepSeek owes its global recognition to low-cost and efficient models, shaping the hardware, which constitutes the bulk of service costs, according to its own models is an attempt to carry the efficiency advantage established in software over to the hardware layer. Therefore, the move is a logical extension of the company’s business model.
From Beijing’s perspective, the crux of the matter is the software ecosystem surrounding the chip rather than the chip itself. Chinese state media presented the full adaptation of DeepSeek’s V4 model, released in April 2026, to Huawei Ascend chips as a sign of an era where domestic chips have ceased to be an “option” and become a “necessity”; emphasizing that dependency on Nvidia’s CUDA ecosystem pushed China into a passive position of “having models but lacking computing power.”[iv] This discourse clearly indicates that Beijing views the competition as a structural problem beyond a mere performance gap. Indeed, without software independence, even if domestic chips become stronger, developers will remain tethered to US standards.
The path DeepSeek is taking consists of two complementary tracks. On one hand, a co-design process is being carried out with Huawei that will reflect the computational challenges encountered in running large language models onto the memory bandwidth and inter-chip connection architecture of Ascend chips. On the other hand, the company is holding talks with chip design, foundry, and memory companies for its own inference chip and is increasing its hiring of chip design engineers. This dual structure is a choice to rely on Huawei’s scale in model training, where thousands of chips operate simultaneously, while establishing its own control in the inference domain, where cost is the determining factor.Thus, DeepSeek is building bargaining power that avoids being condemned to a single supplier in the domestic market.
For the Chinese giant Huawei, this collaboration carries the potential to transform the company from a substitute supplier replacing Nvidia into the load-bearing pillar of the ecosystem. Both the company’s plan to strengthen its CANN software platform to rival CUDA and its Ascend 960 roadmap announced for 2027 serve this goal.[v] The real workloads of a leading model developer serving as input for chip design is the Chinese equivalent of the feedback loop Nvidia has maintained with US labs for years. However, DeepSeek’s own chip is a factor that could, in the long run, turn Huawei’s customer into its competitor.
Consequently, the relationship between the two companies is a structure where cooperation and competition are intertwined. This tension does not contradict Beijing’s preferences. The fact that Alibaba and Baidu are also gaining market share by developing their own AI chips, and that domestic producers like Cambricon, Moore Threads, and Hygon adapted to the V4 model on its release day, shows that China aims for a pluralistic, mutually balancing domestic ecosystem rather than a single national champion. This approach is a resilience strategy that prevents a potential sanction or production disruption from paralyzing the entire system. Therefore, DeepSeek’s chip move can be evaluated as the corporate-level reflection of Beijing’s goal of self-reliance in science and technology.[vi]
In conclusion, DeepSeek’s initiative to develop its own chip and its deepening co-design process with Huawei show that China is building an integrated Techno-Hub that aligns models, hardware, and software. Washington’s response has accordingly extended from restricting inputs to excluding outputs: The DHS decision mandates the removal of third-party products built on DeepSeek models, alongside DeepSeek products themselves, from the systems of federal civilian agencies. On October 20, 2026, Chinese Ministry of Foreign Affairs Spokesperson Mao Ning stated that they oppose the expansion of the national security concept and the politicization of technological and economic issues, adding that they will resolutely protect the legitimate rights of Chinese companies.[vii] The fact that this reaction does not contain a concrete retaliation indicates that Beijing views the ban more as a development validating its domestic technology narrative than a loss. Therefore, in the coming period, the decoupling between the two stacks is expected to deepen, and the question of which ecosystem third countries will prefer is expected to become increasingly strategic.
[i] “China’s DeepSeek developing its own AI chip, sources say”, Reuters, https://www.reuters.com/world/china/chinas-deepseek-developing-its-own-ai-chip-sources-say-2026-07-07/, (Erişim Tarihi: 23.10.2026).
[ii] “DeepSeek and Huawei team up on AI chip design to challenge Nvidia”, Dataconomy, https://dataconomy.com/2026/10/20/deepseek-huawei-ai-chip-design-challenge-nvidia/, (Erişim Tarihi: 23.10.2026); “Notice of Order Regarding Exclusion and Removal of DeepSeek Products from Federal Civilian Executive Branch Information Technology Systems and Procurements”, U.S. Department of Homeland Security – Federal Register, https://www.federalregister.gov/d/2026-21447, (Erişim Tarihi: 23.10.2026).
[iii] “China’s DeepSeek developing its own AI chip, sources say”, Reuters, https://www.reuters.com/world/china/chinas-deepseek-developing-its-own-ai-chip-sources-say-2026-07-07/, (Erişim Tarihi: 23.10.2026).
[iv] “瞭望｜国产算力突围重塑产业竞争逻辑”, 新华社, https://finance.eastmoney.com/a/202605073729524171.html, (Erişim Tarihi: 23.10.2026).
[v] “DeepSeek and Huawei team up on AI chip design to challenge Nvidia”, Dataconomy, https://dataconomy.com/2026/10/20/deepseek-huawei-ai-chip-design-challenge-nvidia/, (Erişim Tarihi: 23.10.2026).
[vi] “瞭望｜国产算力突围重塑产业竞争逻辑”, 新华社, https://finance.eastmoney.com/a/202605073729524171.html, (Erişim Tarihi: 23.10.2026).
[vii] “2026年10月20日外交部发言人毛宁主持例行记者会”, 中华人民共和国驻印度尼西亚共和国大使馆, https://id.china-embassy.gov.cn/fyrth/202610/t20261020_12043799.htm, (Erişim Tarihi: 23.10.2026).
