Without surprise, China ditch US AI chips and decides to go its own way.
US has to to blame itself
US has a poor record on backdoor
The US government has various technical means to create “backdoors” in chips, potentially without the manufacturer’s knowledge.
In 1993, the National Security Agency mandated the installation of the Clipper chip in AT&T’s secure communication equipment, ensuring that US government agencies could access call content through the chip’s built-in key escrow system. However, security researchers later discovered a significant flaw in the system, potentially allowing malicious parties to tamper with the software.
As Nvidia itself stated in its statement, “The Clipper chip’s built-in backdoor proved to be a mistake,” adding to the tainted record of US intelligence agencies.
In 2024, the US government installed not only large tracking devices on the packaging of Dell servers equipped with Nvidia chips, but also smaller, more discreet tracking devices inside the packaging and even inside the servers.
On August 12, 2025, Reuters reported that US authorities have secretly installed tracking devices in shipments of advanced chips they deem high-risk to prevent the illegal transfer of these chips to export-controlled destinations, such as China. These tracking measures primarily target AI chips being transferred to destinations subject to US export restrictions and apply only to a subset of shipments under investigation, including those containing Nvidia and AMD semiconductor chips.
Five individuals involved in the AI server supply chain said they were aware of server shipments from companies including Dell and AMD that contained Nvidia and AMD semiconductor chips, which had been fitted with tracking devices.
US lawmakers have introduced legislation
In May of this year, US lawmakers introduced a bill called the Chip Security Act, which would require all high-end artificial intelligence (AI) GPUs and AI chips to be equipped with location tracking technology within 180 days to ensure they are not destined for certain countries, particularly mainland China.
US officials do intend to do so
Michael Kratsios, director of the White House Office of Science and Technology Policy, stated that the US government is exploring various methods to more effectively track the movement of AI chips and prevent chips from companies like Nvidia, which are subject to export controls, from flowing to mainland China.
To ensure US technological leadership, Washington is developing a wide-ranging plan, one of which involves collaborating with industry to monitor the movement of sensitive components and combat smuggling. Kratsios is one of the key architects of the AI action plan announced by US President Trump. In an interview with Bloomberg Television, he stated that current discussions with the industry involve “whether it’s possible to make the chip itself more traceable through some sort of software or physical adjustments.” Kratsius emphasized, “This is something that’s clearly part of the plan.”
U.S. Commerce Secretary’s mocking and insulting remarks
The Financial Times reported that Chinese regulators are taking action to discourage local companies from purchasing Nvidia H20 chips due to dissatisfaction with “insulting” remarks made by U.S. Commerce Secretary Lutnick. The report, citing sources, stated that relevant agencies, including the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology, have taken action to discourage local tech companies from purchasing Nvidia H20 chips in response to Lutnick’s remarks in July when the U.S. lifted the ban on Nvidia H20 sales to China.
In an interview at the time, Lutnick stated that the H20 was a “fourth-grade” wafer designed for the Chinese market: “We don’t sell them top-grade, second-grade, or even third-grade.” He also emphasized that the policy’s purpose was to “ensure China’s continued dependence on American technology” while not posing a national security threat.
Sources indicate that some senior Chinese officials considered these remarks “insulting” and, as a result, have restricted Chinese tech companies from purchasing H20 wafers. Several Chinese tech companies are reportedly suspending or significantly reducing their orders.
China ditch US AI chips is now in place
China questioned backdoors in US chips
On August 10, 2025, the CCTV official account “Yuyuan Tantian” published an article accusing the US of “backdooring chips,” claiming that the Nvidia H20 chip is fully capable of a “remote shutdown” function, including setting voltage, timing, and other conditions to automatically shut down the chip. The H20 chip was also criticized as “neither environmentally friendly, advanced, nor safe” for China.
Following reports from mainland official media outlets such as People’s Daily and Global Times, China Central Television (CCTV) released a program on August 22, 2025, exposing the “background” risks of criticizing American chips. It also pointed out that Nvidia’s computing power chips already have mature “tracking and positioning” and “remote shutdown” technologies, saying, “It’s like the seller secretly inserting an ‘electronic key’ into your smart door lock, allowing you to open your door at any time without your knowledge.”
The Chinese government also raised security concerns about the H20 chip and summoned Nvidia for an explanation. Officials also discouraged companies from using Nvidia’s H20 chips, forcing Nvidia to halt production of a downgraded version of the H20 chip designed specifically for China.
China’s official policy announcement
Bloomberg reported on August 12, 2025, that relevant Chinese authorities recently issued a notice to domestic companies, recommending that they avoid using Nvidia’s H20 chips, designed specifically for the Chinese market, particularly in government and critical infrastructure sectors.
Sources familiar with the matter revealed that Chinese state-owned data center companies will be required to purchase more than 50% of their chips from domestic manufacturers to support the domestic semiconductor industry.
Shanghai, China, issued guidelines last March requiring “smart computing centers to use at least 50% domestic computing and storage chips by 2025,” a policy that now appears to have been implemented nationwide.
The Financial Times reported that Beijing has required Chinese tech companies to justify their orders for H20 chips, thereby promoting domestic alternatives like Huawei and Cambricon.
Chinese are abandoning US-made chips
DeepSeek will exclusively use Chinese chips
Recent reports indicate that DeepSeek’s DeepSeek R2 model has been delayed due to Beijing’s request to switch to training with Huawei chips instead of NVIDIA products. Mainland China AI startup DeepSeek recently released its large model V3.1, which utilizes the UE8M0 FP8 architecture and claims to be designed specifically for “the upcoming next generation of domestically produced chips,” sparking expectations for the widespread adoption of domestically produced AI chips in mainland China.
Alibaba
Alibaba has a semiconductor design division called T-Head Semiconductor Co., Ltd. The division’s last inference chip was the Hanguang(含光) 800 in 2019. These chips are installed in servers in data centers to handle AI workloads.
Alibaba announced in its second-quarter earnings report at the end of August that it had developed a new AI inference chip to enhance China’s autonomy and reduce its reliance on Nvidia. This new AI inference chip, known as the Zhenwu (真武) processor, is reportedly complete but not yet deployed, and its performance rivals that of the H20.
Note: Alibaba’S T-Head Semiconductor has long relied on its own chip design division, Pingtou Ge, to develop numerous consumer, server, and other chips; Pingtou Ge is notably China’s largest RISC-V chip provider.
Unlike previous chips designed for limited scenarios manufactured by TSMC, this new chip is manufactured by a Chinese manufacturer and is compatible with Nvidia’s platform, allowing engineers to reuse existing software tools. Notably, this chip is not intended for AI training, but rather for the rapidly growing demand for “inference.”
Against the backdrop of US restrictions on China’s access to advanced AI chips and manufacturing technology, Alibaba’s move is undoubtedly a significant step for China to reduce its dependence on the US and strengthen its industrial chain independence. Alibaba’s strategic layout demonstrates its continued commitment to “AI + Cloud,” with plans to invest $53 billion in AI and cloud services over the next three years.
Tencent
Amidst international trade protectionism, uncertainty remains in the international chip supply chain. The release of DeepSeek V3.1, coupled with Tencent’s announcement that it will no longer need to purchase H20 chips and will shift to non-Western AI chips for inference, are both beneficial to the development of domestically produced AI computing power and chips in mainland China.
China Mobile
China Mobile recently completed a tender for its 2025-2026 AI general-purpose computing equipment (inference-based) with a total value of 5.112 billion RMB. The centralized procurement was entirely awarded to mainland-made chips.
Companies are being asked not to use US AI chips
According to a report by the Financial Times on Tuesday, Beijing is pressuring major Chinese tech companies, such as Alibaba and ByteDance, to place orders for H20 chips. The Ministry of Industry and Information Technology (MIIT) has demanded that these companies explain why they are ordering H20 chips rather than using domestic alternatives. Some tech companies are planning to reduce orders due to regulatory concerns. Tencent has already taken the lead, revealing that it no longer needs to purchase H20 chips.
Self-sufficiency target: 70% to 100%
Nikkei Chinese reported on August 25, 2025, that several major local governments in China, home to artificial intelligence (AI) and data centers, recently set high self-sufficiency targets for semiconductors designed or manufactured by Chinese companies. The Beijing Municipal Government plans to raise self-sufficiency to 100% by 2027.
The Shanghai Municipal Government plans to increase China’s self-sufficiency rate for semiconductors used in data centers for AI and other applications to over 70% by 2027. The Gui’an New District in Guiyang, Guizhou Province, home to China’s most important data center, has also mandated that approximately 90% of the semiconductors in its AI-focused data centers under construction be made in China.
Chinese chip companies are benefiting
First of all, please see my post of “Chinese AI progress and top companies” to get the latest list of Chinese AI SoC vendors.
GPU stocks are cheering across the board
On August 22, 2025, the domestic computing power sector saw a significant surge. The Wind GPU Index rose 9.87%, with Cambricon’s stock price rising 20%, reaching a new high and a market capitalization exceeding 500 billion yuan. Hygon Information’s stock price rose 20%, also reaching a new high, with a market capitalization exceeding 400 billion yuan. If the progress of Hygon Information’s merger with Sugon is also taken into account, Hygon Information’s market capitalization is likely to exceed 1 trillion yuan.
Thus, if all publicly listed domestic companies with GPU-related businesses are combined, the combined market capitalization of Cambricon, Hygon Information, Jingjia Microelectronics, and VeriSilicon will officially exceed 1 trillion yuan on August 22, 2025.
The surge was primarily driven by the news that Nvidia had halted H20 production. On August 21, 2025, the technology media outlet The Information, citing sources familiar with the matter, reported that Nvidia had instructed key component suppliers, including Samsung Electronics of South Korea and Amkor Technology of the United States, to suspend production related to the H20 chip.
Cambricon
Following reports that mainland Chinese officials urged domestic companies to avoid using Nvidia’s H20 chips, coupled with market rumors about the impending release of the DeepSeek-R2, semiconductor chip stocks in Shanghai and Shenzhen strengthened on August 12, 2025. Among them, Cambricon, mainland China’s “first domestically produced AI chip stock” and considered a viable alternative to Nvidia, surged 20%, reaching its daily limit and stabilizing above the RMB 800 mark, setting a new record high.
On August 28, 2025, Cambricon surpassed Kweichow Moutai at the close, officially becoming the top-ranked Chinese A-share stock. Its total market capitalization reached 664.3 billion RMB, surpassing even semiconductor giant SMIC’s A-share market capitalization of 580.4 billion RMB. However, Cambricon’s price-to-earnings ratio was a staggering 595.33 times.
Cambricon released its first-half financial report on the evening of the 26th, reporting revenue of RMB 2.881 billion, a 43-fold year-over-year increase, and net profit of RMB 1.038 billion, a significant turnaround from a loss of RMB 530 million in the same period last year. Earnings per share were RMB 2.5.
Baidu
To break Nvidia’s monopoly in the mainland AI market, Kunlun Chip secured a multi-billion RMB chip order from China Mobile’s 2025-2026 centralized procurement program for artificial intelligence general-purpose computing devices (inference-based). The entire centralized procurement project uses only mainland-made products.
Kunlun Core, formerly the semiconductor design company of Baidu’s Smart Chip and Architecture Department, has also become a key option for Baidu’s AI training. Baidu is currently experimenting with using its Kunlun P800 chip to train the new Ernie AI model.
Reuters reported that Kunlun Core issued a statement on August 21, 2025, stating that it would provide chips compatible with the CUDA ecosystem to first-tier suppliers of China Mobile, including H3C Technologies and ZTE. CUDA is the Compute Unified Device Architecture (CUDA) developed by Nvidia. The company claimed the order was worth billions of RMB, but the statement did not disclose the exact amount.
Huawei
Sources familiar with the matter revealed that after releasing the R1 model in January 2025, DeepSeek was encouraged by officials to switch to using the Huawei Ascend processor, which Nvidia considers its biggest competitor in China, rather than Nvidia’s own system. However, due to persistent technical issues during R2 training using Ascend chips, the company switched back to Nvidia chips for training, using Huawei chips only for inference.
Sources familiar with the matter revealed that this was the primary reason for the model release delay from May, causing DeepSeek to fall behind the competition. Huawei is experiencing growing pains with its use of Ascend for training, but it is expected to eventually adapt. Just because we haven’t seen leading models trained on Huawei today doesn’t mean they won’t happen in the future; it’s just a matter of time.
MetaX
Shanghai startup MetaX has also made a name for itself in this competition. In July, the company released a replacement chip for the H20, which improves performance by combining two smaller chips. Despite higher power consumption, the company is now pursuing mass production. Meanwhile, Huawei’s Ascend system has demonstrated its inference computing capabilities through stacking and clustering, achieving performance comparable to Nvidia’s top-tier Blackwell system in some cases, though overall power consumption remains relatively high.
Nvidia’s Fall from Heaven to Hell
US Takes 15% Additional Commission from H20
In early August 2025, Nvidia agreed to pay the US government a 15% commission on sales of its H20 chips to China in exchange for obtaining an export control license to resume shipments to Chinese companies. Trump revealed that he initially offered a 20% price cut during his White House meeting with Jensen Huang, but Huang negotiated a reduction to 15%.
At a press conference, Trump emphasized that the H20 is an older chip designed for the Chinese market after undergoing regulatory oversight. “China already has similar chips, and so does Huawei,” he said, adding that its performance is significantly different from the H100 and H200 chips sold in the US.
Nvidia Suffers Heavy Losses
Related articles
- “China ditch US AI chips and decides to go its own way“
- “AI inference chips vs. training chips“
- “A must-read for Nvidia investors＂The Thinking Machine＂“
- “How GPU farms CoreWeave make money?“
- “Geoffrey Hinton, 2024 Nobel Physics winner, inadvertently helped Nvida transform to AI overlord“
- “Chinese AI progress and top companies“
- “DeepSeek routed the global AI and stock“
- “Top vendors and uses of GPU“
- “How does CUDA strengthen the moat of Nvidia’s monopoly?“
- “Comparison of AI chips GPU and ASIC“
- “ASIC market is getting bigger, and related listed companies in the US and Taiwan“
- “Significant changes in Broadcom’s business approach“
- “The reasons behind Broadcom share price consistantly outperformance“
- “How low-key Marvell makes money?“
- “How does nVidia make money, Nvidia is changing the gaming rules“
- “The reasons for Nvidia’s monopoly and the challenges it faces“
- “Why nVidia failed to acquire ARM?“
- “Revisiting Nvidia: The Absolute Leader in Artificial Intelligence, Data Center, and Graphics“
- “Data center, a rapidly growing semiconductor field“
Disclaimer
- The content of this site is the author’s personal opinions and is for reference only. I am not responsible for the correctness, opinions, and immediacy of the content and information of the article. Readers must make their own judgments.
- I shall not be liable for any damages or other legal liabilities for the direct or indirect losses caused by the readers’ direct or indirect reliance on and reference to the information on this site, or all the responsibilities arising therefrom, as a result of any investment behavior.
