China’s artificial intelligence (AI) ambitions are running up against a hard wall of silicon. Earlier in September 2025, Beijing’s internet regulator instructed some of the country’s largest technology companies – including ByteDance, Alibaba, and Tencent – to halt testing and orders of Nvidia’s AI semiconductor chips. That includes the RTX Pro 6000D, a product custom-designed to meet US restrictions on exports of advanced chips to China.
“AI has become the fulcrum of geopolitical tension between the US and China,” analysts at Deutsche Bank wrote in a research note earlier this month appropriately titled “The Summer AI Turned Ugly.”
The ban marks a new stage in the US–China chip war, and it is affecting how Chinese AI leaders – from DeepSeek to SpikingBrain – build and scale their models.
- Beijing is prohibiting Chinese tech companies from buying Nvidia’s China-specific AI chips, which will force them to accelerate their shift to domestic hardware.
- Huawei’s new AI cluster, powered by its Ascend chips, represents the most ambitious domestic challenge to Nvidia’s dominance.
- Chinese LLMs face transition risk as moving from Nvidia to domestic chips could slow scaling in the short term, although it may drive ecosystem maturity over time.
- Projects like SpikingBrain highlight how China may explore non-traditional AI architectures that can outperform without Nvidia’s hardware.
- The chip war is pushing China toward a self-sufficient AI infrastructure, likely leading to a fragmentation of the global AI ecosystem.
China Bans Nvidia Chips, Accelerating Domestic AI Chip Development
The loss of access to chips from Nvidia (NVDA), the global market leader in AI hardware, is a significant shift for Chinese companies, forcing them to rely on domestic alternatives while potentially slowing their ability to compete in the global AI development arena.
Until the US enforced sweeping export controls in October 2022 to block China’s access to advanced semiconductors and the equipment needed to manufacture them, the world’s largest chip importer was considered to be several years behind its US and Asian counterparts in developing advanced chip technologies.
Washington aims to slow China’s military and technological advances as advanced semiconductors can be used in weapons systems and in training the large-scale AI models that underpin surveillance, cyberwarfare, and autonomous capabilities.
In response, Beijing is channeling vast state resources into local fabrication, packaging, and alternative computing methods. Chipmakers in the country are aiming to triple local production of AI chips next year in response to soaring demand. Buyers have also been able to access some high-end chips through back doors, according to multiple reports, Deutsche noted.
The move by the Cyberspace Administration of China (CAC) to restrict the use of Nvidia’s technology signals the government is stepping up efforts to develop its domestic industry to self-sufficiency and compete with the US.
The ban extends beyond guidance from regulators in July 2025 that focused on Nvidia’s H20 chip, which it had also developed specifically to comply with the US export controls to be able to sell to China.
It comes after Chinese regulators concluded that domestic manufacturers have developed their own semiconductors that can achieve performance comparable to Nvidia’s models available in China, according to the Financial Times.
Domestic Innovation Steps Up
“Nvidia, which designs the advanced chips that are the cornerstone of cutting-edge AI, is at the heart of the issue,” the Deutsche analysts noted.
For years, Chinese tech giants have relied on Nvidia’s graphics processing units (GPUs) to train and deploy large AI models. The company’s H100 and A100 chips became the gold standard for handling complex neural networks, powering everything from ChatGPT-like services to computer vision applications.
When the US introduced sweeping export restrictions in 2022 and 2023, Nvidia responded with stripped-back versions to comply with the directive not to supply advanced chips. These included the A800 and H800, later followed by the RTX Pro 6000D.
But Washington’s tightening rules – and Beijing’s move to block local purchases – have effectively closed the door on this workaround. Chip and equipment suppliers have previously reported that Chinese buyers were purchasing the products available to them in large quantities to meet the needs of domestic industries. The Nvidia chip ban now leaves firms scrambling to secure alternative compute power.
Domestic manufacturers are racing to fill the vacuum. Chinese electronics firm Huawei, long a target of US sanctions, last week introduced its SuperPod system, which strings together more than 15,000 of its Ascend AI chips into a high-performance cluster, its boldest push yet to rival Nvidia.
Huawei claims the platform can deliver performance levels suitable for training frontier-scale models.
- The Ascend 910C chip has been available since the first quarter of this year.
- The company said at the Huawei Connect event that it will release the Ascend 950 series, including the 950PR and 950DT, over the next year.
- Huawei plans to incorporate advanced high-bandwidth memory (HBM), which it has developed in-house, into these chips.
- It expects to release the Ascend 960 and Ascend 970 in 2027 and 2028, making them competitive with AI chips from Nvidia and Advanced Micro Devices (AMD).
“Computing power is – and will continue to be – key to AI. This is especially true in China,” Xu Zhijun, Huawei’s deputy chairman, said at the event.
According to the company’s statement summarizing the keynote, he reiterated that:
“Huawei’s goal is to sustainably meet long-term computing demand by building SuperPoDs and SuperClusters with the semiconductor manufacturing process nodes that are practically available to the Chinese mainland.”
SuperPod is being marketed not only to Chinese cloud providers but also to enterprises seeking local alternatives to Western technology. For companies like Baidu, which runs its own Ernie large language model (LLM), and SenseTime, which focuses on computer vision, this could provide a lifeline.
“Huawei’s Ascend chips have become the national standard, helping the company turn a fourth quarter loss into a first-half profit… while rivals including Alibaba are racing to develop their own substitutes,” the Deutsche analysts wrote. “That has contributed to an investment frenzy, boosted by evidence of real growth in AI and cloud revenues.”
Shares in Alibaba surged recently after the company reported a triple-digit percentage gain in first-quarter AI revenue, and shares in other Chinese tech giants, such as Baidu and Tencent, also rose.
Meanwhile, startups like SpikingBrain boast of building “100x faster AI” on homegrown semiconductors. The narrative is clear: China is doubling down on chip independence, even if the path forward is uneven.
The Impact on Chinese AI Models
How will China’s Nvidia ban and domestic substitution affect the country’s AI projects?
Positioned as one of China’s leading general-purpose AI models, DeepSeek has ambitions to compete with OpenAI’s GPT series. Training such a model requires immense GPU clusters and optimized hardware–software integration.
The DeepSeek model was built using cheaper chips at a fraction of the price of OpenAI and other Western LLMs. DeepSeek’s reliance on homegrown technology may accelerate the development of custom training pipelines tailored to Chinese hardware.
Developed by researchers at the Chinese Academy of Sciences’ Institute of Automation in Beijing, the SpikingBrain 1.0 AI model is designed to consume less energy and perform better than models that run on Nvidia hardware.
The developers claim the system can run up to 100 times faster than traditional models and is focused on neuromorphic computing.
SpikingBrain is based on architectures that imitate how the human brain only fires the neurons it needs. This enables it to selectively respond to input instead of activating the entire network like ChatGPT, and learn from small sets of training data to speed up response times and save power even when processing long blocks of text.
The system runs entirely on Chinese infrastructure, including the homegrown MetaX chip platform. Having another domestically produced and independent LLM is strategically important as the trade war escalates.
Chinese Internet giant Baidu is developing Ernie as a “foundation model” that powers conversational AI chatbots as well as multimodal functions such as document comprehension, visual reasoning, retrieval-augmented generation (RAG), and more. This is part of a larger play to reinvent the company’s core search and cloud businesses.
Alibaba is developing Qwen as a general-purpose AI engine for its extensive ecosystem that spans cloud, e-commerce, logistics, and enterprise software. The company has emphasized open licensing and collaboration to build developer ecosystems around Qwen.
Baidu claims that Ernie 4.5 outperforms GPT-4o (OpenAI’s multimodal version) in multimodal benchmarks, scoring on average 77.77 versus 73.92 for GPT-4o. In January 2025, Alibaba announced Qwen 2.5-Max, claiming it outperforms DeepSeek-V3, GPT-4o, and Llama-3.1 “across the board.”
These numbers suggest that Chinese LLMs can now credibly match or surpass leading Western models in benchmarking – at least in the controlled settings disclosed. But benchmarks are only part of the story; real-world deployment is where it matters.
These household names in China’s AI race are now facing a transitional period. Both companies have staked their reputations on competing with Western frontrunners like OpenAI and Anthropic, and their models are already being embedded into search engines, e-commerce platforms, and enterprise software.
Their competitive edge depends on continuous model scaling, which may slow under hardware constraints. However, as they migrate toward domestic chips, they could become early stress-testers of Huawei’s platforms.
Short-Term Pain, Long-Term Repositioning
Still, challenges remain. Huawei’s chips are less mature than Nvidia’s, with limitations in software ecosystems and developer tools.
The immediate consequence of a ban on Nvidia chips, if it holds, could involve delays in model training, which requires hundreds of billions of parameters and massive compute power, which domestic chips may struggle to provide efficiently. Chinese startups could face delays in model rollouts, higher costs, and technical bottlenecks.
But the long-term trajectory could favor China. By being forced off Nvidia’s roadmap, Chinese companies could invest more heavily in R&D, software ecosystems, and alternative architectures. The rise of neuromorphic AI, domain-specific accelerators, and energy-efficient chips may create breakthroughs less dependent on Western hardware.
The effects are likely to extend beyond China. The chip war risks fragmenting the global AI landscape into two parallel systems – multinational companies operating in China may need to redesign their AI strategies around Huawei or other domestic hardware, creating compatibility challenges.
Meanwhile, Nvidia’s global dominance is being tested. The company continues to thrive in Western markets, but its exclusion from the world’s second-largest economy could dent long-term growth.
The company reported that China accounted for around 13% of its revenue in the financial year that ended in January. Competitors like AMD and Intel, although also restricted by U.S. export controls, may explore indirect opportunities.
The Chinese government’s move to prevent its largest tech companies from buying Nvidia’s AI chips represents more than a procurement setback – it is a turning point in the US-China chip war.
There may potentially be short-term disruption and slower progress for leading Chinese AI models, but over the long term, China could develop a more self-sufficient ecosystem powered by domestic innovations like Huawei AI chips and experimental approaches such as neuromorphic computing.
For DeepSeek, SpikingBrain, and their peers, the transition away from Nvidia will be challenging, but it could forge a uniquely Chinese path to AI leadership that is less dependent on Western silicon and more attuned to homegrown innovation.
What does China’s ban on Nvidia AI chips mean for its tech companies?
The ban prevents Chinese firms from accessing Nvidia’s China-specific GPUs, which have been used for training large AI models. This forces companies like Alibaba, Baidu, and ByteDance to rely on domestic chips from Huawei and others. While the transition may slow short-term progress, it aligns with Beijing’s long-term goal of achieving chip self-sufficiency.
How will Chinese AI models like DeepSeek adapt to domestic chips?
Models such as DeepSeek will need to re-engineer their training pipelines to work efficiently on Huawei’s Ascend chips or other local accelerators. This could create bottlenecks in scaling and performance, but may also lead to innovations specifically designed for domestic hardware.
What is Huawei’s SuperPod?
Huawei’s SuperPod is a high-performance AI cluster built on its Ascend GPU line, linking more than 15,000 chips into a massive compute system. It is marketed as a direct alternative to Nvidia-powered clusters and is intended to support the training of China’s large-scale AI models.
How does the US–China chip war shape global AI competition?
The trade war could fragment the global AI market into parallel ecosystems – one built on Nvidia and Western suppliers, the other on Chinese domestic hardware.
Could alternative chip architectures give China an edge?
Yes. Projects like SpikingBrain, which experiment with neuromorphic or brain-inspired chips, show how China could bypass traditional GPU limitations.