Nvidia's Shanghai Absence Gives Huawei and China's AI Industry the Spotlight

Huawei Takes Center Stage as Nvidia Stays Away

Huawei's CloudMatrix 384 brought 384 Ascend 910C processors together across 16 racks at the World Artificial Intelligence Conference 2025 in Shanghai. Nvidia, the world's most valuable chipmaker, had no booth and did not participate in an official capacity, leaving Chinese competitors to command attention at one of the country's leading technology events.

As of July 29, 2025, Nvidia had not publicly explained its absence. Its position in China remained constrained by United States export controls, which prevent the company from selling its most advanced artificial intelligence chips to Chinese customers. Uncertainty also surrounded shipments of its H20 processor, a less advanced product designed for the Chinese market.

Huawei used the conference, known as WAIC, to showcase an ambitious alternative: a computing system that combines many domestic processors to pursue performance comparable with Nvidia's leading platforms. The approach requires much more electricity, making power consumption a central part of the competition.

The gathering extended beyond chips. Chinese companies announced alliances linking processors, AI models and infrastructure, while introducing products ranging from virtual livestreamers to glasses that accept voice commands. Premier Li Qiang also called for an international institution based in Shanghai to guide AI cooperation and governance.

The conference brought global technology companies, policymakers and investors together at a moment when access to computing equipment and influence over AI rules were increasingly connected. For Chinese businesses, building domestic alternatives has become both a commercial opportunity and a response to restrictions on foreign technology. For Nvidia, the same restrictions are complicating access to a market where demand for its processors remains strong.

Why Nvidia's Absence Mattered

Nvidia's decision to stay away surprised observers because the company had been working to preserve its Chinese business. Chief executive Jensen Huang had recently visited Beijing, and Nvidia had developed the H20 specifically to comply with American export restrictions.

The H20 is a less advanced version of Nvidia's flagship AI chips. Its purpose is to give Chinese customers access to some of the company's technology while staying within United States rules. Nvidia had not disclosed when H20 shipments would resume or provided details about order volumes.

More powerful products, including the H100 and Blackwell series, remained barred from the Chinese market. That left customers facing uncertainty about both the most advanced equipment and the availability of a product designed to meet the restrictions.

Industry analysts suggested that skipping WAIC could have been a strategic decision to avoid drawing attention to Nvidia's constrained position. Nvidia itself gave no explanation for its absence. An industry analyst, whose name was not given, described the competitive effect:

Nvidia's retreat is accelerating China's push for AI self-sufficiency and giving rivals like Huawei a chance to shine.

Chinese chipmakers gained an opportunity to present their products directly to prospective customers without Nvidia operating an official conference presence. The opening was especially valuable to Huawei, which could show how it was addressing the computing needs that American export controls have made harder to meet.

Inside Huawei's CloudMatrix 384

Huawei presented CloudMatrix 384 as a rival to advanced Nvidia systems such as the GB200 NVL72, which is unavailable to Chinese customers because of United States export controls. Its design combines 384 Ascend 910C AI processors across 16 racks.

A single Ascend processor is less powerful than Nvidia's Blackwell or H100 graphics processing units, commonly called GPUs. Huawei's response is to connect more processors and improve how the complete system works. Many chips can share a large computing task, but they need fast connections to exchange information without leaving one another waiting.

CloudMatrix 384 uses a peer-to-peer optical interconnect backbone for fast transfers with little delay. In simpler terms, its processors communicate through a network using optical connections, helping the assembled system operate as a coordinated machine.

The specifications presented for the platform describe substantial computing and memory capacity:

• 384 Ascend 910C processors distributed across 16 racks.
• More than 5.5 petabits per second of internal bandwidth.
• 49.2 terabytes of high bandwidth memory, described as three times the capacity of Nvidia's comparable platform.
• Memory bandwidth exceeding 1,200 terabytes per second, described as twice Nvidia's.
• Estimated computing performance of 300 petaflops, compared with 180 petaflops for Nvidia's GB200 NVL72.
• Power consumption of 559 kilowatts, about four times that of Nvidia's system.

Memory capacity describes how much information the system can hold close to its processors. Memory bandwidth describes how quickly that information can move. Both matter when AI systems must repeatedly process large amounts of data.

More Computing, With a Bigger Electricity Bill

Huawei's estimated 300 petaflops compares with 180 petaflops for Nvidia's GB200 NVL72, a difference of 120 petaflops, or about 67%. A petaflop is a measure of computing speed equal to a thousand trillion calculations per second.

Electricity use creates a different comparison. CloudMatrix 384's stated consumption of 559 kilowatts is about four times that of Nvidia's system. Data centers considering Huawei's platform therefore have to weigh its computing capacity against the power required to operate it.

Huawei's strategy places much of the engineering effort at the level of the complete system. More processors, extensive memory and fast connections can compensate for a weaker individual chip, producing comparable or greater combined performance while requiring more energy.

The system is also a strategic response to American sanctions. A domestic alternative gives Chinese organizations another route to AI computing when leading foreign processors are unavailable. Huawei's progress consequently matters beyond the specifications of a single machine: it supports China's effort to reduce reliance on foreign technology.

Huawei's Market Opportunity Has Limits

Huawei faces restrictions of its own. United States export laws constrain its international business, while the United States Commerce Department has warned that companies using Huawei's Ascend chips around the world could face penalties because the hardware was designed using technology originating in the United States.

That threat sharply narrows Huawei's opportunities abroad and effectively confines its AI hardware business to the Chinese market. The company's domestic opportunity, however, is large.

Analysts expected Huawei's share of China's AI chip market to rise from 23% to 28% in 2025. That would be an increase of five percentage points, as customers seek alternatives to restricted Nvidia products.

China's spending on AI chips was projected to reach as much as $39.5 billion during the year. Broader investment in AI infrastructure was approaching $91 billion. Those figures describe different scopes of spending: chips are one part of the equipment and systems needed to run AI.

Alibaba and ByteDance were among the major Chinese technology companies investing heavily in AI computing. Their demand helps create a market for domestic processors, while giving suppliers a reason to keep improving the hardware and software needed to use them.

New Alliances Connect Chips and AI Models

Chinese companies also used WAIC to organize cooperation across the technology chain. Two new groups aimed to connect businesses that build AI models, supply processors and bring the technology into industrial use.

The Model-Chip Ecosystem Innovation Alliance brings developers of large language models together with chip manufacturers, including Huawei, Biren, Moore Threads and Enflame. Large language models are AI systems that learn patterns in language and use them to generate or process text.

The alliance's goal is to connect chips, models and infrastructure. Those components have to work together: a powerful processor is more useful when AI software can run efficiently on it, and a capable model needs suitable computing equipment to reach users.

The Shanghai General Chamber of Commerce AI Committee focuses on integrating AI with industrial transformation. Participants include SenseTime, StepFun, MiniMax, Metax and Iluvatar CoreX.

Both initiatives respond to sanctions that have blocked Chinese firms from buying advanced technology made with American know-how. Cooperation within China is intended to reduce dependence on foreign suppliers and speed the development of domestic solutions.

The practical challenge goes beyond replacing one imported chip. Chinese businesses are trying to build a connected technology stack, meaning the layers of hardware, software and infrastructure required to turn an AI model into a working service.

From Supernodes to Glasses and Virtual Presenters

Metax demonstrated an AI supernode containing 128 C550 chips for large data centers cooled with liquid. Like Huawei's system, the product showed an approach based on bringing many processors together to handle demanding computing work.

Tencent introduced its open-source Hunyuan3D World Model 1.0, which allows users to generate interactive three dimensional environments from text or image prompts. Open-source development makes technology available for others to use or build upon under its release terms.

Baidu announced a new generation of digital human technology for virtual livestreamers. The system was described as capable of copying a person's voice, tone and body language from just 10 minutes of sample footage.

Alibaba introduced Quark AI Glasses, powered by its Qwen AI model. Users can navigate maps and use Alipay through voice commands and QR code scanning, bringing AI functions into a wearable device.

NetEase's Youdao presented a handheld device using AI for student learning. Its appearance highlighted work on power efficiency and edge computing, where processing happens on or near a user's device rather than relying entirely on distant data centers.

The range of products showed why the conference was about more than a contest between chipmakers. Chinese companies were developing the underlying computing systems alongside applications for entertainment, commerce, navigation and education, serving domestic needs while seeking opportunities to compete internationally.

Washington and Beijing Compete Over AI Rules

United States President Donald Trump announced a new AI Action Plan just days before WAIC. It aimed to promote American technology overseas and reduce bias in AI models, placing the conference against a fresh policy push from Washington.

Premier Li Qiang used his keynote address to call for an international institution guiding AI collaboration and governance, with its headquarters in Shanghai. He argued for common standards and better management of the risks accompanying rapid development.

Li, China's premier, described the need to coordinate national policies:

AI is a powerful driver of economic expansion, but regulatory approaches vary widely. Nations must harmonize their policies to ensure that rapid advancements do not outpace safeguards.

Li also warned against AI development being monopolized by a small number of players. He pointed to export controls and restrictions on the movement of experts as challenges facing international cooperation.

China's support for open-source AI projects and international collaboration forms part of its effort to shape the rules and standards governing the technology. The competition therefore concerns both access to equipment and influence over how AI is developed and used.

The sequence leading into July 29 put commercial announcements alongside government strategy:

• Days before WAIC, Trump announced the United States AI Action Plan.
• At WAIC 2025 in Shanghai, Huawei showcased CloudMatrix 384, Chinese companies announced alliances and products, and Li proposed an international AI institution.
• As of July 29, Nvidia had not explained its absence or disclosed a date for H20 shipments to resume and order volumes.

Repair Shops Keep Restricted Nvidia Chips Working

Nvidia's absence from the conference contrasted with continuing Chinese demand for its advanced processors. In cities including Shenzhen, repair workshops had emerged to service Nvidia AI chips that entered the country through unofficial channels.

The workshops refurbished restricted H100 and A100 GPUs, with servicing volumes reaching as many as 500 units a month. Repairing existing equipment gave customers a way to keep valuable computing hardware operating despite barriers to obtaining new processors.

Nvidia warned that only authorized partners could provide proper service and support. The company also said using restricted products without approved support was not viable.

The repair trade existed alongside the push for domestic alternatives. Chinese customers were seeking new computing capacity from companies such as Huawei while continuing to maintain Nvidia equipment already in circulation.

China's AI Progress Reaches Beyond Hardware

DeepSeek's R1 and V3, Alibaba's Qwen3 and MiniMax M1 illustrated another part of China's AI development. The open-source models had matched or surpassed Western counterparts on some benchmarks despite American restrictions on advanced computing equipment. Benchmarks are tests used to compare particular capabilities of AI systems.

Some Western observers had underestimated that progress by concentrating on China's access to Nvidia GPUs. China's advances also reflected a national strategy developed over a decade, coordination between public and private organizations, and substantial investment in talent and infrastructure.

Chinese developers combined changes in model architecture, greater efficiency and the rapid spread of AI across industries. Architectural innovation means changing how a model is organized and processes information, another route to improving results alongside adding computing power.

China's regulatory infrastructure included a database for registering generative AI algorithms, providing oversight while development continued. Alignment among government, universities and industry helped research move into practical use at scale.

The broader approach connects education, infrastructure, talent development and regulation. Together with domestic hardware and industry alliances, it has helped China close the gap with the United States and pursue a role in setting global AI standards.

Nvidia remained the global leader in AI data center systems, but its Chinese business faced both political restrictions and stronger domestic competition. An industry expert, whose name was not given, described the challenge:

Nvidia's ability to navigate these challenges will determine its long-term value and influence in the global AI ecosystem.

For businesses buying AI equipment, the immediate choices involved availability, performance, electricity requirements and dependable support. WAIC showed Chinese suppliers working on each part of that chain while Nvidia's access to the market remained uncertain.

Key Points

• Nvidia had no booth or official participation at WAIC 2025 in Shanghai and had not explained its absence.
• Huawei's CloudMatrix 384 combines 384 Ascend 910C processors across 16 racks, with estimated performance of 300 petaflops and power consumption of 559 kilowatts.
• Analysts expected Huawei's share of China's AI chip market to rise from 23% to 28% in 2025.
• New industry alliances connected Chinese chipmakers, model developers and businesses adopting AI.
• Conference products included Metax's supernode, Tencent's world model, Baidu's digital humans, Alibaba's AI glasses and Youdao's learning device.
• Li Qiang proposed an international AI governance institution in Shanghai, days after Trump announced an American AI Action Plan.
• Demand for restricted Nvidia processors continued, including through repair workshops, while Chinese companies expanded domestic hardware and open-source models.