Huawei Takes Over the Chinese AI Chip Market Against Nvidia

Huawei Establishes Itself in the Chinese AI Chip Market

In a context of intense technological rivalry, Huawei is striving to assert itself against the American giant Nvidia. However, the Chinese company faces significant obstacles, particularly its inability to access advanced technologies from ASML and TSMC.

Taking advantage of Nvidia's absence in the Chinese artificial intelligence chip market, Huawei is engaging in rapid expansion. This strategy allows it to anticipate a 60% increase in sales this year, primarily driven by large Chinese companies seeking local alternatives to Nvidia.

Huawei expects its AI chips to generate approximately $12 billion in revenue this year, based on already recorded orders. This figure could reach $7.5 billion by 2025. For the first time, Huawei could match or even surpass Nvidia's revenue in China, marking a historic turning point in a market where Nvidia held over 80% market share in 2023. This development signals a gradual de-Americanization of AI infrastructure in China.

American Sanctions as a Catalyst for Huawei

Huawei's growth is partly fueled by the rise of the Chinese AI market, which is projected to reach between $30 and $35 billion by 2026, according to forecasts from TrendForce and SemiAnalysis. However, it is primarily Nvidia's withdrawal, compelled by American sanctions, that is facilitating Huawei's advancement.

"The technological competition between China and the United States has unexpected but predictable consequences. American restrictions act as a catalyst, accelerating China's quest for technological self-sufficiency. Nvidia's exit from the Chinese market is stimulating technological progress for companies like Huawei, SMIC, and CXMT. The Chinese semiconductor sector is advancing more rapidly precisely because it is being forced to," explains Mina Kim, senior economist at MKEcon Insights.

Since 2022, the United States has gradually limited China's access to Nvidia chips, the most powerful for AI, starting with the most advanced models before also targeting the throttled versions designed to circumvent the initial sanctions. This strategy aims to curb China's AI advancements, as the country is the only serious potential rival to the United States in this field, even if it means sacrificing some of Nvidia's revenue. This approach has proven effective: projects like DeepSeek would not have come to fruition without Nvidia chips obtained through indirect means.

In response, the Chinese government has sought to circumvent the sanctions to maintain its lead in AI while officially striving for self-sufficiency. Beijing has notably mandated that all state-funded data centers exclusively use Chinese AI chips and has supported Huawei through various means:

• Subsidies for electricity for data centers using local chips
• Public contracts directed towards Chinese AI
• Competition among local governments with tax exemptions and dedicated computing clusters

Huawei's Production Challenges

To strengthen its position in the Chinese market, Huawei is banking on its new Ascend 950PR chip, designed for inference. Compared to Nvidia's throttled H20 chip, the 950PR boasts superior performance and is approaching Nvidia's Hopper architecture (H100), launched in 2022. However, against Nvidia's latest Blackwell generation, the 950PR lags one generation behind in raw capabilities, particularly in memory bandwidth and training workloads.

Huawei's main challenge lies in its ability to produce at scale. "Huawei aimed to produce one million Ascend chips by 2025, a target it has not met," highlights Antoine Chkaiban, a consultant at New Street Research. The Chinese government's ambition cannot compensate for the lack of access to the best production technologies, particularly ASML's lithography machines and TSMC's 2nm process. SMIC, which manufactures Huawei's chips, is currently limited to a 7nm process.

"Huawei's production volumes remain extremely limited compared to what TSMC can produce for Western chip manufacturers, as Huawei and its Chinese partners do not have access to ASML's advanced lithography tools that TSMC has acquired in large quantities. Therefore, Huawei is currently not even able to supply the domestic Chinese market and is far from being able to export on a significant scale," estimates Chris Miller, author of "Chip War."

According to a study by the Council on Foreign Relations, even in the most optimistic scenarios for Huawei's AI chip production—two million in 2026 and four million in 2027—the company would only produce about 5% of Nvidia's aggregated AI computing power in 2025, dropping to 4% in 2026 and 2% in 2027.

"It is practically impossible for Huawei to close this gap: even a hundredfold increase in AI chip production by 2027 would not allow Huawei to reach half of Nvidia's production. Meanwhile, China's demand for AI computing capacity is growing exponentially as models become more sophisticated, meaning the country's AI chip shortage will worsen over time, not ease," notes the think tank.

To mitigate access to the best production technology, particularly the finest process nodes, Huawei has opted for a monolithic design for the 950PR. All components are concentrated on a single chip, as opposed to multi-chip architectures, which require advanced packaging technologies like TSMC's CoWoS, to which Huawei does not have access. In exchange, Huawei accepts compromises on the maximum chip size and yield rates.

A Building Ecosystem for Huawei

The big question is whether Huawei will be able to progress quickly and strongly enough to build an ecosystem capable of competing with Nvidia's, not only in China but also elsewhere in the world, potentially challenging American hegemony in chips. This is currently, aside from cutting-edge production techniques, what the company lacks the most, according to Antoine Chkaiban. "Huawei's delay is also due to the fact that it is hard to compete with Nvidia's ecosystem, which allows any customer to easily deploy and use its chips. Huawei is forced to build its own ecosystem."

This would have profound implications in the chip and AI world. Models trained on Nvidia hardware may not function optimally on Ascend infrastructure, and vice versa, leading to fragmentation within the global AI research community. Open-source AI frameworks like PyTorch and JAX will need to maintain extensive support for both ecosystems, further taxing already limited maintenance teams.

Huawei's task in this regard appears daunting, as Nvidia has gained a decisive lead. "While Nvidia has lost nearly all of its market share in China for high-end AI GPUs, its global dominance remains firmly entrenched. By vertically integrating across the entire technology stack and horizontally expanding across industries, Nvidia has built a global position that will be extraordinarily difficult to dislodge," assesses Mina Kim.

Nevertheless, geopolitics could one day take precedence over technology. For non-aligned countries or those more or less hostile to the United States, Huawei chips could eventually offer a viable alternative to Nvidia hardware, free from the constraints of American export controls. Countries struggling to obtain Nvidia GPU allocations, either because they are deemed hostile or because their markets are too small, may also find Huawei an attractive supplier.