The Economist: Nvidia's market share in China drops to 8%! Huawei's AI chips capture half of the market, export controls rewrite the landscape

The Economist quoted the latest estimates from research firm Bernstein on Tuesday (14th), revealing for the first time a complete picture of the supply landscape in China's artificial intelligence (AI) chip market. Data shows that the Chinese market has completely reversed in just three years, with NVIDIA (NVDA-US), which once nearly monopolized the market, now seeing its compliant shipment share drop to single-digit percentages; meanwhile, local Chinese manufacturer Huawei has captured half of the market.

According to Bernstein's estimates, Huawei currently holds about 50% of the Chinese AI accelerator market, with annual sales reaching $12.1 billion, maintaining its leading position.

In contrast, NVIDIA's compliant shipment share, which was once close to a monopoly three years ago, has plummeted to just 8%, with annual sales of about $2 billion.

Additionally, Advanced Micro Devices (AMD-US) follows closely with about 12% market share and annual sales of approximately $3 billion; local Chinese companies Cambricon (688256-CN) and Hygon Information Technology (688041-CN) reported sales of $2.1 billion and $2 billion, respectively.

Moreover, tech giants such as Alibaba (09988-HK, $1.2 billion) and Baidu Group (09888-HK, $700 million), which develop their own chips, have also entered the ranks of major suppliers NVIDIA CEO Jensen Huang publicly admitted earlier this year that the company's market share in the Chinese AI chip market has dropped from about 95% at its peak to "basically zero."

Bernstein estimates an 8% remaining market share, which includes low-end products that meet export regulations and a very small amount of shipments released through special permits from the U.S. government.

Both Bloomberg and Yahoo Finance previously reported that a small quantity of advanced AI chips from NVIDIA and AMD have recently entered the Chinese market through special permits issued by the U.S. government, which briefly boosted the stock prices of both companies.

However, the reports also pointed out that such permits are strictly limited to "small quantities," making it difficult to reverse NVIDIA's overall marginalization in the Chinese market.

The reports mentioned that to comply with the review process, NVIDIA has significantly reduced the approval list for Asian customers, with more than half of its Chinese customers being removed, and each transaction must undergo additional compliance review.

According to data from Bernstein and market research firm IDC, by 2025, the combined market share of domestic AI chip suppliers in China's AI accelerator server market is estimated to be between 41% and 46%; by 2026, this proportion is expected to rise to about 56%.

Among them, Huawei's market share is projected to expand from about 39% in 2025 to 50%, with sales nearly equivalent to the total of the other five domestic and foreign competitors.

Analysts believe that the key driver behind this surge in market share is not merely commercial competition, but rather the "forced substitution demand" created by U.S. export controls.

Since the U.S. began ramping up export restrictions on advanced AI chips to China in October 2022, NVIDIA's H100, H200, and even the H20, which was specially downgraded for the Chinese market, have been unable to maintain stable supply, forcing large Chinese internet companies and AI laboratories to fully shift to domestic chips.

It is noteworthy that while the original intent of the controls was to cut off China's access to advanced computing capabilities, the actual effect has unexpectedly leveled the existing market barriers, creating an almost vacuum market space for domestic chips like Huawei's Ascend series.

Despite official data showing that NVIDIA is nearly out of the market, another narrative is circulating. It is reported that some leading Chinese large language model teams, when building large-scale training clusters, adopt a hybrid architecture of "Huawei Ascend clusters paired with NVIDIA chips obtained through informal channels" to seek a balance between compliance risks and training efficiency.

Reports from Yahoo Finance and The Information also indicate that even under strict U.S. controls, NVIDIA's next-generation B300 servers continue to flow into China through informal channels, with each priced at about $1 million, far exceeding normal market prices, yet still in short supply.

Analysts point out that these underground transactions, from which NVIDIA cannot profit, precisely demonstrate that there remains an inelastic demand for NVIDIA chips among Chinese AI operators in the most cutting-edge training computing needs.

Ascend 950PR : Leading the generation in inference, still has gaps in training

It is worth noting that if the focus shifts from training to inference and large-scale deployment, Huawei's Ascend 950PR has already demonstrated compelling alternative capabilities in its specifications.

According to the parameters released by Huawei, the FP4 peak computing power of the Ascend 950PR reaches 1.56 PFLOPS, approximately 2.8 times that of the special version H20; the FP8 computing power reaches 1 PFLOPS, paired with 112GB of high bandwidth memory, and the inter-chip interconnect bandwidth reaches 2TB/s.

In contrast, the H20, which is limited by export control specifications, has an FP4 computing power of only about 0.56 PFLOPS and an interconnect bandwidth of only 900GB/s.

In application scenarios such as inference and multimodal generation, the Ascend 950PR has already achieved generational leadership over the compliant version H20, forming the technical foundation for Huawei to meet the demand for large-scale domestic substitution.

However, when comparing with NVIDIA's global flagship products, the gap remains evident. The FP8 computing power of the NVIDIA H100 reaches 3.96 PFLOPS, while the B200, based on the Blackwell architecture, achieves an FP4 computing power of up to 18 PFLOPS and a memory bandwidth of 8TB/s.

In terms of mixed-precision floating-point computing capabilities and memory bandwidth required for high-density training and ultra-large-scale model pre-training, Huawei still lags about one generation behind.

In the software ecosystem, Huawei's CANN Next software stack claims to have achieved about 80% CUDA compatibility, but for frontline algorithm engineers, the practical costs of model migration and tuning remain real issues in their daily work.

The Dialectical Effect of Regulation: Short-term Suppression, Long-term Assistance

Analysis indicates that the actual effect of the U.S. export control policy is presenting a dialectical structure.

In the short term, restrictions on advanced GPU exports do impose hard constraints on China's training of ultra-large-scale cutting-edge models. Chinese AI teams, including DeepSeek and Zhipu (02513-HK), have listed "computing power bottlenecks" as a primary challenge, and the regulations have indeed produced a lagging effect over the past two to three years.

However, when looking at a ten-year horizon, high-intensity blockades may instead accelerate the formation of a positive cycle of "demand—scale—renewal" for domestic substitution.

A similar situation has played out in the semiconductor equipment field. After being banned from extreme ultraviolet lithography (EUV) machines, SMIC (00981-HK) still advanced the mass production of 7-nanometer process chips using deep ultraviolet multiple exposure technology, demonstrating the industry's ability to adapt and reconstruct under extreme conditions often exceeding expectations based on theoretical projections.

Bernstein's model further extrapolates a curve with long-term indicative significance: the domestic sales of Chinese AI chips are climbing at a compound annual growth rate of 74% Based on this calculation, by 2026, China's domestic chip supply could cover 39% of national demand; by 2028, this ratio will exceed 100%, indicating that domestic production capacity will surpass demand for the first time, resulting in a net surplus.

In other words, the window unexpectedly opened by export controls is gradually evolving from a passive alternative into a thorough reconstruction of capacity and supply chains.

However, analysis also points out that Chinese companies such as Huawei, Cambricon, and Hygon will face not only capacity issues but also challenges in ecosystem construction and profitability.

The premium of NVIDIA chips in the informal market and the ongoing unconventional demand for its products constantly remind us of a fact: this reshuffling of market share is primarily the result of export control policies rather than victories won solely through product competitiveness.

As China's domestic chips are expected to no longer face physical shortages by 2028, whether companies like Huawei can continue to increase their market share will depend on more fundamental product competitiveness indicators, such as floating-point computing density, software toolchain maturity, and the stickiness of the developer ecosystem