Domestic AI chip share approaches Nvidia; major firms shift from competing on models to selling chips
In the first half of 2026, China's data center AI accelerator shipments reached 2.008 million units, with non-GPU accelerators shipping 1.016 million units and surpassing GPUs for the first time; Nvidia's share fell to 34%, while Huawei Ascend reached 28%, a gap of only 6 percentage points. The article argues that the surge in inference demand has made custom chips a means of cost reduction, and that major firms are monetizing through two paths: packaged sales of super nodes and spinning off chip companies for public listing, while a production capacity gap of about 40% has shifted the focus of competition from performance to capacity.
Why read this
AI score 75
The article uses IDC shipment data to break down changes in domestic AI chip market share and compares two monetization paths for major firms selling chips.
Topics
Body · AI translation
[[In China's domestic AI chip market, domestic vendors are rapidly eroding Nvidia's share. Shipments from major-company chips such as Huawei Ascend, Alibaba T-Head, and Baidu Kunlunxin have surged, and selling chips has become a realistic path for big companies to hedge against model involution and achieve commercialization. **Key points** **1. Domestic chip share approaches Nvidia; non-GPU accelerators surpass GPUs for the first time** In the first half of 2026, China's data center AI accelerator shipments reached 2.008 million units, with non-GPU shipments at 1.016 million units, surpassing GPUs for the first time; Nvidia's share fell to 34%, while Huawei reached 28%, a gap of only 6 percentage points. Nvidia's shipments fell 44% year on year. **2. Inference demand explodes; custom chips become a cost-cutting tool** Model call volumes and agent applications are driving a surge in inference demand. Big companies' in-house chips can be optimized around specific models, with energy efficiency reaching 1.5 to 3 times that of general-purpose GPUs and unit computing cost reduced by more than 35%. Deloitte expects inference to account for about two-thirds of AI computing volume in 2026. **3. Full-stack AI capabilities pave the way for chip sales** Major companies such as Alibaba, Baidu, and Tencent first use their own models, clouds, and real traffic from high-frequency applications to hone their chips, then push them to external customers. After comprehensive internal validation, Baidu's Kunlunxin P800 won 70% to 100% of multiple bid packages in China Mobile's centralized procurement. **4. Big companies monetize chips in two ways: hardware bundling and spin-off listings** Huawei sells chips, networking, and storage as a package through the Ascend SuperNode model; Alibaba released Zhenwu V900, planned for mass production in 2027; Baidu is pushing for a spin-off listing of Kunlunxin, seeking a valuation of about $50 billion, and Alibaba has not ruled out an independent listing for T-Head. **5. Production capacity replaces performance as the focus of competition; order schedules extended** Domestic computing power chips are seeing a surge in orders, with demand exceeding supply by more than 10 times and delivery schedules extended to more than a year out. SMIC's N+2 supply is about 2.6 million units, a gap of about 40%; Huawei has locked up about 43% of capacity with long-term contracts, and the HBM memory shortage further intensifies supply tightness. 2026-10-14 15:55 Big companies have woken up: competing on models is not as good as selling chips Letter List Quick View This article is from the WeChat public account: Letter List, author: Sibo In China's AI chip market, two changes are accelerating. Nvidia's share is being eroded by domestic vendors, and shipments of non-GPU accelerators are also rising. Data recently disclosed by IDC precisely confirms this change—in the first half of 2026, total data center AI accelerator shipments in China reached 2.008 million units, of which non-GPU accelerators shipped 1.016 million units, surpassing GPUs in scale for the first time. Nvidia is still No. 1, but its share gap with Huawei Ascend is only 6 percentage points. Behind both changes are major companies. Alibaba CEO Wu Yongming disclosed in March this year that T-Head chips' annualized revenue had reached the tens of billions of yuan level; at the Yunqi Conference in September, Alibaba also released Zhenwu V900, planned for mass production and commercialization in the first quarter of 2027, with annual AI chip shipments expected to grow substantially. Baidu, meanwhile, is advancing the spin-off listing of Kunlunxin; the Tianchi SuperNode has achieved a full domestic stack, and Kunlunxin has completed the R&D and commercialization of three generations of AI chips. Model companies are still looking for a way to profitability, while the bills for training and inference have become the profits of chip giants. According to an IDC report, in the first half of this year, Nvidia's shipments in China's data center AI accelerator chip market fell 44% year on year, with a 34% share; Huawei accounted for 28%, T-Head and Kunlunxin 7% and 4% respectively, and together they have surpassed Nvidia; Huawei, T-Head, Kunlunxin, and Cambricon account for about 95% of shipments in the non-GPU segment. The direct reason for the increase in big companies' chip shipments is that model call volumes have risen, supply of Nvidia's high-end products is constrained, and domestic chips have received both new demand and replacement orders. Huawei is already urging the industrial chain to expand production. Xu Zhijun said that currently "however much can be produced, that much will be supplied"; the 950PR for inference has been launched, and the 950DT SuperNode for training is expected to be supplied at scale by the end of this year or early next year. Alibaba also revealed that Zhenwu M890 has more than 650 external customers, covering more than 20 industries. Domestic large models are open-sourcing while cutting prices, lowering the deployment threshold for enterprises, and agents have brought demand for repeated model calls. At a forum in September, Baidu cited the example of Taikang: more than 3,500 agents online, with average daily Token consumption reaching 20 billion. Once an enterprise begins to use agents at scale, a continuous computing power expense follows. High model call volumes mean strong inference demand. Training requires continuous investment as models iterate, while inference costs money every time it is used. Deloitte expects that in 2026 inference will account for about two-thirds of AI computing volume. This leaves room for self-developed custom chips. When the same type of task runs repeatedly, chips can be optimized around specific models and scenarios to lower the cost of each call. Some institutions estimate that in suitable inference scenarios, the energy efficiency of customized chips can reach 1.5 to 3 times that of general-purpose GPUs, with unit computing cost reduced by more than 35%. Seeing inference chips become a good business, more and more big companies want to enter. Tencent launched Zixiao 2.0 in July, mainly for its own business, and plans to advance to a 100,000-card deployment by the end of the year; according to reports, ByteDance's chip R&D team has exceeded 1,000 people; even the model company DeepSeek has been reported to be developing its own inference chips. However, to sell chips to outsiders, you first have to make them work at home.
Major domestic companies have now also made it work. Shen Dou, executive vice president of Baidu Group and president of Baidu AI Cloud Business Group, said that the Kunlunxin P800 "has been fully validated within Baidu" and was only promoted to external industries after handling the vast majority of Baidu's inference tasks. In China Mobile's billion-level inference procurement, Kunlunxin won 70%, 70%, and 100% shares in three bid packages respectively.
Before the generative AI boom, Amazon had already laid out self-developed chips for AWS, originally to reduce costs for cloud services. In the AI era, cheaper training and inference computing power became its selling point for attracting external customers, and Anthropic and OpenAI successively committed to using Trainium. In the second quarter of this year, the chip business, including Trainium and the general-purpose computing chip Graviton, had an annualized revenue run rate exceeding $25 billion.
Now, Amazon is no longer satisfied with collecting computing service fees only through AWS. In this year's shareholder letter, Amazon CEO Jassy specifically measured the scale of the chip business on the basis of an independent chip company and raised the possibility of selling full-rack chip equipment to third parties.
Compared with Amazon, major domestic companies have gone a step further, putting the business directly on the table and forming two monetization directions. One is hardware monetization, turning chips into various solutions such as super nodes; the other is pushing chip companies separately into the capital market, turning chips into assets.
The first path is to persist with hardware monetization. The representatives are Huawei and Ascend super nodes.
Super nodes are becoming the collective sales method for domestic chips, packaging hundreds of chips into a computing system and selling it to customers, bypassing single-card comparisons with Nvidia and showing customers how large a model the entire system can run and how efficient it is. In this way, supporting products such as networking and storage are also packaged into the chip business.
Huawei is the representative of this path. At the September Huawei Connect conference, Huawei released the Ascend 960 super node, and Huawei also disclosed that more than 1,000 Ascend 910C super nodes had been deployed and that Ascend 950 super nodes had also entered large-scale commercial use. In addition, Huawei is promoting the open sourcing of CANN and improving the Ascend software ecosystem. In September this year, Huawei disclosed that more than 40 models had been natively trained based on Ascend and CANN.
Alibaba is also reaching beyond the cloud. The Panjiu super node server equipped with Zhenwu V900 is planned to launch in the first quarter of 2027. Baidu also revealed that the Kunlunxin P800 30,000-card cluster has been lit up.
Hardware monetization is not just about selling equipment. Major domestic companies are pushing self-developed chips into markets beyond cloud services while doing both cloud service and local delivery businesses.
The second path is to split the chip business into assets that can be separately valued and financed.
Baidu's approach is the most aggressive: since the model business has not yet turned around, I will first let the chip business be priced separately.
Kunlunxin filed with the Hong Kong Stock Exchange in confidential form in January and completed the filing for listing guidance on the STAR Market in May. Baidu explained in an announcement that the spin-off would help investors evaluate Kunlunxin separately and establish independent financing channels. According to The Information, Kunlunxin is seeking a listing valuation of about $50 billion, higher than Baidu's market value of about $36 billion at the time of the report.
The rich returns of the chip business are already reflected in Baidu's financial statements. In the second quarter, Baidu's AI cloud infrastructure revenue grew 50%, of which GPU cloud revenue grew 283%.
Alibaba has similar plans. Since models and applications have not yet made it work, it should first expand the business that is certain to generate revenue.
In its August financial report, Alibaba merged the Cloud Intelligence Group and T-Head into "AI Cloud and Computing Services." In the quarter ended June 30, this segment's revenue was 48.437 billion yuan, up 45% year on year. First merge chips and cloud into one account book, then consider pushing it to the capital market.
Caixin previously cited informed sources as saying that Alibaba had decided to support T-Head's independent listing in the future; Wu Yongming said this possibility could not be ruled out. However, no clear timetable has yet been disclosed.
Big companies simultaneously hold procurement budgets, application traffic, and investment capabilities, and can act as buyers, shareholders, sellers, and competitors in different cooperative relationships.
How complex this relationship can be can be seen from Tencent and Enflame.
Enflame recently listed on the STAR Market. Tencent invested in it for 11 consecutive rounds and holds 20.26%, making it the largest shareholder; Enflame's prospectus shows that in 2025, 83.79% of its revenue came from Tencent, which is both its largest customer and largest shareholder.
At the same time, Tencent is also advancing its own Zixiao chips.
For Enflame, what Tencent provides is not just money, but also real tasks, testing environments, and procurement scale. For Tencent, supporting a chip company can allow products to improve around its own needs and also provide an additional source of supply. The more Tencent uses its own chips, the more orders left for external suppliers will need to be re-won.
The relationship between Alibaba and Huawei is similar. While building T-Head into a business worth tens of billions, Alibaba is still looking for suitable chips in the procurement market.
According to previous reports, after the release of DeepSeek V4, Alibaba, Tencent, and ByteDance were all in contact with Huawei for new orders for Ascend chips. Alibaba and Tencent together locked in more than 450,000 Ascend 950PR chips, with orders exceeding 47.5 billion yuan.
Externally, Alibaba and Baidu pushing self-developed chips to the market will compete with their own suppliers for the next order; internally, their two departments are also keeping separate accounts: the procurement department considers which card is more suitable, while the chip team considers to whom its products can be sold.
However, next, capacity issues may make these relationships harder to handle.
According to a report by CCTV Finance, domestic computing power chips have already "exploded with orders": domestic computing power demand has reached more than 10 times supply, and delivery schedules for some products have been postponed to more than a year later.
According to institutional estimates, demand for domestic AI chips in 2026 will be about 4.2 million units, while SMIC N+2 supply will be about 2.6 million units, leaving a gap of about 40%; Huawei has locked in about 43% of capacity with five-year long-term contracts, and Cambricon has been allocated 9% to 11%.
Beyond wafers, HBM memory is also in shortage. The price increases for domestic chips mentioned earlier include the factor of rising procurement costs for high-bandwidth memory.
Competition in this industry is shifting from "whose chip performance is stronger" to "who can obtain more capacity."
Big companies have orders, capital, and businesses that continuously consume chips, giving them more bargaining chips when competing for capacity. But for every extra share they take, other manufacturers may get one share less.
Will future capacity allocation shift from the highest bidder winning to some kind of coordinated arrangement?
Has published 796 articles on Huxiu. This content comes from the internet. Original link. The views represent only the author and do not represent Huxiu's position. If there is a copyright issue, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]
AI Trends