Kunlunxin M100 Chip to Launch in 2026: In-depth Analysis of Technological Iteration and Market Landscape

In November 2025, Baidu officially released its new generation of AI chip product lines at its annual World Conference, with

becoming the focus of market attention[1]. This chip is specifically optimized for large-scale inference scenarios, especially achieving significant performance improvements in mixed expert (MoE) model inference, and is scheduled to be officially launched in [2]. At the same time, Baidu also released the M300 chip for ultra-large-scale multimodal training and inference, which is expected to be commercially available in early 2027[3].

Notably, Kunlunxin submitted a confidential listing application form (Form A1) to the Hong Kong Stock Exchange on January 1, 2026, planning to list independently on the main board of the Hong Kong Stock Exchange[4]. This move marks a new stage in Baidu’s layout in the AI chip field, accelerating technological research and development and market expansion through capital market financing.

The design concept of Kunlunxin M100 closely revolves around the core demand of

. As large model applications shift from the training phase to the inference phase, the market demand for inference chips has exploded. The M100 chip has been deeply optimized for inference scenarios under the MoE architecture, and this choice has important strategic significance[5]. In the MoE architecture, inter-card communication volume rises sharply, and traditional 8-card node clusters have encountered communication bottlenecks, requiring dozens to hundreds of cards to be integrated into a “super node” to achieve efficient collaboration.

Based on this technical judgment, Baidu simultaneously released two super node products:

(to be launched in the first half of 2026) and (to be launched in the second half of 2026)[6]. Among them, the 256 Super Node supports up to 256-card ultra-fast interconnection. The total inter-card interconnection bandwidth has increased by 4 times compared with the previous generation product, the performance has improved by more than 50%, and the single-card token throughput for mainstream large model inference tasks has increased by 3.5 times[7].

---

China’s AI chip market is currently undergoing profound structural changes. According to Bernstein’s research report,

[8]. This sharp change reflects the profound impact of geopolitical factors on the semiconductor industry and also marks the acceleration of the domestic substitution process.

In contrast,

, becoming the leader in China’s AI chip market[9]. Cambricon’s market share is expected to grow from 9% to 12%, and the combined market share of other domestic chip manufacturers will reach 35%[10].

Facing increasingly tight export controls, Nvidia is adjusting its China strategy. In December 2025, the U.S. government announced that Nvidia would be allowed to export H200 chips to China, but a 25% income tax must be paid[11]. The H200 is based on the Hopper architecture, equipped with 141GB HBM3e high-speed memory, with a bandwidth of up to 4.8TB/s, which is a direct upgrade of the H100[12]. However, this “castrated” high-end chip has main indicators only half or even lower than Nvidia’s latest Blackwell series (B200/B300); but compared with the previous Chinese customized chip H20, the main indicators are 5 to 10 times ahead[13].

Market analysis generally believes that although the entry of H200 has increased variables, it has not shaken the basic market of domestic computing power in government, operators, finance and other fields[14]. Industry chain insiders pointed out that this “boiling frog in warm water” strategy aims to delay the process of domestic substitution in China, but considering that supply chain security has become a harder decision indicator than performance, the long-term competitiveness of domestic chips remains stable[15].

In the domestic AI chip track, Huawei’s Ascend series constitutes Kunlunxin’s most direct competitor. According to public information, the overall localization rate of Huawei’s Ascend 910C chip has exceeded 90%, and the peak computing power under FP16 precision reaches 800 TFLOPS, which is about 80% of Nvidia’s H100 level[16]. More importantly, Huawei builds cluster capabilities through large-scale interconnection at the system level. Its CloudMatrix 384 Super Node released in May 2025 combines 384 Ascend 910C chips into a whole, with a total computing power of 300 PFLOPS. Some institutions believe it is close to or even exceeds the level of Nvidia’s GB200 NVL72 system[17].

At the 2025 All-Connect Conference, Huawei also announced the three-year (2026-2028) roadmap for Ascend AI chips, stating that it will successively launch Ascend 950, 960 and 970 series, with computing power doubling continuously[18]. This clear technical evolution path has brought considerable competitive pressure to Kunlunxin.

---

Kunlunxin P800 has been fully verified inside Baidu—not only undertaking most of the inference tasks, but also successfully training multimodal models based on a single cluster of 5,000 cards. Currently, this cluster has expanded to more than 10,000 cards and is training larger-scale models[19]. This large-scale internal deployment provides strong endorsement for the reliability of M100.

M100 focuses on optimizing large-scale inference scenarios and has achieved significant performance improvements in MoE model inference. This differentiated positioning allows it to avoid direct competition with Nvidia in the training field and instead seek breakthroughs in the fast-growing segment of inference market[20].

By integrating multiple Kunlunxin AI acceleration cards into a unified super node architecture, under the optimization of DeepSeek V3/R1 PD separation inference architecture, single-card performance can be improved by 95%, and single-instance inference performance can be greatly improved by up to 8 times[21].

The first-mover advantage of CUDA ecosystem makes Nvidia have an unparalleled position in the developer community. Although Kunlunxin has established a corresponding software development environment, there is still a gap with CUDA in terms of maturity and compatibility. Many interviewed industry chain insiders pointed out that the improvement of software ecosystem will be the key factor determining the success or failure of domestic AI chips[22].

Currently, Kunlunxin has not announced the specific technical parameters of M100, including key data such as process technology, computing power indicators, and memory configuration. This makes it difficult for the market to conduct accurate performance comparisons and also raises doubts about its actual competitiveness[23].

As a Fabless design company, Kunlunxin’s chip production depends on external wafer foundries. In the current complex international environment, there are uncertainties in obtaining advanced processes. Shawn Yang, an analyst at Arete Research, pointed out that Kunlunxin’s specific advantages over domestic competitors are: on one side, Cambricon faces production capacity bottlenecks, and on the other side, Huawei faces external restrictions, which provides favorable space for Kunlunxin’s market expansion[24].

---

In the

, Kunlunxin M100 is difficult to pose a substantial challenge to Nvidia. Nvidia still maintains a significant lead in AI training and inference performance with its Blackwell architecture’s B200/B300 series chips. According to the report of the Progressive Institute, Blackwell chips are 1.5 times faster than H200 in AI training tasks and 5 times faster in inference tasks[25].

However, it should be noted that Kunlunxin’s target market positioning is significantly different from Nvidia’s. M100 is mainly oriented to

rather than training scenarios, and focuses on . For most enterprises, using Huawei Ascend or Cambricon chips can already meet more than 90% of AI needs, and there is no need to pay 200% cost for 10% performance gap[26]. From this perspective, Kunlunxin is more likely to form a competitive pattern with Huawei and Cambricon in the Chinese market, rather than directly challenging Nvidia.

From a mid-to-long term perspective, Kunlunxin’s competitive potential depends on the following key factors:

Kunlunxin has announced its development roadmap for the next five years: Tianchi 1000-level super node will be launched in 2028, Kunlunxin N series will be launched in 2029, and Baidu Baige 1 million-card Kunlunxin single cluster will be officially activated in 2030[27]. If this roadmap can be realized as scheduled, Kunlunxin is expected to gradually narrow the gap with leading players in system-level performance.

Kunlunxin’s Hong Kong listing will bring continuous financing channels for the company to support R&D investment and market expansion. According to Goldman Sachs, Baidu’s 59% stake in Kunlunxin is worth about $16.5 billion, accounting for about 30% of its target valuation for Baidu[28]. Macquarie Securities estimates that Kunlunxin’s revenue will double to about $1.4 billion next year, which will put the company in the same echelon as Cambricon[29].

Under the macro narrative of semiconductor localization, policy support provides unique development opportunities for Chinese local chip enterprises. Data shows that affected by policy support, Chinese semiconductor stocks have mostly outperformed Internet stocks in the AI boom[30]. According to Frost & Sullivan’s forecast, from 2025 to 2029, the compound annual growth rate of China’s AI chip market will reach 53.7%, and the market size will surge from 142.537 billion yuan in 2024 to 1.34 trillion yuan in 2029[31].

---

Kunlunxin’s development provides investors with diversified industry chain opportunities:

as Kunlunxin’s controlling shareholder will directly benefit from the valuation reconstruction of the chip business. Goldman Sachs pointed out that even without considering spin-off listing, the sales growth of Kunlunxin and the use of the chip by Baidu’s own business will directly increase the company’s performance[32].

such as Cambricon (688256.SH) will benefit from the overall market expansion and domestic substitution trend. Cambricon’s stock price rose by 110% in 2025 and was called “China’s Nvidia” by investors[33].

will also indirectly benefit from the localization process. In the first half of 2025, the localization rate of domestic semiconductor equipment exceeded 20%, and the replacement rate of advanced packaging rose to nearly 40%[34].

The technical iteration speed of AI chips is fast. If Kunlunxin’s technical route deviates from the mainstream direction of the industry, it may face the risk of losing market share.

Under the influence of geopolitical factors, there are uncertainties in the foundry of advanced process chips, which may restrict Kunlunxin’s capacity expansion.

As more players enter the AI chip track, market competition will become increasingly fierce, which may compress the overall profit margin of the industry.

---

Based on the above analysis,

. Considering that Nvidia’s market share in China is expected to drop from 40% in 2025 to 8% in 2026, the space for domestic substitution is still vast[35].

Kunlunxin’s competitive advantages are mainly reflected in:

. However, factors such as software ecosystem gaps, unpublished technical parameters, and supply chain uncertainties still need continuous attention.

From an industry perspective, the launch of Kunlunxin M100 marks an important progress in China’s AI chip design capability. Driven by policy support, capital assistance and market demand, China’s AI chip industry is accelerating its evolution from “usable” to “good”. Although it is difficult to shake Nvidia’s global leading position in the short term, in the long term, technological iteration and ecosystem improvement are expected to gradually narrow the gap with international leading levels.

---

[1] Baidu Responds to Kunlunxin’s Listing Rumors: Traditional Business Under Pressure, Needs New Growth Story

[2] Kunlunxin Spin-off Listing? Baidu’s Latest Response

[3] Beijing AI Chip Leader Secretly Submits IPO Application

[4] Baidu Kunlunxin Confidentially Submits Hong Kong Listing Application, AI Chip Business Spin-off Accelerates

[5] AI Chip 2025: Giants Compete Fiercely, Power Structure Reforms

[6] Baidu Kunlunxin Confidentially Submits Hong Kong Listing Application, AI Chip Business Spin-off Accelerates

[7] Kunlunxin Spin-off Listing? Baidu’s Latest Response

[8] Nvidia’s H200 Release Fails to Stop Share Decline, China’s AI Chip Pattern Being Rewritten by Huawei

[9] AI Chip 2025: Giants Compete Fiercely, Power Structure Reforms

[10] Domestic AI Chips Bid Farewell to the Wild Age

[11] Mindset Observatory: Will H200 Make China “Addicted”?

[12] The Scheme Behind H200’s Release: If It Can’t Be Blocked, Sell It to Slow Down China’s Chip Replacement Rhythm?

[13] Domestic AI Chips Bid Farewell to the Wild Age

[14] AI Chip 2025: Giants Compete Fiercely, Power Structure Reforms

[15] Domestic AI Chips Bid Farewell to the Wild Age

[16] The Scheme Behind H200’s Release: If It Can’t Be Blocked, Sell It to Slow Down China’s Chip Replacement Rhythm?

[17] The Scheme Behind H200’s Release: If It Can’t Be Blocked, Sell It to Slow Down China’s Chip Replacement Rhythm?

[18] AI Chip 2025: Giants Compete Fiercely, Power Structure Reforms

[19] Kunlunxin Spin-off Listing? Baidu’s Latest Response

[20] Beijing AI Chip Leader Secretly Submits IPO Application

[21] Baidu Responds to Kunlunxin’s Listing Rumors: Traditional Business Under Pressure, Needs New Growth Story

[22] Domestic AI Chips Bid Farewell to the Wild Age

[23] Kunlunxin Spin-off Listing? Baidu’s Latest Response

[24] Kunlunxin Gains Momentum! Wall Street Optimistic About Baidu: Expected to Copy Google’s AI Counterattack Path

[25] The Scheme Behind H200’s Release: If It Can’t Be Blocked, Sell It to Slow Down China’s Chip Replacement Rhythm?

[26] Nvidia’s H200 Release Fails to Stop Share Decline, China’s AI Chip Pattern Being Rewritten by Huawei

[27] Baidu Responds to Kunlunxin’s Listing Rumors: Traditional Business Under Pressure, Needs New Growth Story

[28] Kunlunxin Gains Momentum! Wall Street Optimistic About Baidu: Expected to Copy Google’s AI Counterattack Path

[29] Kunlunxin Gains Momentum! Wall Street Optimistic About Baidu: Expected to Copy Google’s AI Counterattack Path

[30] Kunlunxin Gains Momentum! Wall Street Optimistic About Baidu: Expected to Copy Google’s AI Counterattack Path

[31] AI Chip 2025: Giants Compete Fiercely, Power Structure Reforms

[32] Kunlunxin Gains Momentum! Wall Street Optimistic About Baidu: Expected to Copy Google’s AI Counterattack Path

[33] Kunlunxin Gains Momentum! Wall Street Optimistic About Baidu: Expected to Copy Google’s AI Counterattack Path

[34] AI Chip 2025: Giants Compete Fiercely, Power Structure Reforms

[35] Nvidia’s H200 Release Fails to Stop Share Decline, China’s AI Chip Pattern Being Rewritten by Huawei