By the end of 2025, news that Kunlunxin, a company backed by Baidu, was preparing for a Hong Kong IPO became a major talking point in China’s domestic AI chip sector and the broader capital markets.

On December 5th, Reuters and other international media outlets reported that Kunlunxin, an AI chip firm originally spun off from Baidu’s Intelligent Chip and Architecture Division, is gearing up for a Hong Kong listing. The company recently completed a new round of financing, raising $283 million and achieving a post-money valuation of $2.97 billion (approximately RMB 21 billion). Key investors in this round include high-level talent funds under the Guoxin Fund, China Mobile Innovation, Shanzheng Investment, and other notable institutions.

The announcement sent Baidu’s Hong Kong-listed shares surging in afternoon trading, with gains reaching 7.77% at one point. This reflects strong market anticipation for a revaluation of Baidu’s AI computing assets.

AI Chip IPO Plan Takes Shape

Multiple sources indicate that Kunlunxin’s listing preparations have entered an early stage. While the exact fundraising scale and timeline remain under discussion, the company could submit its application to the Hong Kong Stock Exchange as early as the first quarter of 2026.

This isn’t the first time rumors have circulated about Kunlunxin going public, but this time the company appears better prepared.

Kunlunxin’s journey began within Baidu in 2011. Originally known as Baidu’s Intelligent Chip and Architecture Division, it served as the cornerstone of Baidu’s AI strategy. In April 2021, the division completed independent financing at an initial valuation of approximately RMB 13 billion, with Baidu retaining a 59.45% stake. Over four years, its valuation has grown nearly 60%, reflecting shifting market perceptions of scarce domestic AI chip assets.

The latest funding round stands out for its investor lineup. State-backed entities like the Guoxin Fund and Shanzheng Investment joined, adding a “national team” endorsement to Kunlunxin’s path to listing.

Kunlunxin CEO Ouyang Jian brings an impressive background. A chip expert with a bachelor’s degree from Beihang University and a master’s from the University of Science and Technology of China, he previously served as Baidu’s Chief Chip Architect, leading projects in ARM servers, software-defined Flash, and smart network cards.

Ouyang’s transition from technical architect to company leader mirrors Kunlunxin’s evolution from an internal R&D unit to a market-driven enterprise.

For Baidu, spinning off Kunlunxin for listing serves multiple strategic goals. The most immediate is value realization. For years, Baidu’s market cap has been weighed down by its advertising business, with insufficient recognition of its AI investments.

An independent listing would allow Kunlunxin’s core assets to be revalued under a tech stock framework, potentially supporting Baidu’s second growth phase. A deeper objective is ecosystem building. Through capital market financing, Kunlunxin can secure resources to accelerate R&D and market expansion, strengthening Baidu’s position in AI infrastructure.

Kunlunxin’s listing plans also align with a broader trend of domestic AI chip firms rushing to the public markets. Moore Threads has already listed on the STAR Market, Muxi is preparing to go public, and companies like Biren are accelerating their IPO processes.

Kunlunxin’s choice of Hong Kong over A-shares likely reflects the city’s more flexible tech listing rules and its status as an international financial hub, which could attract global capital and pave the way for overseas expansion.

Crossing the ‘30,000-Card Cluster’ Threshold

Behind Kunlunxin’s soaring valuation lies a technological journey from internal R&D to market validation. In 2024, its revenue exceeded RMB 1 billion, outpacing peers like Cambricon and Moore Threads.

At Baidu World Conference in November, Kunlunxin showcased its technological prowess. Shen Dou, Executive Vice President of Baidu Group, unveiled the next-generation Kunlunxin chip and the super node product ‘Tianchi,’ announcing annual product launches for the next five years.

The Kunlunxin M100 and M300 made their debut: the M100, optimized for large-scale inference, will launch in 2026; the M300, designed for training and inference of ultra-large multimodal models, will debut in 2027.

However, the technological foundation supporting Kunlunxin’s high valuation is its third-generation P800 series, which is now in mass production. This AI acceleration chip, featuring a self-developed XPU-P architecture, competes with NVIDIA on several key metrics.

Benchmark tests show the Kunlunxin P800 delivers 345 TFLOPS of FP16 computing power, far surpassing NVIDIA’s China-customized H20 chip (148 TFLOPS), placing it among China’s top-tier chips. Its memory specifications exceed comparable GPUs by 20–50%, giving it an edge in MoE (Mixture of Experts) architectures. A single eight-card machine can run ultra-large models like DeepSeek-V3/R1 with 671 billion parameters, and just 32 machines are needed for full-parameter training.

In software, Kunlunxin has fully adapted to FlashAttention and MLA operators for open-source models like DeepSeek, supporting 8-bit inference to deliver lossless services. An eight-card machine achieves 2,437 tokens/s throughput, demonstrating a software stack that not only integrates with mainstream frameworks but also optimizes for algorithmic trends.

The real technological barrier lies in engineering ‘30,000-card clusters.’ Baidu Intelligent Cloud has activated China’s first fully self-developed ‘30,000-card’ cluster (based on the third-gen Kunlunxin), which goes beyond physical stacking to solve stability challenges at scale.

Leveraging Baidu Baige AI Heterogeneous Computing Platform 4.0, the cluster achieves 99.5% effective training time and over 96% linear speedup, matching international first-tier performance when processing models with hundreds of billions of parameters.

The concurrently released ‘Tianchi’ super node product highlights Kunlunxin’s system-level engineering. A single Tianchi 512 super node can train trillion-parameter models, integrating 64 Kunlunxin AI acceleration cards into one rack.

In terms of performance, one rack replaces hundreds of traditional machines, significantly cutting space and power costs per unit of computing power.

Kunlunxin’s technological evolution reflects China’s AI chip transition from isolated breakthroughs to system integration. The first-gen Kunlunxin (2018) deployed over 20,000 units for Baidu Search and Xiaodu; the second-gen (2021) deployed tens of thousands; the third-gen P800 series achieved a qualitative leap. These iterations mark Kunlunxin’s progress from ‘usable’ to ‘user-friendly’ and finally to ‘leading.’

‘Billion-Level Order’ Validation

The ultimate test of technological prowess is market acceptance. Kunlunxin’s commercialization is expanding beyond Baidu’s internal operations into diverse industries.

In August 2025, Kunlunxin secured a billion-level AI computing equipment procurement deal from China Mobile, a landmark event in the industry. This order not only signifies Kunlunxin’s breakthrough beyond the ‘internet circle’ but also marks its entry into China’s core information infrastructure procurement list.

For capital markets, operator procurement carries far greater endorsement value than the order amount itself, proving Kunlunxin’s products meet telecom-grade standards in stability, reliability, and service.

China Merchants Bank’s AI chip resource project represents Kunlunxin’s financial sector breakthrough. Benchmark tests show the Kunlunxin P800 outperforms comparable domestic chips in supporting the Qwen series, with multimodal model inference leading the industry and rapidly enhancing efficiency in data analysis, customer service, and code assistance. The financial sector’s strict stability and security requirements make this case highly demonstrative.

Kunlunxin’s commercial footprint now spans key sectors: energy (State Grid, China Southern Power Grid), manufacturing (China Steel Research Group), and education (Tongji University).

Through collaboration with hundreds of customers, AI computing power is being delivered to industries like internet services, telecoms, intelligent computing, finance, energy, and automotive, benefiting millions of end-users.

Kunlunxin’s unique advantage lies in having Baidu as a natural ‘testing ground’ and ‘first customer.’ Baidu’s internal businesses—from search engines and ERNIE large models to autonomous driving—provide real-world scenarios for refinement.

This ‘internal validation + external promotion’ model reduces market resistance and accelerates product maturity.

Kunlunxin’s commercialization path mirrors NVIDIA’s early success. NVIDIA initially broke through in gaming graphics before entering data centers. Similarly, Kunlunxin uses Baidu’s internal businesses as its foundation, meeting computing demands for core services like Baidu Search, ERNIE training, and Apollo autonomous driving. It converts internal spending into stable revenues while refining technologies through practical use.

It then leverages this foundation to expand into external markets, entering sectors like telecoms, finance, and energy, achieving a smooth transition from internal to external sales.

For Baidu, continued investment in large model R&D requires Kunlunxin to solidify its computing power base. The exponential growth of model parameters and multimodal capabilities has driven computing demand to new heights. Kunlunxin’s high-performance chips and 10,000-card clusters provide the foundation for Baidu to push intelligence boundaries and consolidate its technological edge.

Balancing the Capital Game

Despite its high valuation, Kunlunxin’s IPO path faces challenges.

The first issue is valuation rationality. Over four years, its valuation has jumped from RMB 13 billion to RMB 21 billion, a 60%+ increase. While its 2024 revenue exceeded RMB 1 billion and J.P. Morgan projects 2026 revenue at RMB 8.3 billion (a sixfold increase) with potential break-even in 2025, recent industry bubble concerns have made markets wary of AI chip ‘dream rates.’

Technological ecosystem challenges are equally severe. NVIDIA’s two-decade-old CUDA ecosystem forms an insurmountable moat. While Kunlunxin has achieved compatibility with some frameworks, gaps remain in developer community size and toolchain maturity.

Compatibility issues in multi-card training scenarios could hinder customer migration. Additionally, Kunlunxin’s deep integration with Baidu’s PaddlePaddle framework is both an advantage and a limitation—it ensures optimal performance within Baidu’s ecosystem but may affect broader open-source adoption.

Commercialization faces hidden concerns. Current orders rely heavily on policy-driven procurement and Baidu ecosystem synergies, raising questions about true gross margins and payment cycles.

If China Mobile and other major clients dominate, pricing power will be limited, making it difficult to sustain high R&D investments. The AI chip industry’s characteristics—high R&D costs and long payback periods—require Kunlunxin to prove its sustainable profit model post-IPO.

Geopolitical and supply chain risks also loom. While Kunlunxin emphasizes its ‘100% self-developed architecture,’ chip manufacturing still depends on the global supply chain in the short term.

Market competition is intensifying. Huawei Ascend has deep government and operator market roots; Moore Threads, Muxi, and others are advancing IPOs, with financing set to intensify rivalry.

Some clients may adopt a ‘dual-track’ approach, purchasing both NVIDIA and domestic chips, subjecting domestic products to constant comparison with the ‘gold standard.’ Fully catching up with international giants in performance, ecosystem, and service will be a long-term challenge for Kunlunxin.

In this critical financing round for Kunlunxin’s Hong Kong IPO, Shanzheng Investment completed a strategic investment of over RMB 100 million, becoming a key capital partner.

Shanzheng Investment explicitly stated that this move aligns with national technology strategy and will continue to invest in core links of ‘computing power, storage capacity, and transport capacity’ in the AGI field, as well as in large models and application scenarios.

This investment not only supplements Kunlunxin’s R&D and commercialization funds but also, through industrial resource synergies, helps solidify its competitive edge in China’s computing power ecosystem.

This situation underscores the state's enduring confidence in the domestic computing - power industry chain. It also implies that Kunlunxin will take on a greater share of "national missions," compelling the company to strike a balance between commercial gains and national strategic objectives.

The IPO of Kunlunxin is far more than a mere capital - raising endeavor for an individual firm; it serves as a barometer for gauging the maturity of the domestic AI chip industry. The outcome of its IPO, whether a success or a failure, will have a profound impact on the financing climate and the development confidence of the entire industry.

Faced with a market opportunity worth trillions of dollars and a multitude of risks, Kunlunxin is tasked with finding an equilibrium among technological innovation, ecosystem construction, and commercialization. This is essential for converting valuation expectations into sustainable corporate value.

For investors, this event is a double - edged sword. On one hand, it signals the ascent of domestic computing power. On the other hand, it serves as a cautionary note about the potential bubble risks.

From Baidu's perspective, spinning off Kunlunxin for listing represents a value - unlocking move for its core AI assets. This move has the potential to liberate Baidu from the constraints imposed by its advertising business on market capitalization and pave the way for a second growth trajectory.

For the domestic computing - power industry, Kunlunxin's listing process will act as a window through which industry maturity can be observed. Its "chip + cluster + ecosystem" full - stack model may potentially reshape the competitive landscape of domestic AI computing power.