Baidu’s AI Ace Kunlunxin Prepares For Hong Kong IPO, Domestic Computing Power Faces Crucial Test

At the close of 2025, the domestic AI chip sector has been thrust into the spotlight as Baidu’s Kunlunxin advances preparations for a Hong Kong initial public offering.

On December 5, international outlets reported that Kunlunxin, the spin‑out from Baidu’s Intelligent Chip and Architecture Division, is preparing to list in Hong Kong. The company completed a new financing round of USD 283 million, valuing it at approximately USD 2.97 billion (about RMB 21 billion) post‑money. Investors in this round include the High‑Level Talent Fund under Guoxin Fund, China Mobile Hechuang, and Shanzheng Investment. Following the announcement, Baidu’s Hong Kong shares rallied intraday, at one point rising 7.77 percent, reflecting market expectations of a revaluation of its AI compute assets.

Insiders indicate that listing preparations have entered an initial phase; while timing and deal size remain under discussion, an application to the Hong Kong Stock Exchange could be submitted as early as the first quarter of 2026. Although Kunlunxin has been linked to IPO plans before, the current round of activity appears more fully organized.

Kunlunxin traces its origins to Baidu in 2011 as the Intelligent Chip and Architecture Division that underpins the company’s AI strategy. In April 2021 the unit completed an independent financing round at an initial valuation of RMB 13 billion, with Baidu retaining a 59.45 percent stake. Over four years the valuation has risen by nearly 60 percent, reflecting a market reassessment of scarce domestic AI compute assets. The participation of state‑backed investors in the latest round adds a notable “national team” endorsement to Kunlunxin’s path to listing.

CEO Ouyang Jian’s technical background is illustrative: a Beihang University alumnus with graduate study at the University of Science and Technology of China, he served as Baidu’s chief chip architect and contributed to projects such as ARM server development, software‑defined flash, and smart NICs. His transition from architect to company leader mirrors Kunlunxin’s shift from an internal R&D unit to a market‑oriented enterprise.

For Baidu, spinning off Kunlunxin serves multiple strategic objectives. The most immediate is value realization: Baidu’s market capitalization has long been constrained by its advertising business, and an independent listing would allow this core AI asset to be revalued under a technology stock framework, potentially supporting a second growth trajectory. Beyond valuation, IPO proceeds would accelerate R&D and market expansion, reinforcing Baidu’s position in AI infrastructure.

Kunlunxin’s listing effort coincides with a broader wave of domestic AI chip companies pursuing capital markets. Moore Threads has listed on the STAR Market, Muxi is preparing to list, and Biren and others are advancing IPO plans. The choice of Hong Kong over mainland A‑shares likely reflects Hong Kong’s more flexible listing regime for technology firms and its role as an international financial center, which can attract global capital and facilitate overseas expansion.

The valuation uplift rests on Kunlunxin’s technical progression from internal research to external validation, with 2024 revenue exceeding RMB 1 billion and outpacing peers such as Cambricon and Moore Threads. At Baidu World in November, Kunlunxin demonstrated new products and system capabilities. Baidu’s executive leadership introduced the next‑generation Kunlunxin chips and the “Tianchi” supernode, and announced a cadence of annual product launches over the next five years. The M100, optimized for large‑scale inference, is slated for 2026, while the M300, designed for ultra‑large multimodal model training and inference, is expected in 2027.

The technical foundation underpinning Kunlunxin’s valuation is the mass‑produced third‑generation P800 series, built on the in‑house XPU‑P architecture. Benchmark data show the P800 delivers FP16 performance of 345 TFLOPS, substantially above the NVIDIA H20 variant tailored for the China market (148 TFLOPS), placing it among the top domestic performers. Memory capacity exceeds comparable GPUs by 20–50 percent, a material advantage for MoE architectures; a single eight‑card node can run 671‑billion‑parameter models such as DeepSeek‑V3/R1, and a 32‑node cluster suffices for full‑parameter training.

On the software side, Kunlunxin’s stack supports FlashAttention and MLA operators for open‑source large models and enables 8‑bit inference with negligible accuracy loss. An eight‑card configuration achieves throughput of 2,437 tokens per second, demonstrating both framework compatibility and low‑level algorithmic optimization.

The decisive engineering achievement is the deployment of a “30,000‑card” cluster. Baidu Intelligent Cloud has commissioned China’s first fully self‑developed 30,000‑card cluster based on Kunlunxin’s third generation, addressing the stability challenges inherent in massive scale. Leveraging Baidu’s Baige AI heterogeneous computing platform 4.0, the cluster attains 99.5 percent effective training time and over 96 percent linear scaling, delivering compute performance on par with international leaders for trillion‑parameter model training. The Tianchi supernode product further showcases system‑level integration: a single Tianchi 512‑node configuration can train trillion‑parameter models and supports integrating 64 Kunlunxin accelerator cards into one cabinet, enabling a single cabinet to replace hundreds of machines and materially reducing space and power cost per unit of compute.

Kunlunxin’s product iterations map the industry’s evolution from single‑component breakthroughs to full‑stack system integration. The first‑generation Kunlunxin (2018) deployed over 20,000 units for Baidu search and Xiaodu; the second generation (2021) added tens of thousands more; the P800 series represents a qualitative leap from “usable” to “competitive” to “leading.”

Commercial traction is the ultimate test of technical capability, and Kunlunxin’s go‑to‑market progress shows widening industry acceptance. In August 2025 Kunlunxin won a RMB 1 billion AI compute equipment procurement contract from China Mobile, a milestone signaling entry into national core information infrastructure procurement lists. For investors, such carrier procurement carries credibility beyond the contract value, indicating telecom‑grade stability and reliability.

In finance, Kunlunxin secured a project with China Merchants Bank. Performance tests indicate the P800 supports Qwen‑series models with performance that exceeds many domestic peers, delivering industry‑leading multimodal inference in scenarios such as data analysis, customer service, and code assistance. Given the financial sector’s stringent stability and security requirements, this case has strong demonstration value.

Kunlunxin’s commercial footprint now extends to energy and power, manufacturing, and academia. Its products have entered procurement systems of State Grid and China Southern Power Grid, serve clients such as the China Iron & Steel Research Institute, and have been deployed at universities including Tongji. Partnerships with hundreds of customers have delivered AI compute to internet, telecom, intelligent computing, finance, energy, and automotive sectors, benefiting hundreds of millions of end users.

A distinctive advantage for Kunlunxin is Baidu itself as a testing ground and anchor customer. Baidu’s internal workloads—from search and Wenxin large models to autonomous driving—provide real scenarios for iterative refinement. This “internal validation plus external expansion” model reduces market entry friction and accelerates product maturity, echoing the early commercialization path that NVIDIA followed from gaming GPUs to data center dominance. Kunlunxin converted internal compute demand into stable internal revenue before expanding into carriers, finance, and energy, enabling a smooth transition from internal use to external sales.

For Baidu, continued investment in frontier large models requires a robust compute foundation. As model parameter counts grow exponentially and multimodal capabilities advance, Kunlunxin’s high‑performance chips and large‑scale clusters form the base layer that supports Baidu’s technological edge.

Despite the favorable narrative, Kunlunxin’s IPO path faces substantive challenges. Valuation rationality is a primary concern: the company’s valuation rose from RMB 13 billion in 2021 to RMB 21 billion today, an increase exceeding 60 percent. With 2024 revenue above RMB 1 billion, JPMorgan projects revenue could reach RMB 8.3 billion in 2026—a sixfold increase—and anticipates breakeven in 2025. Nevertheless, industry‑wide bubble concerns persist, prompting investor caution toward lofty multiples.

Ecosystem challenges remain significant. NVIDIA’s CUDA ecosystem, built over two decades, constitutes a formidable moat. Although Kunlunxin has achieved compatibility with several mainstream frameworks, gaps persist in developer community scale and toolchain maturity. Compatibility issues in complex multi‑card training scenarios could impede customer migration. Deep integration with Baidu’s PaddlePaddle framework is both an advantage and a constraint: it ensures optimal performance within Baidu’s ecosystem but may limit broader adoption across open‑source communities.

Commercial risks include concentration of orders in policy‑driven procurements and within the Baidu ecosystem, raising questions about true gross margins and cash conversion cycles. Heavy reliance on a few large customers such as China Mobile could constrain pricing leverage and challenge sustained R&D investment. Geopolitical and supply‑chain risks also warrant attention: despite a “100 percent self‑developed architecture,” Kunlunxin remains partially dependent on global manufacturing supply chains in the near term.

Competition is intensifying. Huawei Ascend maintains strong positions in government and carrier markets, while Moore Threads and Muxi advance their own market entries, increasing competitive pressure. Some customers may adopt a dual‑sourcing strategy—purchasing both NVIDIA and domestic chips—keeping domestic vendors under continual comparison with the global benchmark. Closing the gap in performance, ecosystem, and service will be a long‑term imperative for Kunlunxin.

In the current financing round, Shanzheng Investment completed a strategic investment exceeding RMB 100 million, becoming a key capital partner. The firm stated the allocation aligns with national technology priorities and signaled plans to continue investing across compute, storage, and networking capabilities for AGI, as well as in large models and application scenarios. This support not only supplements Kunlunxin’s R&D and commercialization funding but also leverages industrial resources to strengthen its position within the domestic compute ecosystem. The participation of state‑linked capital reflects long‑term confidence in the domestic compute value chain while also assigning Kunlunxin a greater role in national strategic objectives, requiring a balance between commercial returns and public mission.

Kunlunxin’s IPO will serve as a barometer for the maturity of China’s AI chip industry. Its outcome will influence sector financing conditions and investor sentiment. Amid trillion‑yuan market opportunities and multiple risks, Kunlunxin must demonstrate sustainable value through continued technical iteration, ecosystem development, and commercial execution to convert valuation expectations into enduring enterprise value.

For investors, the story signals the rise of domestic compute capability but also calls for vigilance against speculative excess. For Baidu, the spin‑off offers a mechanism to unlock core AI asset value and reduce the company’s historical dependence on advertising revenue, potentially establishing a second growth axis. For the domestic compute industry, Kunlunxin’s listing process will provide a window into sector maturity; its full‑stack model of chip, cluster, and ecosystem may reshape competitive dynamics for China’s AI compute landscape.