Kunlun Chip

Company Profile

Kunlun Chip (Beijing) Technology Co., Ltd., formerly known as Baidu's Intelligent Chip and Architecture Division, completed independent financing and commenced formal operations in April 2021 with an initial valuation of approximately 13 billion yuan. As one of China's earliest enterprises to enter the AI acceleration field, the company has dedicated itself to developing general-purpose AI solutions.chip designOver a decade, the company has built deep technical expertise in system architecture, chip implementation, software systems, and scenario applications. Its mission is “to make computing smarter,” with a vision to become the world's leading intelligent computing company. As of December 2025, the company has completed multiple rounds of financing, with its latest valuation reaching RMB 21 billion. Baidu holds a 59.451% stake as the controlling shareholder, while its shareholder structure includes industrial capital such as BYD and China Mobile, alongside state-owned capital platforms like China Reform Fund. In January 2026, Kunlun Chip submitted its listing application to the Hong Kong Stock Exchange, planning to further expand its capital and business boundaries through an independent IPO.

Products & Services

Kunlun Chip focuses on general-purpose AI computing chips and integrated hardware-software systems, forming a complete product line covering chips, accelerator cards, supernodes, and scenario solutions:

• Chip ProductsTwo generations of general-purpose AI chips have been mass-produced.
  • Kunlun Core Generation 1Mass production commenced in 2020, powering Baidu's search engine, Xiaodu, and other services. Tens of thousands of units were deployed, with expansion into internet, industrial manufacturing, smart finance, and other sectors.
  • Kunlun Core 2nd GenerationMass production commenced in 2021, featuring the next-generation XPU-R architecture. Performance has improved by 2-3 times compared to the first generation, supporting GDDR6 memory and PCIe Gen 4 interfaces. Deliveries to internet and industry clients have been completed, with commercialization progressing smoothly.
  • Kunlun Semiconductor P800 SeriesThird-generation product with tens of thousands of units deployed cumulatively, supporting single-cluster scales exceeding 30,000 cards. It accommodates 8-card and 16-card configurations, achieving peak throughput of up to 482 transactions per second. This marks the first domestically developed chip supporting standalone deployment of the full-power DeepSeek V3/R1 671B large language model.
• Acceleration Cards and Super NodesLaunching the K100, K200, R100, and R200 series accelerator cards, along with the RG800 and R480-X8 accelerator clusters, to support large-scale distributed training and inference scenarios.
• Scenario SolutionsCovering data centers, cloud computing, autonomous driving, smart industry, smart cities, and other fields, it provides full-stack support from underlying hardware to upper-layer model conversion. For example, its supernode product can accommodate 64 accelerator cards in a single cabinet, with inter-card connectivity bandwidth increased eightfold, significantly reducing communication bottlenecks in large model training.

core technology

Kunlun Chip builds a highly efficient and versatile technology system centered on its self-developed XPU architecture:

• XPU ArchitectureReleased in 2017, it is designed to meet general AI computing needs, supporting high-concurrency, low-latency access for scalar and vector computations, and accelerating core operations such as convolutions and matrix multiplication.
• Memory and Storage Optimization::
  • Leading the domestic market in supporting GDDR6 memory, it delivers 201 TP4T-501 TP4T higher memory bandwidth than comparable GPUs.
  • The RapidFS computing-storage acceleration technology achieves tight coupling between compute and storage, delivering 1.03 TiB/s throughput across 70 nodes to support concurrent GPU data access at the thousand-card level for 30,000-card clusters.
• Software EcosystemProvides Kunlun Core SDK, a full-stack toolkit spanning from underlying driver environments to upper-layer model conversion. Supports CUDA operator migration, with seamless adaptation for operators exceeding 701 TP4T. In certain scenarios, inference performance approaches NVIDIA T4 levels.
• Energy Efficiency and Cost-EffectivenessThrough its proprietary high-efficiency SIMD instruction set and core tensor computing units, it achieves a balance between performance and power consumption. A single cabinet delivers computing power equivalent to eight traditional servers, reducing the cost of training hyperscale models.

major client

Kunlun Semiconductor serves a diverse client base spanning the internet, finance, energy, research, telecommunications, and state-owned enterprises sectors.

• Internet industryBaidu Intelligent Cloud's 30,000-card cluster powered by P800 accelerators achieves training efficiency for Wenxin Yiyan that rivals international leaders. A major internet company's recommendation system leveraged Kunlun chip storage solutions to reduce latency by 75% and boost computing utilization by 30%.
• Financial sectorChina Merchants Bank has deployed Kunlun Chip servers to support multimodal data analysis and intelligent customer service, with certain model inference performance reaching industry-leading levels. Guangfa Bank's core transaction system utilizes its storage solution, processing tens of millions of requests daily while reducing response times from 120ms to 20ms.
• Energy IndustryState Grid Corporation of China's provincial company leveraged the P800 to build a power equipment failure prediction system, achieving a threefold increase in model training efficiency. China Southern Power Grid utilized its storage technology to enable long-term archiving of real-time equipment status monitoring data, boosting storage utilization to 911TB/4TB.
• research organizationPeking University, Tongji University, and other institutions have leveraged Kunlun Chip supernodes to build AI computing platforms, achieving a 5-10 times increase in computational efficiency for quantum computing simulations.
• Telecom operators and state-owned enterprisesIn 2025, the company secured a billion-yuan AI equipment procurement project from China Mobile, winning 70% of the “CUDA-like ecosystem” segment with a share of %-100%, becoming the primary supplier of inference computing power. Following BYD's investment, the two companies have expanded collaboration in automotive-grade AI chips and intelligent cockpits.

development prospect

Kunlun Semiconductor possesses significant growth potential, leveraging its technological expertise, customer base, and capital backing:

• Technology IterationPlans to launch new products annually over the next five years, with the Kunlun M100 chip for large-scale inference scenarios set to debut in 2026, followed by the M300 for training hyperscale multimodal models in 2027, continuing to build the “most robust AI cloud.”
• Commercialization ExpansionExternal clients account for approximately 40% of TP4T's revenue, with 2024 earnings exceeding 1 billion yuan—surpassing listed companies like Suyuan Technology and Beren Technology. JPMorgan forecasts its revenue to surge from 1.3 billion yuan in 2025 to 8.3 billion yuan in 2026, representing a sixfold increase.
• ecological coordinationDeeply integrated with Baidu's PaddlePaddle framework, achieving over 30% improvement in inference efficiency compared to TP4T; Collaborating with companies like Tongyou Technology to establish technical standards, driving the adoption of PCIe 5.0, CXL 3.0, and other standards in the financial and energy sectors.
• Capital PathAn independent IPO will enhance valuation transparency, attract specialized investors focused on hard technology, and grant it autonomous financing capabilities, thereby benefiting Baidu's overall ecosystem. Should the listing succeed, its valuation model and market performance will become key benchmarks for China's AI chip industry.

Challenges and risksKunlun Chip must continuously demonstrate its commercialization capabilities and profitability, particularly in expanding its customer base beyond the Baidu ecosystem. It must also address international competition and the pressure of technological iteration while maintaining architectural innovation and ecosystem compatibility. However, in the long term, Kunlun Chip is poised to become a key player in the global intelligent computing sector through technological breakthroughs and ecosystem synergy.