Hard Krypton Exclusive | Kuaishou-affiliated AI chip company Lingchuan Technology has completed RMB 1 billion Series B financing, its 3D chips have been taped out and will be shipped at scale next year.
Author | Qiu Xiaofen
Editor | Yuan Silai
Hard Krypton has learned that artificial intelligence chip company "Lingchuan Technology" recently completed a RMB 1 billion Series B financing. This round of financing will be mainly used for R&D and mass production of next-generation chips as well as global market expansion.
This round is jointly led by Yizhuang State Investment Yitang Shenghe Fund, Qiming Venture Partners and Zhaohui Capital, with participation of well-known investment institutions including Guotai Haitong Kaiyuan, Huakong Investment, China Life Yangtze River Delta Science and Innovation Fund, Beigong Investment, CCB Trust (Beijing) Investment, Peikun Capital, Hong Kong Science Park Venture Capital Fund, CDH Baifu, Jinduo Investment, Xiamen Hengxing and Mingshi Capital, industrial investors including Shixi Capital, Loongson Venture Capital and Anhui Cultural Tourism Investment. Old shareholders including Shunxi Fund, Shuimu Investment Group and Jiuzhi Capital continued to increase their capital contribution.
"Lingchuan Technology" was jointly initiated and established by Beijing Artificial Intelligence Fund and Kuaishou Group. Its predecessor was the Heterogeneous Computing and Chip Division of Kuaishou Group, focusing on the underlying computing power of multi-modal large models such as intelligent video and generative video.
Liu Lingzhi, the founder, chairman and CEO of the company, holds a doctorate from Shanghai Jiao Tong University and a postdoctoral degree from the University of Washington. He once served as vice president of Kuaishou Group and held core technical management positions in Huawei, Intel and other enterprises. Since 2018, he has built the Kuaishou heterogeneous computing and chip team from scratch, and is a leading figure with both academic depth and industrial practical experience. The team independently developed the SL200 chip intelligent video SoC chip inside Kuaishou, and was officially split for independent operation in March 2024.
Since the publication of Attention Is All You Need in 2017 and NVIDIA launched the first Volta V100 integrated with Tensor Core, AI computing chips represented by general-purpose GPUs have experienced a golden decade;
Since 2027, the industry will enter a new cycle of Makimoto's Wave, and special computing power devices for vertical tracks such as AIGV (AI Generated Video) will embrace the greatest innovation opportunities and growth space.
Short video is a vertical track where China has obvious leading advantages across the globe, giving rise to large commercial applications such as Kuaishou, ByteDance and Xiaohongshu, and internationally leading large models such as Keling and Seedance are continuously iterating.
With the outbreak of applications such as short dramas, interactive videos and AI advertisements, the demand for processing and transmission of video data has increased sharply, putting forward higher requirements for the efficiency, power consumption and cost of underlying computing power.
Under this trend, AIGV is regarded as an important application scenario of next-generation intelligent computing power. The industry is moving from the stage of "generating pixels" to the world model-driven stage of "simulating physics". The demand for effective computing power has increased exponentially compared with traditional text models, occupying an important part in the Token consumption of all large models.
Facing the surging wave of AIGV, Liu Lingzhi previously judged that this track is currently a hundred-billion-level track, "it will explode to hundreds of billions of levels in the next three to five years, and may exceed one trillion yuan."
However, traditional general-purpose GPUs have obvious efficiency bottlenecks in video processing tasks, resulting in high Token and operating costs.
At the same time, domestic video generation large models are evolving rapidly, and computing power consumption is further amplified. Facing the fundamental problem of computing power supply, the demand for domestic special chips is becoming more and more urgent, which is particularly important for China to maintain its international leading position in the short video field.
In this situation, computing power base chips that are more efficient, more dedicated and easier to use have become rigid demands.
Aiming at the computing power bottleneck of generative large model scenarios such as AIGV, Lingchuan Technology focuses its technology on the deterministic latency computing architecture.
According to the introduction, Lingchuan Technology is deeply integrated into the RISC-V open source ecosystem, launches the innovative deterministic latency DiPU (Deterministic Inference PU) architecture, adopts data flow-driven and spatial computing architecture, fully controls storage and scheduling through the compiler, solves the dependence of multi-modal large models and generative video on high video memory and bandwidth, and achieves triple determinism of latency, timing and bandwidth.
Hard Krypton has learned that the company's pioneering 3D near-memory architecture adopts a multi-chip module (MCM) design, and realizes heterogeneous integration of chips with different functions and processes through hybrid bonding and advanced packaging technologies.
According to the team, the chip of Lingchuan Technology adopting all-domestic 3D stacking technology completed tape-out in April, and the returned chip was received recently, which is expected to be shipped on a large scale next year. This chip is specialized in the AI video production link, aiming to greatly reduce the cost of video generation and shorten the generation time, and together with the SL series, it covers the full link of video production, recommendation, review and distribution.
It is worth mentioning that Lingchuan Technology has also designed the NUMA Fast Load mechanism for 3D near-memory topology, breaking the near-memory access barrier — the Memory Slice forms a layered NUMA domain, the computing core is directly connected to memory access, the compiler pre-arranges the slices of large model weights and activation values, and the local access latency is less than 2ns.
This concept is in line with OpenAI Jalapeño disclosed on Hot Chips (the industry's first inference ASIC natively supporting NUMA HBM4). Under 3D integration, the DiPU architecture is further deepened into a "launch desktop" prefetch mechanism with both extreme bandwidth and controllable power consumption.
Furthermore, at the software level, Lingchuan Technology focuses on the DiPU architecture and independently develops the unified computing platform LUCAS (Lingchuan Unified Computing Architecture Stack). Through a four-layer software architecture from the application adaptation layer to the hardware execution layer, it adapts to different applications, and at the same time supports the mainstream open source ecosystem, graph/Eager mode, TILE-BASED programming and CLUSTER AFFINE, forming a complete technical route of dedicated computing power + open source architecture + in-depth scenario optimization.
In terms of software stack construction, Lingchuan Technology chooses to embrace the open source ecosystem, and automatically generates operators based on Python DSL (such as Triton/TileLang+MLIR), so that customers do not need to build their own ecosystem;
At the same time, the underlying Intrinsic interface is open for extreme optimization of performance-critical code, and a CUDA compatibility layer is provided for in-depth optimization of customers who have a large amount of existing CUDA code and do not want to reconstruct it.
At present, this platform has completed the adaptation and optimization of mainstream open source and video generation models such as Qwen, Tongyi Wanxiang, Hunyuan Video and MiniMax.
Based on the full-stack capability of "chip + software stack", Lingchuan Technology has also launched the AIGV platform Nebula, which takes self-developed chips as the base, connects reasoning, model services to industry applications, covers the full link of text-to-video/image-to-video, realizes software-hardware collaboration and multi-model management, and is oriented to content production and cultural tourism implementation.
While laying out the next-generation architecture, Lingchuan Technology's existing products have achieved large-scale implementation. At present, nearly 100,000 SL200 chips have been sold and deployed to Internet companies such as Kuaishou, Alibaba Cloud, Baidu Cloud and Bilibili, covering 99.7% of Kuaishou's live transcoding business.
In addition, in 2026, SL200 will also realize product implementation through cooperation with partners such as telecommunications, and jointly serve scenarios such as data centers, smart security, cultural entertainment and transportation, saving customers' storage space while ensuring clear image quality, and all relevant technical indicator tests have passed.
In terms of team, the core team of Lingchuan Technology comes from domestic and foreign enterprises such as Kuaishou, Huawei, Alibaba, Baidu, NVIDIA, AMD, Intel and Arm, with an average of more than 15 years of experience in chip design, software algorithm and infrastructure, and has led the tape-out and mass production of more than 10 AI and video chips.
The team has a number of experts selected into the national high-level talent plan and provincial and ministerial talent plans of Beijing, Shanghai, Hubei and other provinces and cities, as well as a number of industry leaders with rich industrialization experience.
Investor Views
Yizhuang State Investment: Zheng Hao, General Manager of the Integrated Circuit Investment Division of Yizhuang State Investment and Managing Director of Yitang Shenghe Fund, said: "As an industry-empowered investment platform of Beijing Economic and Technological Development Area, Yizhuang State Investment invests in Lingchuan Technology, which focuses on its scarce large-scale implementation capability and independent architecture innovation, the ecological niche of deterministic computing power base supporting the large model era, and the strong traction effect of the project landing in the Economic Development Area on improving the intelligent computing power industrial chain. We will also accompany it from large-scale mass production to global competition with patient capital."
Qiming Venture Partners: Zhou Zhifeng, Managing Partner, said: "At the node where the demand for generative video and large model computing power is growing exponentially, Lingchuan has chosen a differentiated path: cutting in with the 'deterministic latency' architecture, deeply integrating into the RISC-V ecosystem, and using dedicated chips to solve the pain points of low efficiency and high cost of general-purpose GPUs in multi-modal large model scenarios. This 'hardware-based, software-defined' approach perfectly fits our judgment of 'efficiency first' for AI infrastructure. We believe that the team has both the chip genes of Huawei and Intel and the large-scale system capability of Kuaishou, and is expected to achieve a 'home run' level opportunity in the intelligent computing power segment track."
Shixi Capital: Shixi Capital has long focused on and deeply cultivated the semiconductor and emerging technology tracks. Lingchuan Technology is an important layout in our cloud-side dedicated large computing power infrastructure link. At present, video generation has been upgraded from a cutting-edge trial function to the core load with the fastest growing cloud computing power, and the energy efficiency boundary of general computing power can no longer keep up with the explosion of demand for large video models. The company's team and industrial partners have long been rooted in the field of video generation processing, have in-depth understanding of large model loads and new industry application trends, and are expected to grow into a core participant in this track. We will move forward with the enterprise and witness the next growth cycle of the computing power industry together.
Old Shareholder Shunxi Fund: Lingchuan Technology was spun off from Kuaishou Group in 2024, focusing on the high-growth blue ocean track of multi-modal large model reasoning, and is a leading domestic enterprise dedicated to multi-modal large model reasoning chips. Relying on the advantages of innovative technology architecture, the company is expected to greatly reduce the unit reasoning cost of chips, and is deeply bound with Kuaishou, Keling and other commercial sides. As the company's first round of investment institution, Shunxi Fund has continuously increased its investment for multiple rounds, and has always been optimistic about the team's technical accumulation and commercialization prospects.
The following is the interview between Hard Krypton and Zhang Wenjing, COO of Lingchuan Technology
Hard Krypton: There have been many changes in the video industry this year, such as the change of Sora. What new requirements does this put forward for computing power? Have you changed your product strategy because of this?
Zhang Wenjing: The company's technical route has always been deeply coordinated with the video scenario. The rapid iteration of the industry this year essentially indicates that video generation is fully shifting from the early "technology show" to "commercial closed loop", and technology realization must be driven by business ROI. This is a clear positive for Lingchuan, which not only does not shake our route, but strongly confirms our early prediction of product definition.
For example, OpenAI's strategic adjustment on Sora this year verifies that the "mode of piling up general computing power regardless of cost" is commercially unsustainable, as well as the importance of scenarios and audience scale. The core pain point of video generation is that the reasoning computing power cost is too high, the unit output ROI is difficult to turn positive, and the more business, the greater the loss, which is essentially the heavy TCO pressure that backfires on business expansion. The rapid catch-up of domestic models in cost performance and engineering has also accelerated the emergence of this trend.
This matter points to an irreversible consensus from another angle: large models are moving into the deep water of application, and manufacturers must spare no effort to reduce the full-cycle TCO. The model is an enabling tool for the industrial chain, and must fall into specific commercial scenarios to generate positive cash flow. Based on this, Lingchuan firmly focuses on "engineering realization and commercial cost reduction", not simply pursuing leadership in technology and indicators, but emphasizing the ability of real implementation to overcome the industrial bottleneck of reasoning end cost.
Hard Krypton: What is the first scenario in the video field that has achieved a commercial closed loop?
Zhang Wenjing: Vertical content forms represented by AI manhua dramas, AI short dramas, AI interactive videos, AI advertisements, etc., are scenarios that have been clearly verified and formed a stable paid closed loop. Not only the traffic realization link is clear, but the penetration rate of AI in its content generation flow is also increasing exponentially with the decrease of cost.
From the perspective of the underlying architecture, when the service scale is pushed to hundreds of millions of users, continuing to use general-purpose GPUs costing tens of thousands of dollars to carry real-time reasoning has a structural imbalance in business. In addition to the high cost, GPUs cannot exert the ultimate utility at the reasoning end of large models, which leads to the problem of partial "computing power idling". For downstream content providers who are extremely sensitive to profits, the reasoning computing power cost directly determines the life and death of the project.
In addition, the first frame latency and end-to-end generation efficiency are also key factors affecting user experience and willingness to pay. In the past, the waiting experience of generating a video for several minutes at a time was unfriendly to users' immediate interaction willingness. Therefore, the requirements of the large model reasoning end for next-generation chips and computing power clusters are concentrated on the ultimate TCO cost per unit, and the low-latency throughput that can support real-time interaction. This is exactly the core battlefield where Lingchuan has laid out in advance.
Hard Krypton: What kind of changes will happen to the industry trend of general-purpose GPUs and dedicated ASICs?
Zhang Wenjing: There is a classic law in the history of semiconductor industry evolution — "Makimoto's Wave", that is, generalization and specialization alternately dominate in a cycle of roughly ten years.
The past decade was the cycle of general-purpose chips. When the efficiency of general computing power encounters a bottleneck, the industry will turn to a dedicated route to pursue extreme energy efficiency. After large models enter the implementation period, scenarios are the first driving force for computing power innovation. Through deep customization of key operators and memory access mechanisms for video reasoning, dedicated chips can bring orders of magnitude improvement in efficiency and greatly optimize TCO. Lingchuan is firmly following this implementation route driven by dedicated efficiency to promote industrial popularization.
Homepage Image Source | Provided by the Enterprise
Welcome for communication