Hardcore Exclusive | Kuaishou-backed AI chip company raises billions, sales near 100,000 units, video compression performance surpasses NVIDIA 硬氪首发 | 快手系AI芯片公司再融资数亿，销量近十万颗，视频压缩性能超英伟达

TL;DR

• Lingchuan Tech raised a multi-hundred million A+ round for its AI video processing chips.
• The company originated from Kuaishou’s chip team, achieving massive deployment there.
• Its SL200 chip claims a 30-35% encoding efficiency lead over NVIDIA’s AV1 solution.
• Funding focuses on next-gen 3D-stacked chips, production scale-up, and overseas expansion.
• The firm targets the emerging AIGC/AIGV market with a specialized, high-efficiency architecture.

Key Data

Deep Analysis

Lingchuan Tech’s massive A+ round isn’t just another AI funding headline—it’s a telling signal of where smart money is betting in the post-ChatGPT craze: the infrastructure bottleneck. While foundation models grab eyeballs, the brutal economics of actually running them, especially for video, are what will determine winners and losers. Lingchuan is positioning itself not as a general-purpose AI chip maker, but as a scalpel for the specific, agonizing pain point of video inference and generation.

Let’s be clear about the competitive claim. A 30-35% efficiency gain over NVIDIA’s AV1 encoder is a staggering assertion, if it holds at scale in real workloads. This isn’t about raw FLOPS; it’s about power, cost, and latency per task. In hyperscale data centers, a 30% efficiency jump translates directly to billions in saved electricity and capital expenditure over a few years. The fact they’ve already won three consecutive world video coding contests gives this some serious technical credibility. It suggests they’ve attacked the problem at a fundamental architectural level, not just optimized a corner.

Their origin story is their greatest asset and potential weakness. Spinning out of Kuaishou means their technology was born and battle-tested in one of the world’s most demanding video processing environments—billions of user-generated videos, live streams, and AI applications. When they say the SL200 handles 99.7% of Kuaishou’s live transcoding, that’s not a lab demo; that’s production-grade validation few startups can claim. It de-risks the technology for investors and customers. However, this same deep tie to Kuaishou could pigeonhole them. The market will wonder: can they truly be a platform player for everyone (Alibaba Cloud, Bilibili, Huawei), or will they always be seen as Kuaishou’s in-house project gone external? Their success hinges on proving the latter.

The strategic pivot to AIGV (AI-generated video) is astute but fraught with peril. Generating video is orders of magnitude more compute-intensive than text. Their “delay deterministic” DiPU architecture and focus on solving memory and bandwidth constraints sounds like the right technical bet for this next wave. They’re essentially betting that the era of brute-forcing problems with generic GPU clusters is ending, and that intelligent, application-specific architectures will win. The development of a 3D-stacked chip shows they’re thinking long-term about physics and manufacturing, not just clever design. But jumping from a video processing SoC to a chip designed for generative model inference is a monumental leap. The compiler and software stack (LUCAS) will make or break this transition.

The investors read like a strategic who’s-who. Qifu Capital, BV Baidu Ventures, New Kingdom (a fintech player), and critical state-backed funds like Beijing Yizhuang. This isn’t just financial capital; it’s a coalition. BV Baidu gives them a direct link to Baidu’s AI models and cloud. The state backing is crucial in an era of geopolitical tech fragmentation and the urgent national push for semiconductor self-sufficiency. They’re not selling a chip; they’re positioning themselves as a piece of critical national AI infrastructure. This political tailwind could accelerate adoption in government and SOE projects, but also invites intense scrutiny and expectation management.

The core question for Lingchuan is scale versus specialization. Can their specialized architecture achieve the cost and yield needed to compete in a market conditioned by NVIDIA’s economies of scale? Their current product, the SL200, is a mature workhorse. Their next-gen 3D chip is a high-risk, high-reward bet on the future of packaging and memory. If it delivers, they could leapfrog competitors. If it stumbles on thermal or yield issues, they could burn through this funding round quickly. In the brutal AI chip game, having brilliant technology is only half the battle; the other half is flawless execution on manufacturing, ecosystem building, and sales—areas where history is littered with failed startups.

Industry Insights

1. The AI chip market is fracturing from general-purpose towards hyper-specialized architectures for video, recommendation, and edge workloads. One-size-fits-all GPUs will cede ground to domain-specific accelerators.
2. Software ecosystems and compiler technology are becoming the primary moat for chip startups. A superior architecture is useless without a seamless developer experience; the LUCAS platform is as important as the DiPU itself.
3. Geopolitical and supply chain pressures are creating a protected, high-stakes market for domestic AI infrastructure. Success requires not just technical merit, but strategic alignment with national initiatives and deep integration with key industrial players.

FAQ

Q: What makes Lingchuan Tech different from other Chinese AI chip startups?
A: Unlike many competitors, Lingchuan’s technology is uniquely forged in a high-stakes, massive-scale consumer internet environment (Kuaishou). Its proven success in handling live video for 700 million users provides a rare, real-world validation of its performance and reliability claims.

Q: How reliant is Lingchuan Tech on its relationship with Kuaishou?
A: This relationship is currently a double-edged sword. It provides an unparalleled launchpad and validation, but the company is actively diversifying by securing other major cloud and internet clients like Alibaba Cloud and Bilibili, and expanding into new application scenarios to prove its platform independence.

Q: What is the biggest risk to Lingchuan’s future?
A: The biggest risk is execution. Scaling a next-generation, advanced packaging (3D-stacked) chip to volume production while simultaneously building a broad software ecosystem is an extremely complex and capital-intensive challenge. Any misstep in manufacturing yield or compiler development could stall its growth trajectory.

TL;DR

• 人工智能芯片公司「凌川科技」完成数亿元A+轮融资，由啟赋资本领投。
• 公司脱胎于快手集团芯片团队，核心产品SL200芯片已在快手部署数万颗。
• 公司推出创新性延迟确定性DiPU架构，专注解决视频与AIGV算力瓶颈。
• 下一代采用全国产3D堆叠技术的芯片已完成流片。
• SL200芯片已销售近十万颗，服务B站、阿里云、百度云等客户。

核心数据

深度解读

一家脱胎于大厂业务部门的芯片公司，一年之内拿到数亿融资，核心产品已在集团内部大规模验证。凌川科技的故事，绝不仅仅是“又一个芯片创业公司”那么简单。它精准地卡在了当前AI算力演进最尖锐的矛盾点上：大模型从文本冲向视频，算力需求呈指数爆炸，而我们依赖的通用GPU（特指英伟达）架构，在面对海量、实时、低延迟的视频处理任务时，已显现出效率与成本上的双重“笨重”。

快手孵化出凌川，这步棋堪称教科书级的“带产孵化”。它把内部面对数亿用户、苛刻业务场景打磨出的芯片团队和产品（SL200）直接“抛向”市场。这不是实验室里的原型，而是已在战场上久经考验的“老兵”。SL200在视频编码大赛上对英伟达的“断崖式领先”，以及在快手核心业务上的全覆盖，为其提供了无可辩驳的商业信任状。这揭示了一个趋势：未来顶尖的垂直算力芯片，很可能最先诞生于拥有海量、复杂真实场景的巨头内部，然后独立出去，成为细分赛道的霸主。

但凌川的野心不止于做一款“视频处理加速卡”。其提出的DiPU（确定性推理单元）架构，才是更值得玩味的棋。它强调的“延迟、时序、带宽三重确定性”，直指当前通用AI芯片在复杂推理任务中的痛点——不确定性。当你驱动一个文生视频大模型时，你最不希望看到的就是输出帧率忽高忽低、延迟抖动。DiPU试图从底层硬件架构上，为这类负载“编程式”地保障性能。这是一种非常“工程化”、却也极其关键的路线。它避开了与英伟达在通用AI训练算力上的正面军备竞赛，转而深挖“AI推理”特别是“视频AI推理”这个即将爆发的场景。

结合其软件栈LUCAS和对RISC-V的拥抱，凌川的蓝图清晰：构建一个从专用硬件架构到开源软件生态的完整垂直堆栈。这不仅仅是“国产替代”，而是在一个细分但庞大的赛道上，试图定义新的游戏规则。投资方的名单里，既有互联网战投（百度风投），也有地方政府基金（亦庄国投），还有产业资本（新国都），说明其故事获得了“技术+产业+政策”的多重认可。

然而，挑战也同样尖锐。首先，从服务快手到全面市场化，如何让阿里的、百度的、B站的工程师习惯并信任一个全新的架构和生态，是巨大的工程和销售挑战。其次，视频与AIGV的算力需求虽然明确，但模型架构和算法仍在高速迭代，专用芯片如何保持足够的灵活性和通用性，以适应未来的变化？最后，当整个行业都在为“国产AI算力”欢呼时，我们更需要冷静：真正的自主可控，不是换个国产标签，而是在核心架构和关键IP上，拥有不被卡脖子的、经得起全球市场检验的真本事。凌川的“韬定律”和3D堆叠，是迈出了这一步，但前路依旧漫长。它的成败，将是中国AI算力能否在应用侧撕开一道真正有竞争力的口子的一块关键试金石。

行业启示

1. 大厂业务部门“带产孵化”，将是未来硬核科技创业的重要模式，能极大缩短技术从验证到商业化的周期。
2. AI算力竞争正从通用训练向专用推理深化，在视频、多模态等高带宽、低延迟场景，专用架构芯片的价值将急剧凸显。
3. 国产AI芯片的“替代”叙事正在升级，从“能用”走向“好用”甚至“引领”，关键在于能否从真实、大规模的业务场景中提炼出真正的架构创新。

FAQ

Q: 凌川科技和快手是什么关系？
A: 凌川科技脱胎于快手集团的芯片团队，于2024年3月正式拆分独立运营，快手是其联合发起方和重要客户。

Q: 凌川的技术路线与英伟达等主流厂商有何不同？
A: 凌川专注于视频与AIGV场景，其DiPU架构强调“延迟确定性”，采用数据流驱动与空间计算，旨在以更专用的设计解决特定场景下的高算力与高能效比问题，与通用GPU路线形成差异化。

Q: 公司的产品主要用在哪些地方？
A: 核心产品SL200芯片主要用于互联网数据中心的视频转码、AIGV推理加速，也延伸至智慧城市、智能制造、边缘计算等场景的实时视频分析。