Kuaishou-affiliated AI chip company Transtreams recently completed a CNY 1 billion Series B financing round. The round was jointly led by Yizhuang State Investment's Yitang Shenghe Fund, Qiming Venture Partners, and Zhaohui Capital, with participation from industrial capital including Shixi Capital, Loongson Venture Capital, and Anhui Cultural Tourism Investment. Existing shareholders such as Shunxi Fund, Tsinghua Unigroup (Shuimu Investment Group), and Jiuzhi Capital continued to increase their holdings. Additionally, multiple institutions participated, including Guotai Haitong Kaiyuan, Huakong Investment, China Life Yangtze River Delta Sci-Tech Innovation Fund, Beijing Industrial Investment, Construction Bank (Beijing) Investment, Peikun Capital, Hong Kong Science Park Venture Capital Fund, CDH BaiFu, Jinduo Investment, Xiamen Hengxing, and Mingshi Capital. The funds will be primarily used for next-generation chip R&D, mass production, and global market expansion.
This financing sends two signals: first, AIGV dedicated computing power is moving from concept to commercial validation; second, the domestic AI chip sector is differentiating into more specific vertical routes beyond just "general-purpose GPU replacement."
| Understanding Transtreams
Transtreams originated from Kuaishou's Heterogeneous Computing and Chip Business Unit and became an independent entity in March 2024. Its first product, the SL200, has sold nearly 100,000 units and is deployed on platforms including Kuaishou, Alibaba Cloud, Baidu Cloud, and Bilibili, covering 99.7% of Kuaishou's live streaming-related business. In the MSU World Video Encoding Competition's commercial track, the SL200 ranked first for three consecutive years, achieving a 30%-35% improvement in compression efficiency compared to NVIDIA's latest AV1 solution.
This is critical.
The issue for many AI chip companies is not unclear technical narratives but a lack of validation under real business loads. Transtreams's advantage lies in having validated its product within a high-concurrency, high-bandwidth, and cost-pressured video business before expanding outward. For the chip industry, this path of "defining scenarios first, then designing chips" is more convincing than purely financing-driven approaches.
| Next-Generation 3D Chip Tape-Out Returned
Transtreams's next-generation chip, which completed tape-out in April this year, has returned wafers and is expected to achieve volume shipments next year. It adopts fully domestic 3D stacking technology, featuring a 3D near-memory architecture that achieves heterogeneous integration through hybrid bonding and advanced packaging.
This direction has high barriers to entry.
3D stacking, near-memory computing, hybrid bonding, and NUMA memory access optimization each involve challenges in packaging, heat dissipation, consistency, yield, and software scheduling. By choosing this path, Transtreams is essentially addressing the most pressing issues in multimodal large models and video generation scenarios: memory bandwidth, data movement, and inference costs.
Architecturally, it does not follow the general-purpose GPU route but proposes a Latency Deterministic DiPU architecture, emphasizing compiler control over storage and scheduling. It attempts to transform workloads like video generation, inference, and transcoding into more controllable dataflow execution. This aligns with the trend of global tech giants developing custom AI chips—dedicated architectures often offer superior energy efficiency for specific workloads compared to general-purpose GPUs.
| The Economics of AIGV Computing Power Have Changed
Video generation differs significantly from traditional text models. Text inference primarily consumes language model computation, whereas video generation involves temporal, spatial, resolution, frame rate, and visual consistency requirements, causing computational demand to rise exponentially. For data centers, this means GPU hour costs, storage costs, and bandwidth costs are rapidly amplified.
Transtreams estimates that the current AIGV-related computing power market remains at the tens of billions of yuan level, potentially expanding to hundreds of billions or even trillions of yuan in the next three to five years. Whether this forecast fully materializes requires observation, but the direction is clear: video generation is shifting from a "demo feature" to a "production workload," and production workloads inevitably drive the need for dedicated chips.
A more realistic factor is that scenarios such as domestic short dramas, interactive videos, AI advertising, and cultural tourism content creation are scaling up. Transtreams has signed strategic cooperation agreements with industry partners like Hubei Changjiang Film Digital Communication to promote its AIGV platform, Nebula, into content creation and cultural tourism scenarios. It should be noted that public information currently mainly confirms signed partnerships, while actual deployment scale and depth of commercialization remain to be disclosed. This indicates that the company aims not only to sell chips but also to bundle chips, software stacks, and industry applications into comprehensive solutions.
| Capital Selects Computing Power Based on Implementation
Looking at the investor lineup for this round, state-backed funds such as Yizhuang State Investment, China Life Yangtze River Delta Sci-Tech Innovation Fund, Beijing Industrial Investment, and Construction Bank Investment participated, followed by industrial capital like Shixi Capital and Loongson Venture Capital, with existing shareholders continuing to add to their positions.
Shixi Capital holds resources in the memory chip industry, Loongson Venture Capital represents the domestic CPU ecosystem, and Yizhuang State Investment emphasizes industrial chain implementation. These investors are not merely betting on the "AI story" but on a more complete chain: memory, packaging, instruction set ecosystems, and data center deployment.
In particular, Shixi Capital has actively invested in the 3D chip sector in recent years, backing companies such as Qianhe Yibang, SUNMMIO, and Transtreams. This suggests that 3D stacking, near-memory computing, and compute-in-memory synergy are becoming new focal points for capital attention.
| Three Key Challenges: Mass Production, Ecosystem, Customers
Transtreams's strength lies in the commercial foundation established by the SL200, but whether the next-generation 3D chip can smoothly enter mass production remains a critical milestone. Yield, heat dissipation, packaging costs, and software adaptation for 3D chips all require validation in real customer environments.
Another challenge is the software ecosystem. The company has developed a unified computing platform called LUCAS, which has already adapted to models such as Qwen, Tongyi Wanxiang, Hunyuan Video, and MiniMax, while providing a CUDA compatibility layer. This is important for reducing customer migration costs, but the general-purpose GPU ecosystem remains powerful. For dedicated chips to truly replace them, relying solely on single-point performance advantages is insufficient; customers must be willing to restructure their workflows. It is worth adding that the CUDA compatibility layer is a double-edged sword: it lowers the entry barrier but may not allow customers to leverage the full performance benefits of the dedicated architecture. If customers continue running original workflows on the compatibility layer long-term, the energy efficiency advantages of the dedicated chip may be diluted. The compatibility layer is a transitional tool, not an ecological moat.
Furthermore, competition in the cloud AI accelerator card market is intensifying. Companies such as Enflame, Moore Threads, MetaX, and Biren have entered capital markets sequentially, and both general-purpose GPU and DSA routes will continue to expand. If Transtreams remains limited to video transcoding, its growth space is constrained; however, if it secures stable orders in scenarios like AIGV inference, video generation, and embodied intelligence vision, its differentiation will become more pronounced.
| AI Chips Enter the "Vertical Workload Definition" Phase
Over the past decade, the main theme of AI chips was general-purpose GPUs. In the coming years, the focus may shift to "who understands a specific class of workloads better." Video generation, voice interaction, recommendation systems, industrial vision, and embodied intelligence each have distinct requirements for bandwidth, latency, precision, and power consumption.
Transtreams's financing round reflects industrial capital's willingness to invest in dedicated chips that possess "real business validation, clear workload boundaries, and viable mass production paths." For domestic chip companies, this approach holds more practical significance than simply chasing general-purpose GPUs.
The key metrics to watch next year are whether the 3D chip achieves scheduled volume shipments, enters data centers of non-Kuaishou customers, and whether the AIGV platform can genuinely reduce video generation costs. These are the hard indicators that determine if an AI chip company transitions from "usable" to "essential."
Disclaimer: This article is compiled based on public information and industry interviews. All financing details, product progress, and market judgments referenced herein are subject to official disclosures by relevant parties and do not constitute investment advice. Assessments of the AIGV computing power market size and analyses of Transtreams's prospects in mass production, ecosystem, and customer expansion are inferences based on available information and carry uncertainty. Readers should verify independently and form their own judgments.