On November 4, Lingchuan Technology announced a 1 billion yuan Series B round co-led by Yizhuang State Investment's Yitang Shenghe Fund, Qiming Venture Partners and Zhaohui Capital. The company was only spun out of Kuaishou in March 2024, carved from what had been the short-video platform's heterogeneous computing and chip division. Founder Liu Lingzhi previously served as a Kuaishou group vice president.

What it does sets it apart from most Chinese chip startups: no general-purpose GPUs, just video.

It Started by Saying What It Won't Build

Lingchuan's first chip, the SL200, has already shipped close to 100,000 units, running in data centers at Kuaishou, Alibaba Cloud, Baidu Cloud and Bilibili. According to an interview with 36Kr's Hard Krypton, it handles 99.7% of Kuaishou's live-streaming transcoding workload — transcoding specifically, not the entire streaming stack. That distinction matters: transcoding is the most compute-hungry and stability-critical link in the live chain.

In 2026 it plans to push through partners such as China Telecom into data centers, smart security, entertainment and transportation. The playbook is clear: prove it in your own backyard first, then sell it outside.

Why avoid general-purpose GPUs? COO Zhang Wenjing laid out the arithmetic in an interview: once you are serving hundreds of millions of users, carrying real-time inference on general-purpose GPUs that cost tens of thousands of dollars apiece is a structural imbalance in business terms. She used a phrase for it — "idling computing power" — GPUs that never reach peak utility on the inference side, money spent on computing power that goes unused.

The argument lands hardest in video generation. She listed several scenarios that have already closed a paying loop: AI comic series, AI short dramas, AI interactive video and AI advertising. What they share is extreme sensitivity to per-clip cost; the bigger the scale, the more inference cost weighs on margins.

Determinism, and Who Is Buying It

Lingchuan's technical pitch is DiPU, short for Deterministic Inference PU, a latency-deterministic inference architecture. The point is not raw compute but making latency, timing and bandwidth all predictable.

The approach combines dataflow-driven design with spatial computing, handing the compiler full control over memory and scheduling. The 3D near-memory design uses a multi-chip module built with hybrid bonding and advanced packaging to stack dies of different functions and different process nodes. An accompanying NUMA Fast Load mechanism sets up hierarchical NUMA domains so compute cores access memory directly, with the compiler slicing and laying out weights and activations in advance to hold local access latency under 2 nanoseconds. The team benchmarks this against OpenAI's Jalapeño, disclosed at Hot Chips as the industry's first inference ASIC with native NUMA HBM4 support. Benchmarking against it is not the same as matching it, but the direction is the same.

Why is determinism worth paying for? Content companies are buying experience. Wait several minutes for a clip to generate and the user is gone; when the first frame appears, and whether it stutters, decides willingness to pay. General-purpose GPU latency is hostage to scheduling and memory jitter, inherently uncertain. A dedicated chip can make a promise, and that promise is itself the product.

On the software side sits LUCAS, which auto-generates operators via a Python DSL plus MLIR, exposes low-level intrinsic interfaces and ships a CUDA compatibility layer so customers with large existing CUDA codebases avoid heavy rewrites. It currently supports Qwen, Tongyi Wanxiang, HunyuanVideo and MiniMax. Above that sits Nebula, an AIGV platform covering the full pipeline for text-to-video and image-to-video.

The Next Generation Waits Until 2027

The chip built entirely on domestic 3D stacking technology taped out in April and came back from the fab recently, with volume shipments planned for next year. Its stated target is the AI video production stage: cutting cost and time, pairing with the SL series to cover the full chain of production, recommendation, moderation and distribution.