Chinese short-video platforms already run some of the world’s densest real-time video pipelines — livestream transcoding, recommendation, moderation, and now generative video. The default answer has been “rent more GPUs.” Lingchuan bets that AIGV (AI-generated video) will not stay there: HBM cost and non-deterministic latency make TCO ugly once short drama, interactive video, and AI ads need to pay for themselves. That is COO 张文静’s thesis to 36氪 — video gen moving from demo to ROI, with OpenAI’s Sora cost lesson as caution*, and AI short-drama / interactive / ads as the first paying loops*.
The raise and the spinout
36氪 (Nov 4 ~09:33 Asia/Shanghai) says Lingchuan closed a ¥10亿 RMB Series B*. Joint leads*: 亦庄国投屹唐盛和基金, 启明创投 (Qiming), 朝晖资本. Co-investors include 国泰海通开元, 华控, 国寿长三角科创, 北工投资, 建信北京, 沛坤, HKSTP VC, 鼎晖百孚, 金舵, 厦门恒兴, 铭时; industry 石溪, 龙芯创投, 安徽文旅投; old 顺禧 / 水木 / 九智 topped up*. Round size and use-of-proceeds are press/company framing*.
Origin*: Beijing AI Fund + Kuaishou co-initiated; formerly Kuaishou’s heterogeneous compute & chip division; spun out March 2024* for smart video + generative-video multimodal compute*. CEO/founder 刘凌志 Liu Lingzhi*: SJTU PhD, UW postdoc; ex-Kuaishou VP; Huawei/Intel tech mgmt; built Kuaishou’s hetero/chip team from 2018*. Team pedigree listed as Kuaishou/Huawei/Ali/Baidu/NVIDIA/AMD/Intel/Arm* — 15+ years avg*, 10+ AI/video tape-outs* — company claims.
SL200 is already shipping into Chinese cloud / platform infra at reported volume — not a Western Digi-Key SKU, not an AWS/GCP instance you can click today. The 3D near-memory MCM* taped out in April on an all-domestic 3D stack*, silicon returned recently*, with volume expected next year*. Overseas access path, if any, is partner/OEM and later global expansion* — not retail boards this week.
Editorial still in the same magenta-coral / tungsten / graphite studio palette. Deployment numbers and latency claims below are vendor / press figures*.
SL200 — the chip already in the pipe*
SL200 is Lingchuan’s smart-video SoC*. Press claims: ~100k chips sold*; deployed at Kuaishou, Alibaba Cloud, Baidu Cloud, Bilibili*; covers 99.7% of Kuaishou livestream transcoding*. In 2026 it is also framed via telco/partners for datacenter, smart security, entertainment/transport*. Extraordinary penetration if accurate — asterisk them; they are China video-infra design-wins, not a Digi-Key catalog.
The useful overseas read is not “buy SL200 tomorrow.” It is that a Kuaishou-born video SoC has been stress-tested on one of the largest livestream/transcode fleets, and the same team is raising into generative-video load. Different maturity from this morning’s Zhizi Guangyuan optical seed story (tape-out rhetoric, no shipping SKU).
DiPU, 3D near-memory, and NUMA Fast Load*
Architecture*: RISC-V-rooted DiPU (Deterministic Inference PU)* — dataflow + spatial compute; compiler fully controls memory/scheduling; claims latency / timing / bandwidth triple-determinism* for multimodal / generative video vs GPU HBM pressure*. First-party 3D near-memory MCM* with hybrid bonding / advanced packaging*: all-domestic 3D stack taped out April*; silicon returned recently*; volume next year*; aims to cut AI video production cost/time*, and with SL series cover produce / recommend / moderate / distribute*.
NUMA Fast Load*: Memory Slice NUMA domains; core-direct memory; compiler pre-lays weights/activations; local access latency claim <2 ns*. Press frames this vs Hot Chips OpenAI Jalapeño NUMA-HBM4 ASIC language* — architecture talk, not an independent bench. Until third parties measure end-to-end video-gen latency and $/minute, treat determinism and nanosecond claims as vendor story.
Software: LUCAS + Nebula*
Software LUCAS* (Lingchuan Unified Computing Architecture Stack): four-layer stack*; Python DSL via Triton / TileLang + MLIR auto ops*; open Intrinsics*; CUDA compatibility layer* for shops that will not rewrite CUDA. Platform Nebula*: chip inference → model service → apps; text/image-to-video*; soft/hard co-design for content and cultural-tourism*. CUDA-compat* and MLIR/Triton paths are what overseas infra people will actually evaluate if a partner SKU ever appears.
Thesis and caveats
Zhang’s thesis*: Makimoto’s Wave toward specialization ~2027* — domain ASICs where video-gen TCO and first-frame latency decide paying loops. Liu’s market-size framing in press is aggressive*; treat it as rhetoric. Concrete watch-items: SL200’s claimed China infra footprint*, Series B capital for volume*, and whether the 3D near-mem part ships next year with a software path that is not a rewrite tax.
- Not Digi-Key / not Western cloud SKU — SL200 is China video-infra shipping*; 3D MCM is post-silicon / next-year volume*.
- ~100k units* / 99.7% Kuaishou livestream transcode* — vendor/press; extraordinary if true; not independently audited here.
- DiPU triple-determinism* / <2 ns local access* — architecture claims; wait for third-party benches.
- ¥10亿 Series B* — 36氪 exclusive; leads and co-investors as reported*; use-of-proceeds company framing*.
- CUDA-compat* / LUCAS* — software story is necessary for adoption; depth of compatibility unknown publicly.
- Distinct from Zhizi Guangyuan — this morning’s optical seed post is pre-silicon photonic; Lingchuan is shipping video SoCs + raising into AIGV specialization.
Bottom line: Lingchuan is the Kuaishou-line video-chip team turning livestream/transcode silicon into a generative-video inference franchise — already inside major Chinese clouds by its own numbers*, capitalized at ¥10亿*, pushing a deterministic RISC-V DiPU + 3D near-memory stack past tape-out but not yet a commodity part. Asterisk the unit counts and nanoseconds; watch whether next year’s volume and LUCAS/CUDA path make it usable outside the founding customer set.