A short-video platform never really finishes with a video. Every clip that gets uploaded is copied, shrunk, sharpened, subtitled, checked and pushed out again to phones with wildly different screens. The work is invisible to viewers and relentless for the machines.
TranStreams (凌川科技) was built on a narrow bet: that this work deserves its own silicon rather than a slice of a general-purpose accelerator.
On 4–5 November the Beijing company said it had closed a RMB 1 billion (≈ US$141M / HK$1.1B) Series B, with the money earmarked for a next-generation chip and for selling abroad.
Key takeaways
- TranStreams raised a RMB 1 billion (≈ US$141M / HK$1.1B) Series B, jointly led by the E-Town Yitang Shenghe fund (亦庄国投屹唐盛和基金), Qiming Venture Partners (启明创投) and Aurora Capital (朝晖资本).
- The company was carved out of Kuaishou's (快手) heterogeneous computing and chip division in March 2024, and is backed by the Beijing AI industry fund and Kuaishou itself.
- Its current SL200 video chip has shipped close to 100,000 units, according to the company, and supports 99.7% of Kuaishou's live-streaming-related workloads; other named customers include Alibaba Cloud, Baidu AI Cloud and Bilibili.
- A next-generation chip built on domestic 3D stacking taped out in April 2026, has returned from the fab, and is targeted for volume shipment in 2027.
The spin-out that kept its first customer
The team did not start as a startup. Liu Lingzhi (刘凌志), now founder, chairman and CEO, joined Kuaishou in 2018 and built its heterogeneous computing and chip group from scratch, after research and engineering roles at Huawei and Intel and a postdoc stint at the University of Washington.
That origin explains the product. A chip team inside a video company does not begin with a benchmark; it begins with a cost line on someone else's P&L. Kuaishou needed to transcode enormous volumes of live video into many formats without buying proportionally more general-purpose GPUs, so the group designed a fixed-function part for the job.
In March 2024 the unit became a standalone company, with the Beijing artificial intelligence industry investment fund and Kuaishou as founding backers. The first market was the parent's own traffic; the second was everyone else's.
What the chip actually does
The SL200 handles video transcoding, compression and analysis. It is sold as a chip and accelerator card, and also bundled through server makers so customers can drop it into an existing data-centre rack.
The commercial evidence is unusually concrete for a Chinese chip startup: close to 100,000 units shipped, deployment at Kuaishou, Alibaba Cloud, Baidu AI Cloud and Bilibili, and coverage of 99.7% of Kuaishou's live-streaming-related workloads. This year the company has also been selling through telecom partners into data centres, security and transport applications.
The second act: generating video, not just moving it
The new round is not really about transcoding. It is about the far more expensive half of the pipeline: making video.
The next chip uses domestically developed 3D stacking — logic and memory bonded vertically through hybrid bonding and advanced packaging — plus the company's own "DiPU" architecture, which it describes as latency-deterministic, and a RISC-V core. The pitch is that packing compute next to memory cuts the data movement that dominates generative video workloads.
The timeline is checkable: tape-out in April 2026, silicon back recently, volume shipment targeted for 2027. On the software side, TranStreams has a unified compute platform, LUCAS, which it says is tuned for open video and language models including Qwen, Wan (通义万相), Hunyuan Video and MiniMax, and an AIGV platform called Nebula for text-to-video and image-to-video production.
If that sounds like a stack rather than a component, that is the point. Selling a video chip means selling a compiler, a model zoo and a deployment path, or customers will simply stay on GPUs.
The round, and what it buys
The Series B was jointly led by E-Town Yitang Shenghe, Qiming and Aurora, with Guotai Haitong Kaiyuan, Huakong, China Life Yangtze River Delta, Beijing Industrial Investment, CCB (Beijing) Investment, Peak Capital, the HKSTP Venture Fund, CDH Baifu, Jindun and Xiamen Hengxing among participants, plus industrial investors including Shixi Capital, Loongson Venture and Anhui Cultural Tourism Investment. Existing shareholders Shunxi, Waterwood and JZ Capital followed on.
By the count in Kechuangban Daily (科创板日报), that is 20 institutions. The company did not disclose how much each put in, nor the valuation — a gap worth noticing, because a headline round size without a price tells you about ambition, not about terms.
This is the second large cheque of the year: in June 2026 the company closed an A+ round of "several hundred million yuan" (数亿元) led by Qifu Capital.
Why video is where the bill is
Text models bill by the token; video models bill by the second, and a second of generated footage can cost orders of magnitude more compute than a paragraph. As Chinese platforms push short drama, advertising and interactive content through generative pipelines, the arithmetic turns from research curiosity into a procurement problem.
Liu has argued publicly that dedicated silicon for generated video is a market in the hundreds of billions of yuan (≈ tens of billions of US dollars / HK$100B-plus) today and could reach trillions of yuan (≈ hundreds of billions of dollars) within three to five years. That is a founder's forecast, not a measured market.
What would make this fail
Three risks stand out. First, customer concentration: a young chip company whose reference customer is also its co-founder is not yet a merchant supplier. Second, timing: 2027 volume shipment means the product arrives after another two years of GPU roadmaps. Third, the software gap — most alternative accelerators stall not on peak throughput but on the years of kernel work needed before a model runs well.
The round buys time for all three. It does not settle any of them.
Honest limitations
The funding announcement was reported by Kechuangban Daily, CNR (央广网), Jiemian and Kankan News from a company briefing; I did not obtain a primary press release, a cap table or any filing, because TranStreams is private and unlisted. The 100,000-unit shipment figure, the 99.7% coverage figure and the customer list are company statements repeated by those outlets, not independently audited numbers. The A+ round is described only as "several hundred million yuan"; the exact amount is unknown. Revenue, gross margin and valuation are undisclosed, so no unit economics can be calculated. The 2027 volume-shipment date is a plan. Liu's market-size forecast is his own estimate.
What readers can do now
- Treat "units shipped" and "revenue" as different facts; ask any private chip company for the second when it only offers the first.
- Watch the 2027 shipment date as the test of the 3D-stacking claim, and watch which models run well on LUCAS rather than which are listed as supported.
- If you buy video infrastructure, price a dedicated video card against a GPU on total cost per streamed hour, including engineering time, not on peak TOPS.