A livestream host sells skincare for 14 hours straight without a bathroom break, a dropped tone, or a bad review rattling her. The telecom caller who renews your plan sounds patient at midnight. Neither of them is human.
These are "digital humans" (数字人) — and in China one company has turned them into a countable product it calls "silicon-based labor" (硅基劳动力), deliberately set against the carbon-based (碳基) human workforce. The firm is Nanjing Silicon Intelligence (南京硅基智能), and its story is a clean window into why the most hyped corner of generative media is also one of the hardest to make money in.
The 100,000-agent milestone
Founded in 2017 by Sima Huapeng (司马华鹏), Silicon Intelligence built its business on the idea that digital humans are not mascots but line items — replacements for repetitive voice and video labor. Its updated prospectus says it had delivered more than 100,000 digital-human agents to clients in telecom, finance, healthcare, education and public services by 2025.
The company leans on its DUIX platform and the Yandi large model (炎帝大模型) to power four product lines: intelligent voice (outbound calls), digital-human video (AI narrators), digital-human livestreaming (24/7 virtual selling), and multimodal interactive agents. A 2025 addition, fully automated content production, turns a text script into a finished video — aimed squarely at the short-drama (短剧) and advertising cost crunch.
Market lead, thin margins
By revenue in China's digital-human agent market, Silicon Intelligence says it held a 32.2% share in 2024 — first place, according to CIC (灼识咨询) — and ranked second globally. The total addressable market is real and growing: iResearch (艾瑞咨询) estimated China's digital-human market passed 100 billion yuan (about US$14 billion) in 2023 and projected roughly 300 billion yuan (about US$42 billion) by 2025, a compound growth rate above 50%.
Yet the company cannot convert leadership into profit. Revenue rose from 531 million yuan (about US$75 million) in 2023 to 655 million yuan (about US$92 million) in 2024 and 789 million yuan (about US$111 million) in 2025. Over the same three years it posted losses of 95.9 million, 112 million and 25.9 million yuan (about US$13.5 million, US$15.8 million and US$3.6 million). It warns 2026 will stay in the red.
The second run at Hong Kong
Silicon Intelligence first filed for a Hong Kong IPO on October 31, 2025; that application lapsed after six months. It refiled on May 21, 2026, with CMB International and DBS as joint sponsors — a second attempt to become the exchange's "first digital-human stock." Backers include Tencent Investment and Sequoia China, and a 2025 funding round valued the company at 3.15 billion yuan (about US$440 million).
Where the model strains
Two cracks show in the filings. First, customer concentration: the largest client accounted for 64.4% of revenue in the first half of 2025, and the top five made up 78.9% in 2024. Second, gross margin fell from 45.8% in 2023 to 34.5% in 2025 as the company cut prices on older products and discounted to win lower-tier markets. Growth, in other words, is being bought.
There is also a demand-side ceiling. Li Mingtao, chief e-commerce expert at the China International Electronic Commerce Center, has noted that digital-human hosts struggle to match the emotional pull of real streamers, and that platforms have begun capping how long AI hosts can broadcast to protect viewer experience. The "always-on" advantage collides with platform rules and audience tolerance.
Not every digital human is a sales clerk, though. The company's own IP character "硅基大司马" (Silicon Da Sima) — a digital-host persona — logged more than 3 billion cumulative plays within a year, a sign that the format can build an audience when the content, not just the labor saving, is the point. The lesson for the category is that distribution and creativity, not headcount replacement, may be where digital humans (数字人) actually earn their keep.
Why the Silicon Intelligence story travels
The Silicon Intelligence case is worth reading outside China because it compresses a tension every generative-media company now faces: the gap between shipping a product and earning from it. The company's own 100,000-agent count is a delivery milestone, not proof of recurring use, and its slide from a 45.8% gross margin toward 34.5% shows what happens when a novel capability becomes a commodity that rivals can undercut. For global readers, the parallel is the wave of overseas avatar and conversational-AI vendors who also discovered that "we can replace a shift of labor" is easier to demo than to monetize. A small e-commerce seller who points a 24/7 digital-human livestreaming line at off-peak hours may recover the cost quickly, but the same seller still needs a human for the moments that decide trust. The lesson is not that digital humans fail, but that labor replacement is a pricing problem long before it is a technology problem.
Honest limitations
- Delivery, market-share and revenue figures are drawn from the company's own prospectus and a third-party (CIC, iResearch) engaged by it; treat them as company-framed, not independently audited here.
- The 100,000-agent count measures units shipped, not active or revenue-retaining deployments; utilization by clients is not disclosed.
- Heavy customer concentration (one client near two-thirds of H1 2025 revenue) makes the top line fragile to a single relationship.
- Platform restrictions on digital-human livestream duration and weak audience engagement are real constraints that the technology alone has not solved.
What readers can do now
- If you run a store or call center, pilot a digital human (数字人) for the repetitive, off-peak shifts first — nights, FAQs, outbound reminders — where 24/7 uptime pays back fastest, and keep a human for high-empathy moments.
- Before buying, demand utilization and retention data from any vendor; "agents delivered" is a vanity number next to "agents still running after 90 days."
- Track the Hong Kong listing and the platform time-cap rules: both will tell you whether silicon-based labor (硅基劳动力) is scaling or being quietly throttled.