At a labeling floor in a provincial capital, a team works through frames of a warehouse robot missing a pallet by ten centimeters. Each frame gets a human judgment: is the gap a labeling error, a sensor fault, or a genuine edge case worth keeping? The work is slow, repetitive, and billed by the hour.
The alternative is a render farm that generates the same scene ten thousand times overnight, with perfect labels attached — the product the synthetic-data industry sells. The company on the labeling floor — Haitian RuiSheng (海天瑞声), a Shanghai-listed pure-play AI training-data provider — has spent two years telling investors that render farm is useful but not sufficient.
Key takeaways
- Haitian RuiSheng (688787.SH) reported 2025 revenue of RMB 377 million (≈ US$53.1 million / HK$414.7 million), up 59%, its fastest growth since listing.
- Gross margin fell 18.38 percentage points to 48.08%, the weakest since it listed in 2021, while operating cash flow dropped 93.83% to RMB 1.77 million (≈ US$249,000 / HK$1.95 million).
- Receivables rose 95.33% to RMB 169 million (≈ US$23.8 million / HK$185.9 million), equal to about 45% of annual revenue.
- In a public investor Q&A, chief executive Li Ke said synthetic data mainly serves rare, high-risk corner cases and still cannot match real-scene data for training most models.
- The company's 2026 half-year revenue reached RMB 191.9 million (≈ US$27.0 million / HK$211.1 million), up 22.45%, with net profit of RMB 9.77 million (≈ US$1.38 million / HK$10.75 million).
A business that grows and thins at once
The 2025 annual report reads like two different companies stapled together.
The top line is a growth story: revenue up 59% in a single year, computer-vision data up 217.66% to RMB 148 million (≈ US$20.8 million / HK$162.8 million), natural-language data up 112.29%, and data-application services up more than thirteenfold as digital-human and voice-model projects landed. The customer list has passed 1,200 companies, including Alibaba, Tencent, ByteDance, Microsoft and Amazon. The catalogue holds 1,877 datasets covering more than 300 languages and dialects — assets built once and resold, the closest thing this industry has to a moat.
The bottom line is a warning. Gross margin fell to 48.08%, down 18.38 points, the weakest since the company listed in 2021. Operating cash flow collapsed from RMB 28.7 million (≈ US$4.0 million / HK$31.6 million) to RMB 1.77 million (≈ US$249,000 / HK$1.95 million). Receivables nearly doubled. In plain terms: it booked far more work than it collected cash for, at a worse price.
The pattern looks like a market where supply arrives faster than pricing power. Seven national data-annotation bases have opened in China, all chasing the same large-model contracts.
What the chief executive actually said about synthetic data
The most useful document on Chinese synthetic data is not a vendor white paper but a transcript of an investor Q&A on the Shanghai exchange's roadshow platform.
Asked whether synthetic data would hurt the business, chief executive Li Ke gave an answer vendors rarely give on the record. Synthetic data, he said, is an inevitable product of the industry's development and can serve as an effective auxiliary to data collection — but it has strong limitations, and it degrades the model's grasp of real-world features. In computer vision, its practical use is simulating dangerous or rare corner cases. Because it is machine-generated, its quality and realism still cannot replace real-scene data, so most enterprises continue training on captured data.
Synthetic data is regularly pitched as the answer to data scarcity: cheaper, faster, infinitely scalable. A company whose entire revenue depends on the scarcity of good data has an obvious incentive to talk it down, which is what makes the admission notable: the person with the most to gain from dismissing synthetic data described its limits in technical, not commercial, terms.
Where synthetic data genuinely wins
Both claims can be true. Synthetic data is excellent at three jobs and weak at a fourth.
It wins where capture is dangerous or impossible: crash scenarios, robot falls, rare medical presentations. It wins on labeling accuracy, because the generator already knows where every object is. It wins on privacy, because generated faces and plates belong to nobody.
It loses on the long tail of physical realism. A rendered warehouse reflects light the way the renderer believes warehouses reflect light. Real footage contains the dust, the wrong pallet wrap, the flickering tube, the camera shake — texture that models quietly learn from and synthetic pipelines smooth away. That is why the industry's most common architecture is hybrid: generate the rare case, capture the ordinary one, mix.
Half-year numbers suggest the hybrid thesis is holding. Revenue rose 22.45% in the first six months of 2026, with overseas revenue at RMB 51.1 million (≈ US$7.2 million / HK$56.2 million), about a quarter of the total, and gross margin stabilised near 49%. The second quarter was weaker than the first: revenue up 9.46% year on year but down 1.73% quarter on quarter, and net profit down 60.1%. Growth arrives in lumps tied to individual contracts, not a smooth curve.
The strategic risk for data vendors is not that synthetic pipelines get better, but that they get good enough at the 80% of ordinary cases and leave humans only the expensive 20% — which is precisely the part most vendors have never had to price on its own.
The part the pitch decks skip
Two numbers explain why the synthetic story is easier to sell than to invoice. The company spent RMB 58.6 million (≈ US$8.25 million / HK$64.5 million) on research in 2025, 15.5% of revenue, and still describes its platform work as an efficiency tool for its own production rather than a product customers buy.
The second is the customer. Buyers of training data are the country's largest technology companies, negotiating multi-year contracts with procurement teams and payment terms that show up as receivables. A vendor that sells to giants at 48% gross margin and collects months later is not a software company, whatever its pitch.
Honest limitations
This article draws on Haitian RuiSheng's 2025 annual report and 2026 half-year report filed with the Shanghai Stock Exchange, coverage from Chinese financial media, and a transcript of a public investor Q&A. Li Ke's remarks on synthetic data date from an earlier reporting period and may not reflect current product plans. The company does not disclose what share of revenue comes from synthetic data, so the split between generated and captured data is unknown. Conversions use RMB 7.1 = US$1 and RMB 1 = HK$1.1.
What readers can do now
- Ask any data vendor what percentage of a delivered dataset is synthetic, and whether that percentage is contractual.
- Budget for hybrid pipelines rather than one source; pure synthetic training fails months later in edge cases.
- Track receivables and cash flow, not revenue, when judging data-service companies; the growth may be bought with working capital.
- If you are buying labels for a robot or vehicle program, keep a real-capture baseline so you can measure what the generated data actually added.