A robot startup needs a million examples of a wrench slipping off a slope, a bank needs fraud patterns it has never actually seen, and a hospital cannot share its patients' scans. None of them can get the data they need the normal way. So they are starting to make it.
Synthetic data (合成数据) — information generated by AI or simulation rather than collected from the real world — has moved from a research curiosity to a line item in China's industrial plan. The shift matters because the bottleneck for many AI systems in 2026 is no longer the model architecture. It is the training data underneath it.
What synthetic data actually is
The simplest analogy is a test set you invent. Instead of filming 10,000 real forklifts, you build a virtual warehouse where a rendered forklift can fail in 10,000 ways overnight. Instead of waiting for rare fraud to happen, you generate plausible fraudulent transactions and label them perfectly, because you created them. The output is data that looks and behaves like the real thing but carries no real person's privacy and no waiting period.
Two forces push this. Real, high-quality data is concentrated in a few giants and is expensive to license. And privacy rules — China's Personal Information Protection Law and sector regulations in finance and healthcare — make moving raw patient or customer records into a training pipeline a legal minefield. Synthetic data sidesteps both.
A Beijing startup just became the first "embodied-data" unicorn
The clearest signal came in March 2026, when 光轮智能 (Guanglun Intelligence), a physics-AI simulation infrastructure company founded in 2023 and based in Beijing's Haidian district, announced it had closed a 10亿元 (≈ US$139 million / HK$1.09 billion) A++ and A+++ funding round. The Beijing municipal science and technology commission confirmed the round and noted the company had become the world's first "embodied-data" unicorn — a business whose product is not robots but the simulated data and evaluation that trains them.
Guanglun's pitch is a "data factory" (数据工厂): a digital parallel plant where robots run thousands of parallel training sessions across extreme conditions, from −40°C cold to 1,000°C heat, on glass and on rubble. According to the company and Beijing government briefings, this compresses a robot's development cycle from the usual three to six months down to two or three weeks and cuts comprehensive training cost by more than tenfold. Its partners reportedly include ByteDance, Alibaba, Unitree and Nvidia.
The government made it policy
On April 28, 2026, the Ministry of Industry and Information Technology (工信部) and the National Data Administration (国家数据局) jointly issued the notice launching the 2026 "model–data resonance" (模数共振) campaign. The name pairs "model" (模, the AI model) with "data" (数, industrial data): the goal is a flywheel where data feeds models, models produce new data in use, and that data improves the next model.
The campaign targets 20 key manufacturing sectors — steel, petrochemicals, automobiles, aerospace, telecom and more — and sets a deadline of late 2026 to form a working "data–model–application" loop. It explicitly calls for building high-quality industry datasets and nurturing "agent factories" (智能体工厂). For synthetic data, that is about as close to a national endorsement as a technique gets.
Why this is happening now
The arithmetic is brutal. Industry analyses estimate China's synthetic-data market passed US$1 billion in 2025 and is growing near 40–50% a year, though these figures are vendor and analyst estimates rather than official statistics. The driver is the data hunger of large models (大模型) themselves: once a model has consumed most available clean text, image and video, the next improvements depend on manufactured or carefully simulated data.
Autonomous driving was the proving ground. Firms such as 51Sim in Beijing built simulation platforms that generate synthetic driving scenes so cars can practice rare "corner cases" — a child running into a road, a sudden fog — that would be unethical or impossible to stage at scale in real life. Embodied intelligence (具身智能) is the next frontier, which is why Guanglun's funding round drew so much attention.
Another Beijing firm, 数创弧光 (Shuchuang Huguang), backed by the IDEA Research Institute and with investment from Huawei's Hubble arm, takes a "knowledge-driven" approach: it builds a context graph of an industry or language first, then generates data against it, using a model-as-judge to filter low-quality output. The bet is that naive rewriting of real data produces repetitive, hallucination-prone samples, while structured generation produces data worth training on.
What readers can do now
- If you train or fine-tune models, treat synthetic data as a normal procurement category and audit providers on diversity and bias, not just volume — repetitive synthetic data can quietly degrade a model.
- If you work in a regulated industry (finance, healthcare, telecom), pilot synthetic data for testing and model training where privacy rules block real-record sharing; it can unlock development without moving sensitive files.
- If you follow China's AI sector, watch the 模数共振 datasets and the National Data Administration's high-quality-data initiatives — they are becoming the de facto public infrastructure that smaller firms build on.
Honest limitations
This article focuses on China and on the industrial and enterprise side of synthetic data; it does not cover the separate debate over synthetic data in consumer image and video generators, nor the growing concern that heavy synthetic-data use can create "model collapse" if real data is crowded out. Market-size figures come from industry and analyst reports rather than official statistics and should be read as directional. Company performance claims (cycle time, cost reduction) are drawn from corporate and government briefings and were not independently benchmarked here.
Key takeaways
- Synthetic data has shifted from research to industrial policy in China, anchored by the April 2026 模数共振 campaign from MIIT and the National Data Administration.
- Guanglun Intelligence's 10亿元 round and "embodied-data unicorn" status show investors are funding the data layer, not just the robots.
- The real constraint on many 2026 AI systems is training data, and manufactured or simulated data is now a mainstream way to supply it.