China Builds 106 Robot Data Centers, 80% Open Yet Profits Lag

Data collection centers for embodied AI are reaching a turning point in China. According to reporting by Yi Yuqin of Huxiu, 106 centralized sites have been built nationwide. As of August 2026, 84 sites were already in operation, a share of 79.25%. The wave of construction has passed, and the industry has entered a phase where survival through operations is being tested.

One trigger was remarks by Mech-Mind founder Shao Tianlan. He publicly criticized patchwork-style moves within the industry. Data collection centers were also among the targets of that criticism. The change is stark for a field that saw a flurry of contracts and construction starts just a year ago.

Huxiu’s survey covered centralized sites with publicly available information. The scope includes collection centers and training grounds with fixed facilities and robot deployments of a certain scale. The findings were corroborated through interviews with more than 10 companies, including robot makers and operators. This article reviews the background and implications based on the facts of that survey.

Background to the Rapid Rise to 106 Sites in

Two Years

Data collection centers began to increase in 2025. From March 2024 to March 2025, launches averaged about 2.7 per month. From April 2025 to January 2026, the figure jumped to about 7.3 per month. Training grounds and collection bases were concentrated in various regions.

Construction peaked in the first quarter of 2026. New projects have since clearly declined. There is a two- to five-month gap between launch and operation. This is because sites moved into operation sequentially after interior work and equipment installation.

The pace of 106 sites in two years cannot be explained by a single factor. Local governments, companies, and training demand moved at the same time. Administrations seeking a base, robot makers seeking orders, and the training side seeking real-world data converged. As a result, concentrated development progressed in a short period.

By region, Jiangsu Province leads with 18 sites. Zhejiang Province follows with 17 sites and Guangdong Province with 11 sites. Sichuan Province with 8 sites and Beijing with 7 sites are also among the top. In addition to the concentration in the Yangtze River Delta, central and western regions are catching up.

With facilities, robots, and personnel in place, it quickly takes shape as a project

The above was introduced as the words of a person involved in attracting local projects. The shorter timeline compared with attracting a finished-machine factory was valued. Greater certainty than attracting a large model company also played a role. It was a format that even regions without a strong industrial base could readily pursue.

Why Public-Private Joint Construction

Became Dominant

Construction methods fall into three broad types. One is in-house construction by robot makers or data companies. They use the sites directly for their own training and product improvement, allocating some capacity to external collaboration. The second is training bases built by universities and higher vocational colleges. They combine research, education, and collection.

The most numerous type is public-private joint construction. Local governments provide facilities, funding, policies, and industrial resources. Companies provide robots, technology, and operating personnel. A considerable number of large winning bids worth 10 million yuan or more over the past two years fall into this category.

The scale differs from a simple procurement of a few units. Dozens to more than 100 robots are introduced in bulk. Facilities, collection equipment, infrastructure, and operating services are booked together. Contract values tend to expand under this structure.

Companies also had clear motives. For robot makers in the early stage of mass production, a concentrated purchase of dozens of units is significant. It builds a track record of deployments and generates real-world data. Usability for subsequent training was also an advantage.

Local governments used the sites as a tool to attract companies. They offered facilities and orders to encourage registration and the establishment of R&D bases. In some cases, robot makers sought joint construction. It was a framework in which the interests of both sides could easily align.

More Than 5,000 Robots and Data Compatibility

With operations underway, the concentration of robots at training grounds has progressed. The 52 sites that disclosed deployment numbers total about 3,972 units. The range spans from 10 to 300 units per site. Nationwide, the total is estimated to exceed at least 5,000 units.

Small and medium-sized bases form the core by size. Sites with 50 or fewer units account for 55.8%. At the same time, large-scale bases with around 300 units have emerged. The Yibin High-Speed Rail South District Southwest Training Center and others tie for first place.

Repetitive work continues on site. Tasks include grasping and transporting parts, opening and closing doors, and tidying homes. Motions are broken down and repeatedly recorded via teleoperation. Video, trajectories, joint states, and force records are captured simultaneously.

The significance of centralized collection lies in the stable production of real-world data. Language models can use documents from the web. Robots require real data involving trajectories and force feedback. Consolidating equipment, scenarios, operators, and processing workflows makes sense.

The challenge is reuse across different models. 65.96% of sites adopted a single brand. Formats are often incompatible across companies. Even for the same task, changing models requires re-collection.

Reworking a large-scale equipment system is not easy. Collection jigs and scenarios are optimized for specific models. Supporting multiple models later requires additional investment. There is a risk that assets will become locked in.

A Business Model That Does Not Circulate and

a Shortage of Buyers

There are three channels for using the data. In-house use, external sales, and public release. In-house bases feed directly into their own training. Third-party operators seek to monetize by providing data externally.

The buyer structure is weak. Large companies often have their own in-house collection systems. Smaller companies lack sustained purchasing power. With time-based billing, verifying training effectiveness becomes difficult.

Buyers scrutinize effectiveness closely. They will not pay for the number of motions alone. The criterion is whether the data transfers to their own models and tasks. Format mismatches become a barrier to adoption.

Contribution through public release also has limits. It helps raise the baseline for common tasks. It does not directly translate into differentiation for individual products. It is unlikely to cover operating costs.

In the second half of 2026, tightening funding is being discussed. An operator of a large-scale base in South China testified to this. Local government funding will not continue forever. The industry has reached a stage where establishing a business model is required.

Persistently High Fixed Costs and the Road to

Polarization

Costs continue to accrue on a fixed basis even after operation begins. Depreciation of robots is heavy. Costs for teleoperators and scenario construction also continue. If revenue fails to keep up, pressure intensifies.

Some bases retain value. In-house bases backed by training demand fall into this category. Third-party operators with stable customers are also more likely to survive. Data quality and repeat orders are key.

Bases dependent on temporary construction funding face difficulties. Even with higher utilization rates, unit prices do not rise. Re-collection and format conversion push up costs. Operating pressure is expected to increase further.

The industry has entered a phase of shakeout and polarization. It is a shift from quantity of construction to quality of operations. Transfer performance, rather than the number of units deployed, will become the yardstick. Consolidation and partnerships among bases may advance.

Editorial Opinion

Over the next three to six months, selection among the 84 operational sites is expected to advance. As local government funding tightens, operators will be forced to shift from time-based billing to performance-linked models. Robot makers will also be judged not on cumulative deployment numbers alone, but on providing data effective for training.

Over the next one to three years, standardization of formats across models will become the focal point. If collection remains closed within single brands, duplicated assets and the burden of re-collection will persist. Players that establish common formats or conversion layers will determine the survival of third-party bases.

The question is who will measure the value of data, and how. Will metrics based on contributions to model performance gains, rather than counts of motions, become established. The industry as a whole needs a path to avoid a situation where only facilities remain without buyers.

References

• “ 遍地开花的数采中心，开始批量露馅 “, by 易玉琴 — 虎嗅网, 2026-09-19T22:00:11.000Z (ARR)
• Source URL: https://www.huxiu.com/article/4892109.html?f=rss