Data output up tenfold, yet only 11.65% of enterprises are willing to pay for data | ToB Industry Watch
The National Data Resource Survey Report (2025) shows that in 2025, China's total data production reached 52.26ZB, up 27.28% year-on-year, but only 11.65% of sampled enterprises had ever purchased data. Yu Ying, deputy director of the National Data Administration, disclosed that the data infrastructure has connected more than 40,000 entities and listed more than 60,000 products, up 10 times and 4 times respectively from the end of 2025. Zhou Yi, head of the Wanlian Yida Data Center, believes that the essence of the difficulty in implementing trusted data spaces lies in data and scenario issues, and that enterprises should first implement scenarios and then accumulate data, using scenarios to work backward to the technical architecture.
Body · AI translation
At the end of 2024, the "Trusted Data Space Development Action Plan (2024–2028)" was released, which stated that by 2028, more than 100 trusted data spaces would be built, forming a widely interconnected data circulation network. Subsequently, the first batch of 63 pilot projects was announced, with 13 cities, 22 industries, and 28 enterprises selected. IDC data shows that in 2025, the narrow-sense market size of trusted data spaces reached 3.04 billion yuan, with city data spaces contributing more than half of the revenue. Everything appears to be progressing according to plan.
But looking at another set of figures together brings one back down to earth. The 63 pilot projects penetrated more than 900 application scenarios and attracted the participation of nearly 70,000 market entities. The participation of 70,000 entities corresponds to 304 million yuan in narrow-sense product and service revenue. This means the vast majority of participants are using the public service capabilities or low-cost access solutions provided by the pilots, rather than paying for a mature commercial system.
During the 2026 Two Sessions, representatives and committee members from multiple fields expressed support for and affirmation of the construction of trusted data spaces, but the related concerns were more concentrated on data compliance requirements, enterprise data openness standards, and progress in commercial implementation.
Faced with these questions, Zhou Yida, head of the data center at Wanlian Yida, told us that the essence is not a technical problem, but a problem of data and scenarios.
Two Sides of Development
When asked where the engineering difficulty lies in deriving the technical architecture backward from scenarios and data assets, Zhou Yida did not first talk about privacy computing, nor connectors or task scheduling, but instead recounted two pieces of his own work experience.
The first was at an internet company. There was already a lot of data there, and the scenarios were ready-made: enterprises on the platform could directly call data for marketing, user operations, user insights, and product optimization, and going one step further, consumer-side demand could in turn drive production scheduling. In financial business, even more possibilities could grow. With both ends in place, the business naturally ran.
The second was at a central state-owned enterprise. This enterprise mainly built data spaces for others and had no business scenarios of its own. Once the project was completed, accepted, and delivered, the team withdrew. There was no operations, and it was unclear what customers were actually struggling with every day. The company had always wanted to shift toward the scenario side, and the shift was painful, because people who leave right after delivery find it hard to answer the question of whom this system later helped save what.
Putting the two experiences side by side, the difference is not in technical capability, but in whether anyone continues to use it after delivery. This difference happens to be the most suitable ruler for observing trusted data spaces over the past two years.
The "Trusted Data Space Development Action Plan" issued by the National Data Administration in November 2024 set a goal: to build more than 100 trusted data spaces by 2028. What such facilities need to handle is how several mutually non-subordinate entities can put data to use under common rules while each still keeps control of its own things. In July 2025, the first batch of 63 pilot projects was announced, with 13 in the city direction, 22 in the industry direction, and 28 in the enterprise direction, including energy, logistics, shipping, new materials, and medical insurance. By the press conference for the Big Data Expo on July 28, 2026, Yu Ying, deputy director of the National Data Administration, reported a set of operating figures: data infrastructure had connected more than 40,000 entities and listed more than 60,000 products, 10 times and 4 times more than at the end of 2025, respectively.
Judging from these figures, the development of trusted data spaces seems fairly good, but the reality is that enterprises with data are either unwilling to share it or, even if willing, do not know how to do so, while the demand side is even less clear on how to obtain data useful to itself. Data from the "National Data Resource Survey Report (2025)" shows that in 2025, the country's total data production reached 52.26 ZB, up 27.28% year on year, accounting for about 27.44% of the global total, with about 90% of the increase in data production coming from enterprises. The scale is still rising. Yet in the same report, the proportion of sample enterprises that had purchased data was 11.65%, with purchase costs up 22.36% year on year; the average amount spent on purchasing data by leading platform enterprises and central enterprises was 60 times that of other enterprises, and the proportion of enterprises in finance and the software and information technology services industries that had purchased data exceeded 30%, clearly above the average line.
In other words, those willing to spend money on external data are mainly the enterprises that already have the ability to process data themselves. Facilities and supply have grown tenfold, but the shape of the demand side has not changed much. Infrastructure calculates whether connection is possible; enterprises calculate whether it is worth it. The two accounts are not kept in the same ledger.
On this point, Zhou Yida's judgment is that not every enterprise needs to build its own data space; for most enterprises, it is enough to connect to an existing collaborative network as a participant. "Wanlian Yida provides connectors and lightweight access methods, so that enterprises can first connect in and then gradually link to the national infrastructure."
Only with Scenarios Is There Data
Starting from product logic, Zhou Yida believes that enterprises should first implement scenarios, then accumulate data, and then derive the technical architecture backward from the scenarios; this construction path will produce better results.
Zhou Yi said he has seen too many attempts that first build a foundation and then look for scenarios end up in deadlock: the product is finished, but no one knows who to sell it to; the data is connected, but no one knows what it can be used for. "Scenario-driven: first have the scenario, then have the data, then derive the technical architecture. The direct benefit of doing this is that we won't blindly develop a product and then look for a scenario; instead, the scenario reverse-drives technology R&D, which then matches the data. This is our core difference," Zhou Yi said. This approach also answers another frequently asked question: how good is the cross-scenario reuse capability of a trusted data space? In vertical scenarios, data capability requirements differ; won't that lead to poor reusability?
In addition, in Zhou Yi's view, industry knowledge can be shared within an industry, and the multiple types of data spaces currently being promoted by the state themselves show that a data space is not a single "one-size-fits-all" form. "A lot of industry data does not fall within the category of big data. For industry applications, what is truly valuable is often not relying solely on large models, but using small models, or combining large models with high-value professional knowledge within the industry," Zhou Yi emphasized.
This judgment also explains why Wanlian Yida dares to put scenarios ahead of technology. Zhou Yi said the company focuses on production-oriented enterprises and production-oriented service enterprises, and is gradually covering 97 major industry categories of the national economy; its four businesses—commodity trading, logistics operations, fintech, and enterprise services—are already running. Data is not first created and then sold to buyers; it is something that already exists in the business. This number shows not how big the market already is, but that it has real scenarios available to use.
Engineering: From Access to Ontology
If scenarios determine the direction of a trusted data space, then engineering implementation determines how fast it can go.
In terms of engineering, the first unavoidable challenge is access. Under the national data infrastructure system, access connectors are divided into three types: full-function, standard, and lightweight. On this, Zhou Yi introduced that Wanlian Yida's lightweight access connector has become very lightweight; enterprise users can even install it directly on a laptop and log in as needed to access the data space network. This year it also launched an APP version of the access connector, making it convenient for individual users, businesspeople, and enterprise users to access at any time.
If a user is only a data user, it is enough to occasionally open a computer to search for and find suitable data products; but if the user is a data provider with high requirements for service SLA and continuous availability, then a standard or even full-function access connector is needed for support. "C-end and B-end users are sometimes not easy to strictly distinguish. A person may serve an enterprise and may also use data services as an individual. Through different versions of access connectors, we can adapt to the differentiated needs of different roles for performance, stability, and convenience," Zhou Yi explained.
Having solved the access challenge, enterprises still need to face the choice of how to balance security and performance. Privacy computing and federated modeling, which were once hot a few years ago, are essentially encrypted computing: data is encrypted throughout and then used in computation. Performance does indeed suffer a noticeable loss compared with plaintext computing—early on it might have been only a few tenths of plaintext computing, and it has improved now, but the gap still exists. Yet in the financial industry, this is precisely the field where privacy computing is most widely used. On this, Zhou Yi explained, "First, because financial industry data is highly sensitive, whether for individual risk control or enterprise risk control, raw data cannot leave the domain; second, the fault tolerance rate is extremely low, and it is better to sacrifice some performance to ensure compliance and security. In financial scenarios, security and compliance take priority over performance; this is a trade-off determined by the characteristics of the industry."
Frankly speaking, this kind of trade-off between performance and security is hard to avoid entirely. But enterprises are all continuously looking for solutions, and related technologies are also constantly iterating. At present, what is relatively favored in the industry is sandbox technology, including data sandboxes and the more cutting-edge agent sandboxes. Zhou Yi took Meta's Muse agent launched this year as an example: it adopts a cloud sandbox architecture and allocates an independent isolated operating environment to each user, ensuring data security on the one hand and isolating dangerous behavior of the agent on the other. "This reflects the direction of current technological evolution—from traditional data sandboxes, it is gradually iterating toward agent sandboxes," Zhou Yi said.
In addition, in the AI era, ontology modeling itself is also an engineering challenge. Facing this, Zhou Yi shared Wanlian Yida's approach with us. He said Wanlian Yida has done three things: first, modularization, separating the core ontology from business extension modules and the data layer from the assertion layer, so that when new categories are connected in the future, the core skeleton does not need to be touched; second, layered reuse, where main tables in the data middle platform, for example, can be directly mapped to entities and attributes of the ontology, while new data sources follow a model of standardized adaptation and customized development; third, versioning, managing governance issues through mechanisms such as version snapshots, difference comparison, and rollback, with rule configuration independent of code, so that future rule changes do not require a release.
Take the capacity matching scenario as an example. It uses five major categories of core ontology to build the skeleton—drivers, goods, vehicles, roads, and orders—and then combines more than 40 rules to support vehicle-cargo matching. Concepts such as "arrival" have vastly different meanings in different business scenarios: in land transport it is called arriving at the station, in water transport it is called arriving at the port, and in some places arriving at the port does not count—unloading and customer acceptance are also required for it to truly count as arrival. "This is a very vague, scenario-dependent concept. We unify this kind of semantics through the ontology," Zhou Yi pointed out.
Beyond the ontology itself, the question asked more often is about reuse: data capability requirements differ across vertical scenarios, so will reusability be poor? Zhou Yi's judgment is that industry knowledge can inherently be shared within an industry, and the various types of data spaces currently being promoted by the state itself show that data spaces are not a "one-size-fits-all" single form. "A lot of industry data does not fall within the category of big data. For industry applications, what is truly valuable is often not relying solely on large models, but using small models, or combining large models with high-value professional knowledge within the industry."
From this perspective, the key to reuse is not copying one set of platforms to all industries, but separating the reusable skeleton from the non-reusable industry knowledge: the skeleton is accumulated through ontology and modularization, while industry knowledge stays within the industry to grow.
How far away is scale-up?
The infrastructure framework has already been built, so why has the trusted data space not yet been truly put to use? In Zhou Yi's view, there are currently two main challenges.
The first is the issue of data ownership. The state is advancing the "separation of three rights" for data, namely the separation of the right to hold data resources, the right to process and use data, and the right to operate data products. This is the property rights institutional framework creatively proposed in the "Twenty Articles on Data" in December 2022, aimed at solving the difficult problem of data rights confirmation. Through data spaces, the separation of the three rights can be realized: data ownership remains with the original holder, but users can use the data without taking away the raw data, and after processing form new data products and operate them.
The direction of the institutional design is clear, and implementation is also accelerating. In September this year, the national data property rights registration service system went online for trial operation, and three data exchanges became the first batch of registration institutions, promoting one-time registration and nationwide universality; the "Guidelines for Data Property Rights Registration Work (Trial)" unified the registration objects, types, processes, review standards, registration effects, and registration system. But Zhou Yi's observation is much more sober. He told us, "In reality, everyone's understanding of the separation of the three rights is still not clear enough, and which data should be opened up and who should open it for use remains insufficiently clear."
The other is the issue of combining data with scenarios. At present, the national level is also promoting solutions to this problem; for example, the newly established National Data Group hopes to break through these problems at the institutional level. But beyond institutions, the threshold on the enterprise side cannot be ignored either.
"The threshold for using data is still relatively high now," Zhou Yi further pointed out. "Looking at it in conjunction with AI, users first need to understand the business scenario, and second need to understand AI, in order to truly use it; at the same time, they also need to understand data, know what data is needed, where to find it, and how to combine and trade it compliantly."
In Zhou Yi's view, for enterprises to truly achieve scale-up, the biggest cost is not technology investment. "The biggest cost is change at the business level, especially the adjustment of organizational structure. We often say that digitalization is a top-leader project, and that is exactly why," Zhou Yi pointed out. He mentioned an enterprise he had once served, which had thousands of offline stores under it. Its chairman said something that left a deep impression on him: in production, sales, marketing, design, and inventory, the data reported by each department was 80 points, but when it reached him, the business as a whole was failing. Why? Because the data was not connected, and the way the business ultimately operated was still fragmented.
Conversely, the way the trusted data space looks when it is put to use is often not grand.
The first is driver onboarding verification. Every day many drivers register on the Wanlian Yida platform, and the platform must first verify the drivers' identities—ID cards, driver's licenses, operating permits—and only compliant ones can take orders. The past problems were that license photos were unclear, expired ones could not be identified, and some even faked them through photo editing. Their approach is: first use their own precise small model to read the license data, and then call authoritative data sources through the data space for verification. Before the transformation, manual verification took about seven or eight minutes; now it has been shortened to about 5 seconds.
The second is freight rate estimation. From city A to city B, the first step is to accurately calculate the transport distance. Truck routes differ from ordinary passenger cars and are also affected by various situations such as road closures. If a quote is 4,000 yuan at the beginning, but the actual trip takes three days to deliver because of traffic jams, traffic controls, landslides, and other reasons, and the price rises to 7,000 yuan, a large number of disputed orders will arise. They introduced multi-dimensional data—weather data, road network data, traffic control data, and multimodal transport data—to narrow the error as much as possible and make transport distance and freight rates more accurate.
Driver verification went from seven or eight minutes to 5 seconds, and freight rate estimation has fewer disputed orders... None of these changes are sexy, but this is what the trusted data space looks like when it is put to use. On the infrastructure's ledger, the numbers of connected entities and listed products are still rising; on enterprises' own ledgers, what is still being calculated is how much money this thing actually saves. Only when the two ledgers can be reconciled will the trusted data space truly complete the first kilometer.