2026 Global Data Merchant Conference: High-value data is scarce—where does the bottleneck lie in data being "used well"?

With the explosive growth of artificial intelligence technology and the surge in daily Token calls, how data can move from being "supplied" to being better "used" has become a topic of industry attention.

"Of the total data volume of 52.26ZB, only about 3% is truly effectively utilized." Su Guangniu, Chief Technology Officer of Shanghai Data Group, shared this set of figures at an industry roundtable dialogue on "Data, Chain, and Intelligence Collaboration" at the 2026 Global Data Merchants Conference, which opened on October 30.

Still lacking high-quality data

"Supplying data is no longer a problem, nor is getting the process to work, but high-value data that supports the industry is still in short supply," Su Guangniu said, noting that the bottleneck in data circulation is shifting from the channel to "whether it is good or not."

In the embodied intelligence sector, the data shortage is particularly acute. Liu Hua, Vice President of Tashi Intelligence, said that the generalization of embodied intelligence faces major obstacles, with the core lying in data. To train a foundation model with strong generalization, the data threshold is "tens of millions of hours," while the industry's high-quality training data currently stands at only the millions-of-hours level, "a gap of about 90%."

This is especially true in terms of collection costs and modality defects. Liu Hua shared that a traditional teleoperation setup for real robots costs 200,000 to 300,000 yuan per set, while the robot itself costs hundreds of thousands of yuan, and a day of debugging and testing often yields only a few hours of qualified data. The deeper flaw lies in the distortion of data modalities: "Many foundation models today rely mainly on text and video for training. They appear to understand instructions and see the world, but they do not understand the real meaning of 'force' and physical actions. When robots perform actions, they lack a sense of touch and are extremely prone to failure."

In Liu Hua's view, to fill in key modalities such as touch and posture, and to cross the tens-of-millions-of-hours threshold through self-collection, robot data feedback, and external procurement, the industry will need at least two to three more years before reaching its "ChatGPT moment."

For telecom operators, the development of agents is also forcing a rebuild of underlying infrastructure. Huang Yihua, Vice President of ZTE Corporation and General Manager of its Corporate Development Department, said the industry is moving from the Internet of People to the Internet of Intelligence, and that by 2030 the global scale of agents is expected to exceed 2.2 billion, a more than 80-fold increase. Along with this comes a complete shift from "bits" to "Tokens," and computing infrastructure is shifting from competing on single-card metrics to testing full-stack engineering with multi-chip collaboration and integrated software and hardware.

In addition, as large models accelerate their evolution toward AI Agents, underlying computing resources are also encountering structural problems.

Ying Zhiwei, Vice President of Hygon Information Technology Co., Ltd., gave an example on site: last year, when running large-model tasks such as text and code generation, computing power was basically shouldered by GPUs; but after Agents were deployed this year, engineers found that after large models generate code, there is a large amount of compilation, running, and testing, and these processes consume a great deal of CPU computing power.

"At this point, GPUs instead sit idle. The surge in programmers' parallel tasks further accelerates CPU consumption." Ying Zhiwei revealed that to avoid idle computing power and bottlenecks, the industry's CPU-to-GPU configuration ratio is being rapidly pushed from the previous 1:1 to 1:4 or even 1:8.

Beyond computing power reorganization, a bigger problem is the security concern of "running naked." By the end of 2025, the country's intelligent computing power scale will reach 1,590 eFlops, but Su Guangniu believes that after technological engineering, the bottleneck has shifted comprehensively to "institutional trust."

When enterprises input Prompts and core materials into cloud-based large models, they are essentially handing over sensitive data in plaintext in its entirety. In response to this problem, Huo Jia, Vice President of Alibaba Cloud Intelligence Group, proposed that trust must be separated into "trust in data elements" and "trust in data applications," and that the industry must currently solve the problem of full-link encryption of chain-of-thought data and Tokens during model reasoning.

Farewell to "Demo shows"

Some guests present believed that the root cause of the difficulty in closing the business model loop for data applications lies in the industry being flooded with many "pseudo-scenarios," and that it should bid farewell to "Demo shows."

Speaking about the implementation of "AI+" at the roundtable, Huo Jia said: "I see that many enterprises' AI+ applications are basically demo-type applications, with production-grade ones nowhere to be seen; many applications are just toys made for a few people and cannot solve the problem of broadly improving productivity."

He believes that data scenarios capable of achieving commercial value must meet strict hard indicators: they must be "24x7 scenarios" that can leverage machines' advantage of never sleeping, must be deeply integrated with cutting-edge foundation models, and must form a closed data flywheel loop of "application to effect feedback."

Citing the atmospheric and vacuum distillation unit of Lanzhou Petrochemical as an example, Huo Jia analyzed that because shutdown costs are extremely high, the unit previously relied on manual 24-hour monitoring of the dashboard; only after building an agent to take over real-time adjustments and run continuously and stably for five months did AI truly enter the core production system for the first time.

Not only do scenarios need to be stripped of bubbles, but technical approaches are also undergoing reassessment. Regarding Su Guangniu's idea of "using blockchain and smart contracts to solve data authorization and revenue distribution," Li Han believes that trusted data spaces are being built everywhere, but problems remain such as fragmented standards and a lack of legal safeguards.

Li Han, General Manager of the Cross-Border Data and AI Application Research Center at China Unicom (Hong Kong) Innovation Research Institute, said that talking about how much raw data is worth in isolation is empty talk; data must be broken down into different chains such as production, governance, productization, and operations. Data only has production value after being integrated with technology and productized, and benefit distribution can only be divided according to the actual value created by each party across the full chain.

Liu Hua also believes that data pricing must ultimately be decided by market mechanisms. When the capabilities of embodied intelligence models bring huge returns due to leaps in high-quality data, the market will naturally offer reasonable consideration for data, and on-chain tracing should serve the refined confirmation of rights for workers, scenario parties, and collection parties.