After two years of big AI models sprinting, the AI giants collectively turn around to fix the data foundation

The engine is installed, but the road still isn’t repaired
First half: everyone races inside the exhibition hall
The AI narrative from the past two years has been highly uniform—bigger models, more parameters, higher leaderboards. Open-source and closed-source teams both fight for attention. Companies’ playbooks are also highly similar: buy compute, hire for algorithms, deploy models, and assume that if the model is strong enough, deployment will follow naturally
Reality cuts in—one story from a large state-owned enterprise CIO is a typical case: 17 business systems, 9 data warehouses, and 3 clouds. Data formats vary widely, and many paper reports have not been digitized. In the end, even the most basic application—equipment failure diagnosis—couldn’t be run. The model can’t even read historical maintenance records
The bottleneck isn’t that the model isn’t good—it’s that the data can’t be fed in
Second half: the car has to drive into real city streets
Factories must keep running, hospitals must stay safe, and governments must remain compliant. No matter how strong the model is, if the road for data isn’t fixed, it can only spin in place
Actions across global industrial supply chains are beginning to converge. Running the same big model naked on enterprise data versus using a complete data engineering system shows a cliff-like accuracy gap. The gap isn’t in the model’s mind—it’s in whether you can eat the right ingredients
So a new architecture emerges: on top is models and capabilities, and below are data engineering, permission auditing, and governance strategies. Models and data are no longer just upstream/downstream in a pipeline, but partners that feed each other
China adds yet another layer of difficulty
Manufacturing supply chains are longer, compliance requirements are stricter, there is more unstructured data, systems are more fragmented, and standards are more inconsistent. Bridging the gap from general intelligence to industry intelligence depends precisely on full-chain data infrastructure
This isn’t about installing a brain for AI—it’s about first fixing the nervous system
What is truly scarce?
This round of turning back to repair the data foundation is, in essence, a correction in cognition. The scalable value of AI doesn’t depend on how many points a model scores today. It depends on whether data can be continuously supplied at quality, whether the system can be continuously governed to be trustworthy, and whether engineering can continuously close the loop in real deployments
China doesn’t lack the 101st open-source large model; what it lacks is first washing the “data pot” clean and then cooking up high-quality “soup”
Once this “getting it to work” runs, AI will move from hot news to a tool on the workbench
DYOR Not investment advice