China's Computing Centers Sit Half-Empty as Data Stays Locked in Provinces

China has built a vast network of intelligent computing centers, the facilities that train and run AI models. But much of that capacity sits idle. At a law and economics conference held at Hebei University on October 31, researchers presented figures that drew immediate attention: national intelligent computing supply is roughly 8 billion card-hours, while actual usage is only about 1.4 billion card-hours. That works out to a utilization rate below 36.8 percent. Some small and mid-sized computing centers fare even worse, running at under 20 percent.

The numbers point to a problem that is easy to miss when counting new data centers. Data is locked inside provincial borders. Computing capacity is treated as something to seize and hold, with local centers competing for position rather than linking into a network. The result is a patchwork of islands, each with its own supply and its own rules, even as the country promotes a unified national market for data and computing power.

The conference, the 2026 annual meeting of the Economic Law Research Society of the China Law Society, brought scholars together to ask a single question from three directions: data, fiscal policy and finance. How can institutions keep up with the movement of production factors? The discussion framed the answer as a matter of opening locks, not pouring more concrete. A center that cannot draw on data from another province cannot fill its racks, no matter how cheap its electricity.

For global readers, the gap matters because China is spending heavily on AI infrastructure. If nearly two-thirds of the advertised capacity goes unused, the returns on those investments depend less on chip supply than on rules that let data and workloads travel. The scholars did not propose a single fix, but their framing suggests that the binding constraint is administrative rather than technical. Data held by one jurisdiction, or computing reserved for local projects, cannot serve a model that needs scale.

The idle capacity also raises questions about how such projects are planned. Local governments have incentives to build centers and count the capacity as an achievement, while the demand that would make them busy may sit elsewhere. When supply is measured in card-hours and demand is measured in real workloads, the two numbers can drift far apart. The conference figures are a rare public look at that drift.

The discussion is unlikely to end at the conference. China’s push to build a unified national market has already touched areas from energy to transport, and data is one of the hardest cases because it crosses both commercial and administrative boundaries. Whether the next step is a national scheduling system, new rules on cross-province data sharing, or simply better accounting of what is actually used, the pressure to connect the islands is growing.

For anyone watching China’s AI buildout, the lesson is that capacity on paper is not the same as capacity at work. A center that runs below 20 percent utilization is not a strategic reserve so much as a stranded asset. The question now is whether the institutional locks open fast enough to make the hardware worth its cost.