China is connecting its AI infrastructure. Who will control the demand?

A data centre can be powered, connected and full of expensive equipment without being a good business. Somewhere between a developer's request and the machines that process it, someone must decide which model to run, where to run it, what it costs and who gets paid. That decision point is becoming an important part of China's AI infrastructure story.

Shanghai offers a glimpse of the missing connection

On 9 October, Shanghai's Putuo district launched the Haina Intelligent Computing (海纳智算) Token platform. The district describes a common interface that brings together resources from Alibaba Cloud, Tencent Cloud, Shanghai's computing infrastructure and providers along the Shanghai–Nanjing corridor. Businesses can select models and services and pay according to tokens consumed.

The starting problem is ordinary rather than futuristic. Putuo has more than 1,700 digital advertising businesses, many too small to negotiate attractive terms or build their own AI infrastructure. Pooling their demand could improve access and purchasing power. The district also described cooperation with regional providers, including a Yangzhou computing business, and proposed subsidies for qualifying computing use. Those subsidies are not yet an enacted entitlement.

This is not proof that distant clusters already operate as a single, seamless computer. It is a more practical intermediate step: a commercial doorway through which businesses can discover and buy services supplied by different providers.

The contest is shifting toward the customer interface

There are several ways to control that doorway. A large cloud can offer its own models alongside others. An independent cloud can compete on model choice, compatibility and inference costs. Or a marketplace can aggregate several clouds and own the relationship with the end customer.

Alibaba illustrates the first model. On 10 October, it scheduled the retirement of specific older DeepSeek and GLM endpoints on Model Studio, recommending Qwen alternatives for some. That is a real example of a cloud provider influencing developers' choices. It is not evidence that Alibaba has expelled every DeepSeek, Kimi, GLM or MiniMax model: its catalogue still lists newer third-party versions. Some Qwen versions also undergo routine retirement. The policy shows the leverage of the platform, not a proven wholesale exclusion strategy.

Kingsoft Cloud offers a different route. Its StarFlow documentation describes access to several model families, including DeepSeek, Kimi, GLM, MiniMax and Xiaomi MiMo, through developer APIs with usage-based pricing. That gives developers a way to use several model families through one provider. They do not establish that developers are migrating from Alibaba, that KC has a commercial contract with every model maker, or that token revenue is profitable.

Shanghai's Haina platform adds a third possibility: an intermediary that aggregates demand across providers. For KC, that could be a route to new customers—or a risk if the intermediary controls prices, model selection and the customer relationship while cloud suppliers compete for lower-margin execution.

National integration has a technical side—and an economic side

China's national computing initiatives are pursuing resource visibility, common interfaces and scheduling across different types of hardware. Projects including CNCOS and OGPU indicate work on cross-cluster orchestration and heterogeneous accelerator compatibility. Huawei's previously disclosed work with China Unicom shows another route from domestic accelerators through cloud services to applications. These are important technical building blocks, but they are not measured evidence of high-volume, profitable cross-region workload migration.

The Shanghai experiment helps distinguish the two questions. Can the technology connect resources? And can a customer buy useful compute at a competitive price, with the provider earning a return? A catalogue of available GPUs answers neither question on its own.

Beijing is also asking whether the investment is productive

Guidance published by China's central authorities on 9 October reinforces the national integrated computing network and East Data–West Compute, while warning against speculative bubbles, blind investment and poorly coordinated expansion. Earlier green-computing standards similarly put greater emphasis on operating conditions and utilisation.

That matters because legacy data-centre oversupply can coexist with shortages of suitable AI-ready capacity. More buildings, more reserved megawatts and more announced chips do not necessarily mean more useful tokens or stronger cash flow. Proposed financing—including Tencent's 12 October note issuance—should likewise be separated from completed funding and actual AI project spending.

What does this mean for VNET, GDS and Kingsoft Cloud?

VNET and GDS may benefit if cloud and marketplace demand turns into more contracted, delivered AI-ready capacity. But none of these developments identifies the customers behind unnamed VNET megawatts or proves higher GDS utilisation. Their economics still depend on delivery, occupancy, pricing and capital efficiency.

Kingsoft Cloud is closer to the point where model choice becomes a product. Its multi-model access is a real capability and a plausible competitive advantage. Whether that translates into paid external demand—or is weakened by independent marketplaces—remains open.

The larger implication is not that China's infrastructure is already unified. It is that the contest to make compute usable, purchasable and profitable is moving into view. The winners may not be the organisations that build the most capacity, but those that can connect real demand to the right resources without destroying the economics along the way.