Back to articles

Qingmiao and Ascend Target a 10,000-Card Token Factory

3 min read

The measure of AI infrastructure is shifting from how many accelerators are installed to how many useful tokens can be delivered reliably. Qingmiao Intelligent has announced a joint-operation initiative with Huawei Ascend ecosystem partners for an Ascend 950DT Token factory. The company plans to advance a 10,000-card operation in 2027 and says it aims to build a 100,000-card network spanning multiple regions and chip types over the next five years. The proposal addresses not only capacity, but also how scattered resources can be organized into measurable inference services.

A shared operating model

Rather than placing construction and operation in the hands of one company, the plan divides responsibilities across participants. Ascend ecosystem partners are to provide chips, supernode products, and technical coordination. Data centers contribute facilities, power, and computing resources. Model and traffic partners bring real workloads and take part in service acceptance, while Qingmiao is responsible for software optimization, cluster scheduling, unified measurement, and operations. The partners intend to define responsibilities and revenue sharing around delivered tokens and service-level agreements. Announced initial partners include the Beijing Academy of Artificial Intelligence, Shanghai Data Center, and China Telecom’s Hubei branch.

Qingmiao says it has worked with more than 20 computing centers, scheduled nearly 20,000 domestic accelerator cards, optimized support for over 15 chip types, and advanced the deployment of nearly 10 Ascend 910C supernodes. It also says multiple Ascend 950 systems have been installed in data centers in Wuhan and Wuhu. These disclosures describe a claimed operating base; they do not mean the proposed 10,000-card factory is already complete.

Installed cards are not the same as useful output

Domestic accelerators come with different architectures and software stacks, so new models often require compatibility work and tuning. Workloads also vary: request length, concurrency, and cache demand can shift, while agent applications may generate repeated model calls. A system that can run a model is not necessarily able to meet a customer’s requirements for throughput, first-token latency, availability, and cost. Scaling therefore depends on connecting model adaptation, resource scheduling, and service-level validation.

Qingmiao describes its software stack as combining an inference engine with a computing-resource management platform. It says dynamic scheduling, a four-level KV cache, kernel optimization, and distributed parallelism improve utilization. The company reports cache-hit rates of up to 99% in certain scenarios, utilization rising from 30–40% to around 80%, an 84-fold gain for a specific MoE operator, and cluster throughput increasing from 40 million to 140 million tokens per minute. It also reports that P99 time to first token fell from 180 seconds to 40 seconds. These are company-reported results; test conditions and workload selection matter, so they should not be treated as universal benchmarks.

What the plan could—and must—prove

Joint operation offers an alternative to expanding solely through direct hardware purchases and data-center construction. By pooling resources, software and operating expertise can help determine how much service a network can deliver. But the 10,000- and 100,000-card figures remain roadmap targets. The test will be whether the network can meet service-level agreements under sustained workloads, control the cost per million useful tokens, and support different chips and models consistently. For domestic compute to become dependable supply, repeatable delivery—not card counts or peak specifications—will be the more meaningful evidence.

Source: QbitAI

Comments

Checking sign-in status...

Loading comments...

Related articles