Give Ten-Thousand-Card Domestic Chips a "Super Plug-in"! Tsinghua Team Aims to Build the Strongest Domestic Token Factory

Partnering with Huawei Ascend, sprinting toward one hundred thousand cards in five years
Jessica, from Feisi Studio, Quantum位 | WeChat official account QbitAI
Quantum位 | WeChat official account QbitAI
In two and a half years, a 5,000-fold increase.
This is the growth scale of China's daily average Token call volume. In early 2024, this call volume was still 100 billion; by June 2026, it had reached approximately 500 trillion.
Agent applications are penetrating every industry, model capabilities are constantly breaking through, and AI's demand for computing power is expanding at an unprecedented rate.

As demand surges, the supply side faces a different picture.
Due to the well-known external environment, uncertainty in the supply of non-domestic advanced computing power persists, and domestic chips are rapidly moving to center stage. Market research firm TrendForce predicts that by 2026, domestic solutions will account for nearly 90% of shipments in China's high-end AI chip market.
There is no doubt that domestic chips and domestic models are transitioning from an important technical choice to a foundational capability that China's AI industry must master.
But this main line is still full of obstacles. Chips from different vendors correspond to different software stacks, and model adaptation and developer ecosystems are still being improved; after equipment is racked, it is difficult to match real traffic, and computing power utilization is insufficient...
The deeper contradiction lies in the fact that the supply side measures resources by number of cards, racks, and peak computing power, while the demand side purchases by effective Token output, response speed, and service quality.
The two ends of supply and demand are speaking two different languages.
An AI infrastructure company with Tsinghua roots—清昴智能—is trying to become the "translator" in between.
On November 17, 清昴, together with the Huawei Ascend ecosystem and multiple industry partners, officially launched the Ascend 950DT ten-thousand-card-level Token Factory joint operation plan.

According to the plan, 清昴 aims to build the largest independent domestic Token factory in China, organizing dispersed and heterogeneous domestic computing power and, through RSI-driven software technology and an operations system, converting it into measurable, acceptance-testable, and stably delivered Token production capacity.
Over the next five years, 清昴 will further build a one-hundred-thousand-card-level Token factory network covering multiple regions and multiple chips, so that domestic chips truly become the productivity supporting the continuous operation of large models and Agents.
Behind the ten-thousand-card plan, the domestic AI industry is incubating a new division of labor.
Turning Dispersed Domestic Computing Power into a Ten-Thousand-Card-Level Token Factory
Building a ten-thousand-card-level computing cluster solves the problem of the scale of computing resources. But to turn this computing power into a factory that continuously produces and stably delivers Tokens, a more complete technology and operations system is still needed.
In addition to chips, data centers, and electricity, Token production also involves model adaptation, resource scheduling, and real business demand. These capabilities and resources are usually dispersed among different companies in the industry chain.
The special thing about 清昴's ten-thousand-card plan this time lies in the four words "joint operation."
According to the latest published roadmap, 清昴 plans to advance Ascend 950DT ten-thousand-card-level joint operation in 2027 and build a leading independent domestic Token factory in China.
They also plan, over the next five years, to expand this system into a one-hundred-thousand-card-level independent and controllable Token factory network, achieving unified scheduling across multiple regions and multiple chips and matching supply and demand of fully domestic computing power nationwide.

Facing such a massive scale of computing power, 清昴 chooses to let different companies leverage their respective strengths and jointly participate in Token production.
Specifically, this plan will be jointly built and operated by four parties:
- The Huawei Ascend ecosystem provides chips, super-node products, and technical collaboration;
- IDCs and intelligent computing centers provide data centers, electricity, and computing power resources;
- Model and traffic parties provide real business demand and participate in service quality acceptance;
- 清昴智能, as the ecosystem organizer, undertakes the intermediate work of software optimization, cluster scheduling, unified metering, and joint operation, enabling unified scheduling of dispersed computing power and then delivering the final output Tokens to customers.
The four parties establish a cooperation mechanism around real Token output and, through service level agreements (SLAs) and a unified metering system, clarify their respective inputs, operational responsibilities, and revenue sharing.
Chips, data centers, and model demand come from different companies, but what is ultimately delivered is a Token service that meets unified standards. The cooperation among the parties also extends from the supply of equipment and resources further into production and operations.
The first batch of operational partners announced by 清昴 includes the Beijing Academy of Artificial Intelligence, Shanghai Data Port, China Telecom Hubei Branch, Hubei Big Data Group, and Beijing International Computing Power Service Co., Ltd., among others.

What supports this ten-thousand-card plan is the domestic computing power operations foundation that 清昴 has accumulated previously.
To date, 清昴 has jointly carried out operations with more than 20 computing centers, accumulated scheduling experience with nearly 20,000 domestic cards, completed adaptation and optimization for more than 15 domestic chips, and advanced the deployment of nearly 10 sets of Ascend 910C super-nodes.

The new generation of hardware has also made deployment progress. As one of Huawei's first batch of 950DT super-node partners, 清昴 disclosed that multiple sets of Ascend 950 computing power have already been deployed in data centers in Wuhan, Wuhu, and other places, and the related full-stack adaptation and operations capabilities are under construction.
These cross-chip and cross-center practical experiences also give 清昴 the opportunity, through the Ascend 950DT platform, to extend Token production capacity to a larger computing power network.
A commercial path different from traditional computing center construction is thus gradually becoming clear.
The expansion of traditional intelligent computing centers is often accompanied by large-scale chip procurement, data center construction, and supporting investment. How much capital an enterprise invests largely determines how much computing resources it can control.
清昴, by contrast, participates in Token production through joint operation, so that the expansion of computing services no longer depends entirely on the number of self-owned devices, and the importance of technical efficiency, resource organization, and customer delivery capabilities increases accordingly.
However, ten-thousand-card scale is only the foundation. With the same ten thousand cards, different model adaptation efficiency and different levels of resource scheduling often lead to vastly different final Token output and quality.
Since the gap truly exists, how should a Token factory establish a comparable standard?
关超宇, founder and CEO of 清昴智能, proposed the core acceptance criteria for measuring the strength of a Token factory: more, faster, better, cheaper.
- More measures the richness of supported models and effective Token throughput;
- Faster focuses on first-Token latency, output speed, and new model adaptation efficiency;
- Better requires consistent model accuracy, stable service, and satisfaction of customer SLAs;
- Cheaper lands on the comprehensive cost per million effective Tokens and output per unit of energy consumption.

These four standards ultimately point to the same goal—under a given service quality, enabling limited computing resources to continuously produce more high-quality Tokens.
And at a time when domestic computing power is expanding rapidly, for a ten-thousand-card factory to truly operate efficiently, there are in fact still a large number of engineering challenges standing in the way.
Having Cards Does Not Equal Having High-Quality Tokens
In the development of domestic computing power today, the first thing encountered is the time gap between models and hardware.
Models advance by leaps and bounds, but hardware has its own growth cycle. A chip may take about three years from planning and R&D to release. But in just over a year, model scale, context length, inference methods, and customer requirements for response speed may undergo tremendous changes.
What is more, domestic computing power is showing a diversified development pattern. Different chips have different architectures and software stacks, and putting a new model into use often still requires compatibility work, operator tuning, and performance verification.
林咏华, vice president and chief engineer of the Beijing Academy of Artificial Intelligence, believes that China already has many excellent chip teams, and what the industry more needs to improve is a software system that allows different hardware to share technological achievements and develop collaboratively.

Therefore, as chip choices become more abundant, software adaptation between models and hardware becomes the first barrier that must be crossed.
The second barrier lies in the actual operation of data centers—equipment rack rate does not equal computing power utilization.
In real business, long and short requests are mostly mixed, concurrency fluctuates constantly, and Agents also repeatedly call models in complex tasks, generating large numbers of long-context and caching requirements.
Fixed model deployment and scheduling schemes find it difficult to always match such load changes. Even after chips are deployed, continuous optimization of parallel strategies, cache management, and resource allocation is still needed to fully unleash computing capability.
Therefore, there is also a temporal mismatch between computing power construction and business demand.
The final barrier to the expansion of domestic computing power lies in actual delivery.
Chip vendors focus on computing performance, and computing centers focus on number of cards and resource utilization; customers who actually purchase model services care more about Token throughput, response latency, service stability, and actual cost.
This is especially true for Agent applications, where a single task may involve multiple rounds of inference and tool calls, and a failure in any intermediate step can affect the entire process.
Tokens that appear cheap, if they frequently cause task failures or repeated calls, may ultimately result in higher business costs.

At this point, all three barriers ultimately point to the production capacity of the Token factory.
Model adaptation determines whether chips can run efficiently, resource scheduling affects how much performance existing computing power can release, and service quality determines whether Tokens can truly enter enterprise production.
If any one of these links has a weakness, it will ultimately be reflected in Token output, cost, and delivery quality.
Article title: A "Super Plug-in" for Ten-Thousand-Card Domestic Chips! Tsinghua Team Aims to Build the Strongest Domestic Token Factory
Main text: A mature Token factory needs to organize these originally scattered engineering capabilities, so that the computing power generated by each chip ultimately becomes stable, predictable Token output.
For Qingmao, this is also the real challenge of building a ten-thousand-card factory.
To Keep Ten-Thousand-Card Systems Continuously Growing Stronger, Qingmao Bets on RSI
Token services are becoming a new battleground in AI infrastructure competition.
Large cloud vendors have already incorporated Token services into their AI infrastructure layouts. They possess enormous computing resources, customer networks, and mature service systems, and some companies have also organized self-developed chips, foundation models, and application traffic within the same ecosystem.
On the other side, many independent companies are also expanding into the market around high-performance inference, heterogeneous computing scheduling, and model services, competing through underlying efficiency and cross-platform capabilities.
Although they all provide Token services, the resources and technical paths each company relies on are not the same.
What Qingmao has chosen is a path closer to the Token production stage.
To establish itself in this competition, Qingmao must rely on software optimization and specialized operations to enable domestic computing power scattered across different companies to create higher actual output.
This path is also consistent with Qingmao's entrepreneurial history.
In October 2022, Guan Chaoyu and his undergraduate partner Yao Hang co-founded Qingmao Intelligence.

Guan Chaoyu studied automated machine learning at Tsinghua University and once led the development of the open-source automated graph learning project AutoGL; Yao Hang was formerly a senior engineer at Huawei, where he led IC hardware R&D and has long focused on domestic chip adaptation.
Most members of Qingmao's founding team also come from the Department of Computer Science and Technology at Tsinghua University, with technical accumulation in directions such as automated machine learning, model compression, and edge computing.
Initially, Qingmao was committed to improving the operating efficiency of AI models on different chips. After the accelerated development of domestic chips and large models, this research gradually expanded to larger-scale computing clusters. In 2024, Hubble Investment under Huawei took a stake in Qingmao, and the company's business further extended toward domestic computing power optimization.
However, as cluster scale expanded, overall efficiency and stable delivery became new challenges. Qingmao's technical system expanded accordingly, gradually forming two core software products.
The MLGuider inference engine is responsible for optimizing the computational efficiency of models on underlying hardware; Xuanwu Intelligent Computing Cloud incorporates different chips and computing centers into unified management, undertaking tasks such as resource scheduling, operation monitoring, and Token metering.
After real business requests enter the factory, they first need to be matched with suitable computing resources, then generate Tokens through efficient inference, and finally complete metering and delivery according to customer requirements. Throughout this process, resource utilization, response speed, and service stability all directly affect final production capacity.
Around these stages, the six technical capabilities Qingmao had previously accumulated happen to form a mutually connected software foundation.
The intelligent gateway handles request distribution, and multi-level KV Cache reduces repeated computation; operator acceleration and distributed parallelism improve inference efficiency and cluster throughput; stability operations ensure continuous delivery, while RSI self-evolving optimization allows experience accumulated during operation to be reused, driving continuous improvement of the entire system.
Technologies originally serving model deployment and computing power optimization are now further applied to Token production and operations. And these capabilities also take effect sequentially along the flow of requests.
The first thing improved is the actual utilization efficiency of computing resources.
Agent applications involve large amounts of long context and continuous calls, and different requests differ significantly in their demand for computing resources. Qingmao dynamically allocates computing power through the intelligent gateway and uses four-level KV Cache to reuse existing computation results and reduce repeated operations.
According to company disclosures, the cache hit rate in related scenarios reaches as high as 99%, and computing utilization has been raised from 30%–40% to about 80%, with problems of repeated computation and idle resources significantly improved.
That is to say, the same batch of chips can handle more real business requests, and computing centers do not need to rely entirely on adding equipment to expand service capacity, which also creates room for reducing the overall cost of Token production.
After resources are fully utilized, the actual output speed of Tokens still needs to be further improved.
Compared with performance changes in individual operators, cluster throughput and response latency are closer to the service capabilities customers can perceive. The former means more Tokens can be processed in the same amount of time, while the latter reflects improvement in the waiting time of high-latency requests.
Through underlying operator and distributed parallelism optimization, Qingmao increased the execution performance of a specific MoE operator by 84 times and the performance of a Linear Attention operator by 63.4 times in related tests.

At the cluster level, Token throughput per minute increased from 40 million to 140 million, and P99 first-Token latency was shortened from 180 seconds to 40 seconds.
Once Tokens are produced, they also need to be delivered stably.
For Agent applications that need to call models continuously, stability directly affects the actual usage cost of Token services, and this is also an important prerequisite for the factory to enter enterprise production environments.
To this end, Qingmao has established an operations system covering operation monitoring, fault isolation, and business recovery, safeguarding service availability of 99.99% as much as possible.
These technologies jointly act on the throughput, response, stability, and cost control of Token production, corresponding exactly to the "more, faster, better, and more economical" standard proposed by Qingmao.
However, the real difficulty of a ten-thousand-card factory still lies in scaling replication.
Faced with tens of thousands of chips, a large number of different model combinations, and changes in business loads, if traditional methods were still relied upon and engineers repeatedly debugged to complete optimization, both cost and difficulty would surge.
For this reason, Qingmao introduced RSI self-evolving technology into the Token factory, attempting to "use AI to optimize AI" and let AI participate in optimizing the computing systems that support its own operation.

At present, this system has accumulated tens of thousands of optimization trajectories, supporting hour-level development and verification of new operators and day-level adaptation and launch of new models.
This means that when new models and new hardware are connected, some of the work originally requiring repeated efforts by engineers has the opportunity to be completed faster with the help of existing experience. The shortened model adaptation and optimization cycle also helps reduce the waiting time before new computing power is put into use.
Qingmao's early accumulation in the field of automated optimization has thus extended into large-scale infrastructure operations, and will be further released as the Token factory's operating network expands.
Moreover, the joint operating network allows Qingmao to enter more partners' computing centers. The operational data generated by different chips, models, and real business operations in turn provides the RSI system with more optimization experience, further improving Token production efficiency.
Technology, operations, and scale thus form a mutually reinforcing relationship.
Along this path, Qingmao is building a domestic Token computing power foundation oriented toward domestic chips and domestic models.
The bottom layer connects computing resources across multiple regions and multiple chips, the upper layer undertakes the real needs of large models, Agents, and industry applications, and the middle layer, through inference optimization, unified scheduling, and continuous operations, transforms scattered computing power into stable, scalable Token supply.
As domestic chip supply continues to expand and model capabilities continue to break through, new software systems, operational services, and business collaboration mechanisms are taking shape. Competition among domestic chips and domestic models is beginning to penetrate the organization and operation of the entire industrial system.
The deeper change is that China's AI industry is establishing its own production order.
The supply of non-domestic advanced computing power is undergoing repeated fluctuations, but the iteration and application demand of domestic models will not stop.
The progress of each generation of chips and the breakthrough of each model require mature software systems and industrial collaboration to carry them, allowing technological achievements to settle into continuous, stable supply capacity.
The rise of domestic chips and domestic models must ultimately go beyond the technological breakthrough of a single product and form an industrial system that operates autonomously and evolves itself.
Only by firmly holding chips, models, and the production system in our own hands can China's AI truly move forward at its own pace.
- Green power in, Tokens out! Lenovo redefines the token factory 2026-11-14
- Yu Chengdong releases the new AITO M8: starting at 380,000 yuan, with L3 architecture pre-embedded across the entire lineup 2026-11-16
- Chinese robotics society "winks through the wall"! Success rate soars from 0 to 100%, teaming up to work without burning out 2026-11-13
- Musk is ready to hire Jensen Huang 2026-11-12
Related Reading

AI Infra enters the self-evolving era! Tsinghua team's AI optimizing AI creates a domestic trillion-Token factory
ALL IN the domestic computing power ecosystem